update

.gitattributes

@@ -33,3 +33,5 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+*.jpg filter=lfs diff=lfs merge=lfs -text
+*.so filter=lfs diff=lfs merge=lfs -text

README.md

@@ -5,10 +5,8 @@ colorFrom: pink
 colorTo: green
 sdk: gradio
 sdk_version: 5.49.1
-app_file: app.py
+app_file: demo/gradio_demo_with_sam3.py
 pinned: false
 license: apache-2.0
 short_description: Complex text label dection using SAM3 with VLM-FO1
 ---
-
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference

demo/gradio_demo.py

@@ -0,0 +1,374 @@
+import gradio as gr
+from PIL import Image, ImageDraw, ImageFont
+import re
+import numpy as np
+from skimage.measure import label, regionprops
+from skimage.morphology import binary_dilation, disk
+from detect_tools.upn import UPNWrapper
+from vlm_fo1.model.builder import load_pretrained_model
+from vlm_fo1.mm_utils import (
+    prepare_inputs,
+    extract_predictions_to_indexes,
+)
+from vlm_fo1.task_templates import *
+import torch
+import os
+from copy import deepcopy
+
+
+TASK_TYPES = {
+    "OD/REC": OD_template,
+    "ODCounting": OD_Counting_template,
+    "Region_OCR": "Please provide the ocr results of these regions in the image.",
+    "Brief_Region_Caption": "Provide a brief description for these regions in the image.",
+    "Detailed_Region_Caption": "Provide a detailed description for these regions in the image.",
+    "Viusal_Region_Reasoning": Viusal_Region_Reasoning_template,
+    "OD_All": OD_All_template,
+    "Grounding": Grounding_template,
+}
+
+EXAMPLES = [
+    ["demo_image.jpg", TASK_TYPES["OD/REC"].format("orange, apple"), "OD/REC"],
+    ["demo_image_01.jpg", TASK_TYPES["ODCounting"].format("airplane with only one propeller"), "ODCounting"],
+    ["demo_image_02.jpg", TASK_TYPES["OD/REC"].format("the ball closest to the bear"), "OD/REC"],
+    ["demo_image_03.jpg", TASK_TYPES["OD_All"].format(""), "OD_All"],
+    ["demo_image_03.jpg", TASK_TYPES["Viusal_Region_Reasoning"].format("What's the brand of this computer?"), "Viusal_Region_Reasoning"],
+]
+
+
+def get_valid_examples():
+    valid_examples = []
+    demo_dir = os.path.dirname(os.path.abspath(__file__))
+    for example in EXAMPLES:
+        img_path = example[0]
+        full_path = os.path.join(demo_dir, img_path)
+        if os.path.exists(full_path):
+            valid_examples.append([
+                full_path,
+                example[1], 
+                example[2]
+            ])
+        elif os.path.exists(img_path):
+            valid_examples.append([
+                img_path,
+                example[1], 
+                example[2]
+            ])
+    return valid_examples
+
+
+def detect_model(image, threshold=0.3):
+    proposals = upn_model.inference(image)
+    filtered_proposals = upn_model.filter(proposals, min_score=threshold)
+    return filtered_proposals['original_xyxy_boxes'][0][:100]
+
+
+def multimodal_model(image, bboxes, text):
+    if '<image>' in text:
+        print(text)
+        parts = [part.replace('\\n', '\n') for part in re.split(rf'(<image>)', text) if part.strip()]
+        print(parts)
+        content = []
+        for part in parts:
+            if part == '<image>':
+                content.append({"type": "image_url", "image_url": {"url": image}})
+            else:
+                content.append({"type": "text", "text": part})
+    else:
+        content = [{
+            "type": "image_url",
+            "image_url": {
+                "url": image
+            }
+            }, {
+                "type": "text",
+                "text": text
+            }]
+
+    messages = [
+        {
+            "role": "user",
+            "content": content,
+            "bbox_list": bboxes
+        }
+    ]
+    generation_kwargs = prepare_inputs(model_path, model, image_processors, tokenizer, messages,
+    max_tokens=4096, top_p=0.05, temperature=0.0, do_sample=False)
+    with torch.inference_mode():
+        output_ids = model.generate(**generation_kwargs)
+        outputs = tokenizer.decode(output_ids[0, generation_kwargs['inputs'].shape[1]:]).strip()
+        print("========output========\n", outputs)
+
+    if '<ground>' in outputs:
+        prediction_dict = extract_predictions_to_indexes(outputs)
+    else:
+        match_pattern = r"<region(\d+)>"
+        matches = re.findall(match_pattern, outputs)
+        prediction_dict = {f"<region{m}>": {int(m)} for m in matches}
+
+    ans_bbox_json = []
+    ans_bbox_list = []
+    for k, v in prediction_dict.items():
+        for box_index in v:
+            box_index = int(box_index)
+            if box_index < len(bboxes):
+                current_bbox = bboxes[box_index]
+                ans_bbox_json.append({
+                    "region_index": f"<region{box_index}>",
+                    "xmin": current_bbox[0],
+                    "ymin": current_bbox[1],
+                    "xmax": current_bbox[2],
+                    "ymax": current_bbox[3],
+                    "label": k
+                })
+                ans_bbox_list.append(current_bbox)
+
+    return outputs, ans_bbox_json, ans_bbox_list
+
+
+def draw_bboxes(image, bboxes, labels=None):
+    image = image.copy()
+    draw = ImageDraw.Draw(image)
+
+    for bbox in bboxes:
+        draw.rectangle(bbox, outline="red", width=3)
+    return image
+
+
+def extract_bbox_and_original_image(edited_image):
+    """Extract original image and bounding boxes from ImageEditor output"""
+    if edited_image is None:
+        return None, []
+    
+    if isinstance(edited_image, dict):
+        original_image = edited_image.get("background")
+        bbox_list = []
+
+        if original_image is None:
+            return None, []
+
+        if edited_image.get("layers") is None or len(edited_image.get("layers", [])) == 0:
+            return original_image, []
+
+        try:
+            drawing_layer = edited_image["layers"][0]
+            alpha_channel = drawing_layer.getchannel('A')
+            alpha_np = np.array(alpha_channel)
+
+            binary_mask = alpha_np > 0
+
+            structuring_element = disk(5)
+            dilated_mask = binary_dilation(binary_mask, structuring_element)
+
+            labeled_image = label(dilated_mask)
+            regions = regionprops(labeled_image)
+
+            for prop in regions:
+                y_min, x_min, y_max, x_max = prop.bbox
+                bbox_list.append((x_min, y_min, x_max, y_max))
+        except Exception as e:
+            print(f"Error extracting bboxes from layers: {e}")
+            return original_image, []
+
+        return original_image, bbox_list
+    elif isinstance(edited_image, Image.Image):
+        return edited_image, []
+    else:
+        print(f"Unknown input type: {type(edited_image)}")
+        return None, []
+
+
+def process(image, example_image, prompt, threshold):
+    image, bbox_list = extract_bbox_and_original_image(image)
+
+    if example_image is not None:
+        image = example_image
+    
+    if image is None:
+        error_msg = "Error: Please upload an image or select a valid example."
+        print(f"Error: image is None, original input type: {type(image)}")
+        return None, None, error_msg, []
+    
+    try:
+        image = image.convert('RGB')
+    except Exception as e:
+        error_msg = f"Error: Cannot process image - {str(e)}"
+        return None, None, error_msg, []
+
+    if len(bbox_list) == 0:
+        bboxes = detect_model(image, threshold)
+    else:
+        bboxes = bbox_list
+        for idx in range(len(bboxes)):
+            prompt += f'<region{idx}>'
+
+    ans, ans_bbox_json, ans_bbox_list = multimodal_model(image, bboxes, prompt)
+
+    image_with_detection = draw_bboxes(image, bboxes)
+
+    annotated_bboxes = []
+    if len(ans_bbox_json) > 0:
+        for item in ans_bbox_json:
+            annotated_bboxes.append(
+                ((int(item['xmin']), int(item['ymin']), int(item['xmax']), int(item['ymax'])), item['label'])
+            )
+    annotated_image = (image, annotated_bboxes)
+
+    return annotated_image, image_with_detection, ans, ans_bbox_json
+
+
+def update_btn(is_processing):
+    if is_processing:
+        return gr.update(value="Processing...", interactive=False)
+    else:
+        return gr.update(value="Submit", interactive=True)
+
+
+def launch_demo():
+    with gr.Blocks() as demo:
+        gr.Markdown("# 🚀 VLM-FO1 Demo")
+        gr.Markdown("""
+        ### 📋 Instructions
+        
+        **Step 1: Prepare Your Image**
+        - Upload an image using the image editor below
+        - *Optional:* Draw circular regions with the red brush to specify areas of interest
+        - *Alternative:* If not drawing regions, the detection model will automatically identify regions
+        
+        **Step 2: Configure Your Task**
+        - Select a task template from the dropdown menu
+        - Replace `[WRITE YOUR INPUT HERE]` with your target objects or query
+        - *Example:* For detecting "person" and "dog", replace with: `person, dog`
+        - *Or:* Write your own custom prompt
+        
+        **Step 3: Fine-tune Detection** *(Optional)*
+        - Adjust the detection threshold slider to control sensitivity
+        
+        **Step 4: Generate Results**
+        - Click the **Submit** button to process your request
+        - View the detection results and model outputs below
+        
+        🔗 [GitHub Repository](https://github.com/om-ai-lab/VLM-FO1)
+        """)
+        
+        with gr.Row():
+            with gr.Column():
+                img_input_draw = gr.ImageEditor(
+                    label="Image Input",
+                    image_mode="RGBA",
+                    type="pil",
+                    sources=['upload'],
+                    brush=gr.Brush(colors=["#FF0000"], color_mode="fixed", default_size=2),
+                    interactive=True
+                )
+
+                gr.Markdown("### Prompt & Parameters")
+
+                def set_prompt_from_template(selected_task):
+                    return gr.update(value=TASK_TYPES[selected_task].format("[WRITE YOUR INPUT HERE]"))
+                
+                def load_example(prompt_input, task_type_input, hidden_image_box):
+                    cached_image = deepcopy(hidden_image_box)
+                    w, h = cached_image.size
+    
+                    transparent_layer = Image.new('RGBA', (w, h), (0, 0, 0, 0))
+
+                    new_editor_value = {
+                        "background": cached_image,
+                        "layers": [transparent_layer],
+                        "composite": None
+                    }
+                    
+                    return new_editor_value, prompt_input, task_type_input
+                
+                def reset_hidden_image_box():
+                    return gr.update(value=None)
+
+                task_type_input = gr.Dropdown(
+                    choices=list(TASK_TYPES.keys()),
+                    value="OD/REC",
+                    label="Prompt Templates",
+                    info="Select the prompt template for the task, or write your own prompt."
+                )
+
+                prompt_input = gr.Textbox(
+                    label="Task Prompt", 
+                    value=TASK_TYPES["OD/REC"].format("[WRITE YOUR INPUT HERE]"),
+                    lines=2,
+                )
+
+                task_type_input.select(
+                    set_prompt_from_template,
+                    inputs=task_type_input,
+                    outputs=prompt_input
+                )
+
+                hidden_image_box = gr.Image(label="Image", type="pil", image_mode="RGBA", visible=False)
+
+                threshold_input = gr.Slider(minimum=0.0, maximum=1.0, value=0.3, step=0.01, label="Detection Model Threshold")
+                submit_btn = gr.Button("Submit", variant="primary")
+                
+                valid_examples = get_valid_examples()
+                if len(valid_examples) > 0:
+                    gr.Markdown("### Examples")
+                    gr.Markdown("Click on the examples below to quickly load images and corresponding prompts:")
+                    
+                    examples_data = [[example[0], example[1], example[2]] for index, example in enumerate(valid_examples)]
+                    
+                    examples = gr.Examples(
+                        examples=examples_data,
+                        inputs=[hidden_image_box, prompt_input, task_type_input],
+                        label="Click to load example",
+                        examples_per_page=5
+                    )
+                    
+                    examples.load_input_event.then(
+                        fn=load_example,
+                        inputs=[prompt_input, task_type_input, hidden_image_box], 
+                        outputs=[img_input_draw, prompt_input, task_type_input]
+                    )
+
+                    img_input_draw.upload(
+                        fn=reset_hidden_image_box,
+                        outputs=[hidden_image_box]
+                    )
+
+            with gr.Column():
+                with gr.Accordion("Detection Result", open=True):
+                    image_with_detection = gr.Image(label="Detection Result", height=200)
+
+                image_output = gr.AnnotatedImage(label="VLM-FO1 Result", height=400)
+
+                result_output = gr.Textbox(label="VLM-FO1 Output", lines=5)
+                ans_bbox_json = gr.JSON(label="Extracted Detection Output")
+
+        submit_btn.click(
+            update_btn, 
+            inputs=[gr.State(True)], 
+            outputs=[submit_btn], 
+            queue=False
+        ).then(
+            process,
+            inputs=[img_input_draw, hidden_image_box, prompt_input, threshold_input],
+            outputs=[image_output, image_with_detection, result_output, ans_bbox_json],
+            queue=True
+        ).then(
+            update_btn, 
+            inputs=[gr.State(False)], 
+            outputs=[submit_btn], 
+            queue=False
+        )
+    
+    return demo
+
+if __name__ == "__main__":
+    model_path = './resources/VLM-FO1_Qwen2.5-VL-3B-v01' 
+    upn_ckpt_path = "./resources/upn_large.pth" 
+    tokenizer, model, image_processors = load_pretrained_model(
+        model_path=model_path,
+        device="cuda:0",
+    )
+    upn_model = UPNWrapper(upn_ckpt_path)
+
+    demo = launch_demo()
+    demo.launch(server_name="0.0.0.0", share=False, server_port=8000, debug=False)

demo/gradio_demo_with_sam3.py

@@ -0,0 +1,323 @@
+import gradio as gr
+import spaces
+from PIL import Image, ImageDraw, ImageFont
+import re
+import numpy as np
+from skimage.measure import label, regionprops
+from skimage.morphology import binary_dilation, disk
+from sam3.model_builder import build_sam3_image_model
+from sam3.model.sam3_image_processor import Sam3Processor
+from sam3.visualization_utils import plot_bbox, plot_mask, COLORS
+import matplotlib.pyplot as plt
+
+from vlm_fo1.model.builder import load_pretrained_model
+from vlm_fo1.mm_utils import (
+    prepare_inputs,
+    extract_predictions_to_indexes,
+)
+from vlm_fo1.task_templates import *
+import torch
+import os
+from copy import deepcopy
+
+
+EXAMPLES = [
+    ["demo/sam3_examples/00000-72.jpg","airplane with letter AE on its body"],
+    ["demo/sam3_examples/00000-32.jpg","the lying cat which is not black"],
+    ["demo/sam3_examples/00000-22.jpg","person wearing a black top"],
+    ["demo/sam3_examples/000000378453.jpg", "zebra inside the mud puddle"],
+    ["demo/sam3_examples/00000-242.jpg", "person who is holding a book"],
+]
+
+
+def get_valid_examples():
+    valid_examples = []
+    demo_dir = os.path.dirname(os.path.abspath(__file__))
+    for example in EXAMPLES:
+        img_path = example[0]
+        full_path = os.path.join(demo_dir, img_path)
+        if os.path.exists(full_path):
+            valid_examples.append([
+                full_path,
+                example[1], 
+                example[2]
+            ])
+        elif os.path.exists(img_path):
+            valid_examples.append([
+                img_path,
+                example[1], 
+                example[2]
+            ])
+    return valid_examples
+
+
+def detect_model(image, text, threshold=0.3):
+    inference_state = sam3_processor.set_image(image)
+    output = sam3_processor.set_text_prompt(state=inference_state, prompt=text)
+    boxes, scores, masks = output["boxes"], output["scores"], output["masks"]   
+    sorted_indices = torch.argsort(scores, descending=True)
+    boxes = boxes[sorted_indices][:100, :]
+    scores = scores[sorted_indices][:100]
+    masks = masks[sorted_indices][:100]
+    # If the highest confidence score is greater than 0.5, filter with 0.3 threshold
+    if len(scores) > 0 and scores[0] > 0.75:
+        conf_threshold = 0.3
+        
+    else:
+        conf_threshold = 0.05
+    mask = scores > conf_threshold
+    boxes = boxes[mask]
+    scores = scores[mask]
+    masks = masks[mask]
+    # Keep boxes with score > 0.8 in a separate list
+    high_conf_mask = scores > 0.8
+    high_conf_boxes = boxes[high_conf_mask]
+    
+    print("========boxes========\n", boxes.tolist())
+    print("========scores========\n", scores.tolist())
+    print("========high_conf_boxes (>0.8)========\n", high_conf_boxes.tolist())
+
+    output = {
+        "boxes": boxes,
+        "scores": scores,
+        "masks": masks,
+    }
+    return boxes.tolist(), scores.tolist(), high_conf_boxes.tolist(), masks.tolist(), output
+
+
+def multimodal_model(image, bboxes, scores, text):
+    if len(bboxes) == 0:
+        return None, {}, []
+    
+    if '<image>' in text:
+        print(text)
+        parts = [part.replace('\\n', '\n') for part in re.split(rf'(<image>)', text) if part.strip()]
+        print(parts)
+        content = []
+        for part in parts:
+            if part == '<image>':
+                content.append({"type": "image_url", "image_url": {"url": image}})
+            else:
+                content.append({"type": "text", "text": part})
+    else:
+        content = [{
+            "type": "image_url",
+            "image_url": {
+                "url": image
+            }
+            }, {
+                "type": "text",
+                "text": text
+            }]
+
+    messages = [
+        {
+            "role": "user",
+            "content": content,
+            "bbox_list": bboxes
+        }
+    ]
+    generation_kwargs = prepare_inputs(model_path, model, image_processors, tokenizer, messages,
+    max_tokens=4096, top_p=0.05, temperature=0.0, do_sample=False, image_size=1024)
+    with torch.inference_mode():
+        output_ids = model.generate(**generation_kwargs)
+        outputs = tokenizer.decode(output_ids[0, generation_kwargs['inputs'].shape[1]:]).strip()
+        print("========output========\n", outputs)
+
+    if '<ground>' in outputs:
+        prediction_dict = extract_predictions_to_indexes(outputs)
+    else:
+        match_pattern = r"<region(\d+)>"
+        matches = re.findall(match_pattern, outputs)
+        prediction_dict = {f"<region{m}>": {int(m)} for m in matches}
+
+    ans_bbox_json = []
+    ans_bbox_list = []
+    for k, v in prediction_dict.items():
+        for box_index in v:
+            box_index = int(box_index)
+            if box_index < len(bboxes):
+                current_bbox = bboxes[box_index]
+                current_score = scores[box_index]
+                ans_bbox_json.append({
+                    "region_index": f"<region{box_index}>",
+                    "xmin": current_bbox[0],
+                    "ymin": current_bbox[1],
+                    "xmax": current_bbox[2],
+                    "ymax": current_bbox[3],
+                    "label": k,
+                    "score": current_score
+                })
+                ans_bbox_list.append(current_bbox)
+
+    return outputs, ans_bbox_json, ans_bbox_list
+
+
+def draw_bboxes(img, results):
+    fig, ax = plt.subplots(figsize=(12, 8))
+    # fig.subplots_adjust(0, 0, 1, 1)
+    ax.imshow(img)
+    nb_objects = len(results["scores"])
+    print(f"found {nb_objects} object(s)")
+    for i in range(nb_objects):
+        color = COLORS[i % len(COLORS)]
+        plot_mask(results["masks"][i].squeeze(0).cpu(), color=color)
+        w, h = img.size
+        prob = results["scores"][i].item()
+        plot_bbox(
+            h,
+            w,
+            results["boxes"][i].cpu(),
+            text=f"(id={i}, {prob=:.2f})",
+            box_format="XYXY",
+            color=color,
+            relative_coords=False,
+        )
+    ax.axis("off")
+    fig.tight_layout(pad=0)
+    
+    # Convert matplotlib figure to PIL Image
+    fig.canvas.draw()
+    buf = fig.canvas.buffer_rgba()
+    pil_img = Image.frombytes('RGBA', fig.canvas.get_width_height(), buf)
+    plt.close(fig)
+    
+    return pil_img
+
+
+@spaces.GPU
+def process(image, prompt, threshold=0):
+    if image is None:
+        error_msg = "Error: Please upload an image or select a valid example."
+        print(f"Error: image is None, original input type: {type(image)}")
+        return None, None, error_msg, []
+    
+    try:
+        image = image.convert('RGB')
+    except Exception as e:
+        error_msg = f"Error: Cannot process image - {str(e)}"
+        return None, None, error_msg, []
+
+    bboxes, scores, high_conf_bboxes, masks, output = detect_model(image, prompt, threshold)
+
+    fo1_prompt = OD_Counting_template.format(prompt)
+    ans, ans_bbox_json, ans_bbox_list = multimodal_model(image, bboxes, scores, fo1_prompt)
+
+    detection_image = draw_bboxes(image, output)
+
+    annotated_bboxes = []
+    if len(ans_bbox_json) > 0:
+        img_width, img_height = image.size
+        for item in ans_bbox_json:
+            xmin = max(0, min(img_width, int(item['xmin'])))
+            ymin = max(0, min(img_height, int(item['ymin'])))
+            xmax = max(0, min(img_width, int(item['xmax'])))
+            ymax = max(0, min(img_height, int(item['ymax'])))
+            annotated_bboxes.append(
+                ((xmin, ymin, xmax, ymax), item['label'])
+            )
+    annotated_image = (image, annotated_bboxes)
+
+    return annotated_image, detection_image, ans_bbox_json
+
+
+def update_btn(is_processing):
+    if is_processing:
+        return gr.update(value="Processing...", interactive=False)
+    else:
+        return gr.update(value="Submit", interactive=True)
+
+
+def launch_demo():
+    with gr.Blocks() as demo:
+        gr.Markdown("# 🚀 VLM-FO1 + SAM3 Demo")
+        gr.Markdown("""
+        ### 📋 Instructions
+        Combine the SAM3 detection results with the VLM-FO1 model to enchance its dectection and segmentation performance on complex label tasks.
+
+        **How it works**
+        1. Upload or pick an example image.
+        2. Describe the target object in natural language.
+        3. Hit **Submit** to run SAM3 + VLM-FO1.
+        
+        **Outputs**
+        - `SAM3 Result`: raw detections with masks/bboxes generated by SAM3.
+        - `VLM-FO1 Result`: filtered detections plus labels generated by VLM-FO1.
+        
+        **Tips**
+        - One prompt at a time is currently supported. Multiple label prompts will be supported soon.
+        - Use the examples below to quickly explore the pipeline.
+        """)
+        
+        gr.Markdown("""
+        ### 🔗 References
+        - [SAM3](https://github.com/facebookresearch/sam3)
+        - [VLM-FO1](https://github.com/om-ai-lab/VLM-FO1)
+        """)
+        
+        with gr.Row():
+            with gr.Column():
+                img_input_draw = gr.Image(
+                    label="Image Input",
+                    type="pil",
+                    sources=['upload'],
+                )
+
+                gr.Markdown("### Prompt")
+
+                prompt_input = gr.Textbox(
+                    label="Label Prompt", 
+                    lines=2,
+                )
+
+                submit_btn = gr.Button("Submit", variant="primary")
+                
+                    
+                examples = gr.Examples(
+                    examples=EXAMPLES,
+                    inputs=[img_input_draw, prompt_input],
+                    label="Click to load example",
+                    examples_per_page=5
+                )
+
+            with gr.Column():
+                with gr.Accordion("SAM3 Result", open=True):
+                    image_output_detection = gr.Image(label="SAM3 Result", height=400)
+
+                image_output = gr.AnnotatedImage(label="VLM-FO1 Result", height=400)
+
+                ans_bbox_json = gr.JSON(label="Extracted Detection Output")
+
+        submit_btn.click(
+            update_btn, 
+            inputs=[gr.State(True)], 
+            outputs=[submit_btn], 
+            queue=False
+        ).then(
+            process,
+            inputs=[img_input_draw, prompt_input],
+            outputs=[image_output, image_output_detection, ans_bbox_json],
+            queue=True
+        ).then(
+            update_btn, 
+            inputs=[gr.State(False)], 
+            outputs=[submit_btn], 
+            queue=False
+        )
+    
+    return demo
+
+if __name__ == "__main__":
+    # model_path = './resources/VLM-FO1_Qwen2.5-VL-3B-v01' 
+    # sam3_model_path = './resources/sam3/sam3.pt'
+
+    model_path = 'omlab/VLM-FO1_Qwen2.5-VL-3B-v01'
+    tokenizer, model, image_processors = load_pretrained_model(
+        model_path=model_path,
+        device="cuda:0",
+    )
+    sam3_model = build_sam3_image_model(device="cuda:0")
+    sam3_processor = Sam3Processor(sam3_model, confidence_threshold=0.0, device="cuda:0")
+
+    demo = launch_demo()
+    demo.launch()

detect_tools/sam3/.gitignore

@@ -0,0 +1,153 @@
+# Byte-compiled / optimized / DLL files
+__pycache__/
+*.py[cod]
+*$py.class
+
+# C extensions
+*.so
+
+# Distribution / packaging
+.Python
+build/
+develop-eggs/
+dist/
+downloads/
+eggs/
+.eggs/
+lib/
+lib64/
+parts/
+sdist/
+var/
+wheels/
+*.egg-info/
+.installed.cfg
+*.egg
+MANIFEST
+
+# PyInstaller
+#  Usually these files are written by a python script from a template
+#  before PyInstaller builds the exe, so as to inject date/other infos into it.
+*.manifest
+*.spec
+
+# Installer logs
+pip-log.txt
+pip-delete-this-directory.txt
+
+# Unit test / coverage reports
+htmlcov/
+.tox/
+.nox/
+.coverage
+.coverage.*
+.cache
+nosetests.xml
+coverage.xml
+*.cover
+.hypothesis/
+.pytest_cache/
+
+# Translations
+*.mo
+*.pot
+
+# Django stuff:
+*.log
+local_settings.py
+db.sqlite3
+
+# Flask stuff:
+instance/
+.webassets-cache
+
+# Scrapy stuff:
+.scrapy
+
+# Sphinx documentation
+docs/_build/
+
+# PyBuilder
+target/
+
+# Jupyter Notebook
+.ipynb_checkpoints
+*-Copy*.ipynb
+
+# IPython
+profile_default/
+ipython_config.py
+
+# pyenv
+.python-version
+
+# celery beat schedule file
+celerybeat-schedule
+
+# SageMath parsed files
+*.sage.py
+
+# Environments
+.env
+.venv
+env/
+venv/
+ENV/
+env.bak/
+venv.bak/
+
+# Spyder project settings
+.spyderproject
+.spyproject
+
+# Rope project settings
+.ropeproject
+
+# mkdocs documentation
+/site
+
+# mypy
+.mypy_cache/
+.dmypy.json
+dmypy.json
+
+# Pyre type checker
+.pyre/
+
+# PyCharm
+.idea/
+
+# VS Code
+.vscode/
+*.code-workspace
+
+# Model weights and checkpoints
+*.pth
+*.pt
+*.bin
+*.ckpt
+*.safetensors
+weights/
+checkpoints/
+sam3_logs/
+
+# Data files
+*.h5
+*.hdf5
+*.pkl
+*.pickle
+*.npy
+*.npz
+
+# Logs
+logs/
+runs/
+tensorboard/
+
+# OS specific
+.DS_Store
+Thumbs.db
+
+# BPE vocabulary files
+*.bpe
+*.vocab

detect_tools/sam3/CODE_OF_CONDUCT.md

@@ -0,0 +1,80 @@
+# Code of Conduct
+
+## Our Pledge
+
+In the interest of fostering an open and welcoming environment, we as
+contributors and maintainers pledge to make participation in our project and
+our community a harassment-free experience for everyone, regardless of age, body
+size, disability, ethnicity, sex characteristics, gender identity and expression,
+level of experience, education, socio-economic status, nationality, personal
+appearance, race, religion, or sexual identity and orientation.
+
+## Our Standards
+
+Examples of behavior that contributes to creating a positive environment
+include:
+
+* Using welcoming and inclusive language
+* Being respectful of differing viewpoints and experiences
+* Gracefully accepting constructive criticism
+* Focusing on what is best for the community
+* Showing empathy towards other community members
+
+Examples of unacceptable behavior by participants include:
+
+* The use of sexualized language or imagery and unwelcome sexual attention or
+advances
+* Trolling, insulting/derogatory comments, and personal or political attacks
+* Public or private harassment
+* Publishing others' private information, such as a physical or electronic
+address, without explicit permission
+* Other conduct which could reasonably be considered inappropriate in a
+professional setting
+
+## Our Responsibilities
+
+Project maintainers are responsible for clarifying the standards of acceptable
+behavior and are expected to take appropriate and fair corrective action in
+response to any instances of unacceptable behavior.
+
+Project maintainers have the right and responsibility to remove, edit, or
+reject comments, commits, code, wiki edits, issues, and other contributions
+that are not aligned to this Code of Conduct, or to ban temporarily or
+permanently any contributor for other behaviors that they deem inappropriate,
+threatening, offensive, or harmful.
+
+## Scope
+
+This Code of Conduct applies within all project spaces, and it also applies when
+an individual is representing the project or its community in public spaces.
+Examples of representing a project or community include using an official
+project e-mail address, posting via an official social media account, or acting
+as an appointed representative at an online or offline event. Representation of
+a project may be further defined and clarified by project maintainers.
+
+This Code of Conduct also applies outside the project spaces when there is a
+reasonable belief that an individual's behavior may have a negative impact on
+the project or its community.
+
+## Enforcement
+
+Instances of abusive, harassing, or otherwise unacceptable behavior may be
+reported by contacting the project team at <[email protected]>. All
+complaints will be reviewed and investigated and will result in a response that
+is deemed necessary and appropriate to the circumstances. The project team is
+obligated to maintain confidentiality with regard to the reporter of an incident.
+Further details of specific enforcement policies may be posted separately.
+
+Project maintainers who do not follow or enforce the Code of Conduct in good
+faith may face temporary or permanent repercussions as determined by other
+members of the project's leadership.
+
+## Attribution
+
+This Code of Conduct is adapted from the [Contributor Covenant][homepage], version 1.4,
+available at https://www.contributor-covenant.org/version/1/4/code-of-conduct.html
+
+[homepage]: https://www.contributor-covenant.org
+
+For answers to common questions about this code of conduct, see
+https://www.contributor-covenant.org/faq

detect_tools/sam3/CONTRIBUTING.md

@@ -0,0 +1,30 @@
+# Contributing to sam3
+We want to make contributing to this project as easy and transparent as
+possible.
+
+## Pull Requests
+We actively welcome your pull requests.
+
+1. Fork the repo and create your branch from `main`.
+2. If you've added code that should be tested, add tests.
+3. If you've changed APIs, update the documentation.
+4. Make sure your code lints.
+5. If you haven't already, complete the Contributor License Agreement ("CLA").
+
+## Contributor License Agreement ("CLA")
+In order to accept your pull request, we need you to submit a CLA. You only need
+to do this once to work on any of Facebook's open source projects.
+
+Complete your CLA here: <https://code.facebook.com/cla>
+
+## Issues
+We use GitHub issues to track public bugs. Please ensure your description is
+clear and has sufficient instructions to be able to reproduce the issue.
+
+Facebook has a [bounty program](https://www.facebook.com/whitehat/) for the safe
+disclosure of security bugs. In those cases, please go through the process
+outlined on that page and do not file a public issue.
+
+## License
+By contributing to sam3, you agree that your contributions will be licensed
+under the LICENSE file in the root directory of this source tree.

detect_tools/sam3/LICENSE

@@ -0,0 +1,61 @@
+SAM License
+Last Updated: November 19, 2025
+
+“Agreement” means the terms and conditions for use, reproduction, distribution and modification of the SAM Materials set forth herein.
+
+
+“SAM Materials” means, collectively, Documentation and the models, software and algorithms, including machine-learning model code, trained model weights, inference-enabling code, training-enabling code, fine-tuning enabling code, and other elements of the foregoing distributed by Meta and made available under this Agreement.
+
+“Documentation” means the specifications, manuals and documentation accompanying
+SAM Materials distributed by Meta.
+
+
+“Licensee” or “you” means you, or your employer or any other person or entity (if you are entering into this Agreement on such person or entity’s behalf), of the age required under applicable laws, rules or regulations to provide legal consent and that has legal authority to bind your employer or such other person or entity if you are entering in this Agreement on their behalf.
+
+
+“Meta” or “we” means Meta Platforms Ireland Limited (if you are located in or, if you are an entity, your principal place of business is in the EEA or Switzerland) or Meta Platforms, Inc. (if you are located outside of the EEA or Switzerland).
+
+
+“Sanctions” means any economic or trade sanctions or restrictions administered or enforced by the United States (including the Office of Foreign Assets Control of the U.S. Department of the Treasury (“OFAC”), the U.S. Department of State and the U.S. Department of Commerce), the United Nations, the European Union, or the United Kingdom.
+
+
+“Trade Controls” means any of the following: Sanctions and applicable export and import controls.
+
+By using or distributing any portion or element of the SAM Materials, you agree to be bound by this Agreement.
+
+
+1. License Rights and Redistribution.
+
+
+a. Grant of Rights. You are granted a non-exclusive, worldwide, non-transferable and royalty-free limited license under Meta’s intellectual property or other rights owned by Meta embodied in the SAM Materials to use, reproduce, distribute, copy, create derivative works of, and make modifications to the SAM Materials.
+
+b. Redistribution and Use.
+i. Distribution of SAM Materials, and any derivative works thereof, are subject to the terms of this Agreement. If you distribute or make the SAM Materials, or any derivative works thereof, available to a third party, you may only do so under the terms of this Agreement and you shall provide a copy of this Agreement with any such SAM Materials.
+
+
+ii.  If you submit for publication the results of research you perform on, using, or otherwise in connection with SAM Materials, you must acknowledge the use of SAM Materials in your publication.
+
+
+iii. Your use of the SAM Materials must comply with applicable laws and regulations, including Trade Control Laws and applicable privacy and data protection laws.
+iv. Your use of the SAM Materials will not involve or encourage others to reverse engineer, decompile or discover the underlying components of the SAM Materials.
+v. You are not the target of Trade Controls and your use of SAM Materials must comply with Trade Controls. You agree not to use, or permit others to use, SAM Materials for any activities subject to the International Traffic in Arms Regulations (ITAR) or end uses prohibited by Trade Controls, including those related to military or warfare purposes, nuclear industries or applications, espionage, or the development or use of guns or illegal weapons.
+2. User Support. Your use of the SAM Materials is done at your own discretion; Meta does not process any information nor provide any service in relation to such use.  Meta is under no obligation to provide any support services for the SAM Materials. Any support provided is “as is”, “with all faults”, and without warranty of any kind.
+
+
+3. Disclaimer of Warranty. UNLESS REQUIRED BY APPLICABLE LAW, THE SAM MATERIALS AND ANY OUTPUT AND RESULTS THEREFROM ARE PROVIDED ON AN “AS IS” BASIS, WITHOUT WARRANTIES OF ANY KIND, AND META DISCLAIMS ALL WARRANTIES OF ANY KIND, BOTH EXPRESS AND IMPLIED, INCLUDING, WITHOUT LIMITATION, ANY WARRANTIES OF TITLE, NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE. YOU ARE SOLELY RESPONSIBLE FOR DETERMINING THE APPROPRIATENESS OF USING OR REDISTRIBUTING THE SAM MATERIALS AND ASSUME ANY RISKS ASSOCIATED WITH YOUR USE OF THE SAM MATERIALS AND ANY OUTPUT AND RESULTS.
+
+4. Limitation of Liability. IN NO EVENT WILL META OR ITS AFFILIATES BE LIABLE UNDER ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, TORT, NEGLIGENCE, PRODUCTS LIABILITY, OR OTHERWISE, ARISING OUT OF THIS AGREEMENT, FOR ANY LOST PROFITS OR ANY DIRECT OR INDIRECT, SPECIAL, CONSEQUENTIAL, INCIDENTAL, EXEMPLARY OR PUNITIVE DAMAGES, EVEN IF META OR ITS AFFILIATES HAVE BEEN ADVISED OF THE POSSIBILITY OF ANY OF THE FOREGOING.
+
+5. Intellectual Property.
+
+
+a. Subject to Meta’s ownership of SAM Materials and derivatives made by or for Meta, with respect to any derivative works and modifications of the SAM Materials that are made by you, as between you and Meta, you are and will be the owner of such derivative works and modifications.
+
+b. If you institute litigation or other proceedings against Meta or any entity (including a cross-claim or counterclaim in a lawsuit) alleging that the SAM Materials, outputs or results, or any portion of any of the foregoing, constitutes infringement of intellectual property or other rights owned or licensable by you, then any licenses granted to you under this Agreement shall terminate as of the date such litigation or claim is filed or instituted. You will indemnify and hold harmless Meta from and against any claim by any third party arising out of or related to your use or distribution of the SAM Materials.
+
+6. Term and Termination. The term of this Agreement will commence upon your acceptance of this Agreement or access to the SAM Materials and will continue in full force and effect until terminated in accordance with the terms and conditions herein. Meta may terminate this Agreement if you are in breach of any term or condition of this Agreement. Upon termination of this Agreement, you shall delete and cease use of the SAM Materials. Sections 3, 4 and 7 shall survive the termination of this Agreement.
+
+7. Governing Law and Jurisdiction. This Agreement will be governed and construed under the laws of the State of California without regard to choice of law principles, and the UN Convention on Contracts for the International Sale of Goods does not apply to this Agreement. The courts of California shall have exclusive jurisdiction of any dispute arising out of this Agreement.
+
+
+8. Modifications and Amendments. Meta may modify this Agreement from time to time; provided that they are similar in spirit to the current version of the Agreement, but may differ in detail to address new problems or concerns. All such changes will be effective immediately. Your continued use of the SAM Materials after any modification to this Agreement constitutes your agreement to such modification. Except as provided in this Agreement, no modification or addition to any provision of this Agreement will be binding unless it is in writing and signed by an authorized representative of both you and Meta.

detect_tools/sam3/MANIFEST.in

@@ -0,0 +1,6 @@
+include LICENSE
+include README.md
+recursive-include examples *.py
+recursive-include examples *.ipynb
+recursive-include examples *.md
+recursive-include tests *.py

detect_tools/sam3/README.md

@@ -0,0 +1,387 @@
+# SAM 3: Segment Anything with Concepts
+
+Meta Superintelligence Labs
+
+[Nicolas Carion](https://www.nicolascarion.com/)\*,
+[Laura Gustafson](https://scholar.google.com/citations?user=c8IpF9gAAAAJ&hl=en)\*,
+[Yuan-Ting Hu](https://scholar.google.com/citations?user=E8DVVYQAAAAJ&hl=en)\*,
+[Shoubhik Debnath](https://scholar.google.com/citations?user=fb6FOfsAAAAJ&hl=en)\*,
+[Ronghang Hu](https://ronghanghu.com/)\*,
+[Didac Suris](https://www.didacsuris.com/)\*,
+[Chaitanya Ryali](https://scholar.google.com/citations?user=4LWx24UAAAAJ&hl=en)\*,
+[Kalyan Vasudev Alwala](https://scholar.google.co.in/citations?user=m34oaWEAAAAJ&hl=en)\*,
+[Haitham Khedr](https://hkhedr.com/)\*, Andrew Huang,
+[Jie Lei](https://jayleicn.github.io/),
+[Tengyu Ma](https://scholar.google.com/citations?user=VeTSl0wAAAAJ&hl=en),
+[Baishan Guo](https://scholar.google.com/citations?user=BC5wDu8AAAAJ&hl=en),
+Arpit Kalla, [Markus Marks](https://damaggu.github.io/),
+[Joseph Greer](https://scholar.google.com/citations?user=guL96CkAAAAJ&hl=en),
+Meng Wang, [Peize Sun](https://peizesun.github.io/),
+[Roman Rädle](https://scholar.google.com/citations?user=Tpt57v0AAAAJ&hl=en),
+[Triantafyllos Afouras](https://www.robots.ox.ac.uk/~afourast/),
+[Effrosyni Mavroudi](https://scholar.google.com/citations?user=vYRzGGEAAAAJ&hl=en),
+[Katherine Xu](https://k8xu.github.io/)°,
+[Tsung-Han Wu](https://patrickthwu.com/)°,
+[Yu Zhou](https://yu-bryan-zhou.github.io/)°,
+[Liliane Momeni](https://scholar.google.com/citations?user=Lb-KgVYAAAAJ&hl=en)°,
+[Rishi Hazra](https://rishihazra.github.io/)°,
+[Shuangrui Ding](https://mark12ding.github.io/)°,
+[Sagar Vaze](https://sgvaze.github.io/)°,
+[Francois Porcher](https://scholar.google.com/citations?user=LgHZ8hUAAAAJ&hl=en)°,
+[Feng Li](https://fengli-ust.github.io/)°,
+[Siyuan Li](https://siyuanliii.github.io/)°,
+[Aishwarya Kamath](https://ashkamath.github.io/)°,
+[Ho Kei Cheng](https://hkchengrex.com/)°,
+[Piotr Dollar](https://pdollar.github.io/)†,
+[Nikhila Ravi](https://nikhilaravi.com/)†,
+[Kate Saenko](https://ai.bu.edu/ksaenko.html)†,
+[Pengchuan Zhang](https://pzzhang.github.io/pzzhang/)†,
+[Christoph Feichtenhofer](https://feichtenhofer.github.io/)†
+
+\* core contributor, ° intern, † project lead, order is random within groups
+
+[[`Paper`](https://ai.meta.com/research/publications/sam-3-segment-anything-with-concepts/)]
+[[`Project`](https://ai.meta.com/sam3)]
+[[`Demo`](https://segment-anything.com/)]
+[[`Blog`](https://ai.meta.com/blog/segment-anything-model-3/)]
+<!-- [[`BibTeX`](#citing-sam-3)] -->
+
+![SAM 3 architecture](assets/model_diagram.png?raw=true) SAM 3 is a unified foundation model for promptable segmentation in images and videos. It can detect, segment, and track objects using text or visual prompts such as points, boxes, and masks. Compared to its predecessor [SAM 2](https://github.com/facebookresearch/sam2), SAM 3 introduces the ability to exhaustively segment all instances of an open-vocabulary concept specified by a short text phrase or exemplars. Unlike prior work, SAM 3 can handle a vastly larger set of open-vocabulary prompts. It achieves 75-80% of human performance on our new [SA-CO benchmark](https://github.com/facebookresearch/sam3/edit/main_readme/README.md#sa-co-dataset) which contains 270K unique concepts, over 50 times more than existing benchmarks.
+
+This breakthrough is driven by an innovative data engine that has automatically annotated over 4 million unique concepts, creating the largest high-quality open-vocabulary segmentation dataset to date. In addition, SAM 3 introduces a new model architecture featuring a presence token that improves discrimination between closely related text prompts (e.g., “a player in white” vs. “a player in red”), as well as a decoupled detector–tracker design that minimizes task interference and scales efficiently with data.
+
+<p align="center">
+  <img src="assets/dog.gif" width=380 />
+  <img src="assets/player.gif" width=380 />
+</p>
+
+## Installation
+
+### Prerequisites
+
+- Python 3.12 or higher
+- PyTorch 2.7 or higher
+- CUDA-compatible GPU with CUDA 12.6 or higher
+
+1. **Create a new Conda environment:**
+
+```bash
+conda create -n sam3 python=3.12
+conda deactivate
+conda activate sam3
+```
+
+2. **Install PyTorch with CUDA support:**
+
+```bash
+pip install torch==2.7.0 torchvision torchaudio --index-url https://download.pytorch.org/whl/cu126
+```
+
+3. **Clone the repository and install the package:**
+
+```bash
+git clone https://github.com/facebookresearch/sam3.git
+cd sam3
+pip install -e .
+```
+
+4. **Install additional dependencies for example notebooks or development:**
+
+```bash
+# For running example notebooks
+pip install -e ".[notebooks]"
+
+# For development
+pip install -e ".[train,dev]"
+```
+
+## Getting Started
+
+⚠️ Before using SAM 3, please request access to the checkpoints on the SAM 3
+Hugging Face [repo](https://huggingface.co/facebook/sam3). Once accepted, you
+need to be authenticated to download the checkpoints. You can do this by running
+the following [steps](https://huggingface.co/docs/huggingface_hub/en/quick-start#authentication)
+(e.g. `hf auth login` after generating an access token.)
+
+### Basic Usage
+
+```python
+import torch
+#################################### For Image ####################################
+from PIL import Image
+from sam3.model_builder import build_sam3_image_model
+from sam3.model.sam3_image_processor import Sam3Processor
+# Load the model
+model = build_sam3_image_model()
+processor = Sam3Processor(model)
+# Load an image
+image = Image.open("<YOUR_IMAGE_PATH.jpg>")
+inference_state = processor.set_image(image)
+# Prompt the model with text
+output = processor.set_text_prompt(state=inference_state, prompt="<YOUR_TEXT_PROMPT>")
+
+# Get the masks, bounding boxes, and scores
+masks, boxes, scores = output["masks"], output["boxes"], output["scores"]
+
+#################################### For Video ####################################
+
+from sam3.model_builder import build_sam3_video_predictor
+
+video_predictor = build_sam3_video_predictor()
+video_path = "<YOUR_VIDEO_PATH>" # a JPEG folder or an MP4 video file
+# Start a session
+response = video_predictor.handle_request(
+    request=dict(
+        type="start_session",
+        resource_path=video_path,
+    )
+)
+response = video_predictor.handle_request(
+    request=dict(
+        type="add_prompt",
+        session_id=response["session_id"],
+        frame_index=0, # Arbitrary frame index
+        text="<YOUR_TEXT_PROMPT>",
+    )
+)
+output = response["outputs"]
+```
+
+## Examples
+
+The `examples` directory contains notebooks demonstrating how to use SAM3 with
+various types of prompts:
+
+- [`sam3_image_predictor_example.ipynb`](examples/sam3_image_predictor_example.ipynb)
+  : Demonstrates how to prompt SAM 3 with text and visual box prompts on images.
+- [`sam3_video_predictor_example.ipynb`](examples/sam3_video_predictor_example.ipynb)
+  : Demonstrates how to prompt SAM 3 with text prompts on videos, and doing
+  further interactive refinements with points.
+- [`sam3_image_batched_inference.ipynb`](examples/sam3_image_batched_inference.ipynb)
+  : Demonstrates how to run batched inference with SAM 3 on images.
+- [`sam3_agent.ipynb`](examples/sam3_agent.ipynb): Demonsterates the use of SAM
+  3 Agent to segment complex text prompt on images.
+- [`saco_gold_silver_vis_example.ipynb`](examples/saco_gold_silver_vis_example.ipynb)
+  : Shows a few examples from SA-Co image evaluation set.
+- [`saco_veval_vis_example.ipynb`](examples/saco_veval_vis_example.ipynb) :
+  Shows a few examples from SA-Co video evaluation set.
+
+There are additional notebooks in the examples directory that demonstrate how to
+use SAM 3 for interactive instance segmentation in images and videos (SAM 1/2
+tasks), or as a tool for an MLLM, and how to run evaluations on the SA-Co
+dataset.
+
+To run the Jupyter notebook examples:
+
+```bash
+# Make sure you have the notebooks dependencies installed
+pip install -e ".[notebooks]"
+
+# Start Jupyter notebook
+jupyter notebook examples/sam3_image_predictor_example.ipynb
+```
+
+## Model
+
+SAM 3 consists of a detector and a tracker that share a vision encoder. It has 848M parameters. The
+detector is a DETR-based model conditioned on text, geometry, and image
+exemplars. The tracker inherits the SAM 2 transformer encoder-decoder
+architecture, supporting video segmentation and interactive refinement.
+
+## Image Results
+
+<div align="center">
+<table style="min-width: 80%; border: 2px solid #ddd; border-collapse: collapse">
+  <thead>
+    <tr>
+      <th rowspan="3" style="border-right: 2px solid #ddd; padding: 12px 20px">Model</th>
+      <th colspan="3" style="text-align: center; border-right: 2px solid #ddd; padding: 12px 20px">Instance Segmentation</th>
+      <th colspan="5" style="text-align: center; padding: 12px 20px">Box Detection</th>
+    </tr>
+    <tr>
+      <th colspan="2" style="text-align: center; border-right: 1px solid #eee; padding: 12px 20px">LVIS</th>
+      <th style="text-align: center; border-right: 2px solid #ddd; padding: 12px 20px">SA-Co/Gold</th>
+      <th colspan="2" style="text-align: center; border-right: 1px solid #eee; padding: 12px 20px">LVIS</th>
+      <th colspan="2" style="text-align: center; border-right: 1px solid #eee; padding: 12px 20px">COCO</th>
+      <th style="text-align: center; padding: 12px 20px">SA-Co/Gold</th>
+    </tr>
+    <tr>
+      <th style="text-align: center; padding: 12px 20px">cgF1</th>
+      <th style="text-align: center; border-right: 1px solid #eee; padding: 12px 20px">AP</th>
+      <th style="text-align: center; border-right: 2px solid #ddd; padding: 12px 20px">cgF1</th>
+      <th style="text-align: center; padding: 12px 20px">cgF1</th>
+      <th style="text-align: center; border-right: 1px solid #eee; padding: 12px 20px">AP</th>
+      <th style="text-align: center; padding: 12px 20px">AP</th>
+      <th style="text-align: center; border-right: 1px solid #eee; padding: 12px 20px">AP<sub>o</sub>
+</th>
+      <th style="text-align: center; padding: 12px 20px">cgF1</th>
+    </tr>
+  </thead>
+  <tbody>
+    <tr>
+      <td style="border-right: 2px solid #ddd; padding: 10px 20px">Human</td>
+      <td style="text-align: center; padding: 10px 20px">-</td>
+      <td style="text-align: center; border-right: 1px solid #eee; padding: 10px 20px">-</td>
+      <td style="text-align: center; border-right: 2px solid #ddd; padding: 10px 20px">72.8</td>
+      <td style="text-align: center; padding: 10px 20px">-</td>
+      <td style="text-align: center; border-right: 1px solid #eee; padding: 10px 20px">-</td>
+      <td style="text-align: center; padding: 10px 20px">-</td>
+      <td style="text-align: center; border-right: 1px solid #eee; padding: 10px 20px">-</td>
+      <td style="text-align: center; padding: 10px 20px">74.0</td>
+    </tr>
+    <tr>
+      <td style="border-right: 2px solid #ddd; padding: 10px 20px">OWLv2*</td>
+      <td style="text-align: center; padding: 10px 20px; color: #999">29.3</td>
+      <td style="text-align: center; border-right: 1px solid #eee; padding: 10px 20px; color: #999">43.4</td>
+      <td style="text-align: center; border-right: 2px solid #ddd; padding: 10px 20px">24.6</td>
+      <td style="text-align: center; padding: 10px 20px; color: #999">30.2</td>
+      <td style="text-align: center; border-right: 1px solid #eee; padding: 10px 20px; color: #999">45.5</td>
+      <td style="text-align: center; padding: 10px 20px">46.1</td>
+      <td style="text-align: center; border-right: 1px solid #eee; padding: 10px 20px">23.9</td>
+      <td style="text-align: center; padding: 10px 20px">24.5</td>
+    </tr>
+    <tr>
+      <td style="border-right: 2px solid #ddd; padding: 10px 20px">DINO-X</td>
+      <td style="text-align: center; padding: 10px 20px">-</td>
+      <td style="text-align: center; border-right: 1px solid #eee; padding: 10px 20px">38.5</td>
+      <td style="text-align: center; border-right: 2px solid #ddd; padding: 10px 20px">21.3</td>
+      <td style="text-align: center; padding: 10px 20px">-</td>
+      <td style="text-align: center; border-right: 1px solid #eee; padding: 10px 20px">52.4</td>
+      <td style="text-align: center; padding: 10px 20px">56.0</td>
+      <td style="text-align: center; border-right: 1px solid #eee; padding: 10px 20px">-</td>
+      <td style="text-align: center; padding: 10px 20px">22.5</td>
+    </tr>
+    <tr>
+      <td style="border-right: 2px solid #ddd; padding: 10px 20px">Gemini 2.5</td>
+      <td style="text-align: center; padding: 10px 20px">13.4</td>
+      <td style="text-align: center; border-right: 1px solid #eee; padding: 10px 20px">-</td>
+      <td style="text-align: center; border-right: 2px solid #ddd; padding: 10px 20px">13.0</td>
+      <td style="text-align: center; padding: 10px 20px">16.1</td>
+      <td style="text-align: center; border-right: 1px solid #eee; padding: 10px 20px">-</td>
+      <td style="text-align: center; padding: 10px 20px">-</td>
+      <td style="text-align: center; border-right: 1px solid #eee; padding: 10px 20px">-</td>
+      <td style="text-align: center; padding: 10px 20px">14.4</td>
+    </tr>
+    <tr style="border-top: 2px solid #b19c9cff">
+      <td style="border-right: 2px solid #ddd; padding: 10px 20px">SAM 3</td>
+      <td style="text-align: center; padding: 10px 20px">37.2</td>
+      <td style="text-align: center; border-right: 1px solid #eee; padding: 10px 20px">48.5</td>
+      <td style="text-align: center; border-right: 2px solid #ddd; padding: 10px 20px">54.1</td>
+      <td style="text-align: center; padding: 10px 20px">40.6</td>
+      <td style="text-align: center; border-right: 1px solid #eee; padding: 10px 20px">53.6</td>
+      <td style="text-align: center; padding: 10px 20px">56.4</td>
+      <td style="text-align: center; border-right: 1px solid #eee; padding: 10px 20px">55.7</td>
+      <td style="text-align: center; padding: 10px 20px">55.7</td>
+    </tr>
+  </tbody>
+</table>
+
+<p style="text-align: center; margin-top: 10px; font-size: 0.9em; color: #ddd;">* Partially trained on LVIS, AP<sub>o</sub> refers to COCO-O accuracy</p>
+
+</div>
+
+## Video Results
+
+<div align="center">
+<table style="min-width: 80%; border: 2px solid #ddd; border-collapse: collapse">
+  <thead>
+    <tr>
+      <th rowspan="2" style="border-right: 2px solid #ddd; padding: 12px 20px">Model</th>
+      <th colspan="2" style="text-align: center; border-right: 1px solid #eee; padding: 12px 20px">SA-V test</th>
+      <th colspan="2" style="text-align: center; border-right: 1px solid #eee; padding: 12px 20px">YT-Temporal-1B test</th>
+      <th colspan="2" style="text-align: center; border-right: 1px solid #eee; padding: 12px 20px">SmartGlasses test</th>
+      <th style="text-align: center; border-right: 1px solid #eee; padding: 12px 20px">LVVIS test</th>
+      <th style="text-align: center; padding: 12px 20px">BURST test</th>
+    </tr>
+    <tr>
+      <th style="text-align: center; padding: 12px 20px">cgF1</th>
+      <th style="text-align: center; border-right: 1px solid #eee; padding: 12px 20px">pHOTA</th>
+      <th style="text-align: center; padding: 12px 20px">cgF1</th>
+      <th style="text-align: center; border-right: 1px solid #eee; padding: 12px 20px">pHOTA</th>
+      <th style="text-align: center; padding: 12px 20px">cgF1</th>
+      <th style="text-align: center; border-right: 1px solid #eee; padding: 12px 20px">pHOTA</th>
+      <th style="text-align: center; border-right: 1px solid #eee; padding: 12px 20px">mAP</th>
+      <th style="text-align: center; padding: 12px 20px">HOTA</th>
+    </tr>
+  </thead>
+  <tbody>
+    <tr>
+      <td style="border-right: 2px solid #ddd; padding: 10px 20px">Human</td>
+      <td style="text-align: center; padding: 10px 20px">53.1</td>
+      <td style="text-align: center; border-right: 1px solid #eee; padding: 10px 20px">70.5</td>
+      <td style="text-align: center; padding: 10px 20px">71.2</td>
+      <td style="text-align: center; border-right: 1px solid #eee; padding: 10px 20px">78.4</td>
+      <td style="text-align: center; padding: 10px 20px">58.5</td>
+      <td style="text-align: center; border-right: 1px solid #eee; padding: 10px 20px">72.3</td>
+      <td style="text-align: center; border-right: 1px solid #eee; padding: 10px 20px">-</td>
+      <td style="text-align: center; padding: 10px 20px">-</td>
+    </tr>
+    <tr style="border-top: 2px solid #b19c9cff">
+      <td style="border-right: 2px solid #ddd; padding: 10px 20px">SAM 3</td>
+      <td style="text-align: center; padding: 10px 20px">30.3</td>
+      <td style="text-align: center; border-right: 1px solid #eee; padding: 10px 20px">58.0</td>
+      <td style="text-align: center; padding: 10px 20px">50.8</td>
+      <td style="text-align: center; border-right: 1px solid #eee; padding: 10px 20px">69.9</td>
+      <td style="text-align: center; padding: 10px 20px">36.4</td>
+      <td style="text-align: center; border-right: 1px solid #eee; padding: 10px 20px">63.6</td>
+      <td style="text-align: center; border-right: 1px solid #eee; padding: 10px 20px">36.3</td>
+      <td style="text-align: center; padding: 10px 20px">44.5</td>
+    </tr>
+  </tbody>
+</table>
+</div>
+
+## SA-Co Dataset
+
+We release 2 image benchmarks, [SA-Co/Gold](scripts/eval/gold/README.md) and
+[SA-Co/Silver](scripts/eval/silver/README.md), and a video benchmark
+[SA-Co/VEval](scripts/eval/veval/README.md). The datasets contain images (or videos) with annotated noun phrases. Each image/video and noun phrase pair is annotated with instance masks and unique IDs of each object matching the phrase. Phrases that have no matching objects (negative prompts) have no masks, shown in red font in the figure. See the linked READMEs for more details on how to download and run evaluations on the datasets.
+
+* HuggingFace host: [SA-Co/Gold](https://huggingface.co/datasets/facebook/SACo-Gold), [SA-Co/Silver](https://huggingface.co/datasets/facebook/SACo-Silver) and [SA-Co/VEval](https://huggingface.co/datasets/facebook/SACo-VEval)
+* Roboflow host: [SA-Co/Gold](https://universe.roboflow.com/sa-co-gold), [SA-Co/Silver](https://universe.roboflow.com/sa-co-silver) and [SA-Co/VEval](https://universe.roboflow.com/sa-co-veval)
+
+![SA-Co dataset](assets/sa_co_dataset.jpg?raw=true)
+
+## Development
+
+To set up the development environment:
+
+```bash
+pip install -e ".[dev,train]"
+```
+
+To format the code:
+
+```bash
+ufmt format .
+```
+
+## Contributing
+
+See [contributing](CONTRIBUTING.md) and the
+[code of conduct](CODE_OF_CONDUCT.md).
+
+## License
+
+This project is licensed under the SAM License - see the [LICENSE](LICENSE) file
+for details.
+
+## Acknowledgements
+
+We would like to thank the following people for their contributions to the SAM 3 project: Alex He, Alexander Kirillov,
+Alyssa Newcomb, Ana Paula Kirschner Mofarrej, Andrea Madotto, Andrew Westbury, Ashley Gabriel, Azita Shokpour,
+Ben Samples, Bernie Huang, Carleigh Wood, Ching-Feng Yeh, Christian Puhrsch, Claudette Ward, Daniel Bolya,
+Daniel Li, Facundo Figueroa, Fazila Vhora, George Orlin, Hanzi Mao, Helen Klein, Hu Xu, Ida Cheng, Jake Kinney,
+Jiale Zhi, Jo Sampaio, Joel Schlosser, Justin Johnson, Kai Brown, Karen Bergan, Karla Martucci, Kenny Lehmann,
+Maddie Mintz, Mallika Malhotra, Matt Ward, Michelle Chan, Michelle Restrepo, Miranda Hartley, Muhammad Maaz,
+Nisha Deo, Peter Park, Phillip Thomas, Raghu Nayani, Rene Martinez Doehner, Robbie Adkins, Ross Girshik, Sasha
+Mitts, Shashank Jain, Spencer Whitehead, Ty Toledano, Valentin Gabeur, Vincent Cho, Vivian Lee, William Ngan,
+Xuehai He, Yael Yungster, Ziqi Pang, Ziyi Dou, Zoe Quake.
+
+<!-- ## Citing SAM 3
+
+If you use SAM 3 or the SA-Co dataset in your research, please use the following BibTeX entry.
+
+```bibtex
+TODO
+``` -->

detect_tools/sam3/README_TRAIN.md

@@ -0,0 +1,190 @@
+# Training
+
+This repository supports finetuning SAM3 models on custom datasets in multi-node setup or local execution. The training script is located at `sam3/train.py` and uses Hydra configuration management to handle complex training setups.
+
+
+## Installation
+
+```bash
+cd sam3
+pip install -e ".[train]"
+```
+
+### Training Script Usage
+
+The main training script is located at `sam3/train.py`. It uses Hydra configuration management to handle complex training setups.
+
+#### Basic Usage
+
+```bash
+# Example: Train on Roboflow dataset
+python sam3/train/train.py -c configs/roboflow_v100/roboflow_v100_full_ft_100_images.yaml
+# Example: Train on ODinW13 dataset
+python sam3/train/train.py -c configs/odinw13/odinw_text_only_train.yaml
+```
+Follow [`Roboflow 100-VL`](https://github.com/roboflow/rf100-vl/) to download the roboflow 100-vl datasets. Follow [`GLIP`](https://github.com/microsoft/GLIP) to download the ODinW datasets. The data folder should be organized as follows, and put your roboflow_vl_100_root and odinw_data_root in the job configs.
+```
+roboflow_vl_100_root:
+  13-lkc01
+    train
+    valid
+    test
+  2024-frc
+  actions
+  ...
+odinw_data_root:
+  AerialMaritimeDrone
+    large
+      train
+      valid
+      test
+  Aquarium
+  ...
+```
+
+#### Command Line Arguments
+
+The training script supports several command line arguments:
+
+```bash
+python sam3/train/train.py \
+    -c CONFIG_NAME \
+    [--use-cluster 0|1] \
+    [--partition PARTITION_NAME] \
+    [--account ACCOUNT_NAME] \
+    [--qos QOS_NAME] \
+    [--num-gpus NUM_GPUS] \
+    [--num-nodes NUM_NODES]
+```
+
+**Arguments:**
+- `-c, --config`: **Required.** Path to the configuration file (e.g., `sam3/train/configs/roboflow_v100_full_ft_100_images.yaml`)
+- `--use-cluster`: Whether to launch on a cluster (0: local, 1: cluster). Default: uses config setting
+- `--partition`: SLURM partition name for cluster execution
+- `--account`: SLURM account name for cluster execution
+- `--qos`: SLURM QOS (Quality of Service) setting
+- `--num-gpus`: Number of GPUs per node. Default: uses config setting
+- `--num-nodes`: Number of nodes for distributed training. Default: uses config setting
+
+#### Local Training Examples
+
+```bash
+# Single GPU training
+python sam3/train/train.py -c configs/roboflow_v100/roboflow_v100_full_ft_100_images.yaml --use-cluster 0 --num-gpus 1
+
+# Multi-GPU training on a single node
+python sam3/train/train.py -c configs/roboflow_v100/roboflow_v100_full_ft_100_images.yaml --use-cluster 0 --num-gpus 4
+
+# Force local execution even if config specifies GPUs
+python sam3/train/train.py -c configs/roboflow_v100/roboflow_v100_full_ft_100_images.yaml --use-cluster 0
+```
+
+#### Cluster Training Examples
+
+```bash
+# Basic cluster training with default settings from config
+python sam3/train/train.py -c configs/roboflow_v100/roboflow_v100_full_ft_100_images.yaml --use-cluster 1
+
+# Cluster training with specific SLURM settings
+python sam3/train/train.py -c configs/roboflow_v100/roboflow_v100_full_ft_100_images.yaml \
+    --use-cluster 1 \
+    --partition gpu_partition \
+    --account my_account \
+    --qos high_priority \
+    --num-gpus 8 \
+    --num-nodes 2
+```
+
+### Configuration Files
+
+Training configurations are stored in `sam3/train/configs/`. The configuration files use Hydra's YAML format and support:
+
+- **Dataset Configuration**: Data paths, transforms, and loading parameters
+- **Model Configuration**: Architecture settings, checkpoint paths, and model parameters
+- **Training Configuration**: Batch sizes, learning rates, optimization settings
+- **Launcher Configuration**: Distributed training and cluster settings
+- **Logging Configuration**: TensorBoard, experiment tracking, and output directories
+
+#### Key Configuration Sections
+
+```yaml
+# Paths to datasets and checkpoints
+paths:
+  bpe_path: /path/to/bpe/file
+  dataset_root: /path/to/dataset
+  experiment_log_dir: /path/to/logs
+
+# Launcher settings for local/cluster execution
+launcher:
+  num_nodes: 1
+  gpus_per_node: 2
+  experiment_log_dir: ${paths.experiment_log_dir}
+
+# Cluster execution settings
+submitit:
+  use_cluster: True
+  timeout_hour: 72
+  cpus_per_task: 10
+  partition: null
+  account: null
+```
+
+### Monitoring Training
+
+The training script automatically sets up logging and saves outputs to the experiment directory:
+
+```bash
+# Logs are saved to the experiment_log_dir specified in config
+experiment_log_dir/
+├── config.yaml              # Original configuration
+├── config_resolved.yaml     # Resolved configuration with all variables expanded
+├── checkpoints/             # Model checkpoints (if skip_checkpointing=False)
+├── tensorboard/             # TensorBoard logs
+├── logs/                    # Text logs
+└── submitit_logs/           # Cluster job logs (if using cluster)
+```
+
+You can monitor training progress using TensorBoard:
+
+```bash
+tensorboard --logdir /path/to/experiment_log_dir/tensorboard
+```
+
+### Job Arrays for Dataset Sweeps
+
+The Roboflow and ODinW configuration supports job arrays for training multiple models on different datasets:
+
+This feature is specifically enabled via,
+```yaml
+submitit:
+  job_array:
+    num_tasks: 100
+    task_index: 0
+```
+
+The configuration includes a complete list of 100 Roboflow supercategories, and the `submitit.job_array.task_index` automatically selects which dataset to use based on the array job index.
+
+```bash
+# Submit job array to train on different Roboflow datasets
+# The job array index selects which dataset from all_roboflow_supercategories
+python sam3/train/train.py -c configs/roboflow_v100/roboflow_v100_full_ft_100_images.yaml \
+    --use-cluster 1
+```
+
+### Reproduce ODinW13 10-shot results
+Running the following job will give the results on the ODinW13 seed 300, see `odinw_train.train_file: fewshot_train_shot10_seed300` in the config file.
+```bash
+# Example: Train on ODinW13 dataset
+python sam3/train/train.py -c configs/odinw13/odinw_text_only_train.yaml
+```
+Change `odinw_train.train_file` to `fewshot_train_shot10_seed30` and `fewshot_train_shot10_seed3` to get the results for the other two seeds. Final results are aggregated from the three seeds. Notice that a small number of jobs may diverge during training, in which case we just use the last checkpoint's result before it diverges.
+
+
+### Eval Script Usage
+With a similar setup as the training config, the training script `sam3/train.py` can also be used for evaluation, too, when setting `trainer.mode = val` in the job config. Run the following job will give the results on the zero-shot results on RF100-VL and ODinW13 datasets.
+```bash
+# Example: Evaluate on Roboflow dataset
+python sam3/train/train.py -c configs/roboflow_v100/roboflow_v100_eval.yaml
+# Example: Evaluate on ODinW13 dataset
+python sam3/train/train.py -c configs/odinw13/odinw_text_only.yaml
+```

detect_tools/sam3/pyproject.toml

@@ -0,0 +1,131 @@
+[build-system]
+requires = ["setuptools>=61", "wheel"]
+build-backend = "setuptools.build_meta"
+
+[project]
+name = "sam3"
+dynamic = ["version"]
+description = "SAM3 (Segment Anything Model 3) implementation"
+readme = "README.md"
+requires-python = ">=3.8"
+license = {file = "LICENSE"}
+authors = [
+    {name = "Meta AI Research"}
+]
+classifiers = [
+    "Development Status :: 4 - Beta",
+    "Intended Audience :: Science/Research",
+    "License :: OSI Approved :: MIT License",
+    "Programming Language :: Python :: 3",
+    "Programming Language :: Python :: 3.8",
+    "Programming Language :: Python :: 3.9",
+    "Programming Language :: Python :: 3.10",
+    "Programming Language :: Python :: 3.11",
+    "Programming Language :: Python :: 3.12",
+    "Topic :: Scientific/Engineering :: Artificial Intelligence",
+]
+dependencies = [
+    "timm>=1.0.17",
+    "numpy==1.26",
+    "tqdm",
+    "ftfy==6.1.1",
+    "regex",
+    "iopath>=0.1.10",
+    "typing_extensions",
+    "huggingface_hub",
+]
+
+[project.optional-dependencies]
+dev = [
+    "pytest",
+    "pytest-cov",
+    "black==24.2.0",
+    "ufmt==2.8.0",
+    "ruff-api==0.1.0",
+    "usort==1.0.2",
+    "gitpython==3.1.31",
+    "yt-dlp",
+    "pandas",
+    "opencv-python",
+    "pycocotools",
+    "numba",
+    "python-rapidjson",
+]
+notebooks = [
+    "matplotlib",
+    "jupyter",
+    "notebook",
+    "ipywidgets",
+    "ipycanvas",
+    "ipympl",
+    "pycocotools",
+    "decord",
+    "opencv-python",
+    "einops",
+    "scikit-image",
+    "scikit-learn",
+]
+train = [
+    "hydra-core",
+    "submitit",
+    "tensorboard",
+    "zstandard",
+    "scipy",
+    "torchmetrics",
+    "fvcore",
+    "fairscale",
+    "scikit-image",
+    "scikit-learn",
+]
+
+[project.urls]
+"Homepage" = "https://github.com/facebookresearch/sam3"
+"Bug Tracker" = "https://github.com/facebookresearch/sam3/issues"
+
+[tool.setuptools]
+packages = ["sam3", "sam3.model"]
+
+[tool.setuptools.dynamic]
+version = {attr = "sam3.__version__"}
+
+[tool.black]
+line-length = 88
+target-version = ['py38', 'py39', 'py310', 'py311', 'py312']
+include = '\.pyi?$'
+
+[tool.isort]
+profile = "black"
+multi_line_output = 3
+
+[tool.usort]
+first_party_detection = false
+
+[tool.ufmt]
+formatter = "ruff-api"
+
+[tool.mypy]
+python_version = "3.12"
+warn_return_any = true
+warn_unused_configs = true
+disallow_untyped_defs = true
+disallow_incomplete_defs = true
+
+[[tool.mypy.overrides]]
+module = [
+    "torch.*",
+    "torchvision.*",
+    "timm.*",
+    "numpy.*",
+    "PIL.*",
+    "tqdm.*",
+    "ftfy.*",
+    "regex.*",
+    "iopath.*",
+]
+ignore_missing_imports = true
+
+[tool.pytest.ini_options]
+testpaths = ["tests"]
+python_files = "test_*.py"
+python_classes = "Test*"
+python_functions = "test_*"

detect_tools/sam3/sam3/__init__.py

@@ -0,0 +1,7 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates. All Rights Reserved
+
+from .model_builder import build_sam3_image_model
+
+__version__ = "0.1.0"
+
+__all__ = ["build_sam3_image_model"]

detect_tools/sam3/sam3/logger.py

@@ -0,0 +1,54 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates. All Rights Reserved
+import logging
+import os
+
+LOG_LEVELS = {
+    "DEBUG": logging.DEBUG,
+    "INFO": logging.INFO,
+    "WARNING": logging.WARNING,
+    "ERROR": logging.ERROR,
+    "CRITICAL": logging.CRITICAL,
+}
+
+
+class ColoredFormatter(logging.Formatter):
+    """A command line formatter with different colors for each level."""
+
+    def __init__(self):
+        super().__init__()
+        reset = "\033[0m"
+        colors = {
+            logging.DEBUG: f"{reset}\033[36m",  # cyan,
+            logging.INFO: f"{reset}\033[32m",  # green
+            logging.WARNING: f"{reset}\033[33m",  # yellow
+            logging.ERROR: f"{reset}\033[31m",  # red
+            logging.CRITICAL: f"{reset}\033[35m",  # magenta
+        }
+        fmt_str = "{color}%(levelname)s %(asctime)s %(process)d %(filename)s:%(lineno)4d:{reset} %(message)s"
+        self.formatters = {
+            level: logging.Formatter(fmt_str.format(color=color, reset=reset))
+            for level, color in colors.items()
+        }
+        self.default_formatter = self.formatters[logging.INFO]
+
+    def format(self, record):
+        formatter = self.formatters.get(record.levelno, self.default_formatter)
+        return formatter.format(record)
+
+
+def get_logger(name, level=logging.INFO):
+    """A command line logger."""
+    if "LOG_LEVEL" in os.environ:
+        level = os.environ["LOG_LEVEL"].upper()
+        assert (
+            level in LOG_LEVELS
+        ), f"Invalid LOG_LEVEL: {level}, must be one of {list(LOG_LEVELS.keys())}"
+        level = LOG_LEVELS[level]
+    logger = logging.getLogger(name)
+    logger.setLevel(level)
+    logger.propagate = False
+    ch = logging.StreamHandler()
+    ch.setLevel(level)
+    ch.setFormatter(ColoredFormatter())
+    logger.addHandler(ch)
+    return logger

detect_tools/sam3/sam3/model/__init__.py

@@ -0,0 +1 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates. All Rights Reserved

detect_tools/sam3/sam3/model/act_ckpt_utils.py

@@ -0,0 +1,114 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates. All Rights Reserved
+
+import inspect
+from functools import wraps
+from typing import Callable, TypeVar, Union
+
+import torch
+import torch.nn as nn
+import torch.utils.checkpoint as checkpoint
+from torch.utils._pytree import tree_map_only
+
+# Type variables for better type hinting
+T = TypeVar("T")
+Module = TypeVar("Module", bound=nn.Module)
+
+
+def activation_ckpt_wrapper(module: Union[nn.Module, Callable]) -> Callable:
+    """
+    Wraps a given module to enable or disable activation checkpointing.
+
+    Activation checkpointing (gradient checkpointing) trades compute for memory by
+    recomputing intermediate activations during the backward pass instead of storing
+    them in memory during the forward pass.
+
+    When activation checkpointing is enabled, the wrapper expects only keyword arguments,
+    and it maps these to positional arguments based on the module's signature.
+
+    Args:
+        module: The module or function to wrap with activation checkpointing
+
+    Returns:
+        A wrapped callable that supports activation checkpointing
+
+    Usage:
+        The returned wrapper function can be called with the same arguments as the
+        original module, with an additional `act_ckpt_enable` keyword argument to control
+        activation checkpointing and optional `use_reentrant` parameter.
+
+    Example:
+        ```python
+        wrapped_module = activation_ckpt_wrapper(my_module)
+        output = wrapped_module(x=input_tensor, y=another_tensor, act_ckpt_enable=True)
+        ```
+    """
+
+    @wraps(module)
+    def act_ckpt_wrapper(
+        *args, act_ckpt_enable: bool = True, use_reentrant: bool = False, **kwargs
+    ):
+        if act_ckpt_enable:
+            if len(args) > 0:
+                raise ValueError(
+                    "This wrapper expects keyword arguments only when `act_ckpt_enable=True`"
+                )
+            # Get the signature of the target function/module
+            callable_fn = module.forward if isinstance(module, nn.Module) else module
+            sig = inspect.signature(callable_fn)
+            # Create a mapping of parameter names to their default values
+            param_defaults = {
+                name: param.default for name, param in sig.parameters.items()
+            }
+            args = []
+            for p_name in param_defaults.keys():
+                if p_name in kwargs:
+                    args.append(kwargs.pop(p_name))
+                elif param_defaults[p_name] is not inspect.Parameter.empty:
+                    # Set arg to default value if it's not in kwargs. Useful for primitive types or args that default to None
+                    args.append(param_defaults[p_name])
+                elif (
+                    sig.parameters[p_name].kind is not inspect.Parameter.VAR_KEYWORD
+                ):  # Skip **kwargs parameter
+                    raise ValueError(f"Missing positional argument: {p_name}")
+
+            # Scan remaining kwargs for torch.Tensor
+            remaining_keys = list(kwargs.keys())
+            for key in remaining_keys:
+                if isinstance(kwargs[key], torch.Tensor):
+                    # Remove the tensor from kwargs, assuming it's not required by the module.
+                    # If it is required, the module's signature should be modified to accept it as a positional or keyword argument.
+                    kwargs[key] = "_REMOVED_BY_ACT_CKPT_WRAPPER_"
+
+            ret = checkpoint.checkpoint(
+                module, *args, use_reentrant=use_reentrant, **kwargs
+            )
+        else:
+            ret = module(*args, **kwargs)
+
+        return ret
+
+    return act_ckpt_wrapper
+
+
+def clone_output_wrapper(f: Callable[..., T]) -> Callable[..., T]:
+    """
+    Clone the CUDA output tensors of a function to avoid in-place operations.
+
+    This wrapper is useful when working with torch.compile to prevent errors
+    related to in-place operations on tensors.
+
+    Args:
+        f: The function whose CUDA tensor outputs should be cloned
+
+    Returns:
+        A wrapped function that clones any CUDA tensor outputs
+    """
+
+    @wraps(f)
+    def wrapped(*args, **kwargs):
+        outputs = f(*args, **kwargs)
+        return tree_map_only(
+            torch.Tensor, lambda t: t.clone() if t.is_cuda else t, outputs
+        )
+
+    return wrapped

detect_tools/sam3/sam3/model/box_ops.py

@@ -0,0 +1,217 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates. All Rights Reserved
+"""
+Utilities for bounding box manipulation and GIoU.
+"""
+
+from typing import Tuple
+
+import torch
+
+
+def box_cxcywh_to_xyxy(x):
+    x_c, y_c, w, h = x.unbind(-1)
+    b = [(x_c - 0.5 * w), (y_c - 0.5 * h), (x_c + 0.5 * w), (y_c + 0.5 * h)]
+    return torch.stack(b, dim=-1)
+
+
+def box_cxcywh_to_xywh(x):
+    x_c, y_c, w, h = x.unbind(-1)
+    b = [(x_c - 0.5 * w), (y_c - 0.5 * h), (w), (h)]
+    return torch.stack(b, dim=-1)
+
+
+def box_xywh_to_xyxy(x):
+    x, y, w, h = x.unbind(-1)
+    b = [(x), (y), (x + w), (y + h)]
+    return torch.stack(b, dim=-1)
+
+
+def box_xywh_to_cxcywh(x):
+    x, y, w, h = x.unbind(-1)
+    b = [(x + 0.5 * w), (y + 0.5 * h), (w), (h)]
+    return torch.stack(b, dim=-1)
+
+
+def box_xyxy_to_xywh(x):
+    x, y, X, Y = x.unbind(-1)
+    b = [(x), (y), (X - x), (Y - y)]
+    return torch.stack(b, dim=-1)
+
+
+def box_xyxy_to_cxcywh(x):
+    x0, y0, x1, y1 = x.unbind(-1)
+    b = [(x0 + x1) / 2, (y0 + y1) / 2, (x1 - x0), (y1 - y0)]
+    return torch.stack(b, dim=-1)
+
+
+def box_area(boxes):
+    """
+    Batched version of box area. Boxes should be in [x0, y0, x1, y1] format.
+
+    Inputs:
+    - boxes: Tensor of shape (..., 4)
+
+    Returns:
+    - areas: Tensor of shape (...,)
+    """
+    x0, y0, x1, y1 = boxes.unbind(-1)
+    return (x1 - x0) * (y1 - y0)
+
+
+def masks_to_boxes(masks):
+    """Compute the bounding boxes around the provided masks
+
+    The masks should be in format [N, H, W] where N is the number of masks, (H, W) are the spatial dimensions.
+
+    Returns a [N, 4] tensors, with the boxes in xyxy format
+    """
+    if masks.numel() == 0:
+        return torch.zeros((0, 4), device=masks.device)
+
+    h, w = masks.shape[-2:]
+
+    y = torch.arange(0, h, dtype=torch.float, device=masks.device)
+    x = torch.arange(0, w, dtype=torch.float, device=masks.device)
+    y, x = torch.meshgrid(y, x)
+
+    x_mask = masks * x.unsqueeze(0)
+    x_max = x_mask.flatten(1).max(-1)[0] + 1
+    x_min = x_mask.masked_fill(~(masks.bool()), 1e8).flatten(1).min(-1)[0]
+
+    y_mask = masks * y.unsqueeze(0)
+    y_max = y_mask.flatten(1).max(-1)[0] + 1
+    y_min = y_mask.masked_fill(~(masks.bool()), 1e8).flatten(1).min(-1)[0]
+
+    boxes = torch.stack([x_min, y_min, x_max, y_max], 1)
+    # Invalidate boxes corresponding to empty masks.
+    boxes = boxes * masks.flatten(-2).any(-1)
+    return boxes
+
+
+def box_iou(boxes1, boxes2):
+    """
+    Batched version of box_iou. Boxes should be in [x0, y0, x1, y1] format.
+
+    Inputs:
+    - boxes1: Tensor of shape (..., N, 4)
+    - boxes2: Tensor of shape (..., M, 4)
+
+    Returns:
+    - iou, union: Tensors of shape (..., N, M)
+    """
+    area1 = box_area(boxes1)
+    area2 = box_area(boxes2)
+
+    # boxes1: (..., N, 4) -> (..., N, 1, 2)
+    # boxes2: (..., M, 4) -> (..., 1, M, 2)
+    lt = torch.max(boxes1[..., :, None, :2], boxes2[..., None, :, :2])
+    rb = torch.min(boxes1[..., :, None, 2:], boxes2[..., None, :, 2:])
+
+    wh = (rb - lt).clamp(min=0)  # (..., N, M, 2)
+    inter = wh[..., 0] * wh[..., 1]  # (..., N, M)
+
+    union = area1[..., None] + area2[..., None, :] - inter
+
+    iou = inter / union
+    return iou, union
+
+
+def generalized_box_iou(boxes1, boxes2):
+    """
+    Batched version of Generalized IoU from https://giou.stanford.edu/
+
+    Boxes should be in [x0, y0, x1, y1] format
+
+    Inputs:
+    - boxes1: Tensor of shape (..., N, 4)
+    - boxes2: Tensor of shape (..., M, 4)
+
+    Returns:
+    - giou: Tensor of shape (..., N, M)
+    """
+    iou, union = box_iou(boxes1, boxes2)
+
+    # boxes1: (..., N, 4) -> (..., N, 1, 2)
+    # boxes2: (..., M, 4) -> (..., 1, M, 2)
+    lt = torch.min(boxes1[..., :, None, :2], boxes2[..., None, :, :2])
+    rb = torch.max(boxes1[..., :, None, 2:], boxes2[..., None, :, 2:])
+
+    wh = (rb - lt).clamp(min=0)  # (..., N, M, 2)
+    area = wh[..., 0] * wh[..., 1]  # (..., N, M)
+
+    return iou - (area - union) / area
+
+
+@torch.jit.script
+def fast_diag_generalized_box_iou(boxes1, boxes2):
+    assert len(boxes1) == len(boxes2)
+    box1_xy = boxes1[:, 2:]
+    box1_XY = boxes1[:, :2]
+    box2_xy = boxes2[:, 2:]
+    box2_XY = boxes2[:, :2]
+    # assert (box1_xy >= box1_XY).all()
+    # assert (box2_xy >= box2_XY).all()
+    area1 = (box1_xy - box1_XY).prod(-1)
+    area2 = (box2_xy - box2_XY).prod(-1)
+
+    lt = torch.max(box1_XY, box2_XY)  # [N,2]
+    lt2 = torch.min(box1_XY, box2_XY)
+    rb = torch.min(box1_xy, box2_xy)  # [N,2]
+    rb2 = torch.max(box1_xy, box2_xy)
+
+    inter = (rb - lt).clamp(min=0).prod(-1)
+    tot_area = (rb2 - lt2).clamp(min=0).prod(-1)
+
+    union = area1 + area2 - inter
+
+    iou = inter / union
+
+    return iou - (tot_area - union) / tot_area
+
+
+@torch.jit.script
+def fast_diag_box_iou(boxes1, boxes2):
+    assert len(boxes1) == len(boxes2)
+    box1_xy = boxes1[:, 2:]
+    box1_XY = boxes1[:, :2]
+    box2_xy = boxes2[:, 2:]
+    box2_XY = boxes2[:, :2]
+    # assert (box1_xy >= box1_XY).all()
+    # assert (box2_xy >= box2_XY).all()
+    area1 = (box1_xy - box1_XY).prod(-1)
+    area2 = (box2_xy - box2_XY).prod(-1)
+
+    lt = torch.max(box1_XY, box2_XY)  # [N,2]
+    rb = torch.min(box1_xy, box2_xy)  # [N,2]
+
+    inter = (rb - lt).clamp(min=0).prod(-1)
+
+    union = area1 + area2 - inter
+
+    iou = inter / union
+
+    return iou
+
+
+def box_xywh_inter_union(
+    boxes1: torch.Tensor, boxes2: torch.Tensor
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    # Asuumes boxes in xywh format
+    assert boxes1.size(-1) == 4 and boxes2.size(-1) == 4
+    boxes1 = box_xywh_to_xyxy(boxes1)
+    boxes2 = box_xywh_to_xyxy(boxes2)
+    box1_tl_xy = boxes1[..., :2]
+    box1_br_xy = boxes1[..., 2:]
+    box2_tl_xy = boxes2[..., :2]
+    box2_br_xy = boxes2[..., 2:]
+    area1 = (box1_br_xy - box1_tl_xy).prod(-1)
+    area2 = (box2_br_xy - box2_tl_xy).prod(-1)
+
+    assert (area1 >= 0).all() and (area2 >= 0).all()
+    tl = torch.max(box1_tl_xy, box2_tl_xy)
+    br = torch.min(box1_br_xy, box2_br_xy)
+
+    inter = (br - tl).clamp(min=0).prod(-1)
+    union = area1 + area2 - inter
+
+    return inter, union

detect_tools/sam3/sam3/model/data_misc.py

@@ -0,0 +1,209 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates. All Rights Reserved
+"""
+Misc functions, including distributed helpers.
+"""
+
+import collections
+import re
+
+from dataclasses import dataclass, field as field_ptr_behaviour, fields, is_dataclass
+from typing import Any, get_args, get_origin, List, Mapping, Optional, Sequence, Union
+
+import torch
+
+
+MyTensor = Union[torch.Tensor, List[Any]]
+
+
+def interpolate(
+    input, size=None, scale_factor=None, mode="nearest", align_corners=None
+):
+    # type: (Tensor, Optional[List[int]], Optional[float], str, Optional[bool]) -> Tensor
+    """
+    Equivalent to nn.functional.interpolate, but with support for empty channel sizes.
+    """
+    if input.numel() > 0:
+        return torch.nn.functional.interpolate(
+            input, size, scale_factor, mode, align_corners
+        )
+
+    assert (
+        input.shape[0] != 0 or input.shape[1] != 0
+    ), "At least one of the two first dimensions must be non zero"
+
+    if input.shape[1] == 0:
+        # Pytorch doesn't support null dimension on the channel dimension, so we transpose to fake a null batch dim
+        return torch.nn.functional.interpolate(
+            input.transpose(0, 1), size, scale_factor, mode, align_corners
+        ).transpose(0, 1)
+
+    # empty batch dimension is now supported in pytorch
+    return torch.nn.functional.interpolate(
+        input, size, scale_factor, mode, align_corners
+    )
+
+
+@dataclass
+class BatchedPointer:
+    stage_ids: MyTensor
+    stage_ids__type = torch.long
+    query_ids: MyTensor
+    query_ids__type = torch.long
+    object_ids: MyTensor
+    object_ids__type = torch.long
+    ptr_mask: MyTensor
+    ptr_mask__type = torch.bool
+    ptr_types: MyTensor
+    ptr_types__type = torch.long
+
+
+@dataclass
+class FindStage:
+    img_ids: MyTensor
+    img_ids__type = torch.long
+    text_ids: MyTensor
+    text_ids__type = torch.long
+
+    input_boxes: MyTensor
+    input_boxes__type = torch.float
+    input_boxes_mask: MyTensor
+    input_boxes_mask__type = torch.bool
+    input_boxes_label: MyTensor
+    input_boxes_label__type = torch.long
+
+    input_points: MyTensor
+    input_points__type = torch.float
+    input_points_mask: MyTensor
+    input_points_mask__type = torch.bool
+
+    # We track the object ids referred to by this query.
+    # This is beneficial for tracking in videos without the need for pointers.
+    object_ids: Optional[List[List]] = None  # List of objects per query
+
+
+@dataclass
+class BatchedFindTarget:
+    # The number of boxes in each find query
+    num_boxes: MyTensor
+    num_boxes__type = torch.long
+
+    # Target boxes in normalized CxCywh format
+    boxes: MyTensor
+    boxes__type = torch.float
+    # Target boxes in normalized CxCywh format but in padded representation
+    # as used in BinaryHungarianMatcherV2 (unlike the packed ones in `boxes`)
+    boxes_padded: MyTensor
+    boxes_padded__type = torch.float
+
+    # For hybrid matching, we repeat the boxes
+    repeated_boxes: MyTensor
+    repeated_boxes__type = torch.float
+
+    # Target Segmentation masks
+    segments: Optional[MyTensor]
+    segments__type = torch.bool
+
+    # Target Semantic Segmentation masks
+    semantic_segments: Optional[MyTensor]
+    semantic_segments__type = torch.bool
+
+    is_valid_segment: Optional[MyTensor]
+    is_valid_segment__type = torch.bool
+
+    # Whether annotations are exhaustive for each query
+    is_exhaustive: MyTensor
+    is_exhaustive__type = torch.bool
+
+    # The object id for each ground-truth box, in both packed and padded representations
+    object_ids: MyTensor
+    object_ids__type = torch.long
+    object_ids_padded: MyTensor
+    object_ids_padded__type = torch.long
+
+
+@dataclass
+class BatchedInferenceMetadata:
+    """All metadata required to post-process a find stage"""
+
+    # Coco id that corresponds to the "image" for evaluation by the coco evaluator
+    coco_image_id: MyTensor
+    coco_image_id__type = torch.long
+
+    # id in the original dataset, such that we can use the original evaluator
+    original_image_id: MyTensor
+    original_image_id__type = torch.long
+
+    # Original category id (if we want to use the original evaluator)
+    original_category_id: MyTensor
+    original_category_id__type = torch.int
+
+    # Size of the raw image (height, width)
+    original_size: MyTensor
+    original_size__type = torch.long
+
+    # id of the object in the media (track_id for a video)
+    object_id: MyTensor
+    object_id__type = torch.long
+
+    # index of the frame in the media (0 in the case of a single-frame media)
+    frame_index: MyTensor
+    frame_index__type = torch.long
+
+    # Adding for relations inference
+    # get_text_input: List[Optional[str]]
+
+    # Adding for TA conditional inference
+    is_conditioning_only: List[Optional[bool]]
+
+
+@dataclass
+class BatchedDatapoint:
+    img_batch: torch.Tensor
+    find_text_batch: List[str]
+    find_inputs: List[FindStage]
+    find_targets: List[BatchedFindTarget]
+    find_metadatas: List[BatchedInferenceMetadata]
+    raw_images: Optional[List[Any]] = None
+
+
+def convert_my_tensors(obj):
+    def is_optional_field(field) -> bool:
+        return get_origin(field) is Union and type(None) in get_args(field)
+
+    for field in fields(obj):
+        if is_dataclass(getattr(obj, field.name)):
+            convert_my_tensors(getattr(obj, field.name))
+            continue
+
+        field_type = field.type
+        if is_optional_field(field.type):
+            field_type = Union[get_args(field.type)[:-1]]  # Get the Optional field type
+
+        if field_type != MyTensor or getattr(obj, field.name) is None:
+            continue
+
+        elif len(getattr(obj, field.name)) and isinstance(
+            getattr(obj, field.name)[0], torch.Tensor
+        ):
+            stack_dim = 0
+            if field.name in [
+                "input_boxes",
+                "input_boxes_label",
+            ]:
+                stack_dim = 1
+            setattr(
+                obj,
+                field.name,
+                torch.stack(getattr(obj, field.name), dim=stack_dim).to(
+                    getattr(obj, field.name + "__type")
+                ),
+            )
+        else:
+            setattr(
+                obj,
+                field.name,
+                torch.as_tensor(
+                    getattr(obj, field.name), dtype=getattr(obj, field.name + "__type")
+                ),
+            )
+    return obj

detect_tools/sam3/sam3/model/decoder.py

@@ -0,0 +1,956 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates. All Rights Reserved
+"""
+Transformer decoder.
+Inspired from Pytorch's version, adds the pre-norm variant
+"""
+
+from typing import Any, Dict, List, Optional
+
+import numpy as np
+
+import torch
+
+from sam3.sam.transformer import RoPEAttention
+
+from torch import nn, Tensor
+from torchvision.ops.roi_align import RoIAlign
+
+from .act_ckpt_utils import activation_ckpt_wrapper
+
+from .box_ops import box_cxcywh_to_xyxy
+
+from .model_misc import (
+    gen_sineembed_for_position,
+    get_activation_fn,
+    get_clones,
+    inverse_sigmoid,
+    MLP,
+)
+
+
+class TransformerDecoderLayer(nn.Module):
+    def __init__(
+        self,
+        activation: str,
+        d_model: int,
+        dim_feedforward: int,
+        dropout: float,
+        cross_attention: nn.Module,
+        n_heads: int,
+        use_text_cross_attention: bool = False,
+    ):
+        super().__init__()
+
+        # cross attention
+        self.cross_attn = cross_attention
+        self.dropout1 = nn.Dropout(dropout) if dropout > 0 else nn.Identity()
+        self.norm1 = nn.LayerNorm(d_model)
+
+        # cross attention text
+        self.use_text_cross_attention = use_text_cross_attention
+        if use_text_cross_attention:
+            self.ca_text = nn.MultiheadAttention(d_model, n_heads, dropout=dropout)
+            self.catext_dropout = nn.Dropout(dropout) if dropout > 0 else nn.Identity()
+            self.catext_norm = nn.LayerNorm(d_model)
+
+        # self attention
+        self.self_attn = nn.MultiheadAttention(d_model, n_heads, dropout=dropout)
+        self.dropout2 = nn.Dropout(dropout) if dropout > 0 else nn.Identity()
+        self.norm2 = nn.LayerNorm(d_model)
+
+        # ffn
+        self.linear1 = nn.Linear(d_model, dim_feedforward)
+        self.activation = get_activation_fn(activation)
+        self.dropout3 = nn.Dropout(dropout) if dropout > 0 else nn.Identity()
+        self.linear2 = nn.Linear(dim_feedforward, d_model)
+        self.dropout4 = nn.Dropout(dropout) if dropout > 0 else nn.Identity()
+        self.norm3 = nn.LayerNorm(d_model)
+
+    @staticmethod
+    def with_pos_embed(tensor, pos):
+        return tensor if pos is None else tensor + pos
+
+    def forward_ffn(self, tgt):
+        with torch.amp.autocast(device_type="cuda", enabled=False):
+            tgt2 = self.linear2(self.dropout3(self.activation(self.linear1(tgt))))
+        tgt = tgt + self.dropout4(tgt2)
+        tgt = self.norm3(tgt)
+        return tgt
+
+    def forward(
+        self,
+        # for tgt
+        tgt: Optional[Tensor],  # nq, bs, d_model
+        tgt_query_pos: Optional[Tensor] = None,  # pos for query. MLP(Sine(pos))
+        tgt_query_sine_embed: Optional[Tensor] = None,  # pos for query. Sine(pos)
+        tgt_key_padding_mask: Optional[Tensor] = None,
+        tgt_reference_points: Optional[Tensor] = None,  # nq, bs, 4
+        memory_text: Optional[Tensor] = None,  # num_token, bs, d_model
+        text_attention_mask: Optional[Tensor] = None,  # bs, num_token
+        # for memory
+        memory: Optional[Tensor] = None,  # hw, bs, d_model
+        memory_key_padding_mask: Optional[Tensor] = None,
+        memory_level_start_index: Optional[Tensor] = None,  # num_levels
+        memory_spatial_shapes: Optional[Tensor] = None,  # bs, num_levels, 2
+        memory_pos: Optional[Tensor] = None,  # pos for memory
+        # sa
+        self_attn_mask: Optional[Tensor] = None,  # mask used for self-attention
+        cross_attn_mask: Optional[Tensor] = None,  # mask used for cross-attention
+        # dac
+        dac=False,
+        dac_use_selfatt_ln=True,
+        presence_token=None,
+        # skip inside deformable attn
+        identity=0.0,
+        **kwargs,  # additional kwargs for compatibility
+    ):
+        """
+        Input:
+            - tgt/tgt_query_pos: nq, bs, d_model
+            -
+        """
+        # self attention
+        if self.self_attn is not None:
+            if dac:
+                # we only apply self attention to the first half of the queries
+                assert tgt.shape[0] % 2 == 0
+                num_o2o_queries = tgt.shape[0] // 2
+                tgt_o2o = tgt[:num_o2o_queries]
+                tgt_query_pos_o2o = tgt_query_pos[:num_o2o_queries]
+                tgt_o2m = tgt[num_o2o_queries:]
+            else:
+                tgt_o2o = tgt
+                tgt_query_pos_o2o = tgt_query_pos
+
+            if presence_token is not None:
+                tgt_o2o = torch.cat([presence_token, tgt_o2o], dim=0)
+                tgt_query_pos_o2o = torch.cat(
+                    [torch.zeros_like(presence_token), tgt_query_pos_o2o], dim=0
+                )
+                tgt_query_pos = torch.cat(
+                    [torch.zeros_like(presence_token), tgt_query_pos], dim=0
+                )
+
+            q = k = self.with_pos_embed(tgt_o2o, tgt_query_pos_o2o)
+            tgt2 = self.self_attn(q, k, tgt_o2o, attn_mask=self_attn_mask)[0]
+            tgt_o2o = tgt_o2o + self.dropout2(tgt2)
+            if dac:
+                if not dac_use_selfatt_ln:
+                    tgt_o2o = self.norm2(tgt_o2o)
+                tgt = torch.cat((tgt_o2o, tgt_o2m), dim=0)  # Recombine
+                if dac_use_selfatt_ln:
+                    tgt = self.norm2(tgt)
+            else:
+                tgt = tgt_o2o
+                tgt = self.norm2(tgt)
+
+        if self.use_text_cross_attention:
+            tgt2 = self.ca_text(
+                self.with_pos_embed(tgt, tgt_query_pos),
+                memory_text,
+                memory_text,
+                key_padding_mask=text_attention_mask,
+            )[0]
+            tgt = tgt + self.catext_dropout(tgt2)
+            tgt = self.catext_norm(tgt)
+
+        if presence_token is not None:
+            presence_token_mask = torch.zeros_like(cross_attn_mask[:, :1, :])
+            cross_attn_mask = torch.cat(
+                [presence_token_mask, cross_attn_mask], dim=1
+            )  # (bs*nheads, 1+nq, hw)
+
+        # Cross attention to image
+        tgt2 = self.cross_attn(
+            query=self.with_pos_embed(tgt, tgt_query_pos),
+            key=self.with_pos_embed(memory, memory_pos),
+            value=memory,
+            attn_mask=cross_attn_mask,
+            key_padding_mask=(
+                memory_key_padding_mask.transpose(0, 1)
+                if memory_key_padding_mask is not None
+                else None
+            ),
+        )[0]
+
+        tgt = tgt + self.dropout1(tgt2)
+        tgt = self.norm1(tgt)
+
+        # ffn
+        tgt = self.forward_ffn(tgt)
+
+        presence_token_out = None
+        if presence_token is not None:
+            presence_token_out = tgt[:1]
+            tgt = tgt[1:]
+
+        return tgt, presence_token_out
+
+
+class TransformerDecoder(nn.Module):
+    def __init__(
+        self,
+        d_model: int,
+        frozen: bool,
+        interaction_layer,
+        layer,
+        num_layers: int,
+        num_queries: int,
+        return_intermediate: bool,
+        box_refine: bool = False,
+        num_o2m_queries: int = 0,
+        dac: bool = False,
+        boxRPB: str = "none",
+        # Experimental: An object query for SAM 2 tasks
+        instance_query: bool = False,
+        # Defines the number of additional instance queries,
+        # 1 or 4 are the most likely for single vs multi mask support
+        num_instances: int = 1,  # Irrelevant if instance_query is False
+        dac_use_selfatt_ln: bool = True,
+        use_act_checkpoint: bool = False,
+        compile_mode=None,
+        presence_token: bool = False,
+        clamp_presence_logits: bool = True,
+        clamp_presence_logit_max_val: float = 10.0,
+        use_normed_output_consistently: bool = True,
+        separate_box_head_instance: bool = False,
+        separate_norm_instance: bool = False,
+        resolution: Optional[int] = None,
+        stride: Optional[int] = None,
+    ):
+        super().__init__()
+        self.d_model = d_model
+        self.layers = get_clones(layer, num_layers)
+        self.fine_layers = (
+            get_clones(interaction_layer, num_layers)
+            if interaction_layer is not None
+            else [None] * num_layers
+        )
+        self.num_layers = num_layers
+        self.num_queries = num_queries
+        self.dac = dac
+        if dac:
+            self.num_o2m_queries = num_queries
+            tot_num_queries = num_queries
+        else:
+            self.num_o2m_queries = num_o2m_queries
+            tot_num_queries = num_queries + num_o2m_queries
+        self.norm = nn.LayerNorm(d_model)
+        self.return_intermediate = return_intermediate
+        self.bbox_embed = MLP(d_model, d_model, 4, 3)
+        self.query_embed = nn.Embedding(tot_num_queries, d_model)
+        self.instance_query_embed = None
+        self.instance_query_reference_points = None
+        self.use_instance_query = instance_query
+        self.num_instances = num_instances
+        self.use_normed_output_consistently = use_normed_output_consistently
+
+        self.instance_norm = nn.LayerNorm(d_model) if separate_norm_instance else None
+        self.instance_bbox_embed = None
+        if separate_box_head_instance:
+            self.instance_bbox_embed = MLP(d_model, d_model, 4, 3)
+        if instance_query:
+            self.instance_query_embed = nn.Embedding(num_instances, d_model)
+        self.box_refine = box_refine
+        if box_refine:
+            nn.init.constant_(self.bbox_embed.layers[-1].weight.data, 0)
+            nn.init.constant_(self.bbox_embed.layers[-1].bias.data, 0)
+
+            self.reference_points = nn.Embedding(num_queries, 4)
+            if instance_query:
+                self.instance_reference_points = nn.Embedding(num_instances, 4)
+
+        assert boxRPB in ["none", "log", "linear", "both"]
+        self.boxRPB = boxRPB
+        if boxRPB != "none":
+            try:
+                nheads = self.layers[0].cross_attn_image.num_heads
+            except AttributeError:
+                nheads = self.layers[0].cross_attn.num_heads
+
+            n_input = 4 if boxRPB == "both" else 2
+            self.boxRPB_embed_x = MLP(n_input, d_model, nheads, 2)
+            self.boxRPB_embed_y = MLP(n_input, d_model, nheads, 2)
+            self.compilable_cord_cache = None
+            self.compilable_stored_size = None
+            self.coord_cache = {}
+
+            if resolution is not None and stride is not None:
+                feat_size = resolution // stride
+                coords_h, coords_w = self._get_coords(
+                    feat_size, feat_size, device="cuda"
+                )
+                self.compilable_cord_cache = (coords_h, coords_w)
+                self.compilable_stored_size = (feat_size, feat_size)
+
+        self.roi_pooler = (
+            RoIAlign(output_size=7, spatial_scale=1, sampling_ratio=-1, aligned=True)
+            if interaction_layer is not None
+            else None
+        )
+        if frozen:
+            for p in self.parameters():
+                p.requires_grad_(False)
+
+        self.presence_token = None
+        self.clamp_presence_logits = clamp_presence_logits
+        self.clamp_presence_logit_max_val = clamp_presence_logit_max_val
+        if presence_token:
+            self.presence_token = nn.Embedding(1, d_model)
+            self.presence_token_head = MLP(d_model, d_model, 1, 3)
+            self.presence_token_out_norm = nn.LayerNorm(d_model)
+
+        self.ref_point_head = MLP(2 * self.d_model, self.d_model, self.d_model, 2)
+        self.dac_use_selfatt_ln = dac_use_selfatt_ln
+        self.use_act_checkpoint = use_act_checkpoint
+
+        nn.init.normal_(self.query_embed.weight.data)
+        if self.instance_query_embed is not None:
+            nn.init.normal_(self.instance_query_embed.weight.data)
+
+        assert self.roi_pooler is None
+        assert self.return_intermediate, "support return_intermediate only"
+        assert self.box_refine, "support box refine only"
+
+        self.compile_mode = compile_mode
+        self.compiled = False
+        # We defer compilation till after the first forward, to first warm-up the boxRPB cache
+
+        # assign layer index to each layer so that some layers can decide what to do
+        # based on which layer index they are (e.g. cross attention to memory bank only
+        # in selected layers)
+        for layer_idx, layer in enumerate(self.layers):
+            layer.layer_idx = layer_idx
+
+    @staticmethod
+    def _get_coords(H, W, device):
+        coords_h = torch.arange(0, H, device=device, dtype=torch.float32) / H
+        coords_w = torch.arange(0, W, device=device, dtype=torch.float32) / W
+        return coords_h, coords_w
+
+    def _get_rpb_matrix(self, reference_boxes, feat_size):
+        H, W = feat_size
+        boxes_xyxy = box_cxcywh_to_xyxy(reference_boxes).transpose(0, 1)
+        bs, num_queries, _ = boxes_xyxy.shape
+        if self.compilable_cord_cache is None:
+            self.compilable_cord_cache = self._get_coords(H, W, reference_boxes.device)
+            self.compilable_stored_size = (H, W)
+
+        if torch.compiler.is_dynamo_compiling() or self.compilable_stored_size == (
+            H,
+            W,
+        ):
+            # good, hitting the cache, will be compilable
+            coords_h, coords_w = self.compilable_cord_cache
+        else:
+            # cache miss, will create compilation issue
+            # In case we're not compiling, we'll still rely on the dict-based cache
+            if feat_size not in self.coord_cache:
+                self.coord_cache[feat_size] = self._get_coords(
+                    H, W, reference_boxes.device
+                )
+            coords_h, coords_w = self.coord_cache[feat_size]
+
+            assert coords_h.shape == (H,)
+            assert coords_w.shape == (W,)
+
+        deltas_y = coords_h.view(1, -1, 1) - boxes_xyxy.reshape(-1, 1, 4)[:, :, 1:4:2]
+        deltas_y = deltas_y.view(bs, num_queries, -1, 2)
+        deltas_x = coords_w.view(1, -1, 1) - boxes_xyxy.reshape(-1, 1, 4)[:, :, 0:3:2]
+        deltas_x = deltas_x.view(bs, num_queries, -1, 2)
+
+        if self.boxRPB in ["log", "both"]:
+            deltas_x_log = deltas_x * 8  # normalize to -8, 8
+            deltas_x_log = (
+                torch.sign(deltas_x_log)
+                * torch.log2(torch.abs(deltas_x_log) + 1.0)
+                / np.log2(8)
+            )
+
+            deltas_y_log = deltas_y * 8  # normalize to -8, 8
+            deltas_y_log = (
+                torch.sign(deltas_y_log)
+                * torch.log2(torch.abs(deltas_y_log) + 1.0)
+                / np.log2(8)
+            )
+            if self.boxRPB == "log":
+                deltas_x = deltas_x_log
+                deltas_y = deltas_y_log
+            else:
+                deltas_x = torch.cat([deltas_x, deltas_x_log], dim=-1)
+                deltas_y = torch.cat([deltas_y, deltas_y_log], dim=-1)
+
+        if self.training:
+            assert self.use_act_checkpoint, "activation ckpt not enabled in decoder"
+        deltas_x = activation_ckpt_wrapper(self.boxRPB_embed_x)(
+            x=deltas_x,
+            act_ckpt_enable=self.training and self.use_act_checkpoint,
+        )  # bs, num_queries, W, n_heads
+        deltas_y = activation_ckpt_wrapper(self.boxRPB_embed_y)(
+            x=deltas_y,
+            act_ckpt_enable=self.training and self.use_act_checkpoint,
+        )  # bs, num_queries, H, n_heads
+
+        if not torch.compiler.is_dynamo_compiling():
+            assert deltas_x.shape[:3] == (bs, num_queries, W)
+            assert deltas_y.shape[:3] == (bs, num_queries, H)
+
+        B = deltas_y.unsqueeze(3) + deltas_x.unsqueeze(
+            2
+        )  # bs, num_queries, H, W, n_heads
+        if not torch.compiler.is_dynamo_compiling():
+            assert B.shape[:4] == (bs, num_queries, H, W)
+        B = B.flatten(2, 3)  # bs, num_queries, H*W, n_heads
+        B = B.permute(0, 3, 1, 2)  # bs, n_heads, num_queries, H*W
+        B = B.contiguous()  # memeff attn likes ordered strides
+        if not torch.compiler.is_dynamo_compiling():
+            assert B.shape[2:] == (num_queries, H * W)
+        return B
+
+    def forward(
+        self,
+        tgt,
+        memory,
+        tgt_mask: Optional[Tensor] = None,
+        memory_mask: Optional[Tensor] = None,
+        tgt_key_padding_mask: Optional[Tensor] = None,
+        memory_key_padding_mask: Optional[Tensor] = None,
+        pos: Optional[Tensor] = None,
+        reference_boxes: Optional[Tensor] = None,  # num_queries, bs, 4
+        # for memory
+        level_start_index: Optional[Tensor] = None,  # num_levels
+        spatial_shapes: Optional[Tensor] = None,  # bs, num_levels, 2
+        valid_ratios: Optional[Tensor] = None,
+        # for text
+        memory_text: Optional[Tensor] = None,
+        text_attention_mask: Optional[Tensor] = None,
+        # if `apply_dac` is None, it will default to `self.dac`
+        apply_dac: Optional[bool] = None,
+        is_instance_prompt=False,
+        decoder_extra_kwargs: Optional[Dict] = None,
+        # ROI memory bank
+        obj_roi_memory_feat=None,
+        obj_roi_memory_mask=None,
+        box_head_trk=None,
+    ):
+        """
+        Input:
+            - tgt: nq, bs, d_model
+            - memory: \\sum{hw}, bs, d_model
+            - pos: \\sum{hw}, bs, d_model
+            - reference_boxes: nq, bs, 4 (after sigmoid)
+            - valid_ratios/spatial_shapes: bs, nlevel, 2
+        """
+        if memory_mask is not None:
+            assert (
+                self.boxRPB == "none"
+            ), "inputting a memory_mask in the presence of boxRPB is unexpected/not implemented"
+
+        apply_dac = apply_dac if apply_dac is not None else self.dac
+        if apply_dac:
+            assert (tgt.shape[0] == self.num_queries) or (
+                self.use_instance_query
+                and (tgt.shape[0] == self.instance_query_embed.num_embeddings)
+            )
+
+            tgt = tgt.repeat(2, 1, 1)
+            # note that we don't tile tgt_mask, since DAC doesn't
+            # use self-attention in o2m queries
+            if reference_boxes is not None:
+                assert (reference_boxes.shape[0] == self.num_queries) or (
+                    self.use_instance_query
+                    and (
+                        reference_boxes.shape[0]
+                        == self.instance_query_embed.num_embeddings
+                    )
+                )
+                reference_boxes = reference_boxes.repeat(2, 1, 1)
+
+        bs = tgt.shape[1]
+        intermediate = []
+        intermediate_presence_logits = []
+        presence_feats = None
+
+        if self.box_refine:
+            if reference_boxes is None:
+                # In this case, we're in a one-stage model, so we generate the reference boxes
+                reference_boxes = self.reference_points.weight.unsqueeze(1)
+                reference_boxes = (
+                    reference_boxes.repeat(2, bs, 1)
+                    if apply_dac
+                    else reference_boxes.repeat(1, bs, 1)
+                )
+                reference_boxes = reference_boxes.sigmoid()
+            intermediate_ref_boxes = [reference_boxes]
+        else:
+            reference_boxes = None
+            intermediate_ref_boxes = None
+
+        output = tgt
+        presence_out = None
+        if self.presence_token is not None and is_instance_prompt is False:
+            # expand to batch dim
+            presence_out = self.presence_token.weight[None].expand(1, bs, -1)
+
+        box_head = self.bbox_embed
+        if is_instance_prompt and self.instance_bbox_embed is not None:
+            box_head = self.instance_bbox_embed
+
+        out_norm = self.norm
+        if is_instance_prompt and self.instance_norm is not None:
+            out_norm = self.instance_norm
+
+        for layer_idx, layer in enumerate(self.layers):
+            reference_points_input = (
+                reference_boxes[:, :, None]
+                * torch.cat([valid_ratios, valid_ratios], -1)[None, :]
+            )  # nq, bs, nlevel, 4
+
+            query_sine_embed = gen_sineembed_for_position(
+                reference_points_input[:, :, 0, :], self.d_model
+            )  # nq, bs, d_model*2
+
+            # conditional query
+            query_pos = self.ref_point_head(query_sine_embed)  # nq, bs, d_model
+
+            if self.boxRPB != "none" and reference_boxes is not None:
+                assert (
+                    spatial_shapes.shape[0] == 1
+                ), "only single scale support implemented"
+                memory_mask = self._get_rpb_matrix(
+                    reference_boxes,
+                    (spatial_shapes[0, 0], spatial_shapes[0, 1]),
+                )
+                memory_mask = memory_mask.flatten(0, 1)  # (bs*n_heads, nq, H*W)
+            if self.training:
+                assert (
+                    self.use_act_checkpoint
+                ), "Activation checkpointing not enabled in the decoder"
+            output, presence_out = activation_ckpt_wrapper(layer)(
+                tgt=output,
+                tgt_query_pos=query_pos,
+                tgt_query_sine_embed=query_sine_embed,
+                tgt_key_padding_mask=tgt_key_padding_mask,
+                tgt_reference_points=reference_points_input,
+                memory_text=memory_text,
+                text_attention_mask=text_attention_mask,
+                memory=memory,
+                memory_key_padding_mask=memory_key_padding_mask,
+                memory_level_start_index=level_start_index,
+                memory_spatial_shapes=spatial_shapes,
+                memory_pos=pos,
+                self_attn_mask=tgt_mask,
+                cross_attn_mask=memory_mask,
+                dac=apply_dac,
+                dac_use_selfatt_ln=self.dac_use_selfatt_ln,
+                presence_token=presence_out,
+                **(decoder_extra_kwargs or {}),
+                act_ckpt_enable=self.training and self.use_act_checkpoint,
+                # ROI memory bank
+                obj_roi_memory_feat=obj_roi_memory_feat,
+                obj_roi_memory_mask=obj_roi_memory_mask,
+            )
+
+            # iter update
+            if self.box_refine:
+                reference_before_sigmoid = inverse_sigmoid(reference_boxes)
+                if box_head_trk is None:
+                    # delta_unsig = self.bbox_embed(output)
+                    if not self.use_normed_output_consistently:
+                        delta_unsig = box_head(output)
+                    else:
+                        delta_unsig = box_head(out_norm(output))
+                else:
+                    # box_head_trk use a separate box head for tracking queries
+                    Q_det = decoder_extra_kwargs["Q_det"]
+                    assert output.size(0) >= Q_det
+                    delta_unsig_det = self.bbox_embed(output[:Q_det])
+                    delta_unsig_trk = box_head_trk(output[Q_det:])
+                    delta_unsig = torch.cat([delta_unsig_det, delta_unsig_trk], dim=0)
+                outputs_unsig = delta_unsig + reference_before_sigmoid
+                new_reference_points = outputs_unsig.sigmoid()
+
+                reference_boxes = new_reference_points.detach()
+                if layer_idx != self.num_layers - 1:
+                    intermediate_ref_boxes.append(new_reference_points)
+            else:
+                raise NotImplementedError("not implemented yet")
+
+            intermediate.append(out_norm(output))
+            if self.presence_token is not None and is_instance_prompt is False:
+                # norm, mlp head
+                intermediate_layer_presence_logits = self.presence_token_head(
+                    self.presence_token_out_norm(presence_out)
+                ).squeeze(-1)
+
+                # clamp to mitigate numerical issues
+                if self.clamp_presence_logits:
+                    intermediate_layer_presence_logits.clamp(
+                        min=-self.clamp_presence_logit_max_val,
+                        max=self.clamp_presence_logit_max_val,
+                    )
+
+                intermediate_presence_logits.append(intermediate_layer_presence_logits)
+                presence_feats = presence_out.clone()
+
+        if not self.compiled and self.compile_mode is not None:
+            self.forward = torch.compile(
+                self.forward, mode=self.compile_mode, fullgraph=True
+            )
+            self.compiled = True
+
+        return (
+            torch.stack(intermediate),
+            torch.stack(intermediate_ref_boxes),
+            (
+                torch.stack(intermediate_presence_logits)
+                if self.presence_token is not None and is_instance_prompt is False
+                else None
+            ),
+            presence_feats,
+        )
+
+
+class TransformerEncoderCrossAttention(nn.Module):
+    def __init__(
+        self,
+        d_model: int,
+        frozen: bool,
+        pos_enc_at_input: bool,
+        layer,
+        num_layers: int,
+        use_act_checkpoint: bool = False,
+        batch_first: bool = False,  # Do layers expect batch first input?
+        # which layers to exclude cross attention? default: None, means all
+        # layers use cross attention
+        remove_cross_attention_layers: Optional[list] = None,
+    ):
+        super().__init__()
+        self.d_model = d_model
+        self.layers = get_clones(layer, num_layers)
+        self.num_layers = num_layers
+        self.norm = nn.LayerNorm(d_model)
+        self.pos_enc_at_input = pos_enc_at_input
+        self.use_act_checkpoint = use_act_checkpoint
+
+        if frozen:
+            for p in self.parameters():
+                p.requires_grad_(False)
+
+        self.batch_first = batch_first
+
+        # remove cross attention layers if specified
+        self.remove_cross_attention_layers = [False] * self.num_layers
+        if remove_cross_attention_layers is not None:
+            for i in remove_cross_attention_layers:
+                self.remove_cross_attention_layers[i] = True
+        assert len(self.remove_cross_attention_layers) == len(self.layers)
+
+        for i, remove_cross_attention in enumerate(self.remove_cross_attention_layers):
+            if remove_cross_attention:
+                self.layers[i].cross_attn_image = None
+                self.layers[i].norm2 = None
+                self.layers[i].dropout2 = None
+
+    def forward(
+        self,
+        src,  # self-attention inputs
+        prompt,  # cross-attention inputs
+        src_mask: Optional[Tensor] = None,  # att.mask for self-attention inputs
+        prompt_mask: Optional[Tensor] = None,  # att.mask for cross-attention inputs
+        src_key_padding_mask: Optional[Tensor] = None,
+        prompt_key_padding_mask: Optional[Tensor] = None,
+        src_pos: Optional[Tensor] = None,  # pos_enc for self-attention inputs
+        prompt_pos: Optional[Tensor] = None,  # pos_enc for cross-attention inputs
+        feat_sizes: Optional[list] = None,
+        num_obj_ptr_tokens: int = 0,  # number of object pointer *tokens*
+    ):
+        if isinstance(src, list):
+            assert isinstance(src_key_padding_mask, list) and isinstance(src_pos, list)
+            assert len(src) == len(src_key_padding_mask) == len(src_pos) == 1
+            src, src_key_padding_mask, src_pos = (
+                src[0],
+                src_key_padding_mask[0],
+                src_pos[0],
+            )
+
+        assert (
+            src.shape[1] == prompt.shape[1]
+        ), "Batch size must be the same for src and prompt"
+
+        output = src
+
+        if self.pos_enc_at_input and src_pos is not None:
+            output = output + 0.1 * src_pos
+
+        if self.batch_first:
+            # Convert to batch first
+            output = output.transpose(0, 1)
+            src_pos = src_pos.transpose(0, 1)
+            prompt = prompt.transpose(0, 1)
+            prompt_pos = prompt_pos.transpose(0, 1)
+
+        for layer in self.layers:
+            kwds = {}
+            if isinstance(layer.cross_attn_image, RoPEAttention):
+                kwds = {"num_k_exclude_rope": num_obj_ptr_tokens}
+
+            output = activation_ckpt_wrapper(layer)(
+                tgt=output,
+                memory=prompt,
+                tgt_mask=src_mask,
+                memory_mask=prompt_mask,
+                tgt_key_padding_mask=src_key_padding_mask,
+                memory_key_padding_mask=prompt_key_padding_mask,
+                pos=prompt_pos,
+                query_pos=src_pos,
+                dac=False,
+                attn_bias=None,
+                act_ckpt_enable=self.training and self.use_act_checkpoint,
+                **kwds,
+            )
+            normed_output = self.norm(output)
+
+        if self.batch_first:
+            # Convert back to seq first
+            normed_output = normed_output.transpose(0, 1)
+            src_pos = src_pos.transpose(0, 1)
+
+        return {
+            "memory": normed_output,
+            "pos_embed": src_pos,
+            "padding_mask": src_key_padding_mask,
+        }
+
+
+class TransformerDecoderLayerv1(nn.Module):
+    def __init__(
+        self,
+        activation: str,
+        cross_attention: nn.Module,
+        d_model: int,
+        dim_feedforward: int,
+        dropout: float,
+        pos_enc_at_attn: bool,
+        pos_enc_at_cross_attn_keys: bool,
+        pos_enc_at_cross_attn_queries: bool,
+        pre_norm: bool,
+        self_attention: nn.Module,
+    ):
+        super().__init__()
+        self.d_model = d_model
+        self.dim_feedforward = dim_feedforward
+        self.dropout_value = dropout
+        self.self_attn = self_attention
+        self.cross_attn_image = cross_attention
+
+        # Implementation of Feedforward model
+        self.linear1 = nn.Linear(d_model, dim_feedforward)
+        self.dropout = nn.Dropout(dropout)
+        self.linear2 = nn.Linear(dim_feedforward, d_model)
+
+        self.norm1 = nn.LayerNorm(d_model)
+        self.norm2 = nn.LayerNorm(d_model)
+        self.norm3 = nn.LayerNorm(d_model)
+        self.dropout1 = nn.Dropout(dropout)
+        self.dropout2 = nn.Dropout(dropout)
+        self.dropout3 = nn.Dropout(dropout)
+
+        self.activation_str = activation
+        self.activation = get_activation_fn(activation)
+        self.pre_norm = pre_norm
+
+        self.pos_enc_at_attn = pos_enc_at_attn
+        self.pos_enc_at_cross_attn_queries = pos_enc_at_cross_attn_queries
+        self.pos_enc_at_cross_attn_keys = pos_enc_at_cross_attn_keys
+
+    def forward_post(
+        self,
+        tgt,
+        memory,
+        tgt_mask: Optional[Tensor] = None,
+        memory_mask: Optional[Tensor] = None,
+        tgt_key_padding_mask: Optional[Tensor] = None,
+        memory_key_padding_mask: Optional[Tensor] = None,
+        pos: Optional[Tensor] = None,
+        query_pos: Optional[Tensor] = None,
+        **kwargs,
+    ):
+        q = k = tgt + query_pos if self.pos_enc_at_attn else tgt
+
+        # Self attention
+        tgt2 = self.self_attn(
+            q,
+            k,
+            value=tgt,
+            attn_mask=tgt_mask,
+            key_padding_mask=tgt_key_padding_mask,
+        )[0]
+        tgt = tgt + self.dropout1(tgt2)
+        tgt = self.norm1(tgt)
+
+        # Cross attention to image
+        tgt2 = self.cross_attn_image(
+            query=tgt + query_pos if self.pos_enc_at_cross_attn_queries else tgt,
+            key=memory + pos if self.pos_enc_at_cross_attn_keys else memory,
+            value=memory,
+            attn_mask=memory_mask,
+            key_padding_mask=memory_key_padding_mask,
+        )[0]
+        tgt = tgt + self.dropout2(tgt2)
+        tgt = self.norm2(tgt)
+
+        # FFN
+        tgt2 = self.linear2(self.dropout(self.activation(self.linear1(tgt))))
+        tgt = tgt + self.dropout3(tgt2)
+        tgt = self.norm3(tgt)
+        return tgt
+
+    def forward_pre(
+        self,
+        tgt,
+        memory,
+        dac: bool = False,
+        tgt_mask: Optional[Tensor] = None,
+        memory_mask: Optional[Tensor] = None,
+        tgt_key_padding_mask: Optional[Tensor] = None,
+        memory_key_padding_mask: Optional[Tensor] = None,
+        pos: Optional[Tensor] = None,
+        query_pos: Optional[Tensor] = None,
+        attn_bias: Optional[Tensor] = None,
+        **kwargs,
+    ):
+        if dac:
+            # we only apply self attention to the first half of the queries
+            assert tgt.shape[0] % 2 == 0
+            other_tgt = tgt[tgt.shape[0] // 2 :]
+            tgt = tgt[: tgt.shape[0] // 2]
+        tgt2 = self.norm1(tgt)
+        q = k = tgt2 + query_pos if self.pos_enc_at_attn else tgt2
+        tgt2 = self.self_attn(
+            q,
+            k,
+            value=tgt2,
+            attn_mask=tgt_mask,
+            key_padding_mask=tgt_key_padding_mask,
+        )[0]
+        tgt = tgt + self.dropout1(tgt2)
+        if dac:
+            # Recombine
+            tgt = torch.cat((tgt, other_tgt), dim=0)
+        tgt2 = self.norm2(tgt)
+        tgt2 = self.cross_attn_image(
+            query=tgt2 + query_pos if self.pos_enc_at_cross_attn_queries else tgt2,
+            key=memory + pos if self.pos_enc_at_cross_attn_keys else memory,
+            value=memory,
+            attn_mask=memory_mask,
+            key_padding_mask=memory_key_padding_mask,
+            attn_bias=attn_bias,
+        )[0]
+        tgt = tgt + self.dropout2(tgt2)
+        tgt2 = self.norm3(tgt)
+        tgt2 = self.linear2(self.dropout(self.activation(self.linear1(tgt2))))
+        tgt = tgt + self.dropout3(tgt2)
+        return tgt
+
+    def forward(
+        self,
+        tgt,
+        memory,
+        dac: bool = False,
+        tgt_mask: Optional[Tensor] = None,
+        memory_mask: Optional[Tensor] = None,
+        tgt_key_padding_mask: Optional[Tensor] = None,
+        memory_key_padding_mask: Optional[Tensor] = None,
+        pos: Optional[Tensor] = None,
+        query_pos: Optional[Tensor] = None,
+        attn_bias: Optional[Tensor] = None,
+        **kwds: Any,
+    ) -> torch.Tensor:
+        fwd_fn = self.forward_pre if self.pre_norm else self.forward_post
+        return fwd_fn(
+            tgt,
+            memory,
+            dac=dac,
+            tgt_mask=tgt_mask,
+            memory_mask=memory_mask,
+            tgt_key_padding_mask=tgt_key_padding_mask,
+            memory_key_padding_mask=memory_key_padding_mask,
+            pos=pos,
+            query_pos=query_pos,
+            attn_bias=attn_bias,
+            **kwds,
+        )
+
+
+class TransformerDecoderLayerv2(TransformerDecoderLayerv1):
+    def __init__(self, cross_attention_first=False, *args: Any, **kwds: Any):
+        super().__init__(*args, **kwds)
+        self.cross_attention_first = cross_attention_first
+
+    def _forward_sa(self, tgt, query_pos):
+        # Self-Attention
+        tgt2 = self.norm1(tgt)
+        q = k = tgt2 + query_pos if self.pos_enc_at_attn else tgt2
+        tgt2 = self.self_attn(q, k, v=tgt2)
+        tgt = tgt + self.dropout1(tgt2)
+        return tgt
+
+    def _forward_ca(self, tgt, memory, query_pos, pos, num_k_exclude_rope=0):
+        if self.cross_attn_image is None:
+            return tgt
+
+        kwds = {}
+        if num_k_exclude_rope > 0:
+            assert isinstance(self.cross_attn_image, RoPEAttention)
+            kwds = {"num_k_exclude_rope": num_k_exclude_rope}
+
+        # Cross-Attention
+        tgt2 = self.norm2(tgt)
+        tgt2 = self.cross_attn_image(
+            q=tgt2 + query_pos if self.pos_enc_at_cross_attn_queries else tgt2,
+            k=memory + pos if self.pos_enc_at_cross_attn_keys else memory,
+            v=memory,
+            **kwds,
+        )
+        tgt = tgt + self.dropout2(tgt2)
+        return tgt
+
+    def forward_pre(
+        self,
+        tgt,
+        memory,
+        dac: bool,
+        tgt_mask: Optional[Tensor] = None,
+        memory_mask: Optional[Tensor] = None,
+        tgt_key_padding_mask: Optional[Tensor] = None,
+        memory_key_padding_mask: Optional[Tensor] = None,
+        pos: Optional[Tensor] = None,
+        query_pos: Optional[Tensor] = None,
+        attn_bias: Optional[Tensor] = None,
+        num_k_exclude_rope: int = 0,
+    ):
+        assert dac is False
+        assert tgt_mask is None
+        assert memory_mask is None
+        assert tgt_key_padding_mask is None
+        assert memory_key_padding_mask is None
+        assert attn_bias is None
+
+        if self.cross_attention_first:
+            tgt = self._forward_ca(tgt, memory, query_pos, pos, num_k_exclude_rope)
+            tgt = self._forward_sa(tgt, query_pos)
+        else:
+            tgt = self._forward_sa(tgt, query_pos)
+            tgt = self._forward_ca(tgt, memory, query_pos, pos, num_k_exclude_rope)
+
+        # MLP
+        tgt2 = self.norm3(tgt)
+        tgt2 = self.linear2(self.dropout(self.activation(self.linear1(tgt2))))
+        tgt = tgt + self.dropout3(tgt2)
+        return tgt
+
+    def forward(self, *args: Any, **kwds: Any) -> torch.Tensor:
+        if self.pre_norm:
+            return self.forward_pre(*args, **kwds)
+        raise NotImplementedError

detect_tools/sam3/sam3/model/edt.py

@@ -0,0 +1,173 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates. All Rights Reserved
+
+"""Triton kernel for euclidean distance transform (EDT)"""
+
+import torch
+import triton
+import triton.language as tl
+
+"""
+Disclaimer: This implementation is not meant to be extremely efficient. A CUDA kernel would likely be more efficient.
+Even in Triton, there may be more suitable algorithms.
+
+The goal of this kernel is to mimic cv2.distanceTransform(input, cv2.DIST_L2, 0).
+Recall that the euclidean distance transform (EDT) calculates the L2 distance to the closest zero pixel for each pixel of the source image.
+
+For images of size NxN, the naive algorithm would be to compute pairwise distances between every pair of points, leading to a O(N^4) algorithm, which is obviously impractical.
+One can do better using the following approach:
+- First, compute the distance to the closest point in the same row. We can write it as Row_EDT[i,j] = min_k (sqrt((k-j)^2) if input[i,k]==0 else +infinity). With a naive implementation, this step has a O(N^3) complexity
+- Then, because of triangular inequality, we notice that the EDT for a given location [i,j] is the min of the row EDTs in the same column. EDT[i,j] = min_k Row_EDT[k, j]. This is also O(N^3)
+
+Overall, this algorithm is quite amenable to parallelization, and has a complexity O(N^3). Can we do better?
+
+It turns out that we can leverage the structure of the L2 distance (nice and convex) to find the minimum in a more efficient way.
+We follow the algorithm from "Distance Transforms of Sampled Functions" (https://cs.brown.edu/people/pfelzens/papers/dt-final.pdf), which is also what's implemented in opencv
+
+For a single dimension EDT, we can compute the EDT of an arbitrary function F, that we discretize over the grid. Note that for the binary EDT that we're interested in, we can set F(i,j) = 0 if input[i,j]==0 else +infinity
+For now, we'll compute the EDT squared, and will take the sqrt only at the very end.
+The basic idea is that each point at location i spawns a parabola around itself, with a bias equal to F(i). So specifically, we're looking at the parabola (x - i)^2 + F(i)
+When we're looking for the row EDT at location j, we're effectively looking for min_i (x-i)^2 + F(i). In other word we want to find the lowest parabola at location j.
+
+To do this efficiently, we need to maintain the lower envelope of the union of parabolas. This can be constructed on the fly using a sort of stack approach:
+ - every time we want to add a new parabola, we check if it may be covering the current right-most parabola. If so, then that parabola was useless, so we can pop it from the stack
+ - repeat until we can't find any more parabola to pop. Then push the new one.
+
+This algorithm runs in O(N) for a single row, so overall O(N^2) when applied to all rows
+Similarly as before, we notice that we can decompose the algorithm for rows and columns, leading to an overall run-time of O(N^2)
+
+This algorithm is less suited for to GPUs, since the one-dimensional EDT computation is quite sequential in nature. However, we can parallelize over batch and row dimensions.
+In Triton, things are particularly bad at the moment, since there is no support for reading/writing to the local memory at a specific index (a local gather is coming soon, see https://github.com/triton-lang/triton/issues/974, but no mention of writing, ie scatter)
+One could emulate these operations with masking, but in initial tests, it proved to be worst than naively reading and writing to the global memory. My guess is that the cache is compensating somewhat for the repeated single-point accesses.
+
+
+The timing obtained on a H100 for a random batch of masks of dimension 256 x 1024 x 1024 are as follows:
+- OpenCV: 1780ms (including round-trip to cpu, but discounting the fact that it introduces a synchronization point)
+- triton, O(N^3) algo: 627ms
+- triton, O(N^2) algo: 322ms
+
+Overall, despite being quite naive, this implementation is roughly 5.5x faster than the openCV cpu implem
+
+"""
+
+
+@triton.jit
+def edt_kernel(inputs_ptr, outputs_ptr, v, z, height, width, horizontal: tl.constexpr):
+    # This is a somewhat verbatim implementation of the efficient 1D EDT algorithm described above
+    # It can be applied horizontally or vertically depending if we're doing the first or second stage.
+    # It's parallelized across batch+row (or batch+col if horizontal=False)
+    # TODO: perhaps the implementation can be revisited if/when local gather/scatter become available in triton
+    batch_id = tl.program_id(axis=0)
+    if horizontal:
+        row_id = tl.program_id(axis=1)
+        block_start = (batch_id * height * width) + row_id * width
+        length = width
+        stride = 1
+    else:
+        col_id = tl.program_id(axis=1)
+        block_start = (batch_id * height * width) + col_id
+        length = height
+        stride = width
+
+    # This will be the index of the right most parabola in the envelope ("the top of the stack")
+    k = 0
+    for q in range(1, length):
+        # Read the function value at the current location. Note that we're doing a singular read, not very efficient
+        cur_input = tl.load(inputs_ptr + block_start + (q * stride))
+        # location of the parabola on top of the stack
+        r = tl.load(v + block_start + (k * stride))
+        # associated boundary
+        z_k = tl.load(z + block_start + (k * stride))
+        # value of the function at the parabola location
+        previous_input = tl.load(inputs_ptr + block_start + (r * stride))
+        # intersection between the two parabolas
+        s = (cur_input - previous_input + q * q - r * r) / (q - r) / 2
+
+        # we'll pop as many parabolas as required
+        while s <= z_k and k - 1 >= 0:
+            k = k - 1
+            r = tl.load(v + block_start + (k * stride))
+            z_k = tl.load(z + block_start + (k * stride))
+            previous_input = tl.load(inputs_ptr + block_start + (r * stride))
+            s = (cur_input - previous_input + q * q - r * r) / (q - r) / 2
+
+        # Store the new one
+        k = k + 1
+        tl.store(v + block_start + (k * stride), q)
+        tl.store(z + block_start + (k * stride), s)
+        if k + 1 < length:
+            tl.store(z + block_start + ((k + 1) * stride), 1e9)
+
+    # Last step, we read the envelope to find the min in every location
+    k = 0
+    for q in range(length):
+        while (
+            k + 1 < length
+            and tl.load(
+                z + block_start + ((k + 1) * stride), mask=(k + 1) < length, other=q
+            )
+            < q
+        ):
+            k += 1
+        r = tl.load(v + block_start + (k * stride))
+        d = q - r
+        old_value = tl.load(inputs_ptr + block_start + (r * stride))
+        tl.store(outputs_ptr + block_start + (q * stride), old_value + d * d)
+
+
+def edt_triton(data: torch.Tensor):
+    """
+    Computes the Euclidean Distance Transform (EDT) of a batch of binary images.
+
+    Args:
+        data: A tensor of shape (B, H, W) representing a batch of binary images.
+
+    Returns:
+        A tensor of the same shape as data containing the EDT.
+        It should be equivalent to a batched version of cv2.distanceTransform(input, cv2.DIST_L2, 0)
+    """
+    assert data.dim() == 3
+    assert data.is_cuda
+    B, H, W = data.shape
+    data = data.contiguous()
+
+    # Allocate the "function" tensor. Implicitly the function is 0 if data[i,j]==0 else +infinity
+    output = torch.where(data, 1e18, 0.0)
+    assert output.is_contiguous()
+
+    # Scratch tensors for the parabola stacks
+    parabola_loc = torch.zeros(B, H, W, dtype=torch.uint32, device=data.device)
+    parabola_inter = torch.empty(B, H, W, dtype=torch.float, device=data.device)
+    parabola_inter[:, :, 0] = -1e18
+    parabola_inter[:, :, 1] = 1e18
+
+    # Grid size (number of blocks)
+    grid = (B, H)
+
+    # Launch initialization kernel
+    edt_kernel[grid](
+        output.clone(),
+        output,
+        parabola_loc,
+        parabola_inter,
+        H,
+        W,
+        horizontal=True,
+    )
+
+    # reset the parabola stacks
+    parabola_loc.zero_()
+    parabola_inter[:, :, 0] = -1e18
+    parabola_inter[:, :, 1] = 1e18
+
+    grid = (B, W)
+    edt_kernel[grid](
+        output.clone(),
+        output,
+        parabola_loc,
+        parabola_inter,
+        H,
+        W,
+        horizontal=False,
+    )
+    # don't forget to take sqrt at the end
+    return output.sqrt()

detect_tools/sam3/sam3/model/encoder.py

@@ -0,0 +1,594 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates. All Rights Reserved
+# Based on https://github.com/IDEA-Research/GroundingDINO
+
+from typing import Any, Dict, List, Optional, Tuple
+
+import torch
+from torch import nn, Tensor
+
+from .act_ckpt_utils import activation_ckpt_wrapper
+from .model_misc import get_activation_fn, get_clones, get_valid_ratio
+
+
+class TransformerEncoderLayer(nn.Module):
+    """
+    Transformer encoder layer that performs self-attention followed by cross-attention.
+
+    This layer was previously called TransformerDecoderLayer but was renamed to better
+    reflect its role in the architecture. It processes input sequences through self-attention
+    and then cross-attention with another input (typically image features).
+
+    The layer supports both pre-norm and post-norm configurations, as well as
+    positional encoding at different stages of the attention mechanism.
+    """
+
+    def __init__(
+        self,
+        activation: str,
+        cross_attention: nn.Module,
+        d_model: int,
+        dim_feedforward: int,
+        dropout: float,
+        pos_enc_at_attn: bool,
+        pos_enc_at_cross_attn_keys: bool,
+        pos_enc_at_cross_attn_queries: bool,
+        pre_norm: bool,
+        self_attention: nn.Module,
+    ):
+        """
+        Initialize a transformer encoder layer.
+
+        Args:
+            activation: Activation function to use in the feedforward network
+            cross_attention: Cross-attention module for attending to image features
+            d_model: Model dimension/hidden size
+            dim_feedforward: Dimension of the feedforward network
+            dropout: Dropout probability
+            pos_enc_at_attn: Whether to add positional encodings at self-attention
+            pos_enc_at_cross_attn_keys: Whether to add positional encodings to keys in cross-attention
+            pos_enc_at_cross_attn_queries: Whether to add positional encodings to queries in cross-attention
+            pre_norm: Whether to use pre-norm (True) or post-norm (False) architecture
+            self_attention: Self-attention module
+        """
+        super().__init__()
+        self.d_model = d_model
+        self.dim_feedforward = dim_feedforward
+        self.dropout_value = dropout
+        self.self_attn = self_attention
+        self.cross_attn_image = cross_attention
+
+        # Implementation of Feedforward model
+        self.linear1 = nn.Linear(d_model, dim_feedforward)
+        self.dropout = nn.Dropout(dropout)
+        self.linear2 = nn.Linear(dim_feedforward, d_model)
+
+        self.norm1 = nn.LayerNorm(d_model)
+        self.norm2 = nn.LayerNorm(d_model)
+        self.norm3 = nn.LayerNorm(d_model)
+        self.dropout1 = nn.Dropout(dropout)
+        self.dropout2 = nn.Dropout(dropout)
+        self.dropout3 = nn.Dropout(dropout)
+
+        self.activation_str = activation
+        self.activation = get_activation_fn(activation)
+        self.pre_norm = pre_norm
+
+        self.pos_enc_at_attn = pos_enc_at_attn
+        self.pos_enc_at_cross_attn_queries = pos_enc_at_cross_attn_queries
+        self.pos_enc_at_cross_attn_keys = pos_enc_at_cross_attn_keys
+
+        self.layer_idx = None
+
+    def forward_post(
+        self,
+        tgt: Tensor,
+        memory: Tensor,
+        tgt_mask: Optional[Tensor] = None,
+        memory_mask: Optional[Tensor] = None,
+        tgt_key_padding_mask: Optional[Tensor] = None,
+        memory_key_padding_mask: Optional[Tensor] = None,
+        pos: Optional[Tensor] = None,
+        query_pos: Optional[Tensor] = None,
+        **kwargs,
+    ) -> Tensor:
+        """
+        Forward pass for post-norm architecture.
+
+        In post-norm architecture, normalization is applied after attention and feedforward operations.
+
+        Args:
+            tgt: Input tensor to be processed
+            memory: Memory tensor for cross-attention
+            tgt_mask: Mask for self-attention
+            memory_mask: Mask for cross-attention
+            tgt_key_padding_mask: Key padding mask for self-attention
+            memory_key_padding_mask: Key padding mask for cross-attention
+            pos: Positional encoding for memory
+            query_pos: Positional encoding for query
+            **kwargs: Additional keyword arguments
+
+        Returns:
+            Processed tensor
+        """
+        q = k = tgt + query_pos if self.pos_enc_at_attn else tgt
+
+        # Self attention
+        tgt2 = self.self_attn(
+            q, k, value=tgt, attn_mask=tgt_mask, key_padding_mask=tgt_key_padding_mask
+        )[0]
+        tgt = tgt + self.dropout1(tgt2)
+        tgt = self.norm1(tgt)
+
+        # Cross attention to image
+        tgt2 = self.cross_attn_image(
+            query=tgt + query_pos if self.pos_enc_at_cross_attn_queries else tgt,
+            key=memory + pos if self.pos_enc_at_cross_attn_keys else memory,
+            value=memory,
+            attn_mask=memory_mask,
+            key_padding_mask=memory_key_padding_mask,
+        )[0]
+        tgt = tgt + self.dropout2(tgt2)
+        tgt = self.norm2(tgt)
+
+        # FFN
+        tgt2 = self.linear2(self.dropout(self.activation(self.linear1(tgt))))
+        tgt = tgt + self.dropout3(tgt2)
+        tgt = self.norm3(tgt)
+        return tgt
+
+    def forward_pre(
+        self,
+        tgt: Tensor,
+        memory: Tensor,
+        dac: bool = False,
+        tgt_mask: Optional[Tensor] = None,
+        memory_mask: Optional[Tensor] = None,
+        tgt_key_padding_mask: Optional[Tensor] = None,
+        memory_key_padding_mask: Optional[Tensor] = None,
+        pos: Optional[Tensor] = None,
+        query_pos: Optional[Tensor] = None,
+        # attn_bias: Optional[Tensor] = None,
+        # **kwargs,
+    ) -> Tensor:
+        """
+        Forward pass for pre-norm architecture.
+
+        In pre-norm architecture, normalization is applied before attention and feedforward operations.
+
+        Args:
+            tgt: Input tensor to be processed
+            memory: Memory tensor for cross-attention
+            dac: Whether to use Divide-and-Conquer attention
+            tgt_mask: Mask for self-attention
+            memory_mask: Mask for cross-attention
+            tgt_key_padding_mask: Key padding mask for self-attention
+            memory_key_padding_mask: Key padding mask for cross-attention
+            pos: Positional encoding for memory
+            query_pos: Positional encoding for query
+            attn_bias: Optional attention bias tensor
+            **kwargs: Additional keyword arguments
+
+        Returns:
+            Processed tensor
+        """
+        if dac:
+            # we only apply self attention to the first half of the queries
+            assert tgt.shape[0] % 2 == 0
+            other_tgt = tgt[tgt.shape[0] // 2 :]
+            tgt = tgt[: tgt.shape[0] // 2]
+        tgt2 = self.norm1(tgt)
+        q = k = tgt2 + query_pos if self.pos_enc_at_attn else tgt2
+        tgt2 = self.self_attn(
+            q, k, value=tgt2, attn_mask=tgt_mask, key_padding_mask=tgt_key_padding_mask
+        )[0]
+        tgt = tgt + self.dropout1(tgt2)
+        if dac:
+            # Recombine
+            tgt = torch.cat((tgt, other_tgt), dim=0)
+        tgt2 = self.norm2(tgt)
+        tgt2 = self.cross_attn_image(
+            query=tgt2 + query_pos if self.pos_enc_at_cross_attn_queries else tgt2,
+            key=memory + pos if self.pos_enc_at_cross_attn_keys else memory,
+            value=memory,
+            attn_mask=memory_mask,
+            key_padding_mask=memory_key_padding_mask,
+            # attn_bias=attn_bias,
+        )[0]
+        tgt = tgt + self.dropout2(tgt2)
+        tgt2 = self.norm3(tgt)
+        tgt2 = self.linear2(self.dropout(self.activation(self.linear1(tgt2))))
+        tgt = tgt + self.dropout3(tgt2)
+        return tgt
+
+    def forward(
+        self,
+        tgt: Tensor,
+        memory: Tensor,
+        dac: bool = False,
+        tgt_mask: Optional[Tensor] = None,
+        memory_mask: Optional[Tensor] = None,
+        tgt_key_padding_mask: Optional[Tensor] = None,
+        memory_key_padding_mask: Optional[Tensor] = None,
+        pos: Optional[Tensor] = None,
+        query_pos: Optional[Tensor] = None,
+        # attn_bias: Optional[Tensor] = None,
+        # **kwds: Any,
+    ) -> torch.Tensor:
+        """
+        Forward pass for the transformer encoder layer.
+
+        Args:
+            tgt: Input tensor to be processed
+            memory: Memory tensor (e.g., image features) for cross-attention
+            dac: Whether to use Divide-and-Conquer attention (only apply self-attention to first half)
+            tgt_mask: Mask for self-attention
+            memory_mask: Mask for cross-attention
+            tgt_key_padding_mask: Key padding mask for self-attention
+            memory_key_padding_mask: Key padding mask for cross-attention
+            pos: Positional encoding for memory
+            query_pos: Positional encoding for query
+            attn_bias: Optional attention bias tensor
+            **kwds: Additional keyword arguments
+
+        Returns:
+            Processed tensor after self-attention, cross-attention, and feedforward network
+        """
+        fwd_fn = self.forward_pre if self.pre_norm else self.forward_post
+        return fwd_fn(
+            tgt,
+            memory,
+            dac=dac,
+            tgt_mask=tgt_mask,
+            memory_mask=memory_mask,
+            tgt_key_padding_mask=tgt_key_padding_mask,
+            memory_key_padding_mask=memory_key_padding_mask,
+            pos=pos,
+            query_pos=query_pos,
+            # attn_bias=attn_bias,
+            # **kwds,
+        )
+
+
+class TransformerEncoder(nn.Module):
+    """
+    Transformer encoder that processes multi-level features.
+
+    This encoder takes multi-level features (e.g., from a backbone network) and processes
+    them through a stack of transformer encoder layers. It supports features from multiple
+    levels (e.g., different resolutions) and can apply activation checkpointing for memory
+    efficiency during training.
+
+    Args:
+        layer: The encoder layer to be stacked multiple times
+        num_layers: Number of encoder layers to stack
+        d_model: Model dimension/hidden size
+        num_feature_levels: Number of feature levels to process
+        frozen: Whether to freeze the parameters of this module
+        use_act_checkpoint: Whether to use activation checkpointing during training
+    """
+
+    def __init__(
+        self,
+        layer: nn.Module,
+        num_layers: int,
+        d_model: int,
+        num_feature_levels: int,
+        frozen: bool = False,
+        use_act_checkpoint: bool = False,
+    ):
+        super().__init__()
+        self.layers = get_clones(layer, num_layers)
+        self.num_layers = num_layers
+
+        self.num_feature_levels = num_feature_levels
+        self.level_embed = None
+        if num_feature_levels > 1:
+            self.level_embed = nn.Parameter(torch.Tensor(num_feature_levels, d_model))
+
+        if frozen:
+            for p in self.parameters():
+                p.requires_grad_(False)
+
+        self.use_act_checkpoint = use_act_checkpoint
+
+        # assign layer index to each layer so that some layers can decide what to do
+        # based on which layer index they are (e.g. cross attention to memory bank only
+        # in selected layers)
+        for layer_idx, layer in enumerate(self.layers):
+            layer.layer_idx = layer_idx
+
+    @staticmethod
+    def get_reference_points(spatial_shapes, valid_ratios, device):
+        with torch.no_grad():
+            reference_points_list = []
+            for lvl, (H_, W_) in enumerate(spatial_shapes):
+                ref_y, ref_x = torch.meshgrid(
+                    torch.linspace(
+                        0.5, H_ - 0.5, H_, dtype=torch.float32, device=device
+                    ),
+                    torch.linspace(
+                        0.5, W_ - 0.5, W_, dtype=torch.float32, device=device
+                    ),
+                )
+                ref_y = ref_y.reshape(-1)[None] / (valid_ratios[:, None, lvl, 1] * H_)
+                ref_x = ref_x.reshape(-1)[None] / (valid_ratios[:, None, lvl, 0] * W_)
+                ref = torch.stack((ref_x, ref_y), -1)
+                reference_points_list.append(ref)
+            reference_points = torch.cat(reference_points_list, 1)
+            reference_points = reference_points[:, :, None] * valid_ratios[:, None]
+
+        return reference_points
+
+    def _prepare_multilevel_features(self, srcs, masks, pos_embeds):
+        assert (
+            len(srcs) == self.num_feature_levels
+        ), "mismatch between expected and received # of feature levels"
+
+        src_flatten = []
+        mask_flatten = []
+        lvl_pos_embed_flatten = []
+        spatial_shapes = []
+        has_mask = masks is not None and masks[0] is not None
+        for lvl, (src, mask, pos_embed) in enumerate(zip(srcs, masks, pos_embeds)):
+            bs, c, h, w = src.shape
+            spatial_shape = (h, w)
+            spatial_shapes.append(spatial_shape)
+
+            src = src.flatten(2).transpose(1, 2)  # bs, hw, c
+            if has_mask:
+                mask = mask.flatten(1)
+            pos_embed = pos_embed.flatten(2).transpose(1, 2)  # bs, hw, c
+            if self.level_embed is not None:
+                lvl_pos_embed = pos_embed + self.level_embed[lvl].view(1, 1, -1)
+            else:
+                lvl_pos_embed = pos_embed
+            lvl_pos_embed_flatten.append(lvl_pos_embed)
+            src_flatten.append(src)
+            if has_mask:
+                mask_flatten.append(mask)
+        src_flatten = torch.cat(src_flatten, 1)  # bs, \sum{hxw}, c
+        mask_flatten = torch.cat(mask_flatten, 1) if has_mask else None  # bs, \sum{hxw}
+        lvl_pos_embed_flatten = torch.cat(lvl_pos_embed_flatten, 1)  # bs, \sum{hxw}, c
+        spatial_shapes = torch.tensor(
+            spatial_shapes, dtype=torch.long, device=src_flatten.device
+        )
+        level_start_index = torch.cat(
+            (
+                spatial_shapes.new_zeros((1,)),
+                spatial_shapes.prod(1).cumsum(0)[:-1],
+            )
+        )
+        if has_mask:
+            valid_ratios = torch.stack([get_valid_ratio(m) for m in masks], 1)
+        else:
+            valid_ratios = torch.ones(
+                (src_flatten.shape[0], self.num_feature_levels, 2),
+                device=src_flatten.device,
+            )
+
+        return (
+            src_flatten,
+            mask_flatten,
+            lvl_pos_embed_flatten,
+            level_start_index,
+            valid_ratios,
+            spatial_shapes,
+        )
+
+    def forward(
+        self,
+        src: List[Tensor],
+        src_key_padding_masks: Optional[List[Tensor]] = None,
+        pos: Optional[List[Tensor]] = None,
+        prompt: Optional[Tensor] = None,
+        prompt_key_padding_mask: Optional[Tensor] = None,
+        encoder_extra_kwargs: Optional[Dict] = None,
+    ) -> Tuple[Tensor, Optional[Tensor], Tensor, Tensor, Tensor, Tensor]:
+        """
+        Process multi-level features through the transformer encoder.
+
+        Args:
+            src: List of multi-level features, each with shape (batch_size, channels, height, width)
+            src_key_padding_masks: List of padding masks for each feature level, each with shape (batch_size, height, width)
+            pos: List of positional embeddings for each feature level, each with shape (batch_size, channels, height, width)
+            prompt: Optional text/prompt features to attend to, with shape (seq_len, batch_size, d_model)
+            prompt_key_padding_mask: Optional padding mask for prompt, with shape (batch_size, seq_len)
+            encoder_extra_kwargs: Optional additional arguments to pass to each encoder layer
+
+        Returns:
+            A tuple containing:
+            - output: Processed features with shape (seq_len, batch_size, d_model)
+            - key_padding_masks_flatten: Flattened padding masks
+            - lvl_pos_embed_flatten: Flattened positional embeddings
+            - level_start_index: Starting indices for each feature level
+            - spatial_shapes: Spatial dimensions of each feature level
+            - valid_ratios: Valid ratios for each feature level
+        """
+        assert (
+            len(src) == self.num_feature_levels
+        ), "must be equal to num_feature_levels"
+        if src_key_padding_masks is not None:
+            assert len(src_key_padding_masks) == self.num_feature_levels
+        if pos is not None:
+            assert len(pos) == self.num_feature_levels
+        # Flatten multilevel feats and add level pos embeds
+        (
+            src_flatten,
+            key_padding_masks_flatten,
+            lvl_pos_embed_flatten,
+            level_start_index,
+            valid_ratios,
+            spatial_shapes,
+        ) = self._prepare_multilevel_features(src, src_key_padding_masks, pos)
+
+        reference_points = self.get_reference_points(
+            spatial_shapes, valid_ratios, device=src_flatten.device
+        )
+
+        output = src_flatten
+        for layer in self.layers:
+            layer_kwargs = {}
+
+            assert isinstance(layer, TransformerEncoderLayer)
+            layer_kwargs["memory"] = prompt
+            layer_kwargs["memory_key_padding_mask"] = prompt_key_padding_mask
+            layer_kwargs["query_pos"] = lvl_pos_embed_flatten
+            layer_kwargs["tgt"] = output
+            layer_kwargs["tgt_key_padding_mask"] = key_padding_masks_flatten
+
+            if self.training:
+                assert self.use_act_checkpoint, "activation ckpt not enabled in encoder"
+            if encoder_extra_kwargs is not None:
+                layer_kwargs.update(encoder_extra_kwargs)
+            output = activation_ckpt_wrapper(layer)(
+                **layer_kwargs,
+                act_ckpt_enable=self.training and self.use_act_checkpoint,
+            )
+        # return as seq first
+        return (
+            output.transpose(0, 1),
+            (
+                key_padding_masks_flatten.transpose(0, 1)
+                if key_padding_masks_flatten is not None
+                else None
+            ),
+            lvl_pos_embed_flatten.transpose(0, 1),
+            level_start_index,
+            spatial_shapes,
+            valid_ratios,
+        )
+
+
+class TransformerEncoderFusion(TransformerEncoder):
+    """
+    Transformer encoder that fuses text and image features.
+
+    This encoder extends TransformerEncoder to handle both text and image features,
+    with the ability to add pooled text features to image features for better
+    cross-modal fusion. It supports torch.compile for performance optimization.
+
+    Args:
+        layer: The encoder layer to be stacked multiple times
+        num_layers: Number of encoder layers to stack
+        d_model: Model dimension/hidden size
+        num_feature_levels: Number of feature levels to process
+        add_pooled_text_to_img_feat: Whether to add pooled text features to image features
+        pool_text_with_mask: Whether to use the mask when pooling text features
+        compile_mode: Mode for torch.compile, or None to disable compilation
+        **kwargs: Additional arguments to pass to the parent class
+    """
+
+    def __init__(
+        self,
+        layer: nn.Module,
+        num_layers: int,
+        d_model: int,
+        num_feature_levels: int,
+        add_pooled_text_to_img_feat: bool = True,
+        pool_text_with_mask: bool = False,
+        compile_mode: Optional[str] = None,
+        **kwargs,
+    ):
+        super().__init__(
+            layer,
+            num_layers,
+            d_model,
+            num_feature_levels,
+            **kwargs,
+        )
+        self.add_pooled_text_to_img_feat = add_pooled_text_to_img_feat
+        if self.add_pooled_text_to_img_feat:
+            self.text_pooling_proj = nn.Linear(d_model, d_model)
+        self.pool_text_with_mask = pool_text_with_mask
+        if compile_mode is not None:
+            self.forward = torch.compile(
+                self.forward, mode=compile_mode, fullgraph=True
+            )
+
+    @staticmethod
+    def get_reference_points(spatial_shapes, valid_ratios, device):
+        # Not needed here
+        return None
+
+    def forward(
+        self,
+        src: List[Tensor],
+        prompt: Tensor,
+        src_key_padding_mask: Optional[List[Tensor]] = None,
+        src_pos: Optional[List[Tensor]] = None,
+        prompt_key_padding_mask: Optional[Tensor] = None,
+        prompt_pos: Optional[Tensor] = None,
+        feat_sizes: Optional[List[int]] = None,
+        encoder_extra_kwargs: Optional[Dict] = None,
+    ):
+        # Restore spatial shapes of vision
+        bs = src[0].shape[1]  # seq first
+        if feat_sizes is not None:
+            assert len(feat_sizes) == len(src)
+            if src_key_padding_mask is None:
+                src_key_padding_mask = [None] * len(src)
+            for i, (h, w) in enumerate(feat_sizes):
+                src[i] = src[i].reshape(h, w, bs, -1).permute(2, 3, 0, 1)
+                src_pos[i] = src_pos[i].reshape(h, w, bs, -1).permute(2, 3, 0, 1)
+                src_key_padding_mask[i] = (
+                    src_key_padding_mask[i].reshape(h, w, bs).permute(2, 0, 1)
+                    if src_key_padding_mask[i] is not None
+                    else None
+                )
+        else:
+            assert all(
+                x.dim == 4 for x in src
+            ), "expected list of (bs, c, h, w) tensors"
+
+        if self.add_pooled_text_to_img_feat:
+            # Fusion: Add mean pooled text to image features
+            pooled_text = pool_text_feat(
+                prompt, prompt_key_padding_mask, self.pool_text_with_mask
+            )
+            pooled_text = self.text_pooling_proj(pooled_text)[
+                ..., None, None
+            ]  # prompt is seq first
+            src = [x.add_(pooled_text) for x in src]
+
+        (
+            out,
+            key_padding_masks_flatten,
+            lvl_pos_embed_flatten,
+            level_start_index,
+            spatial_shapes,
+            valid_ratios,
+        ) = super().forward(
+            src,
+            src_key_padding_masks=src_key_padding_mask,
+            pos=src_pos,
+            prompt=prompt.transpose(0, 1),
+            prompt_key_padding_mask=prompt_key_padding_mask,
+            encoder_extra_kwargs=encoder_extra_kwargs,
+        )
+
+        return {
+            "memory": out,
+            "padding_mask": key_padding_masks_flatten,
+            "pos_embed": lvl_pos_embed_flatten,
+            "memory_text": prompt,
+            "level_start_index": level_start_index,
+            "spatial_shapes": spatial_shapes,
+            "valid_ratios": valid_ratios,
+        }
+
+
+def pool_text_feat(prompt, prompt_mask, pool_with_mask):
+    # prompt has shape (seq, bs, dim)
+    if not pool_with_mask:
+        return prompt.mean(dim=0)
+
+    # prompt_mask has shape (bs, seq), where False is valid and True is padding
+    assert prompt_mask.dim() == 2
+    # is_valid has shape (seq, bs, 1), where 1 is valid and 0 is padding
+    is_valid = (~prompt_mask).float().permute(1, 0)[..., None]
+    # num_valid has shape (bs, 1)
+    num_valid = torch.clamp(torch.sum(is_valid, dim=0), min=1.0)
+
+    # mean pool over all the valid tokens
+    pooled_text = (prompt * is_valid).sum(dim=0) / num_valid
+    return pooled_text

detect_tools/sam3/sam3/model/geometry_encoders.py

@@ -0,0 +1,850 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates. All Rights Reserved
+
+from typing import Tuple
+
+import torch
+import torch.nn as nn
+import torchvision
+from typing_extensions import override
+
+from .act_ckpt_utils import activation_ckpt_wrapper
+from .box_ops import box_cxcywh_to_xyxy
+
+from .model_misc import get_clones
+
+
+def is_right_padded(mask):
+    """Given a padding mask (following pytorch convention, 1s for padded values),
+    returns whether the padding is on the right or not."""
+    return (mask.long() == torch.sort(mask.long(), dim=-1)[0]).all()
+
+
+def concat_padded_sequences(seq1, mask1, seq2, mask2, return_index: bool = False):
+    """
+    Concatenates two right-padded sequences, such that the resulting sequence
+    is contiguous and also right-padded.
+
+    Following pytorch's convention, tensors are sequence first, and the mask are
+    batch first, with 1s for padded values.
+
+    :param seq1: A tensor of shape (seq1_length, batch_size, hidden_size).
+    :param mask1: A tensor of shape (batch_size, seq1_length).
+    :param seq2: A tensor of shape (seq2_length, batch_size,  hidden_size).
+    :param mask2: A tensor of shape (batch_size, seq2_length).
+    :param return_index: If True, also returns the index of the ids of the element of seq2
+        in the concatenated sequence. This can be used to retrieve the elements of seq2
+    :return: A tuple (concatenated_sequence, concatenated_mask) if return_index is False,
+        otherwise (concatenated_sequence, concatenated_mask, index).
+    """
+    seq1_length, batch_size, hidden_size = seq1.shape
+    seq2_length, batch_size, hidden_size = seq2.shape
+
+    assert batch_size == seq1.size(1) == seq2.size(1) == mask1.size(0) == mask2.size(0)
+    assert hidden_size == seq1.size(2) == seq2.size(2)
+    assert seq1_length == mask1.size(1)
+    assert seq2_length == mask2.size(1)
+
+    torch._assert_async(is_right_padded(mask1))
+    torch._assert_async(is_right_padded(mask2))
+
+    actual_seq1_lengths = (~mask1).sum(dim=-1)
+    actual_seq2_lengths = (~mask2).sum(dim=-1)
+
+    final_lengths = actual_seq1_lengths + actual_seq2_lengths
+    max_length = seq1_length + seq2_length
+    concatenated_mask = (
+        torch.arange(max_length, device=seq2.device)[None].repeat(batch_size, 1)
+        >= final_lengths[:, None]
+    )
+
+    # (max_len, batch_size, hidden_size)
+    concatenated_sequence = torch.zeros(
+        (max_length, batch_size, hidden_size), device=seq2.device, dtype=seq2.dtype
+    )
+    concatenated_sequence[:seq1_length, :, :] = seq1
+
+    # At this point, the element of seq1 are in the right place
+    # We just need to shift the elements of seq2
+
+    index = torch.arange(seq2_length, device=seq2.device)[:, None].repeat(1, batch_size)
+    index = index + actual_seq1_lengths[None]
+
+    concatenated_sequence = concatenated_sequence.scatter(
+        0, index[:, :, None].expand(-1, -1, hidden_size), seq2
+    )
+
+    if return_index:
+        return concatenated_sequence, concatenated_mask, index
+
+    return concatenated_sequence, concatenated_mask
+
+
+class Prompt:
+    """Utility class to manipulate geometric prompts.
+
+    We expect the sequences in pytorch convention, that is sequence first, batch second
+    The dimensions are expected as follows:
+    box_embeddings shape: N_boxes x B x C_box
+    box_mask shape: B x N_boxes. Can be None if nothing is masked out
+    point_embeddings shape: N_points x B x C_point
+    point_mask shape: B x N_points. Can be None if nothing is masked out
+    mask_embeddings shape: N_masks x B x 1 x H_mask x W_mask
+    mask_mask shape: B x N_masks. Can be None if nothing is masked out
+
+    We also store positive/negative labels. These tensors are also stored batch-first
+    If they are None, we'll assume positive labels everywhere
+    box_labels: long tensor of shape N_boxes x B
+    point_labels: long tensor of shape N_points x B
+    mask_labels: long tensor of shape N_masks x B
+    """
+
+    def __init__(
+        self,
+        box_embeddings=None,
+        box_mask=None,
+        point_embeddings=None,
+        point_mask=None,
+        box_labels=None,
+        point_labels=None,
+        mask_embeddings=None,
+        mask_mask=None,  # Attention mask for mask prompt
+        mask_labels=None,
+    ):
+        # Check for null prompt
+        if (
+            box_embeddings is None
+            and point_embeddings is None
+            and mask_embeddings is None
+        ):
+            self.box_embeddings = None
+            self.box_labels = None
+            self.box_mask = None
+            self.point_embeddings = None
+            self.point_labels = None
+            self.point_mask = None
+            self.mask_embeddings = None
+            self.mask_mask = None
+            # Masks are assumed positive only for now.
+            self.mask_labels = None
+            return
+        # Get sequence lengths and device
+        box_seq_len, point_seq_len, mask_seq_len, bs, device = (
+            self._init_seq_len_and_device(
+                box_embeddings, point_embeddings, mask_embeddings
+            )
+        )
+
+        # Initialize embeds, labels, attention masks.
+        box_embeddings, box_labels, box_mask = self._init_box(
+            box_embeddings, box_labels, box_mask, box_seq_len, bs, device
+        )
+        point_embeddings, point_labels, point_mask = self._init_point(
+            point_embeddings, point_labels, point_mask, point_seq_len, bs, device
+        )
+        mask_embeddings, mask_labels, mask_mask = self._init_mask(
+            mask_embeddings, mask_labels, mask_mask, mask_seq_len, bs, device
+        )
+
+        # Dimension checks
+        assert (
+            box_embeddings is not None
+            and list(box_embeddings.shape[:2])
+            == [
+                box_seq_len,
+                bs,
+            ]
+        ), f"Wrong dimension for box embeddings. Expected [{box_seq_len}, {bs}, *] got {box_embeddings.shape}"
+        assert (
+            box_mask is not None
+            and list(box_mask.shape)
+            == [
+                bs,
+                box_seq_len,
+            ]
+        ), f"Wrong dimension for box mask. Expected [{bs}, {box_seq_len}] got {box_mask.shape}"
+        assert (
+            point_embeddings is not None
+            and list(point_embeddings.shape[:2])
+            == [
+                point_seq_len,
+                bs,
+            ]
+        ), f"Wrong dimension for point embeddings. Expected [{point_seq_len}, {bs}, *] got {point_embeddings.shape}"
+        assert (
+            point_mask is not None
+            and list(point_mask.shape)
+            == [
+                bs,
+                point_seq_len,
+            ]
+        ), f"Wrong dimension for point mask. Expected [{bs}, {point_seq_len}] got {point_mask.shape}"
+        assert (
+            box_labels is not None
+            and list(box_labels.shape)
+            == [
+                box_seq_len,
+                bs,
+            ]
+        ), f"Wrong dimension for box labels. Expected [{box_seq_len}, {bs}] got {box_labels.shape}"
+        assert (
+            point_labels is not None
+            and list(point_labels.shape)
+            == [
+                point_seq_len,
+                bs,
+            ]
+        ), f"Wrong dimension for point labels. Expected [{point_seq_len}, {bs}] got {point_labels.shape}"
+        assert (
+            # Allowed to be None, we leave it to the encoder to check for validity before encoding.
+            mask_embeddings is None
+            or list(mask_embeddings.shape[:2])
+            == [
+                mask_seq_len,
+                bs,
+            ]
+        ), f"Wrong dimension for mask embeddings. Expected [{mask_seq_len}, {bs}, *] got {mask_embeddings.shape}"
+        assert (
+            mask_mask is None
+            or list(mask_mask.shape)
+            == [
+                bs,
+                mask_seq_len,
+            ]
+        ), f"Wrong dimension for mask attn. mask. Expected [{bs}, {mask_seq_len}] got {mask_mask.shape}"
+
+        # Device checks
+        assert (
+            box_embeddings is not None and box_embeddings.device == device
+        ), f"Expected box embeddings to be on device {device}, got {box_embeddings.device}"
+        assert (
+            box_mask is not None and box_mask.device == device
+        ), f"Expected box mask to be on device {device}, got {box_mask.device}"
+        assert (
+            box_labels is not None and box_labels.device == device
+        ), f"Expected box labels to be on device {device}, got {box_labels.device}"
+        assert (
+            point_embeddings is not None and point_embeddings.device == device
+        ), f"Expected point embeddings to be on device {device}, got {point_embeddings.device}"
+        assert (
+            point_mask is not None and point_mask.device == device
+        ), f"Expected point mask to be on device {device}, got {point_mask.device}"
+        assert (
+            point_labels is not None and point_labels.device == device
+        ), f"Expected point labels to be on device {device}, got {point_labels.device}"
+        assert (
+            mask_embeddings is None or mask_embeddings.device == device
+        ), f"Expected mask embeddings to be on device {device}, got {mask_embeddings.device}"
+        assert (
+            mask_mask is None or mask_mask.device == device
+        ), f"Expected mask attn. mask to be on device {device}, got {mask_mask.device}"
+
+        self.box_embeddings = box_embeddings
+        self.point_embeddings = point_embeddings
+        self.box_mask = box_mask
+        self.point_mask = point_mask
+        self.box_labels = box_labels
+        self.point_labels = point_labels
+        self.mask_embeddings = mask_embeddings
+        self.mask_labels = mask_labels
+        self.mask_mask = mask_mask
+
+    def _init_seq_len_and_device(
+        self, box_embeddings, point_embeddings, mask_embeddings
+    ):
+        box_seq_len = point_seq_len = mask_seq_len = 0
+        bs = None
+        device = None
+        if box_embeddings is not None:
+            bs = box_embeddings.shape[1]
+            box_seq_len = box_embeddings.shape[0]
+            device = box_embeddings.device
+
+        if point_embeddings is not None:
+            point_seq_len = point_embeddings.shape[0]
+            if bs is not None:
+                assert (
+                    bs == point_embeddings.shape[1]
+                ), f"Batch size mismatch between box and point embeddings. Got {bs} and {point_embeddings.shape[1]}."
+            else:
+                bs = point_embeddings.shape[1]
+            if device is not None:
+                assert (
+                    device == point_embeddings.device
+                ), "Device mismatch between box and point embeddings"
+            else:
+                device = point_embeddings.device
+
+        if mask_embeddings is not None:
+            mask_seq_len = mask_embeddings.shape[0]
+            if bs is not None:
+                assert (
+                    bs == mask_embeddings.shape[1]
+                ), f"Batch size mismatch between box/point and mask embedding. Got {bs} and {mask_embeddings.shape[1]}"
+            else:
+                bs = mask_embeddings.shape[1]
+            if device is not None:
+                assert (
+                    device == mask_embeddings.device
+                ), "Device mismatch between box/point and mask embeddings."
+            else:
+                device = mask_embeddings.device
+
+        return box_seq_len, point_seq_len, mask_seq_len, bs, device
+
+    def _init_box(self, box_embeddings, box_labels, box_mask, box_seq_len, bs, device):
+        if box_embeddings is None:
+            box_embeddings = torch.zeros(box_seq_len, bs, 4, device=device)
+        if box_labels is None:
+            box_labels = torch.ones(box_seq_len, bs, device=device, dtype=torch.long)
+        if box_mask is None:
+            box_mask = torch.zeros(bs, box_seq_len, device=device, dtype=torch.bool)
+        return box_embeddings, box_labels, box_mask
+
+    def _init_point(
+        self, point_embeddings, point_labels, point_mask, point_seq_len, bs, device
+    ):
+        """
+        Identical to _init_box. Except that C=2 for points (vs. 4 for boxes).
+        """
+        if point_embeddings is None:
+            point_embeddings = torch.zeros(point_seq_len, bs, 2, device=device)
+        if point_labels is None:
+            point_labels = torch.ones(
+                point_seq_len, bs, device=device, dtype=torch.long
+            )
+        if point_mask is None:
+            point_mask = torch.zeros(bs, point_seq_len, device=device, dtype=torch.bool)
+        return point_embeddings, point_labels, point_mask
+
+    def _init_mask(
+        self, mask_embeddings, mask_labels, mask_mask, mask_seq_len, bs, device
+    ):
+        # NOTE: Mask embeddings can be of arbitrary resolution, so we don't initialize it here.
+        # In case we append new mask, we check that its resolution matches exisiting ones (if any).
+        # In case mask_embeddings is None, we should never encode it.
+        if mask_labels is None:
+            mask_labels = torch.ones(mask_seq_len, bs, device=device, dtype=torch.long)
+        if mask_mask is None:
+            mask_mask = torch.zeros(bs, mask_seq_len, device=device, dtype=torch.bool)
+        return mask_embeddings, mask_labels, mask_mask
+
+    def append_boxes(self, boxes, labels, mask=None):
+        if self.box_embeddings is None:
+            self.box_embeddings = boxes
+            self.box_labels = labels
+            self.box_mask = mask
+            return
+
+        bs = self.box_embeddings.shape[1]
+        assert boxes.shape[1] == labels.shape[1] == bs
+        assert list(boxes.shape[:2]) == list(labels.shape[:2])
+        if mask is None:
+            mask = torch.zeros(
+                bs, boxes.shape[0], dtype=torch.bool, device=boxes.device
+            )
+
+        self.box_labels, _ = concat_padded_sequences(
+            self.box_labels.unsqueeze(-1), self.box_mask, labels.unsqueeze(-1), mask
+        )
+        self.box_labels = self.box_labels.squeeze(-1)
+        self.box_embeddings, self.box_mask = concat_padded_sequences(
+            self.box_embeddings, self.box_mask, boxes, mask
+        )
+
+    def append_points(self, points, labels, mask=None):
+        if self.point_embeddings is None:
+            self.point_embeddings = points
+            self.point_labels = labels
+            self.point_mask = mask
+            return
+
+        bs = self.point_embeddings.shape[1]
+        assert points.shape[1] == labels.shape[1] == bs
+        assert list(points.shape[:2]) == list(labels.shape[:2])
+        if mask is None:
+            mask = torch.zeros(
+                bs, points.shape[0], dtype=torch.bool, device=points.device
+            )
+
+        self.point_labels, _ = concat_padded_sequences(
+            self.point_labels.unsqueeze(-1), self.point_mask, labels.unsqueeze(-1), mask
+        )
+        self.point_labels = self.point_labels.squeeze(-1)
+        self.point_embeddings, self.point_mask = concat_padded_sequences(
+            self.point_embeddings, self.point_mask, points, mask
+        )
+
+    def append_masks(self, masks, labels=None, attn_mask=None):
+        if labels is not None:
+            assert list(masks.shape[:2]) == list(labels.shape[:2])
+        if self.mask_embeddings is None:
+            self.mask_embeddings = masks
+            mask_seq_len, bs = masks.shape[:2]
+            if labels is None:
+                self.mask_labels = torch.ones(
+                    mask_seq_len, bs, device=masks.device, dtype=torch.long
+                )
+            else:
+                self.mask_labels = labels
+            if attn_mask is None:
+                self.mask_mask = torch.zeros(
+                    bs, mask_seq_len, device=masks.device, dtype=torch.bool
+                )
+            else:
+                self.mask_mask = attn_mask
+        else:
+            raise NotImplementedError("Only one mask per prompt is supported.")
+
+    def clone(self):
+        return Prompt(
+            box_embeddings=(
+                None if self.box_embeddings is None else self.box_embeddings.clone()
+            ),
+            box_mask=None if self.box_mask is None else self.box_mask.clone(),
+            point_embeddings=(
+                None if self.point_embeddings is None else self.point_embeddings.clone()
+            ),
+            point_mask=None if self.point_mask is None else self.point_mask.clone(),
+            box_labels=None if self.box_labels is None else self.box_labels.clone(),
+            point_labels=(
+                None if self.point_labels is None else self.point_labels.clone()
+            ),
+        )
+
+
+class MaskEncoder(nn.Module):
+    """
+    Base class for mask encoders.
+    """
+
+    def __init__(
+        self,
+        mask_downsampler: nn.Module,
+        position_encoding: nn.Module,
+    ):
+        super().__init__()
+        self.mask_downsampler = mask_downsampler
+        self.position_encoding = position_encoding
+
+    def forward(self, masks, *args, **kwargs) -> Tuple[torch.Tensor, torch.Tensor]:
+        masks = self.mask_downsampler(masks)
+        masks_pos = self.position_encoding(masks).to(masks.dtype)
+
+        return masks, masks_pos
+
+
+class FusedMaskEncoder(MaskEncoder):
+    """
+    Identical to memory.SimpleMaskEncoder but follows the interface of geometry_encoders.MaskEncoder.
+    We also remove the `skip_mask_sigmoid` option (to be handled outside the MaskEncoder).
+    Fuses backbone image features with mask features.
+    """
+
+    def __init__(
+        self,
+        mask_downsampler: nn.Module,
+        position_encoding: nn.Module,
+        fuser: nn.Module,
+        in_dim: int = 256,
+        out_dim: int = 256,
+    ):
+        super().__init__(mask_downsampler, position_encoding)
+        self.fuser = fuser
+        self.out_proj = nn.Identity()
+        if out_dim != in_dim:
+            self.out_proj = nn.Conv2d(in_dim, out_dim, kernel_size=1)
+        self.pix_feat_proj = nn.Conv2d(in_dim, in_dim, kernel_size=1)
+
+    @override
+    def forward(
+        self,
+        masks: torch.Tensor,
+        pix_feat: torch.Tensor,
+        **kwargs,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        masks = self.mask_downsampler(masks)
+
+        ## Fuse pix_feats and downsampled masks
+        # in case the visual features are on CPU, cast them to CUDA
+        pix_feat = pix_feat.to(masks.device)
+
+        x = self.pix_feat_proj(pix_feat)
+        x = x + masks
+        x = self.fuser(x)
+        x = self.out_proj(x)
+
+        pos = self.position_encoding(x).to(x.dtype)
+
+        return x, pos
+
+
+class SequenceGeometryEncoder(nn.Module):
+    """
+    This a fully fledged encoder for geometric prompts.
+    It assumes boxes are passed in the "normalized CxCyWH" format, and points in normalized xy
+    This allows flexibility in how to encode the features (eg do pooling)
+
+    Points and boxes can be encoded with any of the three possibilities:
+     - direct projection: we just compute a linear from coordinate space to d_model
+     - pooling: pool features from the backbone in the requested location.
+                For boxes, it's a roi align
+                For points it's a grid sample
+     - pos encoder: Take the position encoding of the point or box center
+
+    These three options are mutually compatible. If several are selected, we'll take a simple addition
+
+    As an alternative, we offer the possibility to encode points only.
+    In that case, the boxes are converted to two points for the top left and bottom right corners (with appropriate labels)
+
+    On top of these encodings, we offer the possibility to further encode the prompt sequence with a transformer.
+    """
+
+    def __init__(
+        self,
+        encode_boxes_as_points: bool,
+        points_direct_project: bool,
+        points_pool: bool,
+        points_pos_enc: bool,
+        boxes_direct_project: bool,
+        boxes_pool: bool,
+        boxes_pos_enc: bool,
+        d_model: int,
+        pos_enc,
+        num_layers: int,
+        layer: nn.Module,
+        roi_size: int = 7,  # for boxes pool
+        add_cls: bool = True,
+        add_post_encode_proj: bool = True,
+        mask_encoder: MaskEncoder = None,
+        add_mask_label: bool = False,
+        use_act_ckpt: bool = False,
+    ):
+        super().__init__()
+
+        self.d_model = d_model
+        self.pos_enc = pos_enc
+        self.encode_boxes_as_points = encode_boxes_as_points
+        self.roi_size = roi_size
+        # There usually are two labels: positive and negatives.
+        # If we encode boxes as points, we have 3 types of points: regular, top left, bottom right
+        # These 3 types can be positives or negatives, hence 2*3 = 6 labels
+        num_labels = 6 if self.encode_boxes_as_points else 2
+        self.label_embed = torch.nn.Embedding(num_labels, self.d_model)
+
+        # This is a cls token, can be used for pooling if need be.
+        # It also ensures that the encoded sequences are always non-empty
+        self.cls_embed = None
+        if add_cls:
+            self.cls_embed = torch.nn.Embedding(1, self.d_model)
+
+        assert (
+            points_direct_project or points_pos_enc or points_pool
+        ), "Error: need at least one way to encode points"
+        assert (
+            encode_boxes_as_points
+            or boxes_direct_project
+            or boxes_pos_enc
+            or boxes_pool
+        ), "Error: need at least one way to encode boxes"
+
+        self.points_direct_project = None
+        if points_direct_project:
+            self.points_direct_project = nn.Linear(2, self.d_model)
+        self.points_pool_project = None
+        if points_pool:
+            self.points_pool_project = nn.Linear(self.d_model, self.d_model)
+        self.points_pos_enc_project = None
+        if points_pos_enc:
+            self.points_pos_enc_project = nn.Linear(self.d_model, self.d_model)
+
+        self.boxes_direct_project = None
+        self.boxes_pool_project = None
+        self.boxes_pos_enc_project = None
+        if not encode_boxes_as_points:
+            if boxes_direct_project:
+                self.boxes_direct_project = nn.Linear(4, self.d_model)
+            if boxes_pool:
+                self.boxes_pool_project = nn.Conv2d(
+                    self.d_model, self.d_model, self.roi_size
+                )
+            if boxes_pos_enc:
+                self.boxes_pos_enc_project = nn.Linear(self.d_model + 2, self.d_model)
+
+        self.final_proj = None
+        if add_post_encode_proj:
+            self.final_proj = nn.Linear(self.d_model, self.d_model)
+            self.norm = nn.LayerNorm(self.d_model)
+
+        self.img_pre_norm = nn.Identity()
+        if self.points_pool_project is not None or self.boxes_pool_project is not None:
+            self.img_pre_norm = nn.LayerNorm(self.d_model)
+
+        self.encode = None
+        if num_layers > 0:
+            assert (
+                add_cls
+            ), "It's currently highly recommended to add a CLS when using a transformer"
+            self.encode = get_clones(layer, num_layers)
+            self.encode_norm = nn.LayerNorm(self.d_model)
+
+        if mask_encoder is not None:
+            assert isinstance(
+                mask_encoder, MaskEncoder
+            ), f"Expected mask_encoder of type MaskEncoder. Got {type(mask_encoder)}."
+            if add_mask_label:
+                self.mask_label_embed = torch.nn.Embedding(2, self.d_model)
+        self.add_mask_label = add_mask_label
+        self.mask_encoder = mask_encoder
+        self.use_act_ckpt = use_act_ckpt
+
+    def _encode_points(self, points, points_mask, points_labels, img_feats):
+        points_embed = None
+        n_points, bs = points.shape[:2]
+
+        if self.points_direct_project is not None:
+            proj = self.points_direct_project(points)
+            assert points_embed is None
+            points_embed = proj
+
+        if self.points_pool_project is not None:
+            # points are [Num_points, bs, 2], normalized in [0, 1]
+            # the grid needs to be [Bs, H_out, W_out, 2] normalized in [-1,1]
+            # Will take H_out = num_points, w_out = 1
+            grid = points.transpose(0, 1).unsqueeze(2)
+            # re normalize to [-1, 1]
+            grid = (grid * 2) - 1
+            sampled = torch.nn.functional.grid_sample(
+                img_feats, grid, align_corners=False
+            )
+            assert list(sampled.shape) == [bs, self.d_model, n_points, 1]
+            sampled = sampled.squeeze(-1).permute(2, 0, 1)
+            proj = self.points_pool_project(sampled)
+            if points_embed is None:
+                points_embed = proj
+            else:
+                points_embed = points_embed + proj
+
+        if self.points_pos_enc_project is not None:
+            x, y = points.unbind(-1)
+            enc_x, enc_y = self.pos_enc._encode_xy(x.flatten(), y.flatten())
+            enc_x = enc_x.view(n_points, bs, enc_x.shape[-1])
+            enc_y = enc_y.view(n_points, bs, enc_y.shape[-1])
+            enc = torch.cat([enc_x, enc_y], -1)
+
+            proj = self.points_pos_enc_project(enc)
+            if points_embed is None:
+                points_embed = proj
+            else:
+                points_embed = points_embed + proj
+
+        type_embed = self.label_embed(points_labels.long())
+        return type_embed + points_embed, points_mask
+
+    def _encode_boxes(self, boxes, boxes_mask, boxes_labels, img_feats):
+        boxes_embed = None
+        n_boxes, bs = boxes.shape[:2]
+
+        if self.boxes_direct_project is not None:
+            proj = self.boxes_direct_project(boxes)
+            assert boxes_embed is None
+            boxes_embed = proj
+
+        if self.boxes_pool_project is not None:
+            H, W = img_feats.shape[-2:]
+
+            # boxes are [Num_boxes, bs, 4], normalized in [0, 1]
+            # We need to denormalize, and convert to [x, y, x, y]
+            boxes_xyxy = box_cxcywh_to_xyxy(boxes)
+            scale = torch.tensor([W, H, W, H], dtype=boxes_xyxy.dtype)
+            scale = scale.pin_memory().to(device=boxes_xyxy.device, non_blocking=True)
+            scale = scale.view(1, 1, 4)
+            boxes_xyxy = boxes_xyxy * scale
+            sampled = torchvision.ops.roi_align(
+                img_feats, boxes_xyxy.float().transpose(0, 1).unbind(0), self.roi_size
+            )
+            assert list(sampled.shape) == [
+                bs * n_boxes,
+                self.d_model,
+                self.roi_size,
+                self.roi_size,
+            ]
+            proj = self.boxes_pool_project(sampled)
+            proj = proj.view(bs, n_boxes, self.d_model).transpose(0, 1)
+            if boxes_embed is None:
+                boxes_embed = proj
+            else:
+                boxes_embed = boxes_embed + proj
+
+        if self.boxes_pos_enc_project is not None:
+            cx, cy, w, h = boxes.unbind(-1)
+            enc = self.pos_enc.encode_boxes(
+                cx.flatten(), cy.flatten(), w.flatten(), h.flatten()
+            )
+            enc = enc.view(boxes.shape[0], boxes.shape[1], enc.shape[-1])
+
+            proj = self.boxes_pos_enc_project(enc)
+            if boxes_embed is None:
+                boxes_embed = proj
+            else:
+                boxes_embed = boxes_embed + proj
+
+        type_embed = self.label_embed(boxes_labels.long())
+        return type_embed + boxes_embed, boxes_mask
+
+    def _encode_masks(
+        self,
+        masks: torch.Tensor,
+        attn_mask: torch.Tensor,
+        mask_labels: torch.Tensor,
+        img_feats: torch.Tensor = None,
+    ):
+        n_masks, bs = masks.shape[:2]
+        assert (
+            n_masks == 1
+        ), "We assume one mask per prompt for now. Code should still be functional if this assertion is removed."
+        assert (
+            list(attn_mask.shape)
+            == [
+                bs,
+                n_masks,
+            ]
+        ), f"Expected attn_mask to be of shape {bs}x{n_masks}. Got {list(attn_mask.shape)}."
+        masks, pos = self.mask_encoder(
+            masks=masks.flatten(0, 1).float(),
+            pix_feat=img_feats,
+        )
+        H, W = masks.shape[-2:]
+        n_tokens_per_mask = H * W
+        # NOTE: We directly add pos enc here as we usually don't keep track of pos encoding for the concatenated prompt (text, other geometric prompts). Might need to do some refactoring for more flexibility.
+        masks = masks + pos
+        masks = masks.view(n_masks, bs, *masks.shape[1:]).flatten(
+            -2
+        )  # n_masks x bs x C x H*W
+        masks = masks.permute(0, 3, 1, 2).flatten(0, 1)  # n_masks * H*W x bs x C
+        attn_mask = attn_mask.repeat_interleave(n_tokens_per_mask, dim=1)
+        if self.add_mask_label:
+            masks = masks + self.mask_label_embed(mask_labels.long())
+        return masks, attn_mask
+
+    def forward(self, geo_prompt: Prompt, img_feats, img_sizes, img_pos_embeds=None):
+        points = geo_prompt.point_embeddings
+        points_mask = geo_prompt.point_mask
+        points_labels = geo_prompt.point_labels
+        boxes = geo_prompt.box_embeddings
+        boxes_mask = geo_prompt.box_mask
+        boxes_labels = geo_prompt.box_labels
+        masks = geo_prompt.mask_embeddings
+        masks_mask = geo_prompt.mask_mask
+        masks_labels = geo_prompt.mask_labels
+        seq_first_img_feats = img_feats[-1]  # [H*W, B, C]
+        seq_first_img_pos_embeds = (
+            img_pos_embeds[-1]
+            if img_pos_embeds is not None
+            else torch.zeros_like(seq_first_img_feats)
+        )
+
+        if self.points_pool_project or self.boxes_pool_project:
+            assert len(img_feats) == len(img_sizes)
+            cur_img_feat = img_feats[-1]
+            cur_img_feat = self.img_pre_norm(cur_img_feat)
+            H, W = img_sizes[-1]
+            assert cur_img_feat.shape[0] == H * W
+            N, C = cur_img_feat.shape[-2:]
+            # Put back in NxCxHxW
+            cur_img_feat = cur_img_feat.permute(1, 2, 0)
+            cur_img_feat = cur_img_feat.view(N, C, H, W)
+            img_feats = cur_img_feat
+
+        if self.encode_boxes_as_points:
+            assert boxes is not None
+            assert geo_prompt.box_mask is not None
+            assert geo_prompt.box_labels is not None
+            assert boxes.shape[-1] == 4
+
+            boxes_xyxy = box_cxcywh_to_xyxy(boxes)
+            top_left, bottom_right = boxes_xyxy.split(split_size=2, dim=-1)
+
+            labels_tl = geo_prompt.box_labels + 2
+            labels_br = geo_prompt.box_labels + 4
+
+            # Append to the existing points
+            points, _ = concat_padded_sequences(
+                points, points_mask, top_left, boxes_mask
+            )
+            points_labels, points_mask = concat_padded_sequences(
+                points_labels.unsqueeze(-1),
+                points_mask,
+                labels_tl.unsqueeze(-1),
+                boxes_mask,
+            )
+            points_labels = points_labels.squeeze(-1)
+
+            points, _ = concat_padded_sequences(
+                points, points_mask, bottom_right, boxes_mask
+            )
+            points_labels, points_mask = concat_padded_sequences(
+                points_labels.unsqueeze(-1),
+                points_mask,
+                labels_br.unsqueeze(-1),
+                boxes_mask,
+            )
+            points_labels = points_labels.squeeze(-1)
+
+        final_embeds, final_mask = self._encode_points(
+            points=points,
+            points_mask=points_mask,
+            points_labels=points_labels,
+            img_feats=img_feats,
+        )
+
+        if not self.encode_boxes_as_points:
+            boxes_embeds, boxes_mask = self._encode_boxes(
+                boxes=boxes,
+                boxes_mask=boxes_mask,
+                boxes_labels=boxes_labels,
+                img_feats=img_feats,
+            )
+
+            final_embeds, final_mask = concat_padded_sequences(
+                final_embeds, final_mask, boxes_embeds, boxes_mask
+            )
+
+        if masks is not None and self.mask_encoder is not None:
+            masks_embed, masks_mask = self._encode_masks(
+                masks=masks,
+                attn_mask=masks_mask,
+                mask_labels=masks_labels,
+                img_feats=img_feats,
+            )
+            if points.size(0) == boxes.size(0) == 0:
+                return masks_embed, masks_mask
+        bs = final_embeds.shape[1]
+        assert final_mask.shape[0] == bs
+        if self.cls_embed is not None:
+            cls = self.cls_embed.weight.view(1, 1, self.d_model).repeat(1, bs, 1)
+            cls_mask = torch.zeros(
+                bs, 1, dtype=final_mask.dtype, device=final_mask.device
+            )
+            final_embeds, final_mask = concat_padded_sequences(
+                final_embeds, final_mask, cls, cls_mask
+            )
+
+        if self.final_proj is not None:
+            final_embeds = self.norm(self.final_proj(final_embeds))
+
+        if self.encode is not None:
+            for lay in self.encode:
+                final_embeds = activation_ckpt_wrapper(lay)(
+                    tgt=final_embeds,
+                    memory=seq_first_img_feats,
+                    tgt_key_padding_mask=final_mask,
+                    pos=seq_first_img_pos_embeds,
+                    act_ckpt_enable=self.training and self.use_act_ckpt,
+                )
+            final_embeds = self.encode_norm(final_embeds)
+        # Finally, concat mask embeddings if any
+        if masks is not None and self.mask_encoder is not None:
+            final_embeds, final_mask = concat_padded_sequences(
+                final_embeds, final_mask, masks_embed, masks_mask
+            )
+        return final_embeds, final_mask

detect_tools/sam3/sam3/model/io_utils.py

@@ -0,0 +1,709 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates. All Rights Reserved
+
+import contextlib
+import os
+import queue
+import re
+import time
+from threading import Condition, get_ident, Lock, Thread
+
+import numpy as np
+import torch
+import torch.nn.functional as F
+import torchvision.transforms.functional as TF
+
+from PIL import Image
+
+from sam3.logger import get_logger
+from tqdm import tqdm
+
+logger = get_logger(__name__)
+
+IS_MAIN_PROCESS = os.getenv("IS_MAIN_PROCESS", "1") == "1"
+RANK = int(os.getenv("RANK", "0"))
+
+IMAGE_EXTS = [".jpg", ".jpeg", ".png", ".bmp", ".tiff", ".webp"]
+VIDEO_EXTS = [".mp4", ".mov", ".avi", ".mkv", ".webm"]
+
+
+def load_resource_as_video_frames(
+    resource_path,
+    image_size,
+    offload_video_to_cpu,
+    img_mean=(0.5, 0.5, 0.5),
+    img_std=(0.5, 0.5, 0.5),
+    async_loading_frames=False,
+    video_loader_type="cv2",
+):
+    """
+    Load video frames from either a video or an image (as a single-frame video).
+    Alternatively, if input is a list of PIL images, convert its format
+    """
+    if isinstance(resource_path, list):
+        img_mean = torch.tensor(img_mean, dtype=torch.float16)[:, None, None]
+        img_std = torch.tensor(img_std, dtype=torch.float16)[:, None, None]
+        assert all(isinstance(img_pil, Image.Image) for img_pil in resource_path)
+        assert len(resource_path) is not None
+        orig_height, orig_width = resource_path[0].size
+        orig_height, orig_width = (
+            orig_width,
+            orig_height,
+        )  # For some reason, this method returns these swapped
+        images = []
+        for img_pil in resource_path:
+            img_np = np.array(img_pil.convert("RGB").resize((image_size, image_size)))
+            assert img_np.dtype == np.uint8, "np.uint8 is expected for JPEG images"
+            img_np = img_np / 255.0
+            img = torch.from_numpy(img_np).permute(2, 0, 1)
+            # float16 precision should be sufficient for image tensor storage
+            img = img.to(dtype=torch.float16)
+            # normalize by mean and std
+            img -= img_mean
+            img /= img_std
+            images.append(img)
+        images = torch.stack(images)
+        if not offload_video_to_cpu:
+            images = images.cuda()
+        return images, orig_height, orig_width
+
+    is_image = (
+        isinstance(resource_path, str)
+        and os.path.splitext(resource_path)[-1].lower() in IMAGE_EXTS
+    )
+    if is_image:
+        return load_image_as_single_frame_video(
+            image_path=resource_path,
+            image_size=image_size,
+            offload_video_to_cpu=offload_video_to_cpu,
+            img_mean=img_mean,
+            img_std=img_std,
+        )
+    else:
+        return load_video_frames(
+            video_path=resource_path,
+            image_size=image_size,
+            offload_video_to_cpu=offload_video_to_cpu,
+            img_mean=img_mean,
+            img_std=img_std,
+            async_loading_frames=async_loading_frames,
+            video_loader_type=video_loader_type,
+        )
+
+
+def load_image_as_single_frame_video(
+    image_path,
+    image_size,
+    offload_video_to_cpu,
+    img_mean=(0.5, 0.5, 0.5),
+    img_std=(0.5, 0.5, 0.5),
+):
+    """Load an image as a single-frame video."""
+    images, image_height, image_width = _load_img_as_tensor(image_path, image_size)
+    images = images.unsqueeze(0).half()
+
+    img_mean = torch.tensor(img_mean, dtype=torch.float16)[:, None, None]
+    img_std = torch.tensor(img_std, dtype=torch.float16)[:, None, None]
+    if not offload_video_to_cpu:
+        images = images.cuda()
+        img_mean = img_mean.cuda()
+        img_std = img_std.cuda()
+    # normalize by mean and std
+    images -= img_mean
+    images /= img_std
+    return images, image_height, image_width
+
+
+def load_video_frames(
+    video_path,
+    image_size,
+    offload_video_to_cpu,
+    img_mean=(0.5, 0.5, 0.5),
+    img_std=(0.5, 0.5, 0.5),
+    async_loading_frames=False,
+    video_loader_type="cv2",
+):
+    """
+    Load the video frames from video_path. The frames are resized to image_size as in
+    the model and are loaded to GPU if offload_video_to_cpu=False. This is used by the demo.
+    """
+    assert isinstance(video_path, str)
+    if video_path.startswith("<load-dummy-video"):
+        # Check for pattern <load-dummy-video-N> where N is an integer
+        match = re.match(r"<load-dummy-video-(\d+)>", video_path)
+        num_frames = int(match.group(1)) if match else 60
+        return load_dummy_video(image_size, offload_video_to_cpu, num_frames=num_frames)
+    elif os.path.isdir(video_path):
+        return load_video_frames_from_image_folder(
+            image_folder=video_path,
+            image_size=image_size,
+            offload_video_to_cpu=offload_video_to_cpu,
+            img_mean=img_mean,
+            img_std=img_std,
+            async_loading_frames=async_loading_frames,
+        )
+    elif os.path.splitext(video_path)[-1].lower() in VIDEO_EXTS:
+        return load_video_frames_from_video_file(
+            video_path=video_path,
+            image_size=image_size,
+            offload_video_to_cpu=offload_video_to_cpu,
+            img_mean=img_mean,
+            img_std=img_std,
+            async_loading_frames=async_loading_frames,
+            video_loader_type=video_loader_type,
+        )
+    else:
+        raise NotImplementedError("Only video files and image folders are supported")
+
+
+def load_video_frames_from_image_folder(
+    image_folder,
+    image_size,
+    offload_video_to_cpu,
+    img_mean,
+    img_std,
+    async_loading_frames,
+):
+    """
+    Load the video frames from a directory of image files ("<frame_index>.<img_ext>" format)
+    """
+    frame_names = [
+        p
+        for p in os.listdir(image_folder)
+        if os.path.splitext(p)[-1].lower() in IMAGE_EXTS
+    ]
+    try:
+        frame_names.sort(key=lambda p: int(os.path.splitext(p)[0]))
+    except ValueError:
+        # fallback to lexicographic sort if the format is not "<frame_index>.<img_ext>"
+        logger.warning(
+            f'frame names are not in "<frame_index>.<img_ext>" format: {frame_names[:5]=}, '
+            f"falling back to lexicographic sort."
+        )
+        frame_names.sort()
+    num_frames = len(frame_names)
+    if num_frames == 0:
+        raise RuntimeError(f"no images found in {image_folder}")
+    img_paths = [os.path.join(image_folder, frame_name) for frame_name in frame_names]
+    img_mean = torch.tensor(img_mean, dtype=torch.float16)[:, None, None]
+    img_std = torch.tensor(img_std, dtype=torch.float16)[:, None, None]
+
+    if async_loading_frames:
+        lazy_images = AsyncImageFrameLoader(
+            img_paths, image_size, offload_video_to_cpu, img_mean, img_std
+        )
+        return lazy_images, lazy_images.video_height, lazy_images.video_width
+
+    # float16 precision should be sufficient for image tensor storage
+    images = torch.zeros(num_frames, 3, image_size, image_size, dtype=torch.float16)
+    video_height, video_width = None, None
+    for n, img_path in enumerate(
+        tqdm(img_paths, desc=f"frame loading (image folder) [rank={RANK}]")
+    ):
+        images[n], video_height, video_width = _load_img_as_tensor(img_path, image_size)
+    if not offload_video_to_cpu:
+        images = images.cuda()
+        img_mean = img_mean.cuda()
+        img_std = img_std.cuda()
+    # normalize by mean and std
+    images -= img_mean
+    images /= img_std
+    return images, video_height, video_width
+
+
+def load_video_frames_from_video_file(
+    video_path,
+    image_size,
+    offload_video_to_cpu,
+    img_mean,
+    img_std,
+    async_loading_frames,
+    gpu_acceleration=False,
+    gpu_device=None,
+    video_loader_type="cv2",
+):
+    """Load the video frames from a video file."""
+    if video_loader_type == "cv2":
+        return load_video_frames_from_video_file_using_cv2(
+            video_path=video_path,
+            image_size=image_size,
+            img_mean=img_mean,
+            img_std=img_std,
+            offload_video_to_cpu=offload_video_to_cpu,
+        )
+    elif video_loader_type == "torchcodec":
+        logger.info("Using torchcodec to load video file")
+        lazy_images = AsyncVideoFileLoaderWithTorchCodec(
+            video_path=video_path,
+            image_size=image_size,
+            offload_video_to_cpu=offload_video_to_cpu,
+            img_mean=img_mean,
+            img_std=img_std,
+            gpu_acceleration=gpu_acceleration,
+            gpu_device=gpu_device,
+        )
+        # The `AsyncVideoFileLoaderWithTorchCodec` class always loads the videos asynchronously,
+        # so we just wait for its loading thread to finish if async_loading_frames=False.
+        if not async_loading_frames:
+            async_thread = lazy_images.thread
+            if async_thread is not None:
+                async_thread.join()
+        return lazy_images, lazy_images.video_height, lazy_images.video_width
+    else:
+        raise RuntimeError("video_loader_type must be either 'cv2' or 'torchcodec'")
+
+
+def load_video_frames_from_video_file_using_cv2(
+    video_path: str,
+    image_size: int,
+    img_mean: tuple = (0.5, 0.5, 0.5),
+    img_std: tuple = (0.5, 0.5, 0.5),
+    offload_video_to_cpu: bool = False,
+) -> torch.Tensor:
+    """
+    Load video from path, convert to normalized tensor with specified preprocessing
+
+    Args:
+        video_path: Path to video file
+        image_size: Target size for square frames (height and width)
+        img_mean: Normalization mean (RGB)
+        img_std: Normalization standard deviation (RGB)
+
+    Returns:
+        torch.Tensor: Preprocessed video tensor in shape (T, C, H, W) with float16 dtype
+    """
+    import cv2  # delay OpenCV import to avoid unnecessary dependency
+
+    # Initialize video capture
+    cap = cv2.VideoCapture(video_path)
+    if not cap.isOpened():
+        raise ValueError(f"Could not open video: {video_path}")
+
+    original_height = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
+    original_width = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
+    num_frames = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
+    num_frames = num_frames if num_frames > 0 else None
+
+    frames = []
+    pbar = tqdm(desc=f"frame loading (OpenCV) [rank={RANK}]", total=num_frames)
+    while True:
+        ret, frame = cap.read()
+        if not ret:
+            break
+
+        # Convert BGR to RGB and resize
+        frame_rgb = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
+        frame_resized = cv2.resize(
+            frame_rgb, (image_size, image_size), interpolation=cv2.INTER_CUBIC
+        )
+        frames.append(frame_resized)
+        pbar.update(1)
+    cap.release()
+    pbar.close()
+
+    # Convert to tensor
+    frames_np = np.stack(frames, axis=0).astype(np.float32)  # (T, H, W, C)
+    video_tensor = torch.from_numpy(frames_np).permute(0, 3, 1, 2)  # (T, C, H, W)
+
+    img_mean = torch.tensor(img_mean, dtype=torch.float16).view(1, 3, 1, 1)
+    img_std = torch.tensor(img_std, dtype=torch.float16).view(1, 3, 1, 1)
+    if not offload_video_to_cpu:
+        video_tensor = video_tensor.cuda()
+        img_mean = img_mean.cuda()
+        img_std = img_std.cuda()
+    # normalize by mean and std
+    video_tensor -= img_mean
+    video_tensor /= img_std
+    return video_tensor, original_height, original_width
+
+
+def load_dummy_video(image_size, offload_video_to_cpu, num_frames=60):
+    """
+    Load a dummy video with random frames for testing and compilation warmup purposes.
+    """
+    video_height, video_width = 480, 640  # dummy original video sizes
+    images = torch.randn(num_frames, 3, image_size, image_size, dtype=torch.float16)
+    if not offload_video_to_cpu:
+        images = images.cuda()
+    return images, video_height, video_width
+
+
+def _load_img_as_tensor(img_path, image_size):
+    """Load and resize an image and convert it into a PyTorch tensor."""
+    img = Image.open(img_path).convert("RGB")
+    orig_width, orig_height = img.width, img.height
+    img = TF.resize(img, size=(image_size, image_size))
+    img = TF.to_tensor(img)
+    return img, orig_height, orig_width
+
+
+class AsyncImageFrameLoader:
+    """
+    A list of video frames to be load asynchronously without blocking session start.
+    """
+
+    def __init__(self, img_paths, image_size, offload_video_to_cpu, img_mean, img_std):
+        self.img_paths = img_paths
+        self.image_size = image_size
+        self.offload_video_to_cpu = offload_video_to_cpu
+        self.img_mean = img_mean
+        self.img_std = img_std
+        # items in `self._images` will be loaded asynchronously
+        self.images = [None] * len(img_paths)
+        # catch and raise any exceptions in the async loading thread
+        self.exception = None
+        # video_height and video_width be filled when loading the first image
+        self.video_height = None
+        self.video_width = None
+
+        # load the first frame to fill video_height and video_width and also
+        # to cache it (since it's most likely where the user will click)
+        self.__getitem__(0)
+
+        # load the rest of frames asynchronously without blocking the session start
+        def _load_frames():
+            try:
+                for n in tqdm(
+                    range(len(self.images)),
+                    desc=f"frame loading (image folder) [rank={RANK}]",
+                ):
+                    self.__getitem__(n)
+            except Exception as e:
+                self.exception = e
+
+        self.thread = Thread(target=_load_frames, daemon=True)
+        self.thread.start()
+
+    def __getitem__(self, index):
+        if self.exception is not None:
+            raise RuntimeError("Failure in frame loading thread") from self.exception
+
+        img = self.images[index]
+        if img is not None:
+            return img
+
+        img, video_height, video_width = _load_img_as_tensor(
+            self.img_paths[index], self.image_size
+        )
+        self.video_height = video_height
+        self.video_width = video_width
+        # float16 precision should be sufficient for image tensor storage
+        img = img.to(dtype=torch.float16)
+        # normalize by mean and std
+        img -= self.img_mean
+        img /= self.img_std
+        if not self.offload_video_to_cpu:
+            img = img.cuda()
+        self.images[index] = img
+        return img
+
+    def __len__(self):
+        return len(self.images)
+
+
+class TorchCodecDecoder:
+    """
+    A wrapper to support GPU device and num_threads in TorchCodec decoder,
+    which are not supported by `torchcodec.decoders.SimpleVideoDecoder` yet.
+    """
+
+    def __init__(self, source, dimension_order="NCHW", device="cpu", num_threads=1):
+        from torchcodec import _core as core
+
+        self._source = source  # hold a reference to the source to prevent it from GC
+        if isinstance(source, str):
+            self._decoder = core.create_from_file(source, "exact")
+        elif isinstance(source, bytes):
+            self._decoder = core.create_from_bytes(source, "exact")
+        else:
+            raise TypeError(f"Unknown source type: {type(source)}.")
+        assert dimension_order in ("NCHW", "NHWC")
+
+        device_string = str(device)
+        core.scan_all_streams_to_update_metadata(self._decoder)
+        core.add_video_stream(
+            self._decoder,
+            dimension_order=dimension_order,
+            device=device_string,
+            num_threads=(1 if "cuda" in device_string else num_threads),
+        )
+        video_metadata = core.get_container_metadata(self._decoder)
+        best_stream_index = video_metadata.best_video_stream_index
+        assert best_stream_index is not None
+        self.metadata = video_metadata.streams[best_stream_index]
+        assert self.metadata.num_frames_from_content is not None
+        self._num_frames = self.metadata.num_frames_from_content
+
+    def __len__(self) -> int:
+        return self._num_frames
+
+    def __getitem__(self, key: int):
+        from torchcodec import _core as core
+
+        if key < 0:
+            key += self._num_frames
+        if key >= self._num_frames or key < 0:
+            raise IndexError(
+                f"Index {key} is out of bounds; length is {self._num_frames}"
+            )
+        frame_data, *_ = core.get_frame_at_index(
+            self._decoder,
+            frame_index=key,
+        )
+        return frame_data
+
+
+class FIFOLock:
+    """A lock that ensures FIFO ordering of lock acquisitions."""
+
+    def __init__(self):
+        self._lock = Lock()
+        self._waiters = queue.Queue()
+        self._condition = Condition()
+
+    def acquire(self):
+        ident = get_ident()
+        with self._condition:
+            self._waiters.put(ident)
+            while self._waiters.queue[0] != ident or not self._lock.acquire(
+                blocking=False
+            ):
+                self._condition.wait()
+                # got the lock and it's our turn
+
+    def release(self):
+        with self._condition:
+            self._lock.release()
+            self._waiters.get()
+            self._condition.notify_all()
+
+    def __enter__(self):
+        self.acquire()
+
+    def __exit__(self, t, v, tb):
+        self.release()
+
+
+class AsyncVideoFileLoaderWithTorchCodec:
+    """
+    Loading frames from video files asynchronously without blocking session start.
+
+    Unlike `AsyncVideoFileLoader`, this class uses PyTorch's offical TorchCodec library
+    for video decoding, which is more efficient and supports more video formats.
+    """
+
+    def __init__(
+        self,
+        video_path,
+        image_size,
+        offload_video_to_cpu,
+        img_mean,
+        img_std,
+        gpu_acceleration=True,
+        gpu_device=None,
+        use_rand_seek_in_loading=False,
+    ):
+        # Check and possibly infer the output device (and also get its GPU id when applicable)
+        assert gpu_device is None or gpu_device.type == "cuda"
+        gpu_id = (
+            gpu_device.index
+            if gpu_device is not None and gpu_device.index is not None
+            else torch.cuda.current_device()
+        )
+        if offload_video_to_cpu:
+            out_device = torch.device("cpu")
+        else:
+            out_device = torch.device("cuda") if gpu_device is None else gpu_device
+        self.out_device = out_device
+        self.gpu_acceleration = gpu_acceleration
+        self.gpu_id = gpu_id
+        self.image_size = image_size
+        self.offload_video_to_cpu = offload_video_to_cpu
+        if not isinstance(img_mean, torch.Tensor):
+            img_mean = torch.tensor(img_mean, dtype=torch.float16)[:, None, None]
+        self.img_mean = img_mean
+        if not isinstance(img_std, torch.Tensor):
+            img_std = torch.tensor(img_std, dtype=torch.float16)[:, None, None]
+        self.img_std = img_std
+
+        if gpu_acceleration:
+            self.img_mean = self.img_mean.to(f"cuda:{self.gpu_id}")
+            self.img_std = self.img_std.to(f"cuda:{self.gpu_id}")
+            decoder_option = {"device": f"cuda:{self.gpu_id}"}
+        else:
+            self.img_mean = self.img_mean.cpu()
+            self.img_std = self.img_std.cpu()
+            decoder_option = {"num_threads": 1}  # use a single thread to save memory
+
+        self.rank = int(os.environ.get("RANK", "0"))
+        self.world_size = int(os.environ.get("WORLD_SIZE", "1"))
+        self.async_reader = TorchCodecDecoder(video_path, **decoder_option)
+
+        # `num_frames_from_content` is the true number of frames in the video content
+        # from the scan operation (rather than from the metadata, which could be wrong)
+        self.num_frames = self.async_reader.metadata.num_frames_from_content
+        self.video_height = self.async_reader.metadata.height
+        self.video_width = self.async_reader.metadata.width
+
+        # items in `self._images` will be loaded asynchronously
+        self.images_loaded = [False] * self.num_frames
+        self.images = torch.zeros(
+            self.num_frames,
+            3,
+            self.image_size,
+            self.image_size,
+            dtype=torch.float16,
+            device=self.out_device,
+        )
+        # catch and raise any exceptions in the async loading thread
+        self.exception = None
+        self.use_rand_seek_in_loading = use_rand_seek_in_loading
+        self.rand_seek_idx_queue = queue.Queue()
+        # use a lock to avoid race condition between concurrent access to torchcodec
+        # libs (which are not thread-safe); the lock is replaced with a nullcontext
+        # when the video is fully loaded
+        self.torchcodec_access_lock = FIFOLock()
+        self._start_video_loading()
+
+    def _load_one_frame(self, idx):
+        frame_resized = self._transform_frame(self.async_reader[idx])
+        return frame_resized
+
+    @torch.inference_mode()
+    def _start_video_loading(self):
+        desc = f"frame loading (TorchCodec w/ {'GPU' if self.gpu_acceleration else 'CPU'}) [rank={RANK}]"
+        pbar = tqdm(desc=desc, total=self.num_frames)
+        self.num_loaded_frames = 0
+        # load the first frame synchronously to cache it before the session is opened
+        idx = self.num_loaded_frames
+        self.images[idx] = self._load_one_frame(idx)
+        self.images_loaded[idx] = True
+        self.num_loaded_frames += 1
+        pbar.update(n=1)
+        self.all_frames_loaded = self.num_loaded_frames == self.num_frames
+
+        # load the frames asynchronously without blocking the session start
+        def _load_frames():
+            finished = self.all_frames_loaded
+            chunk_size = 16
+            while not finished:
+                # asynchronously load `chunk_size` frames each time we acquire the lock
+                with self.torchcodec_access_lock, torch.inference_mode():
+                    for _ in range(chunk_size):
+                        try:
+                            idx = self.num_loaded_frames
+                            self.images[idx] = self._load_one_frame(idx)
+                            self.images_loaded[idx] = True
+                            self.num_loaded_frames += 1
+                            pbar.update(n=1)
+                            if self.num_loaded_frames >= self.num_frames:
+                                finished = True
+                                break
+                        except Exception as e:
+                            self.exception = e
+                            raise
+
+                    # also read the frame that is being randomly seeked to
+                    while True:
+                        try:
+                            idx = self.rand_seek_idx_queue.get_nowait()
+                            if not self.images_loaded[idx]:
+                                self.images[idx] = self._load_one_frame(idx)
+                                self.images_loaded[idx] = True
+                        except queue.Empty:
+                            break
+                        except Exception as e:
+                            self.exception = e
+                            raise
+
+            # finished -- check whether we have loaded the total number of frames
+            if self.num_loaded_frames != self.num_frames:
+                raise RuntimeError(
+                    f"There are {self.num_frames} frames in the video, but only "
+                    f"{self.num_loaded_frames} frames can be loaded successfully."
+                )
+            else:
+                self.all_frames_loaded = True
+                pbar.close()
+                with self.torchcodec_access_lock:
+                    import gc
+
+                    # all frames have been loaded, so we can release the readers and free their memory
+                    # also remove pbar and thread (which shouldn't be a part of session saving)
+                    reader = self.async_reader
+                    if reader is not None:
+                        reader._source = None
+                    self.async_reader = None
+                    self.pbar = None
+                    self.thread = None
+                    self.rand_seek_idx_queue = None
+                    gc.collect()
+                # remove the lock (replace it with nullcontext) when the video is fully loaded
+                self.torchcodec_access_lock = contextlib.nullcontext()
+
+        self.thread = Thread(target=_load_frames, daemon=True)
+        self.thread.start()
+
+    def _transform_frame(self, frame):
+        frame = frame.clone()  # make a copy to avoid modifying the original frame bytes
+        frame = frame.float()  # convert to float32 before interpolation
+        frame_resized = F.interpolate(
+            frame[None, :],
+            size=(self.image_size, self.image_size),
+            mode="bicubic",
+            align_corners=False,
+        )[0]
+        # float16 precision should be sufficient for image tensor storage
+        frame_resized = frame_resized.half()  # uint8 -> float16
+        frame_resized /= 255
+        frame_resized -= self.img_mean
+        frame_resized /= self.img_std
+        if self.offload_video_to_cpu:
+            frame_resized = frame_resized.cpu()
+        elif frame_resized.device != self.out_device:
+            frame_resized = frame_resized.to(device=self.out_device, non_blocking=True)
+        return frame_resized
+
+    def __getitem__(self, index):
+        if self.exception is not None:
+            raise RuntimeError("Failure in frame loading thread") from self.exception
+
+        max_tries = 1200
+        for _ in range(max_tries):
+            # use a lock to avoid race condition between concurrent access to torchcodec
+            # libs (which are not thread-safe); the lock is replaced with a nullcontext
+            # when the video is fully loaded
+            with self.torchcodec_access_lock:
+                if self.images_loaded[index]:
+                    return self.images[index]
+
+                if self.use_rand_seek_in_loading:
+                    # async loading hasn't reached this frame yet, so we load this frame individually
+                    # (it will be loaded by in _load_frames thread and added to self.images[index])
+                    self.rand_seek_idx_queue.put(index)
+
+            time.sleep(0.1)
+
+        raise RuntimeError(f"Failed to load frame {index} after {max_tries} tries")
+
+    def __len__(self):
+        return len(self.images)
+
+    def __getstate__(self):
+        """
+        Remove a few attributes during pickling, so that this async video loader can be
+        saved and loaded as a part of the model session.
+        """
+        # wait for async video loading to finish before pickling
+        async_thread = self.thread
+        if async_thread is not None:
+            async_thread.join()
+        # release a few objects that cannot be pickled
+        reader = self.async_reader
+        if reader is not None:
+            reader._source = None
+        self.async_reader = None
+        self.pbar = None
+        self.thread = None
+        self.rand_seek_idx_queue = None
+        self.torchcodec_access_lock = contextlib.nullcontext()
+        return self.__dict__.copy()

detect_tools/sam3/sam3/model/maskformer_segmentation.py

@@ -0,0 +1,323 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates. All Rights Reserved
+
+import math
+from typing import Dict, List, Optional
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import torch.utils.checkpoint as checkpoint
+
+from .model_misc import MLP
+
+
+class LinearPresenceHead(nn.Sequential):
+    def __init__(self, d_model):
+        # a hack to make `LinearPresenceHead` compatible with old checkpoints
+        super().__init__(nn.Identity(), nn.Identity(), nn.Linear(d_model, 1))
+
+    def forward(self, hs, prompt, prompt_mask):
+        return super().forward(hs)
+
+
+class MaskPredictor(nn.Module):
+    def __init__(self, hidden_dim, mask_dim):
+        super().__init__()
+        self.mask_embed = MLP(hidden_dim, hidden_dim, mask_dim, 3)
+
+    def forward(self, obj_queries, pixel_embed):
+        if len(obj_queries.shape) == 3:
+            if pixel_embed.ndim == 3:
+                # batch size was omitted
+                mask_preds = torch.einsum(
+                    "bqc,chw->bqhw", self.mask_embed(obj_queries), pixel_embed
+                )
+            else:
+                mask_preds = torch.einsum(
+                    "bqc,bchw->bqhw", self.mask_embed(obj_queries), pixel_embed
+                )
+        else:
+            # Assumed to have aux masks
+            if pixel_embed.ndim == 3:
+                # batch size was omitted
+                mask_preds = torch.einsum(
+                    "lbqc,chw->lbqhw", self.mask_embed(obj_queries), pixel_embed
+                )
+            else:
+                mask_preds = torch.einsum(
+                    "lbqc,bchw->lbqhw", self.mask_embed(obj_queries), pixel_embed
+                )
+
+        return mask_preds
+
+
+class SegmentationHead(nn.Module):
+    def __init__(
+        self,
+        hidden_dim,
+        upsampling_stages,
+        use_encoder_inputs=False,
+        aux_masks=False,
+        no_dec=False,
+        pixel_decoder=None,
+        act_ckpt=False,
+        shared_conv=False,
+        compile_mode_pixel_decoder=None,
+    ):
+        super().__init__()
+        self.use_encoder_inputs = use_encoder_inputs
+        self.aux_masks = aux_masks
+        if pixel_decoder is not None:
+            self.pixel_decoder = pixel_decoder
+        else:
+            self.pixel_decoder = PixelDecoder(
+                hidden_dim,
+                upsampling_stages,
+                shared_conv=shared_conv,
+                compile_mode=compile_mode_pixel_decoder,
+            )
+        self.no_dec = no_dec
+        if no_dec:
+            self.mask_predictor = nn.Conv2d(
+                hidden_dim, 1, kernel_size=3, stride=1, padding=1
+            )
+        else:
+            self.mask_predictor = MaskPredictor(hidden_dim, mask_dim=hidden_dim)
+
+        self.act_ckpt = act_ckpt
+
+        # used to update the output dictionary
+        self.instance_keys = ["pred_masks"]
+
+    @property
+    def device(self):
+        self._device = getattr(self, "_device", None) or next(self.parameters()).device
+        return self._device
+
+    def to(self, *args, **kwargs):
+        # clear cached _device in case the model is moved to a different device
+        self._device = None
+        return super().to(*args, **kwargs)
+
+    def _embed_pixels(
+        self,
+        backbone_feats: List[torch.Tensor],
+        image_ids,
+        encoder_hidden_states,
+    ) -> torch.Tensor:
+        feature_device = backbone_feats[0].device  # features could be on CPU
+        model_device = self.device
+        image_ids_ = image_ids.to(feature_device)
+        if self.use_encoder_inputs:
+            if backbone_feats[0].shape[0] > 1:
+                # For bs > 1, we construct the per query backbone features
+                backbone_visual_feats = []
+                for feat in backbone_feats:
+                    # Copy the img features per query (pixel decoder won't share img feats)
+                    backbone_visual_feats.append(feat[image_ids_, ...].to(model_device))
+            else:
+                # Bs=1, we rely on broadcasting for query-based processing
+                backbone_visual_feats = [bb_feat.clone() for bb_feat in backbone_feats]
+            # Extract visual embeddings
+            encoder_hidden_states = encoder_hidden_states.permute(1, 2, 0)
+            spatial_dim = math.prod(backbone_feats[-1].shape[-2:])
+            encoder_visual_embed = encoder_hidden_states[..., :spatial_dim].reshape(
+                -1, *backbone_feats[-1].shape[1:]
+            )
+
+            backbone_visual_feats[-1] = encoder_visual_embed
+            if self.act_ckpt:
+                pixel_embed = checkpoint.checkpoint(
+                    self.pixel_decoder, backbone_visual_feats, use_reentrant=False
+                )
+            else:
+                pixel_embed = self.pixel_decoder(backbone_visual_feats)
+        else:
+            backbone_feats = [x.to(model_device) for x in backbone_feats]
+            pixel_embed = self.pixel_decoder(backbone_feats)
+            if pixel_embed.shape[0] == 1:
+                # For batch_size=1 training, we can avoid the indexing to save memory
+                pixel_embed = pixel_embed.squeeze(0)
+            else:
+                pixel_embed = pixel_embed[image_ids, ...]
+        return pixel_embed
+
+    def forward(
+        self,
+        backbone_feats: List[torch.Tensor],
+        obj_queries: torch.Tensor,
+        image_ids,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        **kwargs,
+    ) -> Dict[str, torch.Tensor]:
+        if self.use_encoder_inputs:
+            assert encoder_hidden_states is not None
+
+        pixel_embed = self._embed_pixels(
+            backbone_feats=backbone_feats,
+            image_ids=image_ids,
+            encoder_hidden_states=encoder_hidden_states,
+        )
+
+        if self.no_dec:
+            mask_pred = self.mask_predictor(pixel_embed)
+        elif self.aux_masks:
+            mask_pred = self.mask_predictor(obj_queries, pixel_embed)
+        else:
+            mask_pred = self.mask_predictor(obj_queries[-1], pixel_embed)
+
+        return {"pred_masks": mask_pred}
+
+
+class PixelDecoder(nn.Module):
+    def __init__(
+        self,
+        hidden_dim,
+        num_upsampling_stages,
+        interpolation_mode="nearest",
+        shared_conv=False,
+        compile_mode=None,
+    ):
+        super().__init__()
+        self.hidden_dim = hidden_dim
+        self.num_upsampling_stages = num_upsampling_stages
+        self.interpolation_mode = interpolation_mode
+        conv_layers = []
+        norms = []
+        num_convs = 1 if shared_conv else num_upsampling_stages
+        for _ in range(num_convs):
+            conv_layers.append(nn.Conv2d(self.hidden_dim, self.hidden_dim, 3, 1, 1))
+            norms.append(nn.GroupNorm(8, self.hidden_dim))
+
+        self.conv_layers = nn.ModuleList(conv_layers)
+        self.norms = nn.ModuleList(norms)
+        self.shared_conv = shared_conv
+        self.out_dim = self.conv_layers[-1].out_channels
+        if compile_mode is not None:
+            self.forward = torch.compile(
+                self.forward, mode=compile_mode, dynamic=True, fullgraph=True
+            )
+            # Needed to make checkpointing happy. But we don't know if the module is checkpointed, so we disable it by default.
+            torch._dynamo.config.optimize_ddp = False
+
+    def forward(self, backbone_feats: List[torch.Tensor]):
+        # Assumes backbone features are already projected (C == hidden dim)
+
+        prev_fpn = backbone_feats[-1]
+        fpn_feats = backbone_feats[:-1]
+        for layer_idx, bb_feat in enumerate(fpn_feats[::-1]):
+            curr_fpn = bb_feat
+            prev_fpn = curr_fpn + F.interpolate(
+                prev_fpn, size=curr_fpn.shape[-2:], mode=self.interpolation_mode
+            )
+            if self.shared_conv:
+                # only one conv layer
+                layer_idx = 0
+            prev_fpn = self.conv_layers[layer_idx](prev_fpn)
+            prev_fpn = F.relu(self.norms[layer_idx](prev_fpn))
+
+        return prev_fpn
+
+
+class UniversalSegmentationHead(SegmentationHead):
+    """This module handles semantic+instance segmentation"""
+
+    def __init__(
+        self,
+        hidden_dim,
+        upsampling_stages,
+        pixel_decoder,
+        aux_masks=False,
+        no_dec=False,
+        act_ckpt=False,
+        presence_head: bool = False,
+        dot_product_scorer=None,
+        cross_attend_prompt=None,
+    ):
+        super().__init__(
+            hidden_dim=hidden_dim,
+            upsampling_stages=upsampling_stages,
+            use_encoder_inputs=True,
+            aux_masks=aux_masks,
+            no_dec=no_dec,
+            pixel_decoder=pixel_decoder,
+            act_ckpt=act_ckpt,
+        )
+        self.d_model = hidden_dim
+
+        if dot_product_scorer is not None:
+            assert presence_head, "Specifying a dot product scorer without a presence head is likely a mistake"
+
+        self.presence_head = None
+        if presence_head:
+            self.presence_head = (
+                dot_product_scorer
+                if dot_product_scorer is not None
+                else LinearPresenceHead(self.d_model)
+            )
+
+        self.cross_attend_prompt = cross_attend_prompt
+        if self.cross_attend_prompt is not None:
+            self.cross_attn_norm = nn.LayerNorm(self.d_model)
+
+        self.semantic_seg_head = nn.Conv2d(self.pixel_decoder.out_dim, 1, kernel_size=1)
+        self.instance_seg_head = nn.Conv2d(
+            self.pixel_decoder.out_dim, self.d_model, kernel_size=1
+        )
+
+    def forward(
+        self,
+        backbone_feats: List[torch.Tensor],
+        obj_queries: torch.Tensor,
+        image_ids,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        prompt: Optional[torch.Tensor] = None,
+        prompt_mask: Optional[torch.Tensor] = None,
+        **kwargs,
+    ) -> Dict[str, Optional[torch.Tensor]]:
+        assert encoder_hidden_states is not None
+        bs = encoder_hidden_states.shape[1]
+
+        if self.cross_attend_prompt is not None:
+            tgt2 = self.cross_attn_norm(encoder_hidden_states)
+            tgt2 = self.cross_attend_prompt(
+                query=tgt2,
+                key=prompt,
+                value=prompt,
+                key_padding_mask=prompt_mask,
+            )[0]
+            encoder_hidden_states = tgt2 + encoder_hidden_states
+
+        presence_logit = None
+        if self.presence_head is not None:
+            pooled_enc = encoder_hidden_states.mean(0)
+            presence_logit = (
+                self.presence_head(
+                    pooled_enc.view(1, bs, 1, self.d_model),
+                    prompt=prompt,
+                    prompt_mask=prompt_mask,
+                )
+                .squeeze(0)
+                .squeeze(1)
+            )
+
+        pixel_embed = self._embed_pixels(
+            backbone_feats=backbone_feats,
+            image_ids=image_ids,
+            encoder_hidden_states=encoder_hidden_states,
+        )
+
+        instance_embeds = self.instance_seg_head(pixel_embed)
+
+        if self.no_dec:
+            mask_pred = self.mask_predictor(instance_embeds)
+        elif self.aux_masks:
+            mask_pred = self.mask_predictor(obj_queries, instance_embeds)
+        else:
+            mask_pred = self.mask_predictor(obj_queries[-1], instance_embeds)
+
+        return {
+            "pred_masks": mask_pred,
+            "semantic_seg": self.semantic_seg_head(pixel_embed),
+            "presence_logit": presence_logit,
+        }

detect_tools/sam3/sam3/model/memory.py

@@ -0,0 +1,201 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates. All Rights Reserved
+
+import math
+from typing import Tuple
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+try:
+    from timm.layers import DropPath
+except ModuleNotFoundError:
+    # compatibility for older timm versions
+    from timm.models.layers import DropPath
+
+from .model_misc import get_clones, LayerNorm2d
+
+
+class SimpleMaskDownSampler(nn.Module):
+    """
+    Progressively downsample a mask by total_stride, each time by stride.
+    Note that LayerNorm is applied per *token*, like in ViT.
+
+    With each downsample (by a factor stride**2), channel capacity increases by the same factor.
+    In the end, we linearly project to embed_dim channels.
+    """
+
+    def __init__(
+        self,
+        embed_dim=256,
+        kernel_size=4,
+        stride=4,
+        padding=0,
+        total_stride=16,
+        activation=nn.GELU,
+        # Option to interpolate the input mask first before downsampling using convs. In that case, the total_stride is assumed to be after interpolation.
+        # If set to input resolution or None, we don't interpolate. We default to None to be safe (for older configs or if not explicitly set)
+        interpol_size=None,
+    ):
+        super().__init__()
+        num_layers = int(math.log2(total_stride) // math.log2(stride))
+        assert stride**num_layers == total_stride
+        self.encoder = nn.Sequential()
+        mask_in_chans, mask_out_chans = 1, 1
+        for _ in range(num_layers):
+            mask_out_chans = mask_in_chans * (stride**2)
+            self.encoder.append(
+                nn.Conv2d(
+                    mask_in_chans,
+                    mask_out_chans,
+                    kernel_size=kernel_size,
+                    stride=stride,
+                    padding=padding,
+                )
+            )
+            self.encoder.append(LayerNorm2d(mask_out_chans))
+            self.encoder.append(activation())
+            mask_in_chans = mask_out_chans
+
+        self.encoder.append(nn.Conv2d(mask_out_chans, embed_dim, kernel_size=1))
+        self.interpol_size = interpol_size
+        if self.interpol_size is not None:
+            assert isinstance(
+                self.interpol_size, (list, tuple)
+            ), f"Unsupported type {type(self.interpol_size)}. Should be a list or tuple."
+            self.interpol_size = list(interpol_size)
+            assert len(self.interpol_size) == 2
+
+    def forward(self, x: torch.Tensor):
+        if self.interpol_size is not None and self.interpol_size != list(x.shape[-2:]):
+            x = F.interpolate(
+                x.float(),
+                size=self.interpol_size,
+                align_corners=False,
+                mode="bilinear",
+                antialias=True,
+            )
+        return self.encoder(x)
+
+
+# Lightly adapted from ConvNext (https://github.com/facebookresearch/ConvNeXt)
+class CXBlock(nn.Module):
+    r"""ConvNeXt Block. There are two equivalent implementations:
+    (1) DwConv -> LayerNorm (channels_first) -> 1x1 Conv -> GELU -> 1x1 Conv; all in (N, C, H, W)
+    (2) DwConv -> Permute to (N, H, W, C); LayerNorm (channels_last) -> Linear -> GELU -> Linear; Permute back
+    We use (2) as we find it slightly faster in PyTorch
+
+    Args:
+        dim (int): Number of input channels.
+        drop_path (float): Stochastic depth rate. Default: 0.0
+        layer_scale_init_value (float): Init value for Layer Scale. Default: 1e-6.
+    """
+
+    def __init__(
+        self,
+        dim,
+        kernel_size=7,
+        padding=3,
+        drop_path=0.0,
+        layer_scale_init_value=1e-6,
+        use_dwconv=True,
+    ):
+        super().__init__()
+        self.dwconv = nn.Conv2d(
+            dim,
+            dim,
+            kernel_size=kernel_size,
+            padding=padding,
+            groups=dim if use_dwconv else 1,
+        )  # depthwise conv
+        self.norm = LayerNorm2d(dim, eps=1e-6)
+        self.pwconv1 = nn.Linear(
+            dim, 4 * dim
+        )  # pointwise/1x1 convs, implemented with linear layers
+        self.act = nn.GELU()
+        self.pwconv2 = nn.Linear(4 * dim, dim)
+        self.gamma = (
+            nn.Parameter(layer_scale_init_value * torch.ones((dim)), requires_grad=True)
+            if layer_scale_init_value > 0
+            else None
+        )
+        self.drop_path = DropPath(drop_path) if drop_path > 0.0 else nn.Identity()
+
+    def forward(self, x):
+        input = x
+        x = self.dwconv(x)
+        x = self.norm(x)
+        x = x.permute(0, 2, 3, 1)  # (N, C, H, W) -> (N, H, W, C)
+        x = self.pwconv1(x)
+        x = self.act(x)
+        x = self.pwconv2(x)
+        if self.gamma is not None:
+            x = self.gamma * x
+        x = x.permute(0, 3, 1, 2)  # (N, H, W, C) -> (N, C, H, W)
+
+        x = input + self.drop_path(x)
+        return x
+
+
+class SimpleFuser(nn.Module):
+    def __init__(self, layer, num_layers, dim=None, input_projection=False):
+        super().__init__()
+        self.proj = nn.Identity()
+        self.layers = get_clones(layer, num_layers)
+
+        if input_projection:
+            assert dim is not None
+            self.proj = nn.Conv2d(dim, dim, kernel_size=1)
+
+    def forward(self, x):
+        # normally x: (N, C, H, W)
+        x = self.proj(x)
+        for layer in self.layers:
+            x = layer(x)
+        return x
+
+
+class SimpleMaskEncoder(nn.Module):
+    def __init__(
+        self,
+        out_dim,
+        mask_downsampler,
+        fuser,
+        position_encoding,
+        in_dim=256,  # in_dim of pix_feats
+    ):
+        super().__init__()
+
+        self.mask_downsampler = mask_downsampler
+
+        self.pix_feat_proj = nn.Conv2d(in_dim, in_dim, kernel_size=1)
+        self.fuser = fuser
+        self.position_encoding = position_encoding
+        self.out_proj = nn.Identity()
+        if out_dim != in_dim:
+            self.out_proj = nn.Conv2d(in_dim, out_dim, kernel_size=1)
+
+    def forward(
+        self,
+        pix_feat: torch.Tensor,
+        masks: torch.Tensor,
+        skip_mask_sigmoid: bool = False,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        ## Process masks
+        # sigmoid, so that less domain shift from gt masks which are bool
+        if not skip_mask_sigmoid:
+            masks = F.sigmoid(masks)
+        masks = self.mask_downsampler(masks)
+
+        ## Fuse pix_feats and downsampled masks
+        # in case the visual features are on CPU, cast them to CUDA
+        pix_feat = pix_feat.to(masks.device)
+
+        x = self.pix_feat_proj(pix_feat)
+        x = x + masks
+        x = self.fuser(x)
+        x = self.out_proj(x)
+
+        pos = self.position_encoding(x).to(x.dtype)
+
+        return {"vision_features": x, "vision_pos_enc": [pos]}

detect_tools/sam3/sam3/model/model_misc.py

@@ -0,0 +1,428 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates. All Rights Reserved
+
+"""Various utility models"""
+
+import copy
+import math
+import weakref
+from collections.abc import Iterator
+from contextlib import AbstractContextManager
+from enum import auto, Enum
+from typing import Dict, List, Optional, Union
+
+import numpy as np
+import torch
+import torch.nn.functional as F
+from torch import nn, Tensor
+from typing_extensions import override
+
+
+def inverse_sigmoid(x, eps=1e-3):
+    """
+    The inverse function for sigmoid activation function.
+    Note: It might face numberical issues with fp16 small eps.
+    """
+    x = x.clamp(min=0, max=1)
+    x1 = x.clamp(min=eps)
+    x2 = (1 - x).clamp(min=eps)
+    return torch.log(x1 / x2)
+
+
+class MultiheadAttentionWrapper(nn.MultiheadAttention):
+    def forward(self, *args, **kwargs):
+        kwargs["need_weights"] = False
+        return super().forward(*args, **kwargs)
+
+
+class DotProductScoring(torch.nn.Module):
+    def __init__(
+        self,
+        d_model,
+        d_proj,
+        prompt_mlp=None,
+        clamp_logits=True,
+        clamp_max_val=12.0,
+    ):
+        super().__init__()
+        self.d_proj = d_proj
+        assert isinstance(prompt_mlp, torch.nn.Module) or prompt_mlp is None
+        self.prompt_mlp = prompt_mlp  # an optional MLP projection for prompt
+        self.prompt_proj = torch.nn.Linear(d_model, d_proj)
+        self.hs_proj = torch.nn.Linear(d_model, d_proj)
+        self.scale = float(1.0 / np.sqrt(d_proj))
+        self.clamp_logits = clamp_logits
+        if self.clamp_logits:
+            self.clamp_max_val = clamp_max_val
+
+    def mean_pool_text(self, prompt, prompt_mask):
+        # is_valid has shape (seq, bs, 1), where 1 is valid and 0 is padding
+        is_valid = (~prompt_mask).float().permute(1, 0)[..., None]
+        # num_valid has shape (bs, 1)
+        num_valid = torch.clamp(torch.sum(is_valid, dim=0), min=1.0)
+        # mean pool over all the valid tokens -- pooled_prompt has shape (bs, proj_dim)
+        pooled_prompt = (prompt * is_valid).sum(dim=0) / num_valid
+        return pooled_prompt
+
+    def forward(self, hs, prompt, prompt_mask):
+        # hs has shape (num_layer, bs, num_query, d_model)
+        # prompt has shape (seq, bs, d_model)
+        # prompt_mask has shape (bs, seq), where 1 is valid and 0 is padding
+        assert hs.dim() == 4 and prompt.dim() == 3 and prompt_mask.dim() == 2
+
+        # apply MLP on prompt if specified
+        if self.prompt_mlp is not None:
+            prompt = self.prompt_mlp(prompt)
+
+        # first, get the mean-pooled version of the prompt
+        pooled_prompt = self.mean_pool_text(prompt, prompt_mask)
+
+        # then, project pooled_prompt and hs to d_proj dimensions
+        proj_pooled_prompt = self.prompt_proj(pooled_prompt)  # (bs, d_proj)
+        proj_hs = self.hs_proj(hs)  # (num_layer, bs, num_query, d_proj)
+
+        # finally, get dot-product scores of shape (num_layer, bs, num_query, 1)
+        scores = torch.matmul(proj_hs, proj_pooled_prompt.unsqueeze(-1))
+        scores *= self.scale
+
+        # clamp scores to a max value to avoid numerical issues in loss or matcher
+        if self.clamp_logits:
+            scores.clamp_(min=-self.clamp_max_val, max=self.clamp_max_val)
+
+        return scores
+
+
+class LayerScale(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        init_values: Union[float, Tensor] = 1e-5,
+        inplace: bool = False,
+    ) -> None:
+        super().__init__()
+        self.inplace = inplace
+        self.gamma = nn.Parameter(init_values * torch.ones(dim))
+
+    def forward(self, x: Tensor) -> Tensor:
+        return x.mul_(self.gamma) if self.inplace else x * self.gamma
+
+
+class LayerNorm2d(nn.Module):
+    def __init__(self, num_channels: int, eps: float = 1e-6) -> None:
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(num_channels))
+        self.bias = nn.Parameter(torch.zeros(num_channels))
+        self.eps = eps
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        u = x.mean(1, keepdim=True)
+        s = (x - u).pow(2).mean(1, keepdim=True)
+        x = (x - u) / torch.sqrt(s + self.eps)
+        x = self.weight[:, None, None] * x + self.bias[:, None, None]
+        return x
+
+
+class TransformerWrapper(nn.Module):
+    def __init__(
+        self,
+        encoder,
+        decoder,
+        d_model: int,
+        two_stage_type="none",  # ["none"] only for now
+        pos_enc_at_input_dec=True,
+    ):
+        super().__init__()
+
+        self.encoder = encoder
+        self.decoder = decoder
+        self.num_queries = decoder.num_queries if decoder is not None else None
+        self.pos_enc_at_input_dec = pos_enc_at_input_dec
+
+        # for two stage
+        assert two_stage_type in ["none"], "unknown param {} of two_stage_type".format(
+            two_stage_type
+        )
+        self.two_stage_type = two_stage_type
+
+        self._reset_parameters()
+        self.d_model = d_model
+
+    def _reset_parameters(self):
+        for n, p in self.named_parameters():
+            if p.dim() > 1:
+                if (
+                    "box_embed" not in n
+                    and "query_embed" not in n
+                    and "reference_points" not in n
+                ):
+                    nn.init.xavier_uniform_(p)
+
+
+class MLP(nn.Module):
+    """Very simple multi-layer perceptron (also called FFN)"""
+
+    def __init__(
+        self,
+        input_dim: int,
+        hidden_dim: int,
+        output_dim: int,
+        num_layers: int,
+        dropout: float = 0.0,
+        residual: bool = False,
+        out_norm: Optional[nn.Module] = None,
+    ):
+        super().__init__()
+        self.num_layers = num_layers
+        h = [hidden_dim] * (num_layers - 1)
+        self.layers = nn.ModuleList(
+            nn.Linear(n, k) for n, k in zip([input_dim] + h, h + [output_dim])
+        )
+        self.drop = nn.Dropout(dropout) if dropout > 0 else nn.Identity()
+        # whether to add the output as a residual connection to the input
+        if residual and input_dim != output_dim:
+            raise ValueError("residual is only supported if input_dim == output_dim")
+        self.residual = residual
+        # whether to apply a normalization layer to the output
+        assert isinstance(out_norm, nn.Module) or out_norm is None
+        self.out_norm = out_norm or nn.Identity()
+
+    def forward(self, x):
+        orig_x = x
+        for i, layer in enumerate(self.layers):
+            x = self.drop(F.relu(layer(x))) if i < self.num_layers - 1 else layer(x)
+        if self.residual:
+            x = x + orig_x
+        x = self.out_norm(x)
+        return x
+
+
+def get_clones(module, N):
+    return nn.ModuleList([copy.deepcopy(module) for i in range(N)])
+
+
+def get_clones_seq(module, N):
+    return nn.Sequential(*[copy.deepcopy(module) for i in range(N)])
+
+
+def get_activation_fn(activation):
+    """Return an activation function given a string"""
+    if activation == "relu":
+        return F.relu
+    if activation == "gelu":
+        return F.gelu
+    if activation == "glu":
+        return F.glu
+    raise RuntimeError(f"activation should be relu/gelu, not {activation}.")
+
+
+def get_activation_module(activation):
+    """Return an activation function given a string"""
+    if activation == "relu":
+        return nn.ReLU
+    if activation == "gelu":
+        return nn.GELU
+    if activation == "glu":
+        return nn.GLU
+    raise RuntimeError(f"activation should be relu/gelu, not {activation}.")
+
+
+def get_valid_ratio(mask):
+    _, H, W = mask.shape
+    valid_H = torch.sum(~mask[:, :, 0], 1)
+    valid_W = torch.sum(~mask[:, 0, :], 1)
+    valid_ratio_h = valid_H.float() / H
+    valid_ratio_w = valid_W.float() / W
+    valid_ratio = torch.stack([valid_ratio_w, valid_ratio_h], -1)
+    return valid_ratio
+
+
+def gen_sineembed_for_position(pos_tensor, num_feats=256):
+    assert num_feats % 2 == 0
+    num_feats = num_feats // 2
+    # n_query, bs, _ = pos_tensor.size()
+    # sineembed_tensor = torch.zeros(n_query, bs, 256)
+    scale = 2 * math.pi
+    dim_t = torch.arange(num_feats, dtype=torch.float32, device=pos_tensor.device)
+    dim_t = 10000 ** (2 * (torch.div(dim_t, 2, rounding_mode="floor")) / num_feats)
+    x_embed = pos_tensor[:, :, 0] * scale
+    y_embed = pos_tensor[:, :, 1] * scale
+    pos_x = x_embed[:, :, None] / dim_t
+    pos_y = y_embed[:, :, None] / dim_t
+    pos_x = torch.stack(
+        (pos_x[:, :, 0::2].sin(), pos_x[:, :, 1::2].cos()), dim=3
+    ).flatten(2)
+    pos_y = torch.stack(
+        (pos_y[:, :, 0::2].sin(), pos_y[:, :, 1::2].cos()), dim=3
+    ).flatten(2)
+    if pos_tensor.size(-1) == 2:
+        pos = torch.cat((pos_y, pos_x), dim=2)
+    elif pos_tensor.size(-1) == 4:
+        w_embed = pos_tensor[:, :, 2] * scale
+        pos_w = w_embed[:, :, None] / dim_t
+        pos_w = torch.stack(
+            (pos_w[:, :, 0::2].sin(), pos_w[:, :, 1::2].cos()), dim=3
+        ).flatten(2)
+
+        h_embed = pos_tensor[:, :, 3] * scale
+        pos_h = h_embed[:, :, None] / dim_t
+        pos_h = torch.stack(
+            (pos_h[:, :, 0::2].sin(), pos_h[:, :, 1::2].cos()), dim=3
+        ).flatten(2)
+
+        pos = torch.cat((pos_y, pos_x, pos_w, pos_h), dim=2)
+    else:
+        raise ValueError("Unknown pos_tensor shape(-1):{}".format(pos_tensor.size(-1)))
+    return pos
+
+
+class SAM3Output(list):
+    """
+    A class representing the output of a SAM3 model.
+    It provides an iterable interface that supports different iteration modes, including iterating over all steps per stage,
+    last step per stage, and flattened output.
+    Attributes:
+        output: The output of the SAM3 model, represented as a list of lists.
+        iter_mode: The current iteration mode.
+    Example:
+        >>> output = [[1, 2], [3, 4], [5, 6]]
+        >>> sam3_output = SAM3Output(output)
+        >>> for step in sam3_output:
+        ...     print(step)
+        [1, 2]
+        [3, 4]
+        [5, 6]
+        >>> with SAM3Output.iteration_mode(SAM3Output.IterMode.LAST_STEP_PER_STAGE) as sam3_last_step_out:
+        ...     for step in sam3_last_step_out:
+        ...         print(step)
+        [2]
+        [4]
+        [6]
+        >>> with SAM3Output.iteration_mode(SAM3Output.IterMode.FLATTENED) as sam3_flattened_out:
+        ...     for step in sam3_flattened_out:
+        ...         print(step)
+        1
+        2
+        3
+        4
+        5
+        6
+    """
+
+    class IterMode(Enum):
+        # Defines the type of iterator over ouptuts.
+        ALL_STEPS_PER_STAGE = auto()
+        LAST_STEP_PER_STAGE = auto()
+        FLATTENED = auto()  # Returns each interactivity step as if it is a separate stage (this is used in SAM3Image model)
+
+    def __init__(
+        self,
+        output: List[List[Dict]] = None,
+        iter_mode: IterMode = IterMode.ALL_STEPS_PER_STAGE,
+        loss_stages: Optional[List[int]] = None,
+    ):
+        if output is not None:
+            assert (
+                isinstance(output, list)
+                and len(output) > 0
+                and isinstance(output[0], list)
+            ), "Expected output to be a list of lists"
+            self.output = output
+        else:
+            self.output = []
+        assert isinstance(
+            iter_mode, SAM3Output.IterMode
+        ), f"iter_mode shoulf be of enum type 'SAM3Output.IterMode'. Got {type(iter_mode)}"
+
+        self.iter_mode = iter_mode
+        # We create a weak reference to self to be used in the lambda functions.
+        # This is to avoid cyclic references and let SAM3Output be garabge collected.
+        self_ref = weakref.ref(self)
+        self._mode2iter = {
+            SAM3Output.IterMode.ALL_STEPS_PER_STAGE: lambda: iter(self_ref().output),
+            SAM3Output.IterMode.LAST_STEP_PER_STAGE: lambda: (
+                inner_list[-1] for inner_list in self_ref().output
+            ),
+            SAM3Output.IterMode.FLATTENED: lambda: (
+                element for inner_list in self_ref().output for element in inner_list
+            ),
+        }
+        self.loss_stages = loss_stages
+
+    @override
+    def __iter__(self) -> Iterator:
+        return self._mode2iter[self.iter_mode]()
+
+    def __getitem__(self, index):
+        """
+        Returns the item at the specified index.
+        Args:
+            index (int): The index of the item to return.
+        Returns:
+            list or element: The item at the specified index.
+        """
+        assert isinstance(index, int), f"index should be an integer. Got {type(index)}"
+        if self.iter_mode == SAM3Output.IterMode.ALL_STEPS_PER_STAGE:
+            return self.output[index]
+        elif self.iter_mode == SAM3Output.IterMode.LAST_STEP_PER_STAGE:
+            return self.output[index][-1]
+        elif self.iter_mode == SAM3Output.IterMode.FLATTENED:
+            if index == -1:
+                return self.self.output[-1][-1]
+            else:
+                flattened_output = sum(self.output, [])
+                return flattened_output[index]
+
+    class _IterationMode(AbstractContextManager):
+        """
+        A context manager that temporarily changes the iteration mode of a SAM3Output object.
+        This class is used internally by the SAM3Output.iteration_mode method.
+        """
+
+        def __init__(
+            self, model_output: "SAM3Output", iter_mode: "SAM3Output.IterMode"
+        ):
+            self._model_output = model_output
+            self._orig_iter_mode = model_output.iter_mode
+            self._new_iter_mode = iter_mode
+
+        @override
+        def __enter__(self) -> "SAM3Output":
+            self._model_output.iter_mode = self._new_iter_mode
+            return self._model_output
+
+        @override
+        def __exit__(self, exc_type, exc_value, traceback):
+            self._model_output.iter_mode = self._orig_iter_mode
+            return super().__exit__(exc_type, exc_value, traceback)
+
+    @staticmethod
+    def iteration_mode(
+        model_output: "SAM3Output", iter_mode: IterMode
+    ) -> _IterationMode:
+        """
+        Returns a context manager that allows you to temporarily change the iteration mode of the SAM3Output object.
+        Args:
+            model_output: The SAM3Output object.
+            iter_mode: The new iteration mode.
+        Returns:
+            SAM3Output._IterationMode: A context manager that changes the iteration mode of the SAM3Output object.
+        """
+        return SAM3Output._IterationMode(model_output=model_output, iter_mode=iter_mode)
+
+    def append(self, item: list):
+        assert isinstance(
+            item, list
+        ), f"Only list items are supported. Got {type(item)}"
+        self.output.append(item)
+
+    def __repr__(self):
+        return self.output.__repr__()
+
+    def __len__(self):
+        if self.iter_mode in [
+            SAM3Output.IterMode.ALL_STEPS_PER_STAGE,
+            SAM3Output.IterMode.LAST_STEP_PER_STAGE,
+        ]:
+            return len(self.output)
+        elif self.iter_mode == SAM3Output.IterMode.FLATTENED:
+            flattened_output = sum(self.output, [])
+            return len(flattened_output)

detect_tools/sam3/sam3/model/necks.py

@@ -0,0 +1,125 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates. All Rights Reserved
+
+"""Necks are the interface between a vision backbone and the rest of the detection model"""
+
+from copy import deepcopy
+from typing import List, Optional, Tuple
+
+import torch
+
+import torch.nn as nn
+
+
+class Sam3DualViTDetNeck(nn.Module):
+    def __init__(
+        self,
+        trunk: nn.Module,
+        position_encoding: nn.Module,
+        d_model: int,
+        scale_factors=(4.0, 2.0, 1.0, 0.5),
+        add_sam2_neck: bool = False,
+    ):
+        """
+        SimpleFPN neck a la ViTDet
+        (From detectron2, very lightly adapted)
+        It supports a "dual neck" setting, where we have two identical necks (for SAM3 and SAM2), with different weights
+
+        :param trunk: the backbone
+        :param position_encoding: the positional encoding to use
+        :param d_model: the dimension of the model
+        """
+        super().__init__()
+        self.trunk = trunk
+        self.position_encoding = position_encoding
+        self.convs = nn.ModuleList()
+
+        self.scale_factors = scale_factors
+        use_bias = True
+        dim: int = self.trunk.channel_list[-1]
+
+        for _, scale in enumerate(scale_factors):
+            current = nn.Sequential()
+
+            if scale == 4.0:
+                current.add_module(
+                    "dconv_2x2_0",
+                    nn.ConvTranspose2d(dim, dim // 2, kernel_size=2, stride=2),
+                )
+                current.add_module(
+                    "gelu",
+                    nn.GELU(),
+                )
+                current.add_module(
+                    "dconv_2x2_1",
+                    nn.ConvTranspose2d(dim // 2, dim // 4, kernel_size=2, stride=2),
+                )
+                out_dim = dim // 4
+            elif scale == 2.0:
+                current.add_module(
+                    "dconv_2x2",
+                    nn.ConvTranspose2d(dim, dim // 2, kernel_size=2, stride=2),
+                )
+                out_dim = dim // 2
+            elif scale == 1.0:
+                out_dim = dim
+            elif scale == 0.5:
+                current.add_module(
+                    "maxpool_2x2",
+                    nn.MaxPool2d(kernel_size=2, stride=2),
+                )
+                out_dim = dim
+            else:
+                raise NotImplementedError(f"scale_factor={scale} is not supported yet.")
+
+            current.add_module(
+                "conv_1x1",
+                nn.Conv2d(
+                    in_channels=out_dim,
+                    out_channels=d_model,
+                    kernel_size=1,
+                    bias=use_bias,
+                ),
+            )
+            current.add_module(
+                "conv_3x3",
+                nn.Conv2d(
+                    in_channels=d_model,
+                    out_channels=d_model,
+                    kernel_size=3,
+                    padding=1,
+                    bias=use_bias,
+                ),
+            )
+            self.convs.append(current)
+
+        self.sam2_convs = None
+        if add_sam2_neck:
+            # Assumes sam2 neck is just a clone of the original neck
+            self.sam2_convs = deepcopy(self.convs)
+
+    def forward(
+        self, tensor_list: List[torch.Tensor]
+    ) -> Tuple[
+        List[torch.Tensor],
+        List[torch.Tensor],
+        Optional[List[torch.Tensor]],
+        Optional[List[torch.Tensor]],
+    ]:
+        xs = self.trunk(tensor_list)
+        sam3_out, sam3_pos = [], []
+        sam2_out, sam2_pos = None, None
+        if self.sam2_convs is not None:
+            sam2_out, sam2_pos = [], []
+        x = xs[-1]  # simpleFPN
+        for i in range(len(self.convs)):
+            sam3_x_out = self.convs[i](x)
+            sam3_pos_out = self.position_encoding(sam3_x_out).to(sam3_x_out.dtype)
+            sam3_out.append(sam3_x_out)
+            sam3_pos.append(sam3_pos_out)
+
+            if self.sam2_convs is not None:
+                sam2_x_out = self.sam2_convs[i](x)
+                sam2_pos_out = self.position_encoding(sam2_x_out).to(sam2_x_out.dtype)
+                sam2_out.append(sam2_x_out)
+                sam2_pos.append(sam2_pos_out)
+        return sam3_out, sam3_pos, sam2_out, sam2_pos

detect_tools/sam3/sam3/model/position_encoding.py

@@ -0,0 +1,124 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates. All Rights Reserved
+
+import math
+from typing import Optional
+
+import torch
+from torch import nn
+
+
+class PositionEmbeddingSine(nn.Module):
+    """
+    This is a more standard version of the position embedding, very similar to the one
+    used by the Attention is all you need paper, generalized to work on images.
+    """
+
+    def __init__(
+        self,
+        num_pos_feats,
+        temperature: int = 10000,
+        normalize: bool = True,
+        scale: Optional[float] = None,
+        precompute_resolution: Optional[int] = None,
+    ):
+        super().__init__()
+        assert num_pos_feats % 2 == 0, "Expecting even model width"
+        self.num_pos_feats = num_pos_feats // 2
+        self.temperature = temperature
+        self.normalize = normalize
+        if scale is not None and normalize is False:
+            raise ValueError("normalize should be True if scale is passed")
+        if scale is None:
+            scale = 2 * math.pi
+        self.scale = scale
+
+        self.cache = {}
+        # Precompute positional encodings under `precompute_resolution` to fill the cache
+        # and avoid symbolic shape tracing errors in torch.compile in PyTorch 2.4 nightly.
+        if precompute_resolution is not None:
+            # We precompute pos enc for stride 4, 8, 16 and 32 to fill `self.cache`.
+            precompute_sizes = [
+                (precompute_resolution // 4, precompute_resolution // 4),
+                (precompute_resolution // 8, precompute_resolution // 8),
+                (precompute_resolution // 16, precompute_resolution // 16),
+                (precompute_resolution // 32, precompute_resolution // 32),
+            ]
+            for size in precompute_sizes:
+                tensors = torch.zeros((1, 1) + size, device="cuda")
+                self.forward(tensors)
+                # further clone and detach it in the cache (just to be safe)
+                self.cache[size] = self.cache[size].clone().detach()
+
+    def _encode_xy(self, x, y):
+        # The positions are expected to be normalized
+        assert len(x) == len(y) and x.ndim == y.ndim == 1
+        x_embed = x * self.scale
+        y_embed = y * self.scale
+
+        dim_t = torch.arange(self.num_pos_feats, dtype=torch.float32, device=x.device)
+        dim_t = self.temperature ** (2 * (dim_t // 2) / self.num_pos_feats)
+
+        pos_x = x_embed[:, None] / dim_t
+        pos_y = y_embed[:, None] / dim_t
+        pos_x = torch.stack(
+            (pos_x[:, 0::2].sin(), pos_x[:, 1::2].cos()), dim=2
+        ).flatten(1)
+        pos_y = torch.stack(
+            (pos_y[:, 0::2].sin(), pos_y[:, 1::2].cos()), dim=2
+        ).flatten(1)
+        return pos_x, pos_y
+
+    @torch.no_grad()
+    def encode_boxes(self, x, y, w, h):
+        pos_x, pos_y = self._encode_xy(x, y)
+        pos = torch.cat((pos_y, pos_x, h[:, None], w[:, None]), dim=1)
+        return pos
+
+    encode = encode_boxes  # Backwards compatibility
+
+    @torch.no_grad()
+    def encode_points(self, x, y, labels):
+        (bx, nx), (by, ny), (bl, nl) = x.shape, y.shape, labels.shape
+        assert bx == by and nx == ny and bx == bl and nx == nl
+        pos_x, pos_y = self._encode_xy(x.flatten(), y.flatten())
+        pos_x, pos_y = pos_x.reshape(bx, nx, -1), pos_y.reshape(by, ny, -1)
+        pos = torch.cat((pos_y, pos_x, labels[:, :, None]), dim=2)
+        return pos
+
+    @torch.no_grad()
+    def forward(self, x):
+        cache_key = None
+        cache_key = (x.shape[-2], x.shape[-1])
+        if cache_key in self.cache:
+            return self.cache[cache_key][None].repeat(x.shape[0], 1, 1, 1)
+        y_embed = (
+            torch.arange(1, x.shape[-2] + 1, dtype=torch.float32, device=x.device)
+            .view(1, -1, 1)
+            .repeat(x.shape[0], 1, x.shape[-1])
+        )
+        x_embed = (
+            torch.arange(1, x.shape[-1] + 1, dtype=torch.float32, device=x.device)
+            .view(1, 1, -1)
+            .repeat(x.shape[0], x.shape[-2], 1)
+        )
+
+        if self.normalize:
+            eps = 1e-6
+            y_embed = y_embed / (y_embed[:, -1:, :] + eps) * self.scale
+            x_embed = x_embed / (x_embed[:, :, -1:] + eps) * self.scale
+
+        dim_t = torch.arange(self.num_pos_feats, dtype=torch.float32, device=x.device)
+        dim_t = self.temperature ** (2 * (dim_t // 2) / self.num_pos_feats)
+
+        pos_x = x_embed[:, :, :, None] / dim_t
+        pos_y = y_embed[:, :, :, None] / dim_t
+        pos_x = torch.stack(
+            (pos_x[:, :, :, 0::2].sin(), pos_x[:, :, :, 1::2].cos()), dim=4
+        ).flatten(3)
+        pos_y = torch.stack(
+            (pos_y[:, :, :, 0::2].sin(), pos_y[:, :, :, 1::2].cos()), dim=4
+        ).flatten(3)
+        pos = torch.cat((pos_y, pos_x), dim=3).permute(0, 3, 1, 2)
+        if cache_key is not None:
+            self.cache[cache_key] = pos[0]
+        return pos

detect_tools/sam3/sam3/model/sam1_task_predictor.py

@@ -0,0 +1,458 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates. All Rights Reserved
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import logging
+
+from typing import List, Optional, Tuple, Union
+
+import numpy as np
+import torch
+
+import torch.nn as nn
+from PIL.Image import Image
+
+from sam3.model.sam3_tracker_base import Sam3TrackerBase
+from sam3.model.utils.sam1_utils import SAM2Transforms
+
+
+# Adapted from https://github.com/facebookresearch/sam2/blob/main/sam2/sam2_image_predictor.py
+class SAM3InteractiveImagePredictor(nn.Module):
+    def __init__(
+        self,
+        sam_model: Sam3TrackerBase,
+        mask_threshold=0.0,
+        max_hole_area=256.0,
+        max_sprinkle_area=0.0,
+        **kwargs,
+    ) -> None:
+        """
+        Uses SAM-3 to calculate the image embedding for an image, and then
+        allow repeated, efficient mask prediction given prompts.
+
+        Arguments:
+          sam_model : The model to use for mask prediction.
+          mask_threshold (float): The threshold to use when converting mask logits
+            to binary masks. Masks are thresholded at 0 by default.
+          max_hole_area (int): If max_hole_area > 0, we fill small holes in up to
+            the maximum area of max_hole_area in low_res_masks.
+          max_sprinkle_area (int): If max_sprinkle_area > 0, we remove small sprinkles up to
+            the maximum area of max_sprinkle_area in low_res_masks.
+        """
+        super().__init__()
+        self.model = sam_model
+        self._transforms = SAM2Transforms(
+            resolution=self.model.image_size,
+            mask_threshold=mask_threshold,
+            max_hole_area=max_hole_area,
+            max_sprinkle_area=max_sprinkle_area,
+        )
+
+        # Predictor state
+        self._is_image_set = False
+        self._features = None
+        self._orig_hw = None
+        # Whether the predictor is set for single image or a batch of images
+        self._is_batch = False
+
+        # Predictor config
+        self.mask_threshold = mask_threshold
+
+        # Spatial dim for backbone feature maps
+        self._bb_feat_sizes = [
+            (288, 288),
+            (144, 144),
+            (72, 72),
+        ]
+
+    @torch.no_grad()
+    def set_image(
+        self,
+        image: Union[np.ndarray, Image],
+    ) -> None:
+        """
+        Calculates the image embeddings for the provided image, allowing
+        masks to be predicted with the 'predict' method.
+
+        Arguments:
+          image (np.ndarray or PIL Image): The input image to embed in RGB format. The image should be in HWC format if np.ndarray, or WHC format if PIL Image
+          with pixel values in [0, 255].
+          image_format (str): The color format of the image, in ['RGB', 'BGR'].
+        """
+        self.reset_predictor()
+        # Transform the image to the form expected by the model
+        if isinstance(image, np.ndarray):
+            logging.info("For numpy array image, we assume (HxWxC) format")
+            self._orig_hw = [image.shape[:2]]
+        elif isinstance(image, Image):
+            w, h = image.size
+            self._orig_hw = [(h, w)]
+        else:
+            raise NotImplementedError("Image format not supported")
+
+        input_image = self._transforms(image)
+        input_image = input_image[None, ...].to(self.device)
+
+        assert (
+            len(input_image.shape) == 4 and input_image.shape[1] == 3
+        ), f"input_image must be of size 1x3xHxW, got {input_image.shape}"
+        logging.info("Computing image embeddings for the provided image...")
+        backbone_out = self.model.forward_image(input_image)
+        (
+            _,
+            vision_feats,
+            _,
+            _,
+        ) = self.model._prepare_backbone_features(backbone_out)
+        # Add no_mem_embed, which is added to the lowest rest feat. map during training on videos
+        vision_feats[-1] = vision_feats[-1] + self.model.no_mem_embed
+
+        feats = [
+            feat.permute(1, 2, 0).view(1, -1, *feat_size)
+            for feat, feat_size in zip(vision_feats[::-1], self._bb_feat_sizes[::-1])
+        ][::-1]
+        self._features = {"image_embed": feats[-1], "high_res_feats": feats[:-1]}
+        self._is_image_set = True
+        logging.info("Image embeddings computed.")
+
+    @torch.no_grad()
+    def set_image_batch(
+        self,
+        image_list: List[Union[np.ndarray]],
+    ) -> None:
+        """
+        Calculates the image embeddings for the provided image batch, allowing
+        masks to be predicted with the 'predict_batch' method.
+
+        Arguments:
+          image_list (List[np.ndarray]): The input images to embed in RGB format. The image should be in HWC format if np.ndarray
+          with pixel values in [0, 255].
+        """
+        self.reset_predictor()
+        assert isinstance(image_list, list)
+        self._orig_hw = []
+        for image in image_list:
+            assert isinstance(
+                image, np.ndarray
+            ), "Images are expected to be an np.ndarray in RGB format, and of shape  HWC"
+            self._orig_hw.append(image.shape[:2])
+        # Transform the image to the form expected by the model
+        img_batch = self._transforms.forward_batch(image_list)
+        img_batch = img_batch.to(self.device)
+        batch_size = img_batch.shape[0]
+        assert (
+            len(img_batch.shape) == 4 and img_batch.shape[1] == 3
+        ), f"img_batch must be of size Bx3xHxW, got {img_batch.shape}"
+        logging.info("Computing image embeddings for the provided images...")
+        backbone_out = self.model.forward_image(img_batch)
+        (
+            _,
+            vision_feats,
+            _,
+            _,
+        ) = self.model._prepare_backbone_features(backbone_out)
+        # Add no_mem_embed, which is added to the lowest rest feat. map during training on videos
+        vision_feats[-1] = vision_feats[-1] + self.model.no_mem_embed
+
+        feats = [
+            feat.permute(1, 2, 0).view(batch_size, -1, *feat_size)
+            for feat, feat_size in zip(vision_feats[::-1], self._bb_feat_sizes[::-1])
+        ][::-1]
+        self._features = {"image_embed": feats[-1], "high_res_feats": feats[:-1]}
+        self._is_image_set = True
+        self._is_batch = True
+        logging.info("Image embeddings computed.")
+
+    def predict_batch(
+        self,
+        point_coords_batch: List[np.ndarray] = None,
+        point_labels_batch: List[np.ndarray] = None,
+        box_batch: List[np.ndarray] = None,
+        mask_input_batch: List[np.ndarray] = None,
+        multimask_output: bool = True,
+        return_logits: bool = False,
+        normalize_coords=True,
+    ) -> Tuple[List[np.ndarray], List[np.ndarray], List[np.ndarray]]:
+        """This function is very similar to predict(...), however it is used for batched mode, when the model is expected to generate predictions on multiple images.
+        It returns a tuple of lists of masks, ious, and low_res_masks_logits.
+        """
+        assert self._is_batch, "This function should only be used when in batched mode"
+        if not self._is_image_set:
+            raise RuntimeError(
+                "An image must be set with .set_image_batch(...) before mask prediction."
+            )
+        num_images = len(self._features["image_embed"])
+        all_masks = []
+        all_ious = []
+        all_low_res_masks = []
+        for img_idx in range(num_images):
+            # Transform input prompts
+            point_coords = (
+                point_coords_batch[img_idx] if point_coords_batch is not None else None
+            )
+            point_labels = (
+                point_labels_batch[img_idx] if point_labels_batch is not None else None
+            )
+            box = box_batch[img_idx] if box_batch is not None else None
+            mask_input = (
+                mask_input_batch[img_idx] if mask_input_batch is not None else None
+            )
+            mask_input, unnorm_coords, labels, unnorm_box = self._prep_prompts(
+                point_coords,
+                point_labels,
+                box,
+                mask_input,
+                normalize_coords,
+                img_idx=img_idx,
+            )
+            masks, iou_predictions, low_res_masks = self._predict(
+                unnorm_coords,
+                labels,
+                unnorm_box,
+                mask_input,
+                multimask_output,
+                return_logits=return_logits,
+                img_idx=img_idx,
+            )
+            masks_np = masks.squeeze(0).float().detach().cpu().numpy()
+            iou_predictions_np = (
+                iou_predictions.squeeze(0).float().detach().cpu().numpy()
+            )
+            low_res_masks_np = low_res_masks.squeeze(0).float().detach().cpu().numpy()
+            all_masks.append(masks_np)
+            all_ious.append(iou_predictions_np)
+            all_low_res_masks.append(low_res_masks_np)
+
+        return all_masks, all_ious, all_low_res_masks
+
+    def predict(
+        self,
+        point_coords: Optional[np.ndarray] = None,
+        point_labels: Optional[np.ndarray] = None,
+        box: Optional[np.ndarray] = None,
+        mask_input: Optional[np.ndarray] = None,
+        multimask_output: bool = True,
+        return_logits: bool = False,
+        normalize_coords=True,
+    ) -> Tuple[np.ndarray, np.ndarray, np.ndarray]:
+        """
+        Predict masks for the given input prompts, using the currently set image.
+
+        Arguments:
+          point_coords (np.ndarray or None): A Nx2 array of point prompts to the
+            model. Each point is in (X,Y) in pixels.
+          point_labels (np.ndarray or None): A length N array of labels for the
+            point prompts. 1 indicates a foreground point and 0 indicates a
+            background point.
+          box (np.ndarray or None): A length 4 array given a box prompt to the
+            model, in XYXY format.
+          mask_input (np.ndarray): A low resolution mask input to the model, typically
+            coming from a previous prediction iteration. Has form 1xHxW, where
+            for SAM, H=W=256.
+          multimask_output (bool): If true, the model will return three masks.
+            For ambiguous input prompts (such as a single click), this will often
+            produce better masks than a single prediction. If only a single
+            mask is needed, the model's predicted quality score can be used
+            to select the best mask. For non-ambiguous prompts, such as multiple
+            input prompts, multimask_output=False can give better results.
+          return_logits (bool): If true, returns un-thresholded masks logits
+            instead of a binary mask.
+          normalize_coords (bool): If true, the point coordinates will be normalized to the range [0,1] and point_coords is expected to be wrt. image dimensions.
+
+        Returns:
+          (np.ndarray): The output masks in CxHxW format, where C is the
+            number of masks, and (H, W) is the original image size.
+          (np.ndarray): An array of length C containing the model's
+            predictions for the quality of each mask.
+          (np.ndarray): An array of shape CxHxW, where C is the number
+            of masks and H=W=256. These low resolution logits can be passed to
+            a subsequent iteration as mask input.
+        """
+        if not self._is_image_set:
+            raise RuntimeError(
+                "An image must be set with .set_image(...) before mask prediction."
+            )
+
+        # Transform input prompts
+
+        mask_input, unnorm_coords, labels, unnorm_box = self._prep_prompts(
+            point_coords, point_labels, box, mask_input, normalize_coords
+        )
+
+        masks, iou_predictions, low_res_masks = self._predict(
+            unnorm_coords,
+            labels,
+            unnorm_box,
+            mask_input,
+            multimask_output,
+            return_logits=return_logits,
+        )
+
+        masks_np = masks.squeeze(0).float().detach().cpu().numpy()
+        iou_predictions_np = iou_predictions.squeeze(0).float().detach().cpu().numpy()
+        low_res_masks_np = low_res_masks.squeeze(0).float().detach().cpu().numpy()
+        return masks_np, iou_predictions_np, low_res_masks_np
+
+    def _prep_prompts(
+        self, point_coords, point_labels, box, mask_logits, normalize_coords, img_idx=-1
+    ):
+        unnorm_coords, labels, unnorm_box, mask_input = None, None, None, None
+        if point_coords is not None:
+            assert (
+                point_labels is not None
+            ), "point_labels must be supplied if point_coords is supplied."
+            point_coords = torch.as_tensor(
+                point_coords, dtype=torch.float, device=self.device
+            )
+            unnorm_coords = self._transforms.transform_coords(
+                point_coords, normalize=normalize_coords, orig_hw=self._orig_hw[img_idx]
+            )
+            labels = torch.as_tensor(point_labels, dtype=torch.int, device=self.device)
+            if len(unnorm_coords.shape) == 2:
+                unnorm_coords, labels = unnorm_coords[None, ...], labels[None, ...]
+        if box is not None:
+            box = torch.as_tensor(box, dtype=torch.float, device=self.device)
+            unnorm_box = self._transforms.transform_boxes(
+                box, normalize=normalize_coords, orig_hw=self._orig_hw[img_idx]
+            )  # Bx2x2
+        if mask_logits is not None:
+            mask_input = torch.as_tensor(
+                mask_logits, dtype=torch.float, device=self.device
+            )
+            if len(mask_input.shape) == 3:
+                mask_input = mask_input[None, :, :, :]
+        return mask_input, unnorm_coords, labels, unnorm_box
+
+    @torch.no_grad()
+    def _predict(
+        self,
+        point_coords: Optional[torch.Tensor],
+        point_labels: Optional[torch.Tensor],
+        boxes: Optional[torch.Tensor] = None,
+        mask_input: Optional[torch.Tensor] = None,
+        multimask_output: bool = True,
+        return_logits: bool = False,
+        img_idx: int = -1,
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        """
+        Predict masks for the given input prompts, using the currently set image.
+        Input prompts are batched torch tensors and are expected to already be
+        transformed to the input frame using SAM2Transforms.
+
+        Arguments:
+          point_coords (torch.Tensor or None): A BxNx2 array of point prompts to the
+            model. Each point is in (X,Y) in pixels.
+          point_labels (torch.Tensor or None): A BxN array of labels for the
+            point prompts. 1 indicates a foreground point and 0 indicates a
+            background point.
+          boxes (np.ndarray or None): A Bx4 array given a box prompt to the
+            model, in XYXY format.
+          mask_input (np.ndarray): A low resolution mask input to the model, typically
+            coming from a previous prediction iteration. Has form Bx1xHxW, where
+            for SAM, H=W=256. Masks returned by a previous iteration of the
+            predict method do not need further transformation.
+          multimask_output (bool): If true, the model will return three masks.
+            For ambiguous input prompts (such as a single click), this will often
+            produce better masks than a single prediction. If only a single
+            mask is needed, the model's predicted quality score can be used
+            to select the best mask. For non-ambiguous prompts, such as multiple
+            input prompts, multimask_output=False can give better results.
+          return_logits (bool): If true, returns un-thresholded masks logits
+            instead of a binary mask.
+
+        Returns:
+          (torch.Tensor): The output masks in BxCxHxW format, where C is the
+            number of masks, and (H, W) is the original image size.
+          (torch.Tensor): An array of shape BxC containing the model's
+            predictions for the quality of each mask.
+          (torch.Tensor): An array of shape BxCxHxW, where C is the number
+            of masks and H=W=256. These low res logits can be passed to
+            a subsequent iteration as mask input.
+        """
+        if not self._is_image_set:
+            raise RuntimeError(
+                "An image must be set with .set_image(...) before mask prediction."
+            )
+
+        if point_coords is not None:
+            concat_points = (point_coords, point_labels)
+        else:
+            concat_points = None
+
+        # Embed prompts
+        if boxes is not None:
+            box_coords = boxes.reshape(-1, 2, 2)
+            box_labels = torch.tensor([[2, 3]], dtype=torch.int, device=boxes.device)
+            box_labels = box_labels.repeat(boxes.size(0), 1)
+            # we merge "boxes" and "points" into a single "concat_points" input (where
+            # boxes are added at the beginning) to sam_prompt_encoder
+            if concat_points is not None:
+                concat_coords = torch.cat([box_coords, concat_points[0]], dim=1)
+                concat_labels = torch.cat([box_labels, concat_points[1]], dim=1)
+                concat_points = (concat_coords, concat_labels)
+            else:
+                concat_points = (box_coords, box_labels)
+
+        sparse_embeddings, dense_embeddings = self.model.sam_prompt_encoder(
+            points=concat_points,
+            boxes=None,
+            masks=mask_input,
+        )
+
+        # Predict masks
+        batched_mode = (
+            concat_points is not None and concat_points[0].shape[0] > 1
+        )  # multi object prediction
+        high_res_features = [
+            feat_level[img_idx].unsqueeze(0)
+            for feat_level in self._features["high_res_feats"]
+        ]
+        low_res_masks, iou_predictions, _, _ = self.model.sam_mask_decoder(
+            image_embeddings=self._features["image_embed"][img_idx].unsqueeze(0),
+            image_pe=self.model.sam_prompt_encoder.get_dense_pe(),
+            sparse_prompt_embeddings=sparse_embeddings,
+            dense_prompt_embeddings=dense_embeddings,
+            multimask_output=multimask_output,
+            repeat_image=batched_mode,
+            high_res_features=high_res_features,
+        )
+
+        # Upscale the masks to the original image resolution
+        masks = self._transforms.postprocess_masks(
+            low_res_masks, self._orig_hw[img_idx]
+        )
+        low_res_masks = torch.clamp(low_res_masks, -32.0, 32.0)
+        if not return_logits:
+            masks = masks > self.mask_threshold
+
+        return masks, iou_predictions, low_res_masks
+
+    def get_image_embedding(self) -> torch.Tensor:
+        """
+        Returns the image embeddings for the currently set image, with
+        shape 1xCxHxW, where C is the embedding dimension and (H,W) are
+        the embedding spatial dimension of SAM (typically C=256, H=W=64).
+        """
+        if not self._is_image_set:
+            raise RuntimeError(
+                "An image must be set with .set_image(...) to generate an embedding."
+            )
+        assert (
+            self._features is not None
+        ), "Features must exist if an image has been set."
+        return self._features["image_embed"]
+
+    @property
+    def device(self) -> torch.device:
+        return self.model.device
+
+    def reset_predictor(self) -> None:
+        """
+        Resets the image embeddings and other state variables.
+        """
+        self._is_image_set = False
+        self._features = None
+        self._orig_hw = None
+        self._is_batch = False

detect_tools/sam3/sam3/model/sam3_image.py

@@ -0,0 +1,883 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates. All Rights Reserved
+
+import os
+from copy import deepcopy
+from typing import Dict, List, Optional, Tuple
+
+import numpy as np
+import torch
+
+from sam3.model.model_misc import SAM3Output
+
+from sam3.model.sam1_task_predictor import SAM3InteractiveImagePredictor
+from sam3.model.vl_combiner import SAM3VLBackbone
+from sam3.perflib.nms import nms_masks
+
+from sam3.train.data.collator import BatchedDatapoint
+
+from .act_ckpt_utils import activation_ckpt_wrapper
+
+from .box_ops import box_cxcywh_to_xyxy
+
+from .geometry_encoders import Prompt
+from .model_misc import inverse_sigmoid
+
+
+def _update_out(out, out_name, out_value, auxiliary=True, update_aux=True):
+    out[out_name] = out_value[-1] if auxiliary else out_value
+    if auxiliary and update_aux:
+        if "aux_outputs" not in out:
+            out["aux_outputs"] = [{} for _ in range(len(out_value) - 1)]
+        assert len(out["aux_outputs"]) == len(out_value) - 1
+        for aux_output, aux_value in zip(out["aux_outputs"], out_value[:-1]):
+            aux_output[out_name] = aux_value
+
+
+class Sam3Image(torch.nn.Module):
+    TEXT_ID_FOR_TEXT = 0
+    TEXT_ID_FOR_VISUAL = 1
+    TEXT_ID_FOR_GEOMETRIC = 2
+
+    def __init__(
+        self,
+        backbone: SAM3VLBackbone,
+        transformer,
+        input_geometry_encoder,
+        segmentation_head=None,
+        num_feature_levels=1,
+        o2m_mask_predict=True,
+        dot_prod_scoring=None,
+        use_instance_query: bool = True,
+        multimask_output: bool = True,
+        use_act_checkpoint_seg_head: bool = True,
+        interactivity_in_encoder: bool = True,
+        matcher=None,
+        use_dot_prod_scoring=True,
+        supervise_joint_box_scores: bool = False,  # only relevant if using presence token/score
+        detach_presence_in_joint_score: bool = False,  # only relevant if using presence token/score
+        separate_scorer_for_instance: bool = False,
+        num_interactive_steps_val: int = 0,
+        inst_interactive_predictor: SAM3InteractiveImagePredictor = None,
+        **kwargs,
+    ):
+        super().__init__()
+        self.backbone = backbone
+        self.geometry_encoder = input_geometry_encoder
+        self.transformer = transformer
+        self.hidden_dim = transformer.d_model
+        self.num_feature_levels = num_feature_levels
+        self.segmentation_head = segmentation_head
+
+        self.o2m_mask_predict = o2m_mask_predict
+
+        self.dot_prod_scoring = dot_prod_scoring
+        self.use_act_checkpoint_seg_head = use_act_checkpoint_seg_head
+        self.interactivity_in_encoder = interactivity_in_encoder
+        self.matcher = matcher
+
+        self.num_interactive_steps_val = num_interactive_steps_val
+        self.use_dot_prod_scoring = use_dot_prod_scoring
+
+        if self.use_dot_prod_scoring:
+            assert dot_prod_scoring is not None
+            self.dot_prod_scoring = dot_prod_scoring
+            self.instance_dot_prod_scoring = None
+            if separate_scorer_for_instance:
+                self.instance_dot_prod_scoring = deepcopy(dot_prod_scoring)
+        else:
+            self.class_embed = torch.nn.Linear(self.hidden_dim, 1)
+            self.instance_class_embed = None
+            if separate_scorer_for_instance:
+                self.instance_class_embed = deepcopy(self.class_embed)
+
+        self.supervise_joint_box_scores = supervise_joint_box_scores
+        self.detach_presence_in_joint_score = detach_presence_in_joint_score
+
+        # verify the number of queries for O2O and O2M
+        num_o2o_static = self.transformer.decoder.num_queries
+        num_o2m_static = self.transformer.decoder.num_o2m_queries
+        assert num_o2m_static == (num_o2o_static if self.transformer.decoder.dac else 0)
+        self.dac = self.transformer.decoder.dac
+
+        self.use_instance_query = use_instance_query
+        self.multimask_output = multimask_output
+
+        self.inst_interactive_predictor = inst_interactive_predictor
+
+    @property
+    def device(self):
+        self._device = getattr(self, "_device", None) or next(self.parameters()).device
+        return self._device
+
+    def to(self, *args, **kwargs):
+        # clear cached _device in case the model is moved to a different device
+        self._device = None
+        return super().to(*args, **kwargs)
+
+    def _get_img_feats(self, backbone_out, img_ids):
+        """Retrieve correct image features from backbone output."""
+        if "backbone_fpn" in backbone_out:
+            if "id_mapping" in backbone_out and backbone_out["id_mapping"] is not None:
+                img_ids = backbone_out["id_mapping"][img_ids]
+                # If this assert fails, it likely means we're requesting different img_ids (perhaps a different frame?)
+                # We currently don't expect this to happen. We could technically trigger a recompute here,
+                # but likely at the cost of a cpu<->gpu sync point, which would deteriorate perf
+                torch._assert_async((img_ids >= 0).all())
+
+            vis_feats = backbone_out["backbone_fpn"][-self.num_feature_levels :]
+            vis_pos_enc = backbone_out["vision_pos_enc"][-self.num_feature_levels :]
+            vis_feat_sizes = [x.shape[-2:] for x in vis_pos_enc]  # (H, W) shapes
+            # index and flatten visual features NxCxHxW => HWxNxC (batch-first => seq-first)
+            img_feats = [x[img_ids].flatten(2).permute(2, 0, 1) for x in vis_feats]
+            img_pos_embeds = [
+                x[img_ids].flatten(2).permute(2, 0, 1) for x in vis_pos_enc
+            ]
+            return backbone_out, img_feats, img_pos_embeds, vis_feat_sizes
+
+        # Image features not available in backbone output, so we compute them on the fly
+        # This case likely occurs for video. In that case, we want to forward only the current frame
+        img_batch = backbone_out["img_batch_all_stages"]
+        if img_ids.numel() > 1:
+            # Only forward backbone on unique image ids to avoid repetitive computation
+            unique_ids, _ = torch.unique(img_ids, return_inverse=True)
+        else:
+            unique_ids, _ = img_ids, slice(None)
+        # Compute the image features on those unique image ids
+        # note: we allow using a list (or other indexable types) of tensors as img_batch
+        # (e.g. for async frame loading in demo). In this case we index img_batch.tensors directly
+        if isinstance(img_batch, torch.Tensor):
+            image = img_batch[unique_ids]
+        elif unique_ids.numel() == 1:
+            image = img_batch[unique_ids.item()].unsqueeze(0)
+        else:
+            image = torch.stack([img_batch[i] for i in unique_ids.tolist()])
+        # `img_batch` might be fp16 and offloaded to CPU
+        image = image.to(dtype=torch.float32, device=self.device)
+        # Next time we call this function, we want to remember which indices we computed
+        id_mapping = torch.full(
+            (len(img_batch),), -1, dtype=torch.long, device=self.device
+        )
+        id_mapping[unique_ids] = torch.arange(len(unique_ids), device=self.device)
+        backbone_out = {
+            **backbone_out,
+            **self.backbone.forward_image(image),
+            "id_mapping": id_mapping,
+        }
+        assert "backbone_fpn" in backbone_out
+        return self._get_img_feats(backbone_out, img_ids=img_ids)
+
+    def _encode_prompt(
+        self,
+        backbone_out,
+        find_input,
+        geometric_prompt,
+        visual_prompt_embed=None,
+        visual_prompt_mask=None,
+        encode_text=True,
+        prev_mask_pred=None,
+    ):
+        # index text features (note that regardless of early or late fusion, the batch size of
+        # `txt_feats` is always the number of *prompts* in the encoder)
+        txt_ids = find_input.text_ids
+        txt_feats = backbone_out["language_features"][:, txt_ids]
+        txt_masks = backbone_out["language_mask"][txt_ids]
+
+        feat_tuple = self._get_img_feats(backbone_out, find_input.img_ids)
+        backbone_out, img_feats, img_pos_embeds, vis_feat_sizes = feat_tuple
+
+        if prev_mask_pred is not None:
+            img_feats = [img_feats[-1] + prev_mask_pred]
+        # Encode geometry
+        geo_feats, geo_masks = self.geometry_encoder(
+            geo_prompt=geometric_prompt,
+            img_feats=img_feats,
+            img_sizes=vis_feat_sizes,
+            img_pos_embeds=img_pos_embeds,
+        )
+        if visual_prompt_embed is None:
+            visual_prompt_embed = torch.zeros(
+                (0, *geo_feats.shape[1:]), device=geo_feats.device
+            )
+            visual_prompt_mask = torch.zeros(
+                (*geo_masks.shape[:-1], 0),
+                device=geo_masks.device,
+                dtype=geo_masks.dtype,
+            )
+        if encode_text:
+            prompt = torch.cat([txt_feats, geo_feats, visual_prompt_embed], dim=0)
+            prompt_mask = torch.cat([txt_masks, geo_masks, visual_prompt_mask], dim=1)
+        else:
+            prompt = torch.cat([geo_feats, visual_prompt_embed], dim=0)
+            prompt_mask = torch.cat([geo_masks, visual_prompt_mask], dim=1)
+        return prompt, prompt_mask, backbone_out
+
+    def _run_encoder(
+        self,
+        backbone_out,
+        find_input,
+        prompt,
+        prompt_mask,
+        encoder_extra_kwargs: Optional[Dict] = None,
+    ):
+        feat_tuple = self._get_img_feats(backbone_out, find_input.img_ids)
+        backbone_out, img_feats, img_pos_embeds, vis_feat_sizes = feat_tuple
+
+        # Run the encoder
+        prompt_pos_embed = torch.zeros_like(prompt)
+        # make a copy of the image feature lists since the encoder may modify these lists in-place
+        memory = self.transformer.encoder(
+            src=img_feats.copy(),
+            src_key_padding_mask=None,
+            src_pos=img_pos_embeds.copy(),
+            prompt=prompt,
+            prompt_pos=prompt_pos_embed,
+            prompt_key_padding_mask=prompt_mask,
+            feat_sizes=vis_feat_sizes,
+            encoder_extra_kwargs=encoder_extra_kwargs,
+        )
+        encoder_out = {
+            # encoded image features
+            "encoder_hidden_states": memory["memory"],
+            "pos_embed": memory["pos_embed"],
+            "padding_mask": memory["padding_mask"],
+            "level_start_index": memory["level_start_index"],
+            "spatial_shapes": memory["spatial_shapes"],
+            "valid_ratios": memory["valid_ratios"],
+            "vis_feat_sizes": vis_feat_sizes,
+            # encoded text features (or other prompts)
+            "prompt_before_enc": prompt,
+            "prompt_after_enc": memory.get("memory_text", prompt),
+            "prompt_mask": prompt_mask,
+        }
+        return backbone_out, encoder_out, feat_tuple
+
+    def _run_decoder(
+        self,
+        pos_embed,
+        memory,
+        src_mask,
+        out,
+        prompt,
+        prompt_mask,
+        encoder_out,
+    ):
+        bs = memory.shape[1]
+        query_embed = self.transformer.decoder.query_embed.weight
+        tgt = query_embed.unsqueeze(1).repeat(1, bs, 1)
+
+        apply_dac = self.transformer.decoder.dac and self.training
+        hs, reference_boxes, dec_presence_out, dec_presence_feats = (
+            self.transformer.decoder(
+                tgt=tgt,
+                memory=memory,
+                memory_key_padding_mask=src_mask,
+                pos=pos_embed,
+                reference_boxes=None,
+                level_start_index=encoder_out["level_start_index"],
+                spatial_shapes=encoder_out["spatial_shapes"],
+                valid_ratios=encoder_out["valid_ratios"],
+                tgt_mask=None,
+                memory_text=prompt,
+                text_attention_mask=prompt_mask,
+                apply_dac=apply_dac,
+            )
+        )
+        hs = hs.transpose(1, 2)  # seq-first to batch-first
+        reference_boxes = reference_boxes.transpose(1, 2)  # seq-first to batch-first
+        if dec_presence_out is not None:
+            # seq-first to batch-first
+            dec_presence_out = dec_presence_out.transpose(1, 2)
+
+        out["presence_feats"] = dec_presence_feats
+        self._update_scores_and_boxes(
+            out,
+            hs,
+            reference_boxes,
+            prompt,
+            prompt_mask,
+            dec_presence_out=dec_presence_out,
+        )
+        return out, hs
+
+    def _update_scores_and_boxes(
+        self,
+        out,
+        hs,
+        reference_boxes,
+        prompt,
+        prompt_mask,
+        dec_presence_out=None,
+        is_instance_prompt=False,
+    ):
+        apply_dac = self.transformer.decoder.dac and self.training
+        num_o2o = (hs.size(2) // 2) if apply_dac else hs.size(2)
+        num_o2m = hs.size(2) - num_o2o
+        assert num_o2m == (num_o2o if apply_dac else 0)
+        out["queries"] = hs[-1][:, :num_o2o]  # remove o2m queries if there are any
+        # score prediction
+        if self.use_dot_prod_scoring:
+            dot_prod_scoring_head = self.dot_prod_scoring
+            if is_instance_prompt and self.instance_dot_prod_scoring is not None:
+                dot_prod_scoring_head = self.instance_dot_prod_scoring
+            outputs_class = dot_prod_scoring_head(hs, prompt, prompt_mask)
+        else:
+            class_embed_head = self.class_embed
+            if is_instance_prompt and self.instance_class_embed is not None:
+                class_embed_head = self.instance_class_embed
+            outputs_class = class_embed_head(hs)
+
+        # box prediction
+        box_head = self.transformer.decoder.bbox_embed
+        if (
+            is_instance_prompt
+            and self.transformer.decoder.instance_bbox_embed is not None
+        ):
+            box_head = self.transformer.decoder.instance_bbox_embed
+        anchor_box_offsets = box_head(hs)
+        reference_boxes_inv_sig = inverse_sigmoid(reference_boxes)
+        outputs_coord = (reference_boxes_inv_sig + anchor_box_offsets).sigmoid()
+        outputs_boxes_xyxy = box_cxcywh_to_xyxy(outputs_coord)
+
+        if dec_presence_out is not None:
+            _update_out(
+                out, "presence_logit_dec", dec_presence_out, update_aux=self.training
+            )
+
+        if self.supervise_joint_box_scores:
+            assert dec_presence_out is not None
+            prob_dec_presence_out = dec_presence_out.clone().sigmoid()
+            if self.detach_presence_in_joint_score:
+                prob_dec_presence_out = prob_dec_presence_out.detach()
+
+            outputs_class = inverse_sigmoid(
+                outputs_class.sigmoid() * prob_dec_presence_out.unsqueeze(2)
+            ).clamp(min=-10.0, max=10.0)
+
+        _update_out(
+            out, "pred_logits", outputs_class[:, :, :num_o2o], update_aux=self.training
+        )
+        _update_out(
+            out, "pred_boxes", outputs_coord[:, :, :num_o2o], update_aux=self.training
+        )
+        _update_out(
+            out,
+            "pred_boxes_xyxy",
+            outputs_boxes_xyxy[:, :, :num_o2o],
+            update_aux=self.training,
+        )
+        if num_o2m > 0 and self.training:
+            _update_out(
+                out,
+                "pred_logits_o2m",
+                outputs_class[:, :, num_o2o:],
+                update_aux=self.training,
+            )
+            _update_out(
+                out,
+                "pred_boxes_o2m",
+                outputs_coord[:, :, num_o2o:],
+                update_aux=self.training,
+            )
+            _update_out(
+                out,
+                "pred_boxes_xyxy_o2m",
+                outputs_boxes_xyxy[:, :, num_o2o:],
+                update_aux=self.training,
+            )
+
+    def _run_segmentation_heads(
+        self,
+        out,
+        backbone_out,
+        img_ids,
+        vis_feat_sizes,
+        encoder_hidden_states,
+        prompt,
+        prompt_mask,
+        hs,
+    ):
+        apply_dac = self.transformer.decoder.dac and self.training
+        if self.segmentation_head is not None:
+            num_o2o = (hs.size(2) // 2) if apply_dac else hs.size(2)
+            num_o2m = hs.size(2) - num_o2o
+            obj_queries = hs if self.o2m_mask_predict else hs[:, :, :num_o2o]
+            seg_head_outputs = activation_ckpt_wrapper(self.segmentation_head)(
+                backbone_feats=backbone_out["backbone_fpn"],
+                obj_queries=obj_queries,
+                image_ids=img_ids,
+                encoder_hidden_states=encoder_hidden_states,
+                act_ckpt_enable=self.training and self.use_act_checkpoint_seg_head,
+                prompt=prompt,
+                prompt_mask=prompt_mask,
+            )
+            aux_masks = False  # self.aux_loss and self.segmentation_head.aux_masks
+            for k, v in seg_head_outputs.items():
+                if k in self.segmentation_head.instance_keys:
+                    _update_out(out, k, v[:, :num_o2o], auxiliary=aux_masks)
+                    if (
+                        self.o2m_mask_predict and num_o2m > 0
+                    ):  # handle o2m mask prediction
+                        _update_out(
+                            out, f"{k}_o2m", v[:, num_o2o:], auxiliary=aux_masks
+                        )
+                else:
+                    out[k] = v
+        else:
+            backbone_out.pop("backbone_fpn", None)
+
+    def _get_best_mask(self, out):
+        prev_mask_idx = out["pred_logits"].argmax(dim=1).squeeze(1)
+        batch_idx = torch.arange(
+            out["pred_logits"].shape[0], device=prev_mask_idx.device
+        )
+        prev_mask_pred = out["pred_masks"][batch_idx, prev_mask_idx][:, None]
+        # Downsample mask to match image resolution.
+        prev_mask_pred = self.geometry_encoder.mask_encoder.mask_downsampler(
+            prev_mask_pred
+        )
+        prev_mask_pred = prev_mask_pred.flatten(-2).permute(2, 0, 1)
+
+        return prev_mask_pred
+
+    def forward_grounding(
+        self,
+        backbone_out,
+        find_input,
+        find_target,
+        geometric_prompt: Prompt,
+    ):
+        with torch.profiler.record_function("SAM3Image._encode_prompt"):
+            prompt, prompt_mask, backbone_out = self._encode_prompt(
+                backbone_out, find_input, geometric_prompt
+            )
+        # Run the encoder
+        with torch.profiler.record_function("SAM3Image._run_encoder"):
+            backbone_out, encoder_out, _ = self._run_encoder(
+                backbone_out, find_input, prompt, prompt_mask
+            )
+        out = {
+            "encoder_hidden_states": encoder_out["encoder_hidden_states"],
+            "prev_encoder_out": {
+                "encoder_out": encoder_out,
+                "backbone_out": backbone_out,
+            },
+        }
+
+        # Run the decoder
+        with torch.profiler.record_function("SAM3Image._run_decoder"):
+            out, hs = self._run_decoder(
+                memory=out["encoder_hidden_states"],
+                pos_embed=encoder_out["pos_embed"],
+                src_mask=encoder_out["padding_mask"],
+                out=out,
+                prompt=prompt,
+                prompt_mask=prompt_mask,
+                encoder_out=encoder_out,
+            )
+
+        # Run segmentation heads
+        with torch.profiler.record_function("SAM3Image._run_segmentation_heads"):
+            self._run_segmentation_heads(
+                out=out,
+                backbone_out=backbone_out,
+                img_ids=find_input.img_ids,
+                vis_feat_sizes=encoder_out["vis_feat_sizes"],
+                encoder_hidden_states=out["encoder_hidden_states"],
+                prompt=prompt,
+                prompt_mask=prompt_mask,
+                hs=hs,
+            )
+
+        if self.training or self.num_interactive_steps_val > 0:
+            self._compute_matching(out, self.back_convert(find_target))
+        return out
+
+    def _postprocess_out(self, out: Dict, multimask_output: bool = False):
+        # For multimask output, during eval we return the single best mask with the dict keys expected by the evaluators, but also return the multimasks output with new keys.
+        num_mask_boxes = out["pred_boxes"].size(1)
+        if not self.training and multimask_output and num_mask_boxes > 1:
+            out["multi_pred_logits"] = out["pred_logits"]
+            if "pred_masks" in out:
+                out["multi_pred_masks"] = out["pred_masks"]
+            out["multi_pred_boxes"] = out["pred_boxes"]
+            out["multi_pred_boxes_xyxy"] = out["pred_boxes_xyxy"]
+
+            best_mask_idx = out["pred_logits"].argmax(1).squeeze(1)
+            batch_idx = torch.arange(len(best_mask_idx), device=best_mask_idx.device)
+
+            out["pred_logits"] = out["pred_logits"][batch_idx, best_mask_idx].unsqueeze(
+                1
+            )
+            if "pred_masks" in out:
+                out["pred_masks"] = out["pred_masks"][
+                    batch_idx, best_mask_idx
+                ].unsqueeze(1)
+            out["pred_boxes"] = out["pred_boxes"][batch_idx, best_mask_idx].unsqueeze(1)
+            out["pred_boxes_xyxy"] = out["pred_boxes_xyxy"][
+                batch_idx, best_mask_idx
+            ].unsqueeze(1)
+
+        return out
+
+    def _get_dummy_prompt(self, num_prompts=1):
+        device = self.device
+        geometric_prompt = Prompt(
+            box_embeddings=torch.zeros(0, num_prompts, 4, device=device),
+            box_mask=torch.zeros(num_prompts, 0, device=device, dtype=torch.bool),
+        )
+        return geometric_prompt
+
+    def forward(self, input: BatchedDatapoint):
+        device = self.device
+        backbone_out = {"img_batch_all_stages": input.img_batch}
+        backbone_out.update(self.backbone.forward_image(input.img_batch))
+        num_frames = len(input.find_inputs)
+        assert num_frames == 1
+
+        text_outputs = self.backbone.forward_text(input.find_text_batch, device=device)
+        backbone_out.update(text_outputs)
+
+        previous_stages_out = SAM3Output(
+            iter_mode=SAM3Output.IterMode.LAST_STEP_PER_STAGE
+        )
+
+        find_input = input.find_inputs[0]
+        find_target = input.find_targets[0]
+
+        if find_input.input_points is not None and find_input.input_points.numel() > 0:
+            print("Warning: Point prompts are ignored in PCS.")
+
+        num_interactive_steps = 0 if self.training else self.num_interactive_steps_val
+        geometric_prompt = Prompt(
+            box_embeddings=find_input.input_boxes,
+            box_mask=find_input.input_boxes_mask,
+            box_labels=find_input.input_boxes_label,
+        )
+
+        # Init vars that are shared across the loop.
+        stage_outs = []
+        for cur_step in range(num_interactive_steps + 1):
+            if cur_step > 0:
+                # We sample interactive geometric prompts (boxes, points)
+                geometric_prompt, _ = self.interactive_prompt_sampler.sample(
+                    geo_prompt=geometric_prompt,
+                    find_target=find_target,
+                    previous_out=stage_outs[-1],
+                )
+            out = self.forward_grounding(
+                backbone_out=backbone_out,
+                find_input=find_input,
+                find_target=find_target,
+                geometric_prompt=geometric_prompt.clone(),
+            )
+            stage_outs.append(out)
+
+        previous_stages_out.append(stage_outs)
+        return previous_stages_out
+
+    def _compute_matching(self, out, targets):
+        out["indices"] = self.matcher(out, targets)
+        for aux_out in out.get("aux_outputs", []):
+            aux_out["indices"] = self.matcher(aux_out, targets)
+
+    def back_convert(self, targets):
+        batched_targets = {
+            "boxes": targets.boxes.view(-1, 4),
+            "boxes_xyxy": box_cxcywh_to_xyxy(targets.boxes.view(-1, 4)),
+            "boxes_padded": targets.boxes_padded,
+            "positive_map": targets.boxes.new_ones(len(targets.boxes), 1),
+            "num_boxes": targets.num_boxes,
+            "masks": targets.segments,
+            "semantic_masks": targets.semantic_segments,
+            "is_valid_mask": targets.is_valid_segment,
+            "is_exhaustive": targets.is_exhaustive,
+            "object_ids_packed": targets.object_ids,
+            "object_ids_padded": targets.object_ids_padded,
+        }
+        return batched_targets
+
+    def predict_inst(
+        self,
+        inference_state,
+        **kwargs,
+    ) -> Tuple[np.ndarray, np.ndarray, np.ndarray]:
+        orig_h, orig_w = (
+            inference_state["original_height"],
+            inference_state["original_width"],
+        )
+        backbone_out = inference_state["backbone_out"]["sam2_backbone_out"]
+        (
+            _,
+            vision_feats,
+            _,
+            _,
+        ) = self.inst_interactive_predictor.model._prepare_backbone_features(
+            backbone_out
+        )
+        # Add no_mem_embed, which is added to the lowest rest feat. map during training on videos
+        vision_feats[-1] = (
+            vision_feats[-1] + self.inst_interactive_predictor.model.no_mem_embed
+        )
+        feats = [
+            feat.permute(1, 2, 0).view(1, -1, *feat_size)
+            for feat, feat_size in zip(
+                vision_feats[::-1], self.inst_interactive_predictor._bb_feat_sizes[::-1]
+            )
+        ][::-1]
+        self.inst_interactive_predictor._features = {
+            "image_embed": feats[-1],
+            "high_res_feats": feats[:-1],
+        }
+        self.inst_interactive_predictor._is_image_set = True
+        self.inst_interactive_predictor._orig_hw = [(orig_h, orig_w)]
+        res = self.inst_interactive_predictor.predict(**kwargs)
+        self.inst_interactive_predictor._features = None
+        self.inst_interactive_predictor._is_image_set = False
+        return res
+
+    def predict_inst_batch(
+        self,
+        inference_state,
+        *args,
+        **kwargs,
+    ) -> Tuple[List[np.ndarray], List[np.ndarray], List[np.ndarray]]:
+        backbone_out = inference_state["backbone_out"]["sam2_backbone_out"]
+        (
+            _,
+            vision_feats,
+            _,
+            _,
+        ) = self.inst_interactive_predictor.model._prepare_backbone_features(
+            backbone_out
+        )
+        # Add no_mem_embed, which is added to the lowest res feat. map during training on videos
+        vision_feats[-1] = (
+            vision_feats[-1] + self.inst_interactive_predictor.model.no_mem_embed
+        )
+        batch_size = vision_feats[-1].shape[1]
+        orig_heights, orig_widths = (
+            inference_state["original_heights"],
+            inference_state["original_widths"],
+        )
+        assert (
+            batch_size == len(orig_heights) == len(orig_widths)
+        ), f"Batch size mismatch in predict_inst_batch. Got {batch_size}, {len(orig_heights)}, {len(orig_widths)}"
+        feats = [
+            feat.permute(1, 2, 0).view(batch_size, -1, *feat_size)
+            for feat, feat_size in zip(
+                vision_feats[::-1], self.inst_interactive_predictor._bb_feat_sizes[::-1]
+            )
+        ][::-1]
+        self.inst_interactive_predictor._features = {
+            "image_embed": feats[-1],
+            "high_res_feats": feats[:-1],
+        }
+        self.inst_interactive_predictor._is_image_set = True
+        self.inst_interactive_predictor._is_batch = True
+        self.inst_interactive_predictor._orig_hw = [
+            (orig_h, orig_w) for orig_h, orig_w in zip(orig_heights, orig_widths)
+        ]
+        res = self.inst_interactive_predictor.predict_batch(*args, **kwargs)
+        self.inst_interactive_predictor._features = None
+        self.inst_interactive_predictor._is_image_set = False
+        self.inst_interactive_predictor._is_batch = False
+        return res
+
+
+class Sam3ImageOnVideoMultiGPU(Sam3Image):
+    def __init__(
+        self, *args, async_all_gather=True, gather_backbone_out=None, **kwargs
+    ):
+        super().__init__(*args, **kwargs)
+        self.rank = int(os.getenv("RANK", "0"))
+        self.world_size = int(os.getenv("WORLD_SIZE", "1"))
+        self.async_all_gather = async_all_gather
+
+        # if gather_backbone is not set, default to gathering only for `SAM3VLBackbone`
+        if gather_backbone_out is None:
+            gather_backbone_out = isinstance(self.backbone, SAM3VLBackbone)
+        self.gather_backbone_out = gather_backbone_out
+
+    def forward_video_grounding_multigpu(
+        self,
+        backbone_out,
+        find_inputs,
+        geometric_prompt: Prompt,
+        frame_idx,
+        num_frames,
+        # `multigpu_buffer` is a dict to cache detector's outputs in a chunk between different calls
+        multigpu_buffer,
+        track_in_reverse=False,
+        # whether to also return the SAM2 backbone features
+        return_sam2_backbone_feats=False,
+        # whether to perform NMS and suppress the scores of those detections removed by NMS
+        run_nms=False,
+        nms_prob_thresh=None,
+        nms_iou_thresh=None,
+        **kwargs,
+    ):
+        """
+        Compute the detector's detection outputs in a distributed manner, where all GPUs process
+        a chunk of frames (equal to the number of GPUs) at once and store them in cache.
+        """
+        # Step 1: fetch the detector outputs in the current chunk from buffer
+        frame_idx_curr_b = frame_idx - frame_idx % self.world_size
+        frame_idx_curr_e = min(frame_idx_curr_b + self.world_size, num_frames)
+        # in case the current frame's detection results are not in the buffer yet, build the current chunk
+        # (this should only happen on the first chunk, since we are also building the next chunk below)
+        if frame_idx not in multigpu_buffer:
+            with torch.profiler.record_function("build_multigpu_buffer_next_chunk1"):
+                self._build_multigpu_buffer_next_chunk(
+                    backbone_out=backbone_out,
+                    find_inputs=find_inputs,
+                    geometric_prompt=geometric_prompt,
+                    frame_idx_begin=frame_idx_curr_b,
+                    frame_idx_end=frame_idx_curr_e,
+                    num_frames=num_frames,
+                    multigpu_buffer=multigpu_buffer,
+                    run_nms=run_nms,
+                    nms_prob_thresh=nms_prob_thresh,
+                    nms_iou_thresh=nms_iou_thresh,
+                )
+
+        # read out the current frame's results from `multigpu_buffer`
+        out = {}
+        for k, (v, handle) in multigpu_buffer[frame_idx].items():
+            if k.startswith("sam2_backbone_") and not return_sam2_backbone_feats:
+                continue
+            if handle is not None:
+                handle.wait()  # wait for async all-gather to finish
+            out[k] = v
+
+        # Step 2: remove detection outputs of the previous chunk from cache to save GPU memory
+        if not track_in_reverse and frame_idx_curr_b - self.world_size >= 0:
+            frame_idx_prev_e = frame_idx_curr_b
+            frame_idx_prev_b = frame_idx_curr_b - self.world_size
+        elif track_in_reverse and frame_idx_curr_e < num_frames:
+            frame_idx_prev_b = frame_idx_curr_e
+            frame_idx_prev_e = min(frame_idx_prev_b + self.world_size, num_frames)
+        else:
+            frame_idx_prev_b = frame_idx_prev_e = None
+        if frame_idx_prev_b is not None:
+            for frame_idx_rm in range(frame_idx_prev_b, frame_idx_prev_e):
+                multigpu_buffer.pop(frame_idx_rm, None)
+
+        # Step 3: compute and cache detection outputs of the next chunk ahead of time
+        # (so that we can overlap computation with all-gather transfer)
+        if not track_in_reverse and frame_idx_curr_e < num_frames:
+            frame_idx_next_b = frame_idx_curr_e
+            frame_idx_next_e = min(frame_idx_next_b + self.world_size, num_frames)
+        elif track_in_reverse and frame_idx_curr_b - self.world_size >= 0:
+            frame_idx_next_e = frame_idx_curr_b
+            frame_idx_next_b = frame_idx_curr_b - self.world_size
+        else:
+            frame_idx_next_b = frame_idx_next_e = None
+        if frame_idx_next_b is not None and frame_idx_next_b not in multigpu_buffer:
+            with torch.profiler.record_function("build_multigpu_buffer_next_chunk2"):
+                self._build_multigpu_buffer_next_chunk(
+                    backbone_out=backbone_out,
+                    find_inputs=find_inputs,
+                    geometric_prompt=geometric_prompt,
+                    frame_idx_begin=frame_idx_next_b,
+                    frame_idx_end=frame_idx_next_e,
+                    num_frames=num_frames,
+                    multigpu_buffer=multigpu_buffer,
+                    run_nms=run_nms,
+                    nms_prob_thresh=nms_prob_thresh,
+                    nms_iou_thresh=nms_iou_thresh,
+                )
+
+        return out, backbone_out
+
+    def _build_multigpu_buffer_next_chunk(
+        self,
+        backbone_out,
+        find_inputs,
+        geometric_prompt: Prompt,
+        frame_idx_begin,
+        frame_idx_end,
+        num_frames,
+        multigpu_buffer,
+        run_nms=False,
+        nms_prob_thresh=None,
+        nms_iou_thresh=None,
+    ):
+        """Compute detection outputs on a chunk of frames and store their results in multigpu_buffer."""
+        # each GPU computes detections on one frame in the chunk (in a round-robin manner)
+        frame_idx_local_gpu = min(frame_idx_begin + self.rank, frame_idx_end - 1)
+        # `forward_grounding` (from base class `Sam3ImageOnVideo`) runs the detector on a single frame
+        with torch.profiler.record_function("forward_grounding"):
+            out_local = self.forward_grounding(
+                backbone_out=backbone_out,
+                find_input=find_inputs[frame_idx_local_gpu],
+                find_target=None,
+                geometric_prompt=geometric_prompt,
+            )
+        if run_nms:
+            with torch.profiler.record_function("nms_masks"):
+                # run NMS as a post-processing step on top of the detection outputs
+                assert nms_prob_thresh is not None and nms_iou_thresh is not None
+                pred_probs = out_local["pred_logits"].squeeze(-1).sigmoid()
+                pred_masks = out_local["pred_masks"]
+                # loop over text prompts (not an overhead for demo where there's only 1 prompt)
+                for prompt_idx in range(pred_probs.size(0)):
+                    keep = nms_masks(
+                        pred_probs=pred_probs[prompt_idx],
+                        pred_masks=pred_masks[prompt_idx],
+                        prob_threshold=nms_prob_thresh,
+                        iou_threshold=nms_iou_thresh,
+                    )
+                    # set a very low threshold for those detections removed by NMS
+                    out_local["pred_logits"][prompt_idx, :, 0] -= 1e4 * (~keep).float()
+
+        if self.gather_backbone_out:
+            # gather the SAM 2 backbone features across GPUs
+            feats = out_local["prev_encoder_out"]["backbone_out"]["sam2_backbone_out"]
+            assert len(feats["backbone_fpn"]) == 3  # SAM2 backbone always have 3 levels
+            # cast the SAM2 backbone features to bfloat16 for all-gather (this is usually
+            # a no-op, SAM2 backbone features are likely already in bfloat16 due to AMP)
+            backbone_fpn_bf16 = [x.to(torch.bfloat16) for x in feats["backbone_fpn"]]
+            fpn0, fpn_handle0 = self._gather_tensor(backbone_fpn_bf16[0])
+            fpn1, fpn_handle1 = self._gather_tensor(backbone_fpn_bf16[1])
+            fpn2, fpn_handle2 = self._gather_tensor(backbone_fpn_bf16[2])
+            # vision_pos_enc is the same on all frames, so no need to all-gather them
+            vision_pos_enc = feats["vision_pos_enc"]
+
+        # trim the detector output to only include the necessary keys
+        out_local = {
+            "pred_logits": out_local["pred_logits"],
+            "pred_boxes": out_local["pred_boxes"],
+            "pred_boxes_xyxy": out_local["pred_boxes_xyxy"],
+            "pred_masks": out_local["pred_masks"],
+        }
+
+        # gather the results: after this step, each GPU will receive detector outputs on
+        # all frames in the chunk and store them in `multigpu_buffer`
+        out_gathered = {k: self._gather_tensor(v) for k, v in out_local.items()}
+        for rank in range(self.world_size):
+            frame_idx_to_save = frame_idx_begin + rank
+            if frame_idx_to_save >= num_frames:
+                continue
+            frame_buffer = {
+                k: (v[rank], handle) for k, (v, handle) in out_gathered.items()
+            }
+            if self.gather_backbone_out:
+                # also add gathered SAM 2 backbone features to frame_buffer
+                frame_buffer["tracker_backbone_fpn_0"] = (fpn0[rank], fpn_handle0)
+                frame_buffer["tracker_backbone_fpn_1"] = (fpn1[rank], fpn_handle1)
+                frame_buffer["tracker_backbone_fpn_2"] = (fpn2[rank], fpn_handle2)
+                frame_buffer["tracker_backbone_pos_enc"] = (vision_pos_enc, None)
+
+            multigpu_buffer[frame_idx_to_save] = frame_buffer
+
+    def _gather_tensor(self, x):
+        if self.world_size == 1:
+            return [x], None
+
+        async_op = self.async_all_gather
+        # here `.contiguous()` is required -- otherwise NCCL all_gather
+        # sometimes gives wrong results
+        x = x.contiguous()  # ensure contiguous memory for NCCL
+        output_list = [torch.empty_like(x) for _ in range(self.world_size)]
+        handle = torch.distributed.all_gather(output_list, x, async_op=async_op)
+        return output_list, handle

detect_tools/sam3/sam3/model/sam3_image_processor.py

@@ -0,0 +1,222 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates. All Rights Reserved
+from typing import Dict, List
+
+import numpy as np
+import PIL
+import torch
+
+from sam3.model import box_ops
+
+from sam3.model.data_misc import FindStage, interpolate
+from torchvision.transforms import v2
+
+
+class Sam3Processor:
+    """ """
+
+    def __init__(self, model, resolution=1008, device="cuda", confidence_threshold=0.5):
+        self.model = model
+        self.resolution = resolution
+        self.device = device
+        self.transform = v2.Compose(
+            [
+                v2.ToDtype(torch.uint8, scale=True),
+                v2.Resize(size=(resolution, resolution)),
+                v2.ToDtype(torch.float32, scale=True),
+                v2.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]),
+            ]
+        )
+        self.confidence_threshold = confidence_threshold
+
+        self.find_stage = FindStage(
+            img_ids=torch.tensor([0], device=device, dtype=torch.long),
+            text_ids=torch.tensor([0], device=device, dtype=torch.long),
+            input_boxes=None,
+            input_boxes_mask=None,
+            input_boxes_label=None,
+            input_points=None,
+            input_points_mask=None,
+        )
+
+    @torch.inference_mode()
+    def set_image(self, image, state=None):
+        """Sets the image on which we want to do predictions."""
+        if state is None:
+            state = {}
+
+        if isinstance(image, PIL.Image.Image):
+            width, height = image.size
+        elif isinstance(image, (torch.Tensor, np.ndarray)):
+            height, width = image.shape[-2:]
+        else:
+            raise ValueError("Image must be a PIL image or a tensor")
+
+        image = v2.functional.to_image(image).to(self.device)
+        image = self.transform(image).unsqueeze(0)
+
+        state["original_height"] = height
+        state["original_width"] = width
+        state["backbone_out"] = self.model.backbone.forward_image(image)
+        inst_interactivity_en = self.model.inst_interactive_predictor is not None
+        if inst_interactivity_en and "sam2_backbone_out" in state["backbone_out"]:
+            sam2_backbone_out = state["backbone_out"]["sam2_backbone_out"]
+            sam2_backbone_out["backbone_fpn"][0] = (
+                self.model.inst_interactive_predictor.model.sam_mask_decoder.conv_s0(
+                    sam2_backbone_out["backbone_fpn"][0]
+                )
+            )
+            sam2_backbone_out["backbone_fpn"][1] = (
+                self.model.inst_interactive_predictor.model.sam_mask_decoder.conv_s1(
+                    sam2_backbone_out["backbone_fpn"][1]
+                )
+            )
+        return state
+
+    @torch.inference_mode()
+    def set_image_batch(self, images: List[np.ndarray], state=None):
+        """Sets the image batch on which we want to do predictions."""
+        if state is None:
+            state = {}
+
+        if not isinstance(images, list):
+            raise ValueError("Images must be a list of PIL images or tensors")
+        assert len(images) > 0, "Images list must not be empty"
+        assert isinstance(
+            images[0], PIL.Image.Image
+        ), "Images must be a list of PIL images"
+
+        state["original_heights"] = [image.height for image in images]
+        state["original_widths"] = [image.width for image in images]
+
+        images = [
+            self.transform(v2.functional.to_image(image).to(self.device))
+            for image in images
+        ]
+        images = torch.stack(images, dim=0)
+        state["backbone_out"] = self.model.backbone.forward_image(images)
+        inst_interactivity_en = self.model.inst_interactive_predictor is not None
+        if inst_interactivity_en and "sam2_backbone_out" in state["backbone_out"]:
+            sam2_backbone_out = state["backbone_out"]["sam2_backbone_out"]
+            sam2_backbone_out["backbone_fpn"][0] = (
+                self.model.inst_interactive_predictor.model.sam_mask_decoder.conv_s0(
+                    sam2_backbone_out["backbone_fpn"][0]
+                )
+            )
+            sam2_backbone_out["backbone_fpn"][1] = (
+                self.model.inst_interactive_predictor.model.sam_mask_decoder.conv_s1(
+                    sam2_backbone_out["backbone_fpn"][1]
+                )
+            )
+        return state
+
+    @torch.inference_mode()
+    def set_text_prompt(self, prompt: str, state: Dict):
+        """Sets the text prompt and run the inference"""
+
+        if "backbone_out" not in state:
+            raise ValueError("You must call set_image before set_text_prompt")
+
+        text_outputs = self.model.backbone.forward_text([prompt], device=self.device)
+        # will erase the previous text prompt if any
+        state["backbone_out"].update(text_outputs)
+        if "geometric_prompt" not in state:
+            state["geometric_prompt"] = self.model._get_dummy_prompt()
+
+        return self._forward_grounding(state)
+
+    @torch.inference_mode()
+    def add_geometric_prompt(self, box: List, label: bool, state: Dict):
+        """Adds a box prompt and run the inference.
+        The image needs to be set, but not necessarily the text prompt.
+        The box is assumed to be in [center_x, center_y, width, height] format and normalized in [0, 1] range.
+        The label is True for a positive box, False for a negative box.
+        """
+        if "backbone_out" not in state:
+            raise ValueError("You must call set_image before set_text_prompt")
+
+        if "language_features" not in state["backbone_out"]:
+            # Looks like we don't have a text prompt yet. This is allowed, but we need to set the text prompt to "visual" for the model to rely only on the geometric prompt
+            dummy_text_outputs = self.model.backbone.forward_text(
+                ["visual"], device=self.device
+            )
+            state["backbone_out"].update(dummy_text_outputs)
+
+        if "geometric_prompt" not in state:
+            state["geometric_prompt"] = self.model._get_dummy_prompt()
+
+        # adding a batch and sequence dimension
+        boxes = torch.tensor(box, device=self.device, dtype=torch.float32).view(1, 1, 4)
+        labels = torch.tensor([label], device=self.device, dtype=torch.bool).view(1, 1)
+        state["geometric_prompt"].append_boxes(boxes, labels)
+
+        return self._forward_grounding(state)
+
+    def reset_all_prompts(self, state: Dict):
+        """Removes all the prompts and results"""
+        if "backbone_out" in state:
+            backbone_keys_to_del = [
+                "language_features",
+                "language_mask",
+                "language_embeds",
+            ]
+            for key in backbone_keys_to_del:
+                if key in state["backbone_out"]:
+                    del state["backbone_out"][key]
+
+        keys_to_del = ["geometric_prompt", "boxes", "masks", "masks_logits", "scores"]
+        for key in keys_to_del:
+            if key in state:
+                del state[key]
+
+    @torch.inference_mode()
+    def set_confidence_threshold(self, threshold: float, state=None):
+        """Sets the confidence threshold for the masks"""
+        self.confidence_threshold = threshold
+        if state is not None and "boxes" in state:
+            # we need to filter the boxes again
+            # In principle we could do this more efficiently since we would only need
+            # to rerun the heads. But this is simpler and not too inefficient
+            return self._forward_grounding(state)
+        return state
+
+    @torch.inference_mode()
+    def _forward_grounding(self, state: Dict):
+        outputs = self.model.forward_grounding(
+            backbone_out=state["backbone_out"],
+            find_input=self.find_stage,
+            geometric_prompt=state["geometric_prompt"],
+            find_target=None,
+        )
+
+        out_bbox = outputs["pred_boxes"]
+        out_logits = outputs["pred_logits"]
+        out_masks = outputs["pred_masks"]
+        out_probs = out_logits.sigmoid()
+        presence_score = outputs["presence_logit_dec"].sigmoid().unsqueeze(1)
+        out_probs = (out_probs * presence_score).squeeze(-1)
+
+        keep = out_probs > self.confidence_threshold
+        out_probs = out_probs[keep]
+        out_masks = out_masks[keep]
+        out_bbox = out_bbox[keep]
+
+        # convert to [x0, y0, x1, y1] format
+        boxes = box_ops.box_cxcywh_to_xyxy(out_bbox)
+
+        img_h = state["original_height"]
+        img_w = state["original_width"]
+        scale_fct = torch.tensor([img_w, img_h, img_w, img_h]).to(self.device)
+        boxes = boxes * scale_fct[None, :]
+
+        out_masks = interpolate(
+            out_masks.unsqueeze(1),
+            (img_h, img_w),
+            mode="bilinear",
+            align_corners=False,
+        ).sigmoid()
+
+        state["masks_logits"] = out_masks
+        state["masks"] = out_masks > 0.5
+        state["boxes"] = boxes
+        state["scores"] = out_probs
+        return state

detect_tools/sam3/sam3/model/sam3_tracker_base.py

@@ -0,0 +1,1188 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates. All Rights Reserved
+
+import logging
+
+import torch
+import torch.nn.functional as F
+
+from sam3.model.memory import SimpleMaskEncoder
+
+from sam3.model.sam3_tracker_utils import get_1d_sine_pe, select_closest_cond_frames
+
+from sam3.sam.mask_decoder import MaskDecoder, MLP
+from sam3.sam.prompt_encoder import PromptEncoder
+from sam3.sam.transformer import TwoWayTransformer
+from sam3.train.data.collator import BatchedDatapoint
+
+try:
+    from timm.layers import trunc_normal_
+except ModuleNotFoundError:
+    # compatibility for older timm versions
+    from timm.models.layers import trunc_normal_
+
+# a large negative value as a placeholder score for missing objects
+NO_OBJ_SCORE = -1024.0
+
+
+class Sam3TrackerBase(torch.nn.Module):
+    def __init__(
+        self,
+        backbone,
+        transformer,
+        maskmem_backbone,
+        num_maskmem=7,  # default 1 input frame + 6 previous frames as in CAE
+        image_size=1008,
+        backbone_stride=14,  # stride of the image backbone output
+        # The maximum number of conditioning frames to participate in the memory attention (-1 means no limit; if there are more conditioning frames than this limit,
+        # we only cross-attend to the temporally closest `max_cond_frames_in_attn` conditioning frames in the encoder when tracking each frame). This gives the model
+        # a temporal locality when handling a large number of annotated frames (since closer frames should be more important) and also avoids GPU OOM.
+        max_cond_frames_in_attn=-1,
+        # Whether to always keep the first conditioning frame in case we exceed the maximum number of conditioning frames allowed
+        keep_first_cond_frame=False,
+        # whether to output multiple (3) masks for the first click on initial conditioning frames
+        multimask_output_in_sam=False,
+        # the minimum and maximum number of clicks to use multimask_output_in_sam (only relevant when `multimask_output_in_sam=True`;
+        # default is 1 for both, meaning that only the first click gives multimask output; also note that a box counts as two points)
+        multimask_min_pt_num=1,
+        multimask_max_pt_num=1,
+        # whether to also use multimask output for tracking (not just for the first click on initial conditioning frames; only relevant when `multimask_output_in_sam=True`)
+        multimask_output_for_tracking=False,
+        # whether to forward image features per frame (as it's being tracked) during evaluation, instead of forwarding image features
+        # of all frames at once. This avoids backbone OOM errors on very long videos in evaluation, but could be slightly slower.
+        forward_backbone_per_frame_for_eval=False,
+        # The memory bank's temporal stride during evaluation (i.e. the `r` parameter in XMem and Cutie; XMem and Cutie use r=5).
+        # For r>1, the (self.num_maskmem - 1) non-conditioning memory frames consist of
+        # (self.num_maskmem - 2) nearest frames from every r-th frames, plus the last frame.
+        memory_temporal_stride_for_eval=1,
+        # whether to offload outputs to CPU memory during evaluation, to avoid GPU OOM on very long videos or very large resolutions or too many objects
+        # (it's recommended to use `forward_backbone_per_frame_for_eval=True` first before setting this option to True)
+        offload_output_to_cpu_for_eval=False,
+        # whether to trim the output of past non-conditioning frames (num_maskmem frames before the current frame) during evaluation
+        # (this helps save GPU or CPU memory on very long videos for semi-supervised VOS eval, where only the first frame receives prompts)
+        trim_past_non_cond_mem_for_eval=False,
+        # whether to apply non-overlapping constraints on the object masks in the memory encoder during evaluation (to avoid/alleviate superposing masks)
+        non_overlap_masks_for_mem_enc=False,
+        # the maximum number of object pointers from other frames in encoder cross attention
+        max_obj_ptrs_in_encoder=16,
+        # extra arguments used to construct the SAM mask decoder; if not None, it should be a dict of kwargs to be passed into `MaskDecoder` class.
+        sam_mask_decoder_extra_args=None,
+        # whether to compile all the model compoents
+        compile_all_components=False,
+        # select the frame with object existence
+        use_memory_selection=False,
+        # when using memory selection, the threshold to determine if the frame is good
+        mf_threshold=0.01,
+    ):
+        super().__init__()
+
+        # Part 1: the image backbone
+        self.backbone = backbone
+        self.num_feature_levels = 3
+        self.max_obj_ptrs_in_encoder = max_obj_ptrs_in_encoder
+        # A conv layer to downsample the GT mask prompt to stride 4 (the same stride as
+        # low-res SAM mask logits) and to change its scales from 0~1 to SAM logit scale,
+        # so that it can be fed into the SAM mask decoder to generate a pointer.
+        self.mask_downsample = torch.nn.Conv2d(1, 1, kernel_size=4, stride=4)
+
+        # Part 2: encoder-only transformer to fuse current frame's visual features
+        # with memories from past frames
+        assert transformer.decoder is None, "transformer should be encoder-only"
+        self.transformer = transformer
+        self.hidden_dim = transformer.d_model
+
+        # Part 3: memory encoder for the previous frame's outputs
+        self.maskmem_backbone = maskmem_backbone
+        self.mem_dim = self.hidden_dim
+        if hasattr(self.maskmem_backbone, "out_proj") and hasattr(
+            self.maskmem_backbone.out_proj, "weight"
+        ):
+            # if there is compression of memories along channel dim
+            self.mem_dim = self.maskmem_backbone.out_proj.weight.shape[0]
+        self.num_maskmem = num_maskmem  # Number of memories accessible
+
+        # Temporal encoding of the memories
+        self.maskmem_tpos_enc = torch.nn.Parameter(
+            torch.zeros(num_maskmem, 1, 1, self.mem_dim)
+        )
+        trunc_normal_(self.maskmem_tpos_enc, std=0.02)
+
+        # a single token to indicate no memory embedding from previous frames
+        self.no_mem_embed = torch.nn.Parameter(torch.zeros(1, 1, self.hidden_dim))
+        self.no_mem_pos_enc = torch.nn.Parameter(torch.zeros(1, 1, self.hidden_dim))
+        trunc_normal_(self.no_mem_embed, std=0.02)
+        trunc_normal_(self.no_mem_pos_enc, std=0.02)
+        # Apply sigmoid to the output raw mask logits (to turn them from
+        # range (-inf, +inf) to range (0, 1)) before feeding them into the memory encoder
+        self.sigmoid_scale_for_mem_enc = 20.0
+        self.sigmoid_bias_for_mem_enc = -10.0
+        self.non_overlap_masks_for_mem_enc = non_overlap_masks_for_mem_enc
+        self.memory_temporal_stride_for_eval = memory_temporal_stride_for_eval
+        # On frames with mask input, whether to directly output the input mask without
+        # using a SAM prompt encoder + mask decoder
+        self.multimask_output_in_sam = multimask_output_in_sam
+        self.multimask_min_pt_num = multimask_min_pt_num
+        self.multimask_max_pt_num = multimask_max_pt_num
+        self.multimask_output_for_tracking = multimask_output_for_tracking
+
+        # Part 4: SAM-style prompt encoder (for both mask and point inputs)
+        # and SAM-style mask decoder for the final mask output
+        self.image_size = image_size
+        self.backbone_stride = backbone_stride
+        self.low_res_mask_size = self.image_size // self.backbone_stride * 4
+        # we resize the mask if it doesn't match `self.input_mask_size` (which is always 4x
+        # the low-res mask size, regardless of the actual input image size); this is because
+        # `_use_mask_as_output` always downsamples the input masks by 4x
+        self.input_mask_size = self.low_res_mask_size * 4
+        self.forward_backbone_per_frame_for_eval = forward_backbone_per_frame_for_eval
+        self.offload_output_to_cpu_for_eval = offload_output_to_cpu_for_eval
+        self.trim_past_non_cond_mem_for_eval = trim_past_non_cond_mem_for_eval
+        self.sam_mask_decoder_extra_args = sam_mask_decoder_extra_args
+        self.no_obj_ptr = torch.nn.Parameter(torch.zeros(1, self.hidden_dim))
+        trunc_normal_(self.no_obj_ptr, std=0.02)
+        self.no_obj_embed_spatial = torch.nn.Parameter(torch.zeros(1, self.mem_dim))
+        trunc_normal_(self.no_obj_embed_spatial, std=0.02)
+
+        self._build_sam_heads()
+        self.max_cond_frames_in_attn = max_cond_frames_in_attn
+        self.keep_first_cond_frame = keep_first_cond_frame
+
+        # Use frame filtering according to SAM2Long
+        self.use_memory_selection = use_memory_selection
+        self.mf_threshold = mf_threshold
+
+        # Compile all components of the model
+        self.compile_all_components = compile_all_components
+        if self.compile_all_components:
+            self._compile_all_components()
+
+    @property
+    def device(self):
+        return next(self.parameters()).device
+
+    def _get_tpos_enc(self, rel_pos_list, device, max_abs_pos=None, dummy=False):
+        if dummy:
+            return torch.zeros(len(rel_pos_list), self.mem_dim, device=device)
+
+        t_diff_max = max_abs_pos - 1 if max_abs_pos is not None else 1
+        pos_enc = (
+            torch.tensor(rel_pos_list).pin_memory().to(device=device, non_blocking=True)
+            / t_diff_max
+        )
+        tpos_dim = self.hidden_dim
+        pos_enc = get_1d_sine_pe(pos_enc, dim=tpos_dim)
+        pos_enc = self.obj_ptr_tpos_proj(pos_enc)
+
+        return pos_enc
+
+    def _build_sam_heads(self):
+        """Build SAM-style prompt encoder and mask decoder."""
+        self.sam_prompt_embed_dim = self.hidden_dim
+        self.sam_image_embedding_size = self.image_size // self.backbone_stride
+
+        # build PromptEncoder and MaskDecoder from SAM
+        # (their hyperparameters like `mask_in_chans=16` are from SAM code)
+        self.sam_prompt_encoder = PromptEncoder(
+            embed_dim=self.sam_prompt_embed_dim,
+            image_embedding_size=(
+                self.sam_image_embedding_size,
+                self.sam_image_embedding_size,
+            ),
+            input_image_size=(self.image_size, self.image_size),
+            mask_in_chans=16,
+        )
+        self.sam_mask_decoder = MaskDecoder(
+            num_multimask_outputs=3,
+            transformer=TwoWayTransformer(
+                depth=2,
+                embedding_dim=self.sam_prompt_embed_dim,
+                mlp_dim=2048,
+                num_heads=8,
+            ),
+            transformer_dim=self.sam_prompt_embed_dim,
+            iou_head_depth=3,
+            iou_head_hidden_dim=256,
+            use_high_res_features=True,
+            iou_prediction_use_sigmoid=True,
+            pred_obj_scores=True,
+            pred_obj_scores_mlp=True,
+            use_multimask_token_for_obj_ptr=True,
+            **(self.sam_mask_decoder_extra_args or {}),
+        )
+        # a linear projection on SAM output tokens to turn them into object pointers
+        self.obj_ptr_proj = torch.nn.Linear(self.hidden_dim, self.hidden_dim)
+        self.obj_ptr_proj = MLP(self.hidden_dim, self.hidden_dim, self.hidden_dim, 3)
+        # a linear projection on temporal positional encoding in object pointers to
+        # avoid potential interference with spatial positional encoding
+        self.obj_ptr_tpos_proj = torch.nn.Linear(self.hidden_dim, self.mem_dim)
+
+    def _forward_sam_heads(
+        self,
+        backbone_features,
+        point_inputs=None,
+        mask_inputs=None,
+        high_res_features=None,
+        multimask_output=False,
+        gt_masks=None,
+    ):
+        """
+        Forward SAM prompt encoders and mask heads.
+
+        Inputs:
+        - backbone_features: image features of [B, C, H, W] shape
+        - point_inputs: a dictionary with "point_coords" and "point_labels", where
+          1) "point_coords" has [B, P, 2] shape and float32 dtype and contains the
+             absolute pixel-unit coordinate in (x, y) format of the P input points
+          2) "point_labels" has shape [B, P] and int32 dtype, where 1 means
+             positive clicks, 0 means negative clicks, and -1 means padding
+        - mask_inputs: a mask of [B, 1, H*16, W*16] shape, float or bool, with the
+          same spatial size as the image.
+        - high_res_features: either 1) None or 2) or a list of length 2 containing
+          two feature maps of [B, C, 4*H, 4*W] and [B, C, 2*H, 2*W] shapes respectively,
+          which will be used as high-resolution feature maps for SAM decoder.
+        - multimask_output: if it's True, we output 3 candidate masks and their 3
+          corresponding IoU estimates, and if it's False, we output only 1 mask and
+          its corresponding IoU estimate.
+
+        Outputs:
+        - low_res_multimasks: [B, M, H*4, W*4] shape (where M = 3 if
+          `multimask_output=True` and M = 1 if `multimask_output=False`), the SAM
+          output mask logits (before sigmoid) for the low-resolution masks, with 4x
+          the resolution (1/4 stride) of the input backbone_features.
+        - high_res_multimasks: [B, M, H*16, W*16] shape (where M = 3
+          if `multimask_output=True` and M = 1 if `multimask_output=False`),
+          upsampled from the low-resolution masks, with shape size as the image
+          (stride is 1 pixel).
+        - ious, [B, M] shape, where (where M = 3 if `multimask_output=True` and M = 1
+          if `multimask_output=False`), the estimated IoU of each output mask.
+        - low_res_masks: [B, 1, H*4, W*4] shape, the best mask in `low_res_multimasks`.
+          If `multimask_output=True`, it's the mask with the highest IoU estimate.
+          If `multimask_output=False`, it's the same as `low_res_multimasks`.
+        - high_res_masks: [B, 1, H*16, W*16] shape, the best mask in `high_res_multimasks`.
+          If `multimask_output=True`, it's the mask with the highest IoU estimate.
+          If `multimask_output=False`, it's the same as `high_res_multimasks`.
+        - obj_ptr: [B, C] shape, the object pointer vector for the output mask, extracted
+          based on the output token from the SAM mask decoder.
+        """
+        B = backbone_features.size(0)
+        device = backbone_features.device
+        assert backbone_features.size(1) == self.sam_prompt_embed_dim
+        assert backbone_features.size(2) == self.sam_image_embedding_size
+        assert backbone_features.size(3) == self.sam_image_embedding_size
+
+        # a) Handle point prompts
+        if point_inputs is not None:
+            sam_point_coords = point_inputs["point_coords"]
+            sam_point_labels = point_inputs["point_labels"]
+            assert sam_point_coords.size(0) == B and sam_point_labels.size(0) == B
+        else:
+            # If no points are provide, pad with an empty point (with label -1)
+            sam_point_coords = torch.zeros(B, 1, 2, device=device)
+            sam_point_labels = -torch.ones(B, 1, dtype=torch.int32, device=device)
+
+        # b) Handle mask prompts
+        if mask_inputs is not None:
+            # If mask_inputs is provided, downsize it into low-res mask input if needed
+            # and feed it as a dense mask prompt into the SAM mask encoder
+            assert len(mask_inputs.shape) == 4 and mask_inputs.shape[:2] == (B, 1)
+            if mask_inputs.shape[-2:] != self.sam_prompt_encoder.mask_input_size:
+                sam_mask_prompt = F.interpolate(
+                    mask_inputs.float(),
+                    size=self.sam_prompt_encoder.mask_input_size,
+                    align_corners=False,
+                    mode="bilinear",
+                    antialias=True,  # use antialias for downsampling
+                )
+            else:
+                sam_mask_prompt = mask_inputs
+        else:
+            # Otherwise, simply feed None (and SAM's prompt encoder will add
+            # a learned `no_mask_embed` to indicate no mask input in this case).
+            sam_mask_prompt = None
+
+        sparse_embeddings, dense_embeddings = self.sam_prompt_encoder(
+            points=(sam_point_coords, sam_point_labels),
+            boxes=None,
+            masks=sam_mask_prompt,
+        )
+        # Clone image_pe and the outputs of sam_prompt_encoder
+        # to enable compilation
+        sparse_embeddings = self._maybe_clone(sparse_embeddings)
+        dense_embeddings = self._maybe_clone(dense_embeddings)
+        image_pe = self._maybe_clone(self.sam_prompt_encoder.get_dense_pe())
+        with torch.profiler.record_function("sam_mask_decoder"):
+            (
+                low_res_multimasks,
+                ious,
+                sam_output_tokens,
+                object_score_logits,
+            ) = self.sam_mask_decoder(
+                image_embeddings=backbone_features,
+                image_pe=image_pe,
+                sparse_prompt_embeddings=sparse_embeddings,
+                dense_prompt_embeddings=dense_embeddings,
+                multimask_output=multimask_output,
+                repeat_image=False,  # the image is already batched
+                high_res_features=high_res_features,
+            )
+        # Clone the output of sam_mask_decoder
+        # to enable compilation
+        low_res_multimasks = self._maybe_clone(low_res_multimasks)
+        ious = self._maybe_clone(ious)
+        sam_output_tokens = self._maybe_clone(sam_output_tokens)
+        object_score_logits = self._maybe_clone(object_score_logits)
+
+        if self.training and self.teacher_force_obj_scores_for_mem:
+            # we use gt to detect if there is an object or not to
+            # select no obj ptr and use an empty mask for spatial memory
+            is_obj_appearing = torch.any(gt_masks.float().flatten(1) > 0, dim=1)
+            is_obj_appearing = is_obj_appearing[..., None]
+        else:
+            is_obj_appearing = object_score_logits > 0
+
+        # Mask used for spatial memories is always a *hard* choice between obj and no obj,
+        # consistent with the actual mask prediction
+        low_res_multimasks = torch.where(
+            is_obj_appearing[:, None, None],
+            low_res_multimasks,
+            NO_OBJ_SCORE,
+        )
+
+        # convert masks from possibly bfloat16 (or float16) to float32
+        # (older PyTorch versions before 2.1 don't support `interpolate` on bf16)
+        low_res_multimasks = low_res_multimasks.float()
+        high_res_multimasks = F.interpolate(
+            low_res_multimasks,
+            size=(self.image_size, self.image_size),
+            mode="bilinear",
+            align_corners=False,
+        )
+
+        sam_output_token = sam_output_tokens[:, 0]
+        if multimask_output:
+            # take the best mask prediction (with the highest IoU estimation)
+            best_iou_inds = torch.argmax(ious, dim=-1)
+            batch_inds = torch.arange(B, device=device)
+            low_res_masks = low_res_multimasks[batch_inds, best_iou_inds].unsqueeze(1)
+            high_res_masks = high_res_multimasks[batch_inds, best_iou_inds].unsqueeze(1)
+            if sam_output_tokens.size(1) > 1:
+                sam_output_token = sam_output_tokens[batch_inds, best_iou_inds]
+        else:
+            low_res_masks, high_res_masks = low_res_multimasks, high_res_multimasks
+
+        # Extract object pointer from the SAM output token (with occlusion handling)
+        obj_ptr = self.obj_ptr_proj(sam_output_token)
+        lambda_is_obj_appearing = is_obj_appearing.float()
+
+        obj_ptr = lambda_is_obj_appearing * obj_ptr
+        obj_ptr = obj_ptr + (1 - lambda_is_obj_appearing) * self.no_obj_ptr
+
+        return (
+            low_res_multimasks,
+            high_res_multimasks,
+            ious,
+            low_res_masks,
+            high_res_masks,
+            obj_ptr,
+            object_score_logits,
+        )
+
+    def _use_mask_as_output(self, backbone_features, high_res_features, mask_inputs):
+        """
+        Directly turn binary `mask_inputs` into a output mask logits without using SAM.
+        (same input and output shapes as in _forward_sam_heads above).
+        """
+        # Use -10/+10 as logits for neg/pos pixels (very close to 0/1 in prob after sigmoid).
+        out_scale, out_bias = 20.0, -10.0  # sigmoid(-10.0)=4.5398e-05
+        mask_inputs_float = mask_inputs.float()
+        high_res_masks = mask_inputs_float * out_scale + out_bias
+        low_res_masks = F.interpolate(
+            high_res_masks,
+            size=(
+                high_res_masks.size(-2) // self.backbone_stride * 4,
+                high_res_masks.size(-1) // self.backbone_stride * 4,
+            ),
+            align_corners=False,
+            mode="bilinear",
+            antialias=True,  # use antialias for downsampling
+        )
+        # a dummy IoU prediction of all 1's under mask input
+        ious = mask_inputs.new_ones(mask_inputs.size(0), 1).float()
+        # produce an object pointer using the SAM decoder from the mask input
+        _, _, _, _, _, obj_ptr, _ = self._forward_sam_heads(
+            backbone_features=backbone_features,
+            mask_inputs=self.mask_downsample(mask_inputs_float),
+            high_res_features=high_res_features,
+            gt_masks=mask_inputs,
+        )
+        # In this method, we are treating mask_input as output, e.g. using it directly to create spatial mem;
+        # Below, we follow the same design axiom to use mask_input to decide if obj appears or not instead of relying
+        # on the object_scores from the SAM decoder.
+        is_obj_appearing = torch.any(mask_inputs.flatten(1).float() > 0.0, dim=1)
+        is_obj_appearing = is_obj_appearing[..., None]
+        lambda_is_obj_appearing = is_obj_appearing.float()
+        object_score_logits = out_scale * lambda_is_obj_appearing + out_bias
+        obj_ptr = lambda_is_obj_appearing * obj_ptr
+        obj_ptr = obj_ptr + (1 - lambda_is_obj_appearing) * self.no_obj_ptr
+
+        return (
+            low_res_masks,
+            high_res_masks,
+            ious,
+            low_res_masks,
+            high_res_masks,
+            obj_ptr,
+            object_score_logits,
+        )
+
+    def forward(self, input: BatchedDatapoint, is_inference=False):
+        raise NotImplementedError(
+            "Please use the corresponding methods in SAM3VideoPredictor for inference."
+            "See examples/sam3_dense_video_tracking.ipynb for an inference example."
+        )
+
+    def forward_image(self, img_batch):
+        """Get the image feature on the input batch."""
+        # This line is the only change from the parent class
+        # to use the SAM3 backbone instead of the SAM2 backbone.
+        backbone_out = self.backbone.forward_image(img_batch)["sam2_backbone_out"]
+        # precompute projected level 0 and level 1 features in SAM decoder
+        # to avoid running it again on every SAM click
+        backbone_out["backbone_fpn"][0] = self.sam_mask_decoder.conv_s0(
+            backbone_out["backbone_fpn"][0]
+        )
+        backbone_out["backbone_fpn"][1] = self.sam_mask_decoder.conv_s1(
+            backbone_out["backbone_fpn"][1]
+        )
+        # Clone to help torch.compile
+        for i in range(len(backbone_out["backbone_fpn"])):
+            backbone_out["backbone_fpn"][i] = self._maybe_clone(
+                backbone_out["backbone_fpn"][i]
+            )
+            backbone_out["vision_pos_enc"][i] = self._maybe_clone(
+                backbone_out["vision_pos_enc"][i]
+            )
+        return backbone_out
+
+    def _prepare_backbone_features(self, backbone_out):
+        """Prepare and flatten visual features (same as in MDETR_API model)."""
+        backbone_out = backbone_out.copy()
+        assert len(backbone_out["backbone_fpn"]) == len(backbone_out["vision_pos_enc"])
+        assert len(backbone_out["backbone_fpn"]) >= self.num_feature_levels
+
+        feature_maps = backbone_out["backbone_fpn"][-self.num_feature_levels :]
+        vision_pos_embeds = backbone_out["vision_pos_enc"][-self.num_feature_levels :]
+
+        feat_sizes = [(x.shape[-2], x.shape[-1]) for x in vision_pos_embeds]
+        # flatten NxCxHxW to HWxNxC
+        vision_feats = [x.flatten(2).permute(2, 0, 1) for x in feature_maps]
+        vision_pos_embeds = [x.flatten(2).permute(2, 0, 1) for x in vision_pos_embeds]
+
+        return backbone_out, vision_feats, vision_pos_embeds, feat_sizes
+
+    def _prepare_backbone_features_per_frame(self, img_batch, img_ids):
+        """Compute the image backbone features on the fly for the given img_ids."""
+        # Only forward backbone on unique image ids to avoid repeatitive computation
+        # (if `img_ids` has only one element, it's already unique so we skip this step).
+        if img_ids.numel() > 1:
+            unique_img_ids, inv_ids = torch.unique(img_ids, return_inverse=True)
+        else:
+            unique_img_ids, inv_ids = img_ids, None
+
+        # Compute the image features on those unique image ids
+        image = img_batch[unique_img_ids]
+        backbone_out = self.forward_image(image)
+        (
+            _,
+            vision_feats,
+            vision_pos_embeds,
+            feat_sizes,
+        ) = self._prepare_backbone_features(backbone_out)
+        # Inverse-map image features for `unique_img_ids` to the final image features
+        # for the original input `img_ids`.
+        if inv_ids is not None:
+            image = image[inv_ids]
+            vision_feats = [x[:, inv_ids] for x in vision_feats]
+            vision_pos_embeds = [x[:, inv_ids] for x in vision_pos_embeds]
+
+        return image, vision_feats, vision_pos_embeds, feat_sizes
+
+    def cal_mem_score(self, object_score_logits, iou_score):
+        object_score_norm = torch.where(
+            object_score_logits > 0,
+            object_score_logits.sigmoid() * 2 - 1,  ## rescale to [0, 1]
+            torch.zeros_like(object_score_logits),
+        )
+        score_per_frame = (object_score_norm * iou_score).mean()
+        return score_per_frame
+
+    def frame_filter(self, output_dict, track_in_reverse, frame_idx, num_frames, r):
+        if (frame_idx == 0 and not track_in_reverse) or (
+            frame_idx == num_frames - 1 and track_in_reverse
+        ):
+            return []
+
+        max_num = min(
+            num_frames, self.max_obj_ptrs_in_encoder
+        )  ## maximum number of pointer memory frames to consider
+
+        if not track_in_reverse:
+            start = frame_idx - 1
+            end = 0
+            step = -r
+            must_include = frame_idx - 1
+        else:
+            start = frame_idx + 1
+            end = num_frames
+            step = r
+            must_include = frame_idx + 1
+
+        valid_indices = []
+        for i in range(start, end, step):
+            if (
+                i not in output_dict["non_cond_frame_outputs"]
+                or "eff_iou_score" not in output_dict["non_cond_frame_outputs"][i]
+            ):
+                continue
+
+            score_per_frame = output_dict["non_cond_frame_outputs"][i]["eff_iou_score"]
+
+            if score_per_frame > self.mf_threshold:  # threshold
+                valid_indices.insert(0, i)
+
+            if len(valid_indices) >= max_num - 1:
+                break
+
+        if must_include not in valid_indices:
+            valid_indices.append(must_include)
+
+        return valid_indices
+
+    def _prepare_memory_conditioned_features(
+        self,
+        frame_idx,
+        is_init_cond_frame,
+        current_vision_feats,
+        current_vision_pos_embeds,
+        feat_sizes,
+        output_dict,
+        num_frames,
+        track_in_reverse=False,  # tracking in reverse time order (for demo usage)
+        use_prev_mem_frame=True,
+    ):
+        """Fuse the current frame's visual feature map with previous memory."""
+        B = current_vision_feats[-1].size(1)  # batch size on this frame
+        C = self.hidden_dim
+        H, W = feat_sizes[-1]  # top-level (lowest-resolution) feature size
+        device = current_vision_feats[-1].device
+        # The case of `self.num_maskmem == 0` below is primarily used for reproducing SAM on images.
+        # In this case, we skip the fusion with any memory.
+        if self.num_maskmem == 0:  # Disable memory and skip fusion
+            pix_feat = current_vision_feats[-1].permute(1, 2, 0).view(B, C, H, W)
+            return pix_feat
+
+        num_obj_ptr_tokens = 0
+        tpos_sign_mul = -1 if track_in_reverse else 1
+        # Step 1: condition the visual features of the current frame on previous memories
+        if not is_init_cond_frame and use_prev_mem_frame:
+            # Retrieve the memories encoded with the maskmem backbone
+            to_cat_prompt, to_cat_prompt_mask, to_cat_prompt_pos_embed = [], [], []
+            # Add conditioning frames's output first (all cond frames have t_pos=0 for
+            # when getting temporal positional embedding below)
+            assert len(output_dict["cond_frame_outputs"]) > 0
+            # Select a maximum number of temporally closest cond frames for cross attention
+            cond_outputs = output_dict["cond_frame_outputs"]
+            selected_cond_outputs, unselected_cond_outputs = select_closest_cond_frames(
+                frame_idx,
+                cond_outputs,
+                self.max_cond_frames_in_attn,
+                keep_first_cond_frame=self.keep_first_cond_frame,
+            )
+            t_pos_and_prevs = [
+                ((frame_idx - t) * tpos_sign_mul, out, True)
+                for t, out in selected_cond_outputs.items()
+            ]
+            # Add last (self.num_maskmem - 1) frames before current frame for non-conditioning memory
+            # the earliest one has t_pos=1 and the latest one has t_pos=self.num_maskmem-1
+            # We also allow taking the memory frame non-consecutively (with r>1), in which case
+            # we take (self.num_maskmem - 2) frames among every r-th frames plus the last frame.
+            r = 1 if self.training else self.memory_temporal_stride_for_eval
+
+            if self.use_memory_selection:
+                valid_indices = self.frame_filter(
+                    output_dict, track_in_reverse, frame_idx, num_frames, r
+                )
+
+            for t_pos in range(1, self.num_maskmem):
+                t_rel = self.num_maskmem - t_pos  # how many frames before current frame
+                if self.use_memory_selection:
+                    if t_rel > len(valid_indices):
+                        continue
+                    prev_frame_idx = valid_indices[-t_rel]
+                else:
+                    if t_rel == 1:
+                        # for t_rel == 1, we take the last frame (regardless of r)
+                        if not track_in_reverse:
+                            # the frame immediately before this frame (i.e. frame_idx - 1)
+                            prev_frame_idx = frame_idx - t_rel
+                        else:
+                            # the frame immediately after this frame (i.e. frame_idx + 1)
+                            prev_frame_idx = frame_idx + t_rel
+                    else:
+                        # for t_rel >= 2, we take the memory frame from every r-th frames
+                        if not track_in_reverse:
+                            # first find the nearest frame among every r-th frames before this frame
+                            # for r=1, this would be (frame_idx - 2)
+                            prev_frame_idx = ((frame_idx - 2) // r) * r
+                            # then seek further among every r-th frames
+                            prev_frame_idx = prev_frame_idx - (t_rel - 2) * r
+                        else:
+                            # first find the nearest frame among every r-th frames after this frame
+                            # for r=1, this would be (frame_idx + 2)
+                            prev_frame_idx = -(-(frame_idx + 2) // r) * r
+                            # then seek further among every r-th frames
+                            prev_frame_idx = prev_frame_idx + (t_rel - 2) * r
+
+                out = output_dict["non_cond_frame_outputs"].get(prev_frame_idx, None)
+                if out is None:
+                    # If an unselected conditioning frame is among the last (self.num_maskmem - 1)
+                    # frames, we still attend to it as if it's a non-conditioning frame.
+                    out = unselected_cond_outputs.get(prev_frame_idx, None)
+                t_pos_and_prevs.append((t_pos, out, False))
+
+            for t_pos, prev, is_selected_cond_frame in t_pos_and_prevs:
+                if prev is None:
+                    continue  # skip padding frames
+                # "maskmem_features" might have been offloaded to CPU in demo use cases,
+                # so we load it back to GPU (it's a no-op if it's already on GPU).
+                feats = prev["maskmem_features"].cuda(non_blocking=True)
+                seq_len = feats.shape[-2] * feats.shape[-1]
+                to_cat_prompt.append(feats.flatten(2).permute(2, 0, 1))
+                to_cat_prompt_mask.append(
+                    torch.zeros(B, seq_len, device=device, dtype=bool)
+                )
+                # Spatial positional encoding (it might have been offloaded to CPU in eval)
+                maskmem_enc = prev["maskmem_pos_enc"][-1].cuda()
+                maskmem_enc = maskmem_enc.flatten(2).permute(2, 0, 1)
+
+                if (
+                    is_selected_cond_frame
+                    and getattr(self, "cond_frame_spatial_embedding", None) is not None
+                ):
+                    # add a spatial embedding for the conditioning frame
+                    maskmem_enc = maskmem_enc + self.cond_frame_spatial_embedding
+
+                # Temporal positional encoding
+                t = t_pos if not is_selected_cond_frame else 0
+                maskmem_enc = (
+                    maskmem_enc + self.maskmem_tpos_enc[self.num_maskmem - t - 1]
+                )
+                to_cat_prompt_pos_embed.append(maskmem_enc)
+
+            # Construct the list of past object pointers
+            # Optionally, select only a subset of spatial memory frames during trainining
+            if (
+                self.training
+                and self.prob_to_dropout_spatial_mem > 0
+                and self.rng.random() < self.prob_to_dropout_spatial_mem
+            ):
+                num_spatial_mem_keep = self.rng.integers(len(to_cat_prompt) + 1)
+                keep = self.rng.choice(
+                    range(len(to_cat_prompt)), num_spatial_mem_keep, replace=False
+                ).tolist()
+                to_cat_prompt = [to_cat_prompt[i] for i in keep]
+                to_cat_prompt_mask = [to_cat_prompt_mask[i] for i in keep]
+                to_cat_prompt_pos_embed = [to_cat_prompt_pos_embed[i] for i in keep]
+
+            max_obj_ptrs_in_encoder = min(num_frames, self.max_obj_ptrs_in_encoder)
+            # First add those object pointers from selected conditioning frames
+            # (optionally, only include object pointers in the past during evaluation)
+            if not self.training:
+                ptr_cond_outputs = {
+                    t: out
+                    for t, out in selected_cond_outputs.items()
+                    if (t >= frame_idx if track_in_reverse else t <= frame_idx)
+                }
+            else:
+                ptr_cond_outputs = selected_cond_outputs
+            pos_and_ptrs = [
+                # Temporal pos encoding contains how far away each pointer is from current frame
+                (
+                    (frame_idx - t) * tpos_sign_mul,
+                    out["obj_ptr"],
+                    True,  # is_selected_cond_frame
+                )
+                for t, out in ptr_cond_outputs.items()
+            ]
+
+            # Add up to (max_obj_ptrs_in_encoder - 1) non-conditioning frames before current frame
+            for t_diff in range(1, max_obj_ptrs_in_encoder):
+                if not self.use_memory_selection:
+                    t = frame_idx + t_diff if track_in_reverse else frame_idx - t_diff
+                    if t < 0 or (num_frames is not None and t >= num_frames):
+                        break
+                else:
+                    if -t_diff <= -len(valid_indices):
+                        break
+                    t = valid_indices[-t_diff]
+
+                out = output_dict["non_cond_frame_outputs"].get(
+                    t, unselected_cond_outputs.get(t, None)
+                )
+                if out is not None:
+                    pos_and_ptrs.append((t_diff, out["obj_ptr"], False))
+
+            # If we have at least one object pointer, add them to the across attention
+            if len(pos_and_ptrs) > 0:
+                pos_list, ptrs_list, is_selected_cond_frame_list = zip(*pos_and_ptrs)
+                # stack object pointers along dim=0 into [ptr_seq_len, B, C] shape
+                obj_ptrs = torch.stack(ptrs_list, dim=0)
+                if getattr(self, "cond_frame_obj_ptr_embedding", None) is not None:
+                    obj_ptrs = (
+                        obj_ptrs
+                        + self.cond_frame_obj_ptr_embedding
+                        * torch.tensor(is_selected_cond_frame_list, device=device)[
+                            ..., None, None
+                        ].float()
+                    )
+                # a temporal positional embedding based on how far each object pointer is from
+                # the current frame (sine embedding normalized by the max pointer num).
+                obj_pos = self._get_tpos_enc(
+                    pos_list,
+                    max_abs_pos=max_obj_ptrs_in_encoder,
+                    device=device,
+                )
+                # expand to batch size
+                obj_pos = obj_pos.unsqueeze(1).expand(-1, B, -1)
+
+                if self.mem_dim < C:
+                    # split a pointer into (C // self.mem_dim) tokens for self.mem_dim < C
+                    obj_ptrs = obj_ptrs.reshape(-1, B, C // self.mem_dim, self.mem_dim)
+                    obj_ptrs = obj_ptrs.permute(0, 2, 1, 3).flatten(0, 1)
+                    obj_pos = obj_pos.repeat_interleave(C // self.mem_dim, dim=0)
+                to_cat_prompt.append(obj_ptrs)
+                to_cat_prompt_mask.append(None)  # "to_cat_prompt_mask" is not used
+                to_cat_prompt_pos_embed.append(obj_pos)
+                num_obj_ptr_tokens = obj_ptrs.shape[0]
+            else:
+                num_obj_ptr_tokens = 0
+        else:
+            # directly add no-mem embedding (instead of using the transformer encoder)
+            pix_feat_with_mem = current_vision_feats[-1] + self.no_mem_embed
+            pix_feat_with_mem = pix_feat_with_mem.permute(1, 2, 0).view(B, C, H, W)
+            return pix_feat_with_mem
+
+            # Use a dummy token on the first grame (to avoid emtpy memory input to tranformer encoder)
+            to_cat_prompt = [self.no_mem_embed.expand(1, B, self.mem_dim)]
+            to_cat_prompt_mask = [torch.zeros(B, 1, device=device, dtype=bool)]
+            to_cat_prompt_pos_embed = [self.no_mem_pos_enc.expand(1, B, self.mem_dim)]
+
+        # Step 2: Concatenate the memories and forward through the transformer encoder
+        prompt = torch.cat(to_cat_prompt, dim=0)
+        prompt_mask = None  # For now, we always masks are zeros anyways
+        prompt_pos_embed = torch.cat(to_cat_prompt_pos_embed, dim=0)
+        encoder_out = self.transformer.encoder(
+            src=current_vision_feats,
+            src_key_padding_mask=[None],
+            src_pos=current_vision_pos_embeds,
+            prompt=prompt,
+            prompt_pos=prompt_pos_embed,
+            prompt_key_padding_mask=prompt_mask,
+            feat_sizes=feat_sizes,
+            num_obj_ptr_tokens=num_obj_ptr_tokens,
+        )
+        # reshape the output (HW)BC => BCHW
+        pix_feat_with_mem = encoder_out["memory"].permute(1, 2, 0).view(B, C, H, W)
+        return pix_feat_with_mem
+
+    def _encode_new_memory(
+        self,
+        image,
+        current_vision_feats,
+        feat_sizes,
+        pred_masks_high_res,
+        object_score_logits,
+        is_mask_from_pts,
+        output_dict=None,
+        is_init_cond_frame=False,
+    ):
+        """Encode the current image and its prediction into a memory feature."""
+        B = current_vision_feats[-1].size(1)  # batch size on this frame
+        C = self.hidden_dim
+        H, W = feat_sizes[-1]  # top-level (lowest-resolution) feature size
+        # top-level feature, (HW)BC => BCHW
+        pix_feat = current_vision_feats[-1].permute(1, 2, 0).view(B, C, H, W)
+        if self.non_overlap_masks_for_mem_enc and not self.training:
+            # optionally, apply non-overlapping constraints to the masks (it's applied
+            # in the batch dimension and should only be used during eval, where all
+            # the objects come from the same video under batch size 1).
+            pred_masks_high_res = self._apply_non_overlapping_constraints(
+                pred_masks_high_res
+            )
+        # scale the raw mask logits with a temperature before applying sigmoid
+        if is_mask_from_pts and not self.training:
+            mask_for_mem = (pred_masks_high_res > 0).float()
+        else:
+            # apply sigmoid on the raw mask logits to turn them into range (0, 1)
+            mask_for_mem = torch.sigmoid(pred_masks_high_res)
+        # apply scale and bias terms to the sigmoid probabilities
+        if self.sigmoid_scale_for_mem_enc != 1.0:
+            mask_for_mem = mask_for_mem * self.sigmoid_scale_for_mem_enc
+        if self.sigmoid_bias_for_mem_enc != 0.0:
+            mask_for_mem = mask_for_mem + self.sigmoid_bias_for_mem_enc
+
+        if isinstance(self.maskmem_backbone, SimpleMaskEncoder):
+            pix_feat = pix_feat.view_as(pix_feat)
+            maskmem_out = self.maskmem_backbone(
+                pix_feat, mask_for_mem, skip_mask_sigmoid=True
+            )
+        else:
+            maskmem_out = self.maskmem_backbone(image, pix_feat, mask_for_mem)
+        # Clone the feats and pos_enc to enable compilation
+        maskmem_features = self._maybe_clone(maskmem_out["vision_features"])
+        maskmem_pos_enc = [self._maybe_clone(m) for m in maskmem_out["vision_pos_enc"]]
+        # add a no-object embedding to the spatial memory to indicate that the frame
+        # is predicted to be occluded (i.e. no object is appearing in the frame)
+        is_obj_appearing = (object_score_logits > 0).float()
+        maskmem_features += (
+            1 - is_obj_appearing[..., None, None]
+        ) * self.no_obj_embed_spatial[..., None, None].expand(*maskmem_features.shape)
+
+        return maskmem_features, maskmem_pos_enc
+
+    def forward_tracking(self, backbone_out, input, return_dict=False):
+        """Forward video tracking on each frame (and sample correction clicks)."""
+        img_feats_already_computed = backbone_out["backbone_fpn"] is not None
+        if img_feats_already_computed:
+            # Prepare the backbone features
+            # - vision_feats and vision_pos_embeds are in (HW)BC format
+            (
+                _,
+                vision_feats,
+                vision_pos_embeds,
+                feat_sizes,
+            ) = self._prepare_backbone_features(backbone_out)
+
+        # Starting the stage loop
+        num_frames = backbone_out["num_frames"]
+        init_cond_frames = backbone_out["init_cond_frames"]
+        frames_to_add_correction_pt = backbone_out["frames_to_add_correction_pt"]
+        # first process all the initial conditioning frames to encode them as memory,
+        # and then conditioning on them to track the remaining frames
+        processing_order = init_cond_frames + backbone_out["frames_not_in_init_cond"]
+        output_dict = {
+            "cond_frame_outputs": {},  # dict containing {frame_idx: <out>}
+            "non_cond_frame_outputs": {},  # dict containing {frame_idx: <out>}
+        }
+        for stage_id in processing_order:
+            # Get the image features for the current frames
+            img_ids = input.find_inputs[stage_id].img_ids
+            if img_feats_already_computed:
+                # Retrieve image features according to img_ids (if they are already computed).
+                current_image = input.img_batch[img_ids]
+                current_vision_feats = [x[:, img_ids] for x in vision_feats]
+                current_vision_pos_embeds = [x[:, img_ids] for x in vision_pos_embeds]
+            else:
+                # Otherwise, compute the image features on the fly for the given img_ids
+                # (this might be used for evaluation on long videos to avoid backbone OOM).
+                (
+                    current_image,
+                    current_vision_feats,
+                    current_vision_pos_embeds,
+                    feat_sizes,
+                ) = self._prepare_backbone_features_per_frame(input.img_batch, img_ids)
+            # Get output masks based on this frame's prompts and previous memory
+            current_out = self.track_step(
+                frame_idx=stage_id,
+                is_init_cond_frame=stage_id in init_cond_frames,
+                current_vision_feats=current_vision_feats,
+                current_vision_pos_embeds=current_vision_pos_embeds,
+                feat_sizes=feat_sizes,
+                image=current_image,
+                point_inputs=backbone_out["point_inputs_per_frame"].get(stage_id, None),
+                mask_inputs=backbone_out["mask_inputs_per_frame"].get(stage_id, None),
+                gt_masks=backbone_out["gt_masks_per_frame"].get(stage_id, None),
+                frames_to_add_correction_pt=frames_to_add_correction_pt,
+                output_dict=output_dict,
+                num_frames=num_frames,
+            )
+            # Append the output, depending on whether it's a conditioning frame
+            add_output_as_cond_frame = stage_id in init_cond_frames or (
+                self.add_all_frames_to_correct_as_cond
+                and stage_id in frames_to_add_correction_pt
+            )
+            if add_output_as_cond_frame:
+                output_dict["cond_frame_outputs"][stage_id] = current_out
+            else:
+                output_dict["non_cond_frame_outputs"][stage_id] = current_out
+
+        if return_dict:
+            return output_dict
+        # turn `output_dict` into a list for loss function
+        all_frame_outputs = {}
+        all_frame_outputs.update(output_dict["cond_frame_outputs"])
+        all_frame_outputs.update(output_dict["non_cond_frame_outputs"])
+        all_frame_outputs = [all_frame_outputs[t] for t in range(num_frames)]
+        # Make DDP happy with activation checkpointing by removing unused keys
+        all_frame_outputs = [
+            {k: v for k, v in d.items() if k != "obj_ptr"} for d in all_frame_outputs
+        ]
+
+        return all_frame_outputs
+
+    def track_step(
+        self,
+        frame_idx,
+        is_init_cond_frame,
+        current_vision_feats,
+        current_vision_pos_embeds,
+        feat_sizes,
+        image,
+        point_inputs,
+        mask_inputs,
+        output_dict,
+        num_frames,
+        track_in_reverse=False,  # tracking in reverse time order (for demo usage)
+        # Whether to run the memory encoder on the predicted masks. Sometimes we might want
+        # to skip the memory encoder with `run_mem_encoder=False`. For example,
+        # in demo we might call `track_step` multiple times for each user click,
+        # and only encode the memory when the user finalizes their clicks. And in ablation
+        # settings like SAM training on static images, we don't need the memory encoder.
+        run_mem_encoder=True,
+        # The previously predicted SAM mask logits (which can be fed together with new clicks in demo).
+        prev_sam_mask_logits=None,
+        use_prev_mem_frame=True,
+    ):
+        current_out = {"point_inputs": point_inputs, "mask_inputs": mask_inputs}
+        # High-resolution feature maps for the SAM head, reshape (HW)BC => BCHW
+        if len(current_vision_feats) > 1:
+            high_res_features = [
+                x.permute(1, 2, 0).view(x.size(1), x.size(2), *s)
+                for x, s in zip(current_vision_feats[:-1], feat_sizes[:-1])
+            ]
+        else:
+            high_res_features = None
+        if mask_inputs is not None:
+            # (see it as a GT mask) without using a SAM prompt encoder + mask decoder.
+            pix_feat = current_vision_feats[-1].permute(1, 2, 0)
+            pix_feat = pix_feat.view(-1, self.hidden_dim, *feat_sizes[-1])
+            sam_outputs = self._use_mask_as_output(
+                pix_feat, high_res_features, mask_inputs
+            )
+        else:
+            # fused the visual feature with previous memory features in the memory bank
+            pix_feat_with_mem = self._prepare_memory_conditioned_features(
+                frame_idx=frame_idx,
+                is_init_cond_frame=is_init_cond_frame,
+                current_vision_feats=current_vision_feats[-1:],
+                current_vision_pos_embeds=current_vision_pos_embeds[-1:],
+                feat_sizes=feat_sizes[-1:],
+                output_dict=output_dict,
+                num_frames=num_frames,
+                track_in_reverse=track_in_reverse,
+                use_prev_mem_frame=use_prev_mem_frame,
+            )
+            # apply SAM-style segmentation head
+            # here we might feed previously predicted low-res SAM mask logits into the SAM mask decoder,
+            # e.g. in demo where such logits come from earlier interaction instead of correction sampling
+            # (in this case, the SAM mask decoder should have `self.iter_use_prev_mask_pred=True`, and
+            # any `mask_inputs` shouldn't reach here as they are sent to _use_mask_as_output instead)
+            if prev_sam_mask_logits is not None:
+                assert self.iter_use_prev_mask_pred
+                assert point_inputs is not None and mask_inputs is None
+                mask_inputs = prev_sam_mask_logits
+            multimask_output = self._use_multimask(is_init_cond_frame, point_inputs)
+            sam_outputs = self._forward_sam_heads(
+                backbone_features=pix_feat_with_mem,
+                point_inputs=point_inputs,
+                mask_inputs=mask_inputs,
+                high_res_features=high_res_features,
+                multimask_output=multimask_output,
+            )
+        (
+            _,
+            high_res_multimasks,
+            ious,
+            low_res_masks,
+            high_res_masks,
+            obj_ptr,
+            object_score_logits,
+        ) = sam_outputs
+        # Use the final prediction (after all correction steps for output and eval)
+        current_out["pred_masks"] = low_res_masks
+        current_out["pred_masks_high_res"] = high_res_masks
+        current_out["obj_ptr"] = obj_ptr
+        if self.use_memory_selection:
+            current_out["object_score_logits"] = object_score_logits
+            iou_score = ious.max(-1)[0]
+            current_out["iou_score"] = iou_score
+            current_out["eff_iou_score"] = self.cal_mem_score(
+                object_score_logits, iou_score
+            )
+        if not self.training:
+            # Only add this in inference (to avoid unused param in activation checkpointing;
+            # it's mainly used in the demo to encode spatial memories w/ consolidated masks)
+            current_out["object_score_logits"] = object_score_logits
+
+        # Finally run the memory encoder on the predicted mask to encode
+        # it into a new memory feature (that can be used in future frames)
+        # (note that `self.num_maskmem == 0` is primarily used for reproducing SAM on
+        # images, in which case we'll just skip memory encoder to save compute).
+        if run_mem_encoder and self.num_maskmem > 0:
+            high_res_masks_for_mem_enc = high_res_masks
+            maskmem_features, maskmem_pos_enc = self._encode_new_memory(
+                image=image,
+                current_vision_feats=current_vision_feats,
+                feat_sizes=feat_sizes,
+                pred_masks_high_res=high_res_masks_for_mem_enc,
+                object_score_logits=object_score_logits,
+                is_mask_from_pts=(point_inputs is not None),
+                output_dict=output_dict,
+                is_init_cond_frame=is_init_cond_frame,
+            )
+            current_out["maskmem_features"] = maskmem_features
+            current_out["maskmem_pos_enc"] = maskmem_pos_enc
+        else:
+            current_out["maskmem_features"] = None
+            current_out["maskmem_pos_enc"] = None
+
+        # Optionally, offload the outputs to CPU memory during evaluation to avoid
+        # GPU OOM on very long videos or very large resolution or too many objects
+        if self.offload_output_to_cpu_for_eval and not self.training:
+            # Here we only keep those keys needed for evaluation to get a compact output
+            trimmed_out = {
+                "pred_masks": current_out["pred_masks"].cpu(),
+                "pred_masks_high_res": current_out["pred_masks_high_res"].cpu(),
+                # other items for evaluation (these are small tensors so we keep them on GPU)
+                "obj_ptr": current_out["obj_ptr"],
+                "object_score_logits": current_out["object_score_logits"],
+            }
+            if run_mem_encoder and self.num_maskmem > 0:
+                trimmed_out["maskmem_features"] = maskmem_features.cpu()
+                trimmed_out["maskmem_pos_enc"] = [x.cpu() for x in maskmem_pos_enc]
+            if self.use_memory_selection:
+                trimmed_out["iou_score"] = current_out["iou_score"].cpu()
+                trimmed_out["eff_iou_score"] = current_out["eff_iou_score"].cpu()
+            current_out = trimmed_out
+
+        # Optionally, trim the output of past non-conditioning frame (r * num_maskmem frames
+        # before the current frame) during evaluation. This is intended to save GPU or CPU
+        # memory for semi-supervised VOS eval, where only the first frame receives prompts.
+        def _trim_past_out(past_out, current_out):
+            if past_out is None:
+                return None
+            return {
+                "pred_masks": past_out["pred_masks"],
+                "obj_ptr": past_out["obj_ptr"],
+                "object_score_logits": past_out["object_score_logits"],
+            }
+
+        if self.trim_past_non_cond_mem_for_eval and not self.training:
+            r = self.memory_temporal_stride_for_eval
+            past_frame_idx = frame_idx - r * self.num_maskmem
+            past_out = output_dict["non_cond_frame_outputs"].get(past_frame_idx, None)
+
+            if past_out is not None:
+                print(past_out.get("eff_iou_score", 0))
+                if (
+                    self.use_memory_selection
+                    and past_out.get("eff_iou_score", 0) < self.mf_threshold
+                ) or not self.use_memory_selection:
+                    output_dict["non_cond_frame_outputs"][past_frame_idx] = (
+                        _trim_past_out(past_out, current_out)
+                    )
+
+            if (
+                self.use_memory_selection and not self.offload_output_to_cpu_for_eval
+            ):  ## design for memory selection, trim too old frames to save memory
+                far_old_frame_idx = frame_idx - 20 * self.max_obj_ptrs_in_encoder
+                past_out = output_dict["non_cond_frame_outputs"].get(
+                    far_old_frame_idx, None
+                )
+                if past_out is not None:
+                    output_dict["non_cond_frame_outputs"][far_old_frame_idx] = (
+                        _trim_past_out(past_out, current_out)
+                    )
+
+        return current_out
+
+    def _use_multimask(self, is_init_cond_frame, point_inputs):
+        """Whether to use multimask output in the SAM head."""
+        num_pts = 0 if point_inputs is None else point_inputs["point_labels"].size(1)
+        multimask_output = (
+            self.multimask_output_in_sam
+            and (is_init_cond_frame or self.multimask_output_for_tracking)
+            and (self.multimask_min_pt_num <= num_pts <= self.multimask_max_pt_num)
+        )
+        return multimask_output
+
+    def _apply_non_overlapping_constraints(self, pred_masks):
+        """
+        Apply non-overlapping constraints to the object scores in pred_masks. Here we
+        keep only the highest scoring object at each spatial location in pred_masks.
+        """
+        batch_size = pred_masks.size(0)
+        if batch_size == 1:
+            return pred_masks
+
+        device = pred_masks.device
+        # "max_obj_inds": object index of the object with the highest score at each location
+        max_obj_inds = torch.argmax(pred_masks, dim=0, keepdim=True)
+        # "batch_obj_inds": object index of each object slice (along dim 0) in `pred_masks`
+        batch_obj_inds = torch.arange(batch_size, device=device)[:, None, None, None]
+        keep = max_obj_inds == batch_obj_inds
+        # suppress overlapping regions' scores below -10.0 so that the foreground regions
+        # don't overlap (here sigmoid(-10.0)=4.5398e-05)
+        pred_masks = torch.where(keep, pred_masks, torch.clamp(pred_masks, max=-10.0))
+        return pred_masks
+
+    def _compile_all_components(self):
+        """Compile all model components for faster inference."""
+        # a larger cache size to hold varying number of shapes for torch.compile
+        # see https://github.com/pytorch/pytorch/blob/v2.5.1/torch/_dynamo/config.py#L42-L49
+        torch._dynamo.config.cache_size_limit = 64
+        torch._dynamo.config.accumulated_cache_size_limit = 2048
+        from sam3.perflib.compile import compile_wrapper
+
+        logging.info("Compiling all components. First time may be very slow.")
+
+        self.maskmem_backbone.forward = compile_wrapper(
+            self.maskmem_backbone.forward,
+            mode="max-autotune",
+            fullgraph=True,
+            dynamic=False,
+        )
+        self.transformer.encoder.forward = compile_wrapper(
+            self.transformer.encoder.forward,
+            mode="max-autotune",
+            fullgraph=True,
+            dynamic=True,  # Num. of memories varies
+        )
+        # We disable compilation of sam_prompt_encoder as it sometimes gives a large accuracy regression,
+        # especially when sam_mask_prompt (previous mask logits) is not None
+        # self.sam_prompt_encoder.forward = torch.compile(
+        #     self.sam_prompt_encoder.forward,
+        #     mode="max-autotune",
+        #     fullgraph=True,
+        #     dynamic=False,  # Accuracy regression on True
+        # )
+        self.sam_mask_decoder.forward = compile_wrapper(
+            self.sam_mask_decoder.forward,
+            mode="max-autotune",
+            fullgraph=True,
+            dynamic=False,  # Accuracy regression on True
+        )
+
+    def _maybe_clone(self, x):
+        """Clone a tensor if and only if `self.compile_all_components` is True."""
+        return x.clone() if self.compile_all_components else x
+
+
+def concat_points(old_point_inputs, new_points, new_labels):
+    """Add new points and labels to previous point inputs (add at the end)."""
+    if old_point_inputs is None:
+        points, labels = new_points, new_labels
+    else:
+        points = torch.cat([old_point_inputs["point_coords"], new_points], dim=1)
+        labels = torch.cat([old_point_inputs["point_labels"], new_labels], dim=1)
+
+    return {"point_coords": points, "point_labels": labels}

detect_tools/sam3/sam3/model/sam3_tracker_utils.py

@@ -0,0 +1,427 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates. All Rights Reserved
+
+import numpy as np
+import torch
+import torch.nn.functional as F
+from numpy.typing import NDArray
+
+from sam3.model.edt import edt_triton
+
+
+def sample_box_points(
+    masks: torch.Tensor,
+    noise: float = 0.1,  # SAM default
+    noise_bound: int = 20,  # SAM default
+    top_left_label: int = 2,
+    bottom_right_label: int = 3,
+) -> tuple[NDArray, NDArray]:
+    """
+    Sample a noised version of the top left and bottom right corners of a given `bbox`
+
+    Inputs:
+    - masks: [B, 1, H, W] tensor
+    - noise: noise as a fraction of box width and height, dtype=float
+    - noise_bound: maximum amount of noise (in pure pixels), dtype=int
+
+    Returns:
+    - box_coords: [B, num_pt, 2], contains (x, y) coordinates of top left and bottom right box corners, dtype=torch.float
+    - box_labels: [B, num_pt], label 2 is reserverd for top left and 3 for bottom right corners, dtype=torch.int32
+    """
+    device = masks.device
+    box_coords = mask_to_box(masks)
+    B, _, H, W = masks.shape
+    box_labels = torch.tensor(
+        [top_left_label, bottom_right_label], dtype=torch.int, device=device
+    ).repeat(B)
+    if noise > 0.0:
+        if not isinstance(noise_bound, torch.Tensor):
+            noise_bound = torch.tensor(noise_bound, device=device)
+        bbox_w = box_coords[..., 2] - box_coords[..., 0]
+        bbox_h = box_coords[..., 3] - box_coords[..., 1]
+        max_dx = torch.min(bbox_w * noise, noise_bound)
+        max_dy = torch.min(bbox_h * noise, noise_bound)
+        box_noise = 2 * torch.rand(B, 1, 4, device=device) - 1
+        box_noise = box_noise * torch.stack((max_dx, max_dy, max_dx, max_dy), dim=-1)
+
+        box_coords = box_coords + box_noise
+        img_bounds = (
+            torch.tensor([W, H, W, H], device=device) - 1
+        )  # uncentered pixel coords
+        box_coords.clamp_(torch.zeros_like(img_bounds), img_bounds)  # In place clamping
+
+    box_coords = box_coords.reshape(-1, 2, 2)  # always 2 points
+    box_labels = box_labels.reshape(-1, 2)
+    return box_coords, box_labels
+
+
+def mask_to_box(masks: torch.Tensor):
+    """
+    compute bounding box given an input mask
+
+    Inputs:
+    - masks: [B, 1, H, W] tensor
+
+    Returns:
+    - box_coords: [B, 1, 4], contains (x, y) coordinates of top left and bottom right box corners, dtype=torch.Tensor
+    """
+    B, _, h, w = masks.shape
+    device = masks.device
+    mask_area = masks.sum(dim=(-1, -2))
+    xs = torch.arange(w, device=device, dtype=torch.int32)
+    ys = torch.arange(h, device=device, dtype=torch.int32)
+    grid_xs, grid_ys = torch.meshgrid(xs, ys, indexing="xy")
+    grid_xs = grid_xs[None, None, ...].expand(B, 1, h, w)
+    grid_ys = grid_ys[None, None, ...].expand(B, 1, h, w)
+    min_xs, _ = torch.min(torch.where(masks, grid_xs, w).flatten(-2), dim=-1)
+    max_xs, _ = torch.max(torch.where(masks, grid_xs, -1).flatten(-2), dim=-1)
+    min_ys, _ = torch.min(torch.where(masks, grid_ys, h).flatten(-2), dim=-1)
+    max_ys, _ = torch.max(torch.where(masks, grid_ys, -1).flatten(-2), dim=-1)
+    bbox_coords = torch.stack((min_xs, min_ys, max_xs, max_ys), dim=-1)
+    bbox_coords = torch.where(
+        mask_area[..., None] > 0, bbox_coords, torch.zeros_like(bbox_coords)
+    )
+    return bbox_coords
+
+
+def sample_random_points_from_errors(gt_masks, pred_masks, num_pt=1):
+    """
+    Sample `num_pt` random points (along with their labels) independently from the error regions.
+
+    Inputs:
+    - gt_masks: [B, 1, H_im, W_im] masks, dtype=torch.bool
+    - pred_masks: [B, 1, H_im, W_im] masks, dtype=torch.bool or None
+    - num_pt: int, number of points to sample independently for each of the B error maps
+
+    Outputs:
+    - points: [B, num_pt, 2], dtype=torch.float, contains (x, y) coordinates of each sampled point
+    - labels: [B, num_pt], dtype=torch.int32, where 1 means positive clicks and 0 means
+      negative clicks
+    """
+    if pred_masks is None:  # if pred_masks is not provided, treat it as empty
+        pred_masks = torch.zeros_like(gt_masks)
+    assert gt_masks.dtype == torch.bool and gt_masks.size(1) == 1
+    assert pred_masks.dtype == torch.bool and pred_masks.shape == gt_masks.shape
+    assert num_pt >= 0
+
+    B, _, H_im, W_im = gt_masks.shape
+    device = gt_masks.device
+
+    # false positive region, a new point sampled in this region should have
+    # negative label to correct the FP error
+    fp_masks = ~gt_masks & pred_masks
+    # false negative region, a new point sampled in this region should have
+    # positive label to correct the FN error
+    fn_masks = gt_masks & ~pred_masks
+    # whether the prediction completely match the ground-truth on each mask
+    all_correct = torch.all((gt_masks == pred_masks).flatten(2), dim=2)
+    all_correct = all_correct[..., None, None]
+
+    # channel 0 is FP map, while channel 1 is FN map
+    pts_noise = torch.rand(B, num_pt, H_im, W_im, 2, device=device)
+    # sample a negative new click from FP region or a positive new click
+    # from FN region, depend on where the maximum falls,
+    # and in case the predictions are all correct (no FP or FN), we just
+    # sample a negative click from the background region
+    pts_noise[..., 0] *= fp_masks | (all_correct & ~gt_masks)
+    pts_noise[..., 1] *= fn_masks
+    pts_idx = pts_noise.flatten(2).argmax(dim=2)
+    labels = (pts_idx % 2).to(torch.int32)
+    pts_idx = pts_idx // 2
+    pts_x = pts_idx % W_im
+    pts_y = pts_idx // W_im
+    points = torch.stack([pts_x, pts_y], dim=2).to(torch.float)
+    return points, labels
+
+
+def sample_one_point_from_error_center(gt_masks, pred_masks, padding=True):
+    """
+    Sample 1 random point (along with its label) from the center of each error region,
+    that is, the point with the largest distance to the boundary of each error region.
+    This is the RITM sampling method from https://github.com/saic-vul/ritm_interactive_segmentation/blob/master/isegm/inference/clicker.py
+
+    Inputs:
+    - gt_masks: [B, 1, H_im, W_im] masks, dtype=torch.bool
+    - pred_masks: [B, 1, H_im, W_im] masks, dtype=torch.bool or None
+    - padding: if True, pad with boundary of 1 px for distance transform
+
+    Outputs:
+    - points: [B, 1, 2], dtype=torch.float, contains (x, y) coordinates of each sampled point
+    - labels: [B, 1], dtype=torch.int32, where 1 means positive clicks and 0 means negative clicks
+    """
+    if pred_masks is None:
+        pred_masks = torch.zeros_like(gt_masks)
+    assert gt_masks.dtype == torch.bool and gt_masks.size(1) == 1
+    assert pred_masks.dtype == torch.bool and pred_masks.shape == gt_masks.shape
+
+    B, _, H, W = gt_masks.shape
+
+    # false positive region, a new point sampled in this region should have
+    # negative label to correct the FP error
+    fp_masks = (~gt_masks & pred_masks).squeeze(1)
+    # false negative region, a new point sampled in this region should have
+    # positive label to correct the FN error
+    fn_masks = (gt_masks & ~pred_masks).squeeze(1)
+
+    if padding:
+        padded_fp_masks = torch.zeros(
+            B, H + 2, W + 2, dtype=fp_masks.dtype, device=fp_masks.device
+        )
+        padded_fp_masks[:, 1 : H + 1, 1 : W + 1] = fp_masks
+        padded_fn_masks = torch.zeros(
+            B, H + 2, W + 2, dtype=fp_masks.dtype, device=fp_masks.device
+        )
+        padded_fn_masks[:, 1 : H + 1, 1 : W + 1] = fn_masks
+    else:
+        padded_fp_masks = fp_masks
+        padded_fn_masks = fn_masks
+
+    fn_mask_dt = edt_triton(padded_fn_masks)
+    fp_mask_dt = edt_triton(padded_fp_masks)
+    if padding:
+        fn_mask_dt = fn_mask_dt[:, 1:-1, 1:-1]
+        fp_mask_dt = fp_mask_dt[:, 1:-1, 1:-1]
+
+    fn_max, fn_argmax = fn_mask_dt.reshape(B, -1).max(dim=-1)
+    fp_max, fp_argmax = fp_mask_dt.reshape(B, -1).max(dim=-1)
+    is_positive = fn_max > fp_max
+    chosen = torch.where(is_positive, fn_argmax, fp_argmax)
+    points_x = chosen % W
+    points_y = chosen // W
+
+    labels = is_positive.long()
+    points = torch.stack([points_x, points_y], -1)
+    return points.unsqueeze(1), labels.unsqueeze(1)
+
+
+def sample_one_point_from_error_center_slow(gt_masks, pred_masks, padding=True):
+    """
+    Sample 1 random point (along with its label) from the center of each error region,
+    that is, the point with the largest distance to the boundary of each error region.
+    This is the RITM sampling method from https://github.com/saic-vul/ritm_interactive_segmentation/blob/master/isegm/inference/clicker.py
+
+    Inputs:
+    - gt_masks: [B, 1, H_im, W_im] masks, dtype=torch.bool
+    - pred_masks: [B, 1, H_im, W_im] masks, dtype=torch.bool or None
+    - padding: if True, pad with boundary of 1 px for distance transform
+
+    Outputs:
+    - points: [B, 1, 2], dtype=torch.float, contains (x, y) coordinates of each sampled point
+    - labels: [B, 1], dtype=torch.int32, where 1 means positive clicks and 0 means negative clicks
+    """
+    import cv2  # delay OpenCV import to avoid unnecessary dependency
+
+    if pred_masks is None:
+        pred_masks = torch.zeros_like(gt_masks)
+    assert gt_masks.dtype == torch.bool and gt_masks.size(1) == 1
+    assert pred_masks.dtype == torch.bool and pred_masks.shape == gt_masks.shape
+
+    B, _, _, W_im = gt_masks.shape
+    device = gt_masks.device
+
+    # false positive region, a new point sampled in this region should have
+    # negative label to correct the FP error
+    fp_masks = ~gt_masks & pred_masks
+    # false negative region, a new point sampled in this region should have
+    # positive label to correct the FN error
+    fn_masks = gt_masks & ~pred_masks
+
+    fp_masks = fp_masks.cpu().numpy()
+    fn_masks = fn_masks.cpu().numpy()
+    points = torch.zeros(B, 1, 2, dtype=torch.float)
+    labels = torch.ones(B, 1, dtype=torch.int32)
+    for b in range(B):
+        fn_mask = fn_masks[b, 0]
+        fp_mask = fp_masks[b, 0]
+        if padding:
+            fn_mask = np.pad(fn_mask, ((1, 1), (1, 1)), "constant")
+            fp_mask = np.pad(fp_mask, ((1, 1), (1, 1)), "constant")
+        # compute the distance of each point in FN/FP region to its boundary
+        fn_mask_dt = cv2.distanceTransform(fn_mask.astype(np.uint8), cv2.DIST_L2, 0)
+        fp_mask_dt = cv2.distanceTransform(fp_mask.astype(np.uint8), cv2.DIST_L2, 0)
+        if padding:
+            fn_mask_dt = fn_mask_dt[1:-1, 1:-1]
+            fp_mask_dt = fp_mask_dt[1:-1, 1:-1]
+
+        # take the point in FN/FP region with the largest distance to its boundary
+        fn_mask_dt_flat = fn_mask_dt.reshape(-1)
+        fp_mask_dt_flat = fp_mask_dt.reshape(-1)
+        fn_argmax = np.argmax(fn_mask_dt_flat)
+        fp_argmax = np.argmax(fp_mask_dt_flat)
+        is_positive = fn_mask_dt_flat[fn_argmax] > fp_mask_dt_flat[fp_argmax]
+        pt_idx = fn_argmax if is_positive else fp_argmax
+        points[b, 0, 0] = pt_idx % W_im  # x
+        points[b, 0, 1] = pt_idx // W_im  # y
+        labels[b, 0] = int(is_positive)
+
+    points = points.to(device)
+    labels = labels.to(device)
+    return points, labels
+
+
+def get_next_point(gt_masks, pred_masks, method):
+    if method == "uniform":
+        return sample_random_points_from_errors(gt_masks, pred_masks)
+    elif method == "center":
+        return sample_one_point_from_error_center(gt_masks, pred_masks)
+    else:
+        raise ValueError(f"unknown sampling method {method}")
+
+
+def select_closest_cond_frames(
+    frame_idx, cond_frame_outputs, max_cond_frame_num, keep_first_cond_frame=False
+):
+    """
+    Select up to `max_cond_frame_num` conditioning frames from `cond_frame_outputs`
+    that are temporally closest to the current frame at `frame_idx`. Here, we take
+    - a) the closest conditioning frame before `frame_idx` (if any);
+    - b) the closest conditioning frame after `frame_idx` (if any);
+    - c) any other temporally closest conditioning frames until reaching a total
+         of `max_cond_frame_num` conditioning frames.
+
+    Outputs:
+    - selected_outputs: selected items (keys & values) from `cond_frame_outputs`.
+    - unselected_outputs: items (keys & values) not selected in `cond_frame_outputs`.
+    """
+    if max_cond_frame_num == -1 or len(cond_frame_outputs) <= max_cond_frame_num:
+        selected_outputs = cond_frame_outputs
+        unselected_outputs = {}
+    else:
+        assert max_cond_frame_num >= 2, "we should allow using 2+ conditioning frames"
+        selected_outputs = {}
+        if keep_first_cond_frame:
+            idx_first = min(
+                (t for t in cond_frame_outputs if t < frame_idx), default=None
+            )
+            if idx_first is None:
+                # Maybe we are tracking in reverse
+                idx_first = max(
+                    (t for t in cond_frame_outputs if t > frame_idx), default=None
+                )
+            if idx_first is not None:
+                selected_outputs[idx_first] = cond_frame_outputs[idx_first]
+        # the closest conditioning frame before `frame_idx` (if any)
+        idx_before = max((t for t in cond_frame_outputs if t < frame_idx), default=None)
+        if idx_before is not None:
+            selected_outputs[idx_before] = cond_frame_outputs[idx_before]
+
+        # the closest conditioning frame after `frame_idx` (if any)
+        idx_after = min((t for t in cond_frame_outputs if t >= frame_idx), default=None)
+        if idx_after is not None:
+            selected_outputs[idx_after] = cond_frame_outputs[idx_after]
+
+        # add other temporally closest conditioning frames until reaching a total
+        # of `max_cond_frame_num` conditioning frames.
+        num_remain = max_cond_frame_num - len(selected_outputs)
+        inds_remain = sorted(
+            (t for t in cond_frame_outputs if t not in selected_outputs),
+            key=lambda x: abs(x - frame_idx),
+        )[:num_remain]
+        selected_outputs.update((t, cond_frame_outputs[t]) for t in inds_remain)
+        unselected_outputs = {
+            t: v for t, v in cond_frame_outputs.items() if t not in selected_outputs
+        }
+
+    return selected_outputs, unselected_outputs
+
+
+def get_1d_sine_pe(pos_inds, dim, temperature=10000):
+    """
+    Get 1D sine positional embedding as in the original Transformer paper.
+    """
+    pe_dim = dim // 2
+    dim_t = torch.arange(pe_dim, dtype=torch.float32, device=pos_inds.device)
+    dim_t = temperature ** (2 * (dim_t // 2) / pe_dim)
+
+    pos_embed = pos_inds.unsqueeze(-1) / dim_t
+    pos_embed = torch.cat([pos_embed.sin(), pos_embed.cos()], dim=-1)
+    return pos_embed
+
+
+def get_best_gt_match_from_multimasks(pred_multimasks, gt_masks, pred_scores=None):
+    """
+    Get the mask with the best match to GT masks (based on IoU) from pred_multimasks.
+    Optionally, use `pred_scores` to break ties in case all IoUs are zeros.
+    """
+    assert pred_multimasks.ndim == 4 and gt_masks.ndim == 4
+    if pred_multimasks.size(1) == 1:
+        return pred_multimasks  # only a single mask channel, nothing to select
+
+    pred_multimasks_binary = pred_multimasks > 0
+    area_i = torch.sum(pred_multimasks_binary & gt_masks, dim=(2, 3)).float()
+    area_u = torch.sum(pred_multimasks_binary | gt_masks, dim=(2, 3)).float()
+    ious = area_i / torch.clamp(area_u, min=1.0)
+
+    # In case all IoUs are zeros (e.g. because the GT mask is empty), use pred_scores
+    # to break ties and select the best mask
+    if pred_scores is not None:
+        has_nonzero_ious = torch.any(ious > 0).expand_as(ious)
+        scores = torch.where(has_nonzero_ious, ious, pred_scores)
+    else:
+        scores = ious
+
+    # Finally, take the best mask prediction (with the highest score)
+    best_scores_inds = torch.argmax(scores, dim=-1)
+    batch_inds = torch.arange(scores.size(0), device=scores.device)
+    best_pred_mask = pred_multimasks[batch_inds, best_scores_inds].unsqueeze(1)
+    return best_pred_mask
+
+
+def fill_holes_in_mask_scores(mask, max_area, fill_holes=True, remove_sprinkles=True):
+    """
+    A post processor to fill small holes in mask scores with area under `max_area`.
+    Holes are those small connected components in either background or foreground.
+
+    Note that it relies on the "cc_torch" package to find connected components fast. You can
+    install it via the following command (`TORCH_CUDA_ARCH_LIST=8.0` is for A100 GPUs):
+    ```
+    pip uninstall -y cc_torch; TORCH_CUDA_ARCH_LIST=8.0 9.0 pip install git+https://github.com/ronghanghu/cc_torch
+    ```
+    Otherwise, it will fallback to a slightly slower triton implementation, or skimage if the tensor is on cpu
+    """
+
+    if max_area <= 0:
+        return mask  # nothing to fill in this case
+
+    if fill_holes:
+        # We remove small connected components in background by changing them to foreground
+        # with a small positive mask score (0.1).
+        mask_bg = mask <= 0
+        bg_area_thresh = max_area
+        _, areas_bg = _get_connected_components_with_padding(mask_bg)
+        small_components_bg = mask_bg & (areas_bg <= bg_area_thresh)
+        mask = torch.where(small_components_bg, 0.1, mask)
+
+    if remove_sprinkles:
+        # We remove small connected components in foreground by changing them to background
+        # with a small negative mask score (-0.1). Here we only remove connected components
+        # whose areas are under both `max_area` and half of the entire mask's area. This
+        # removes sprinkles while avoids filtering out tiny objects that we want to track.
+        mask_fg = mask > 0
+        fg_area_thresh = torch.sum(mask_fg, dim=(2, 3), keepdim=True, dtype=torch.int32)
+        fg_area_thresh.floor_divide_(2).clamp_(max=max_area)
+        _, areas_fg = _get_connected_components_with_padding(mask_fg)
+        small_components_fg = mask_fg & (areas_fg <= fg_area_thresh)
+        mask = torch.where(small_components_fg, -0.1, mask)
+    return mask
+
+
+def _get_connected_components_with_padding(mask):
+    """Get connected components from masks (possibly padding them to an even size)."""
+    from sam3.perflib.connected_components import connected_components
+
+    mask = mask.to(torch.uint8)
+    _, _, H, W = mask.shape
+    # make sure both height and width are even (to be compatible with cc_torch)
+    pad_h = H % 2
+    pad_w = W % 2
+    if pad_h == 0 and pad_w == 0:
+        labels, counts = connected_components(mask)
+    else:
+        # pad the mask to make its height and width even
+        # padding format is (padding_left,padding_right,padding_top,padding_bottom)
+        mask_pad = F.pad(mask, (0, pad_w, 0, pad_h), mode="constant", value=0)
+        labels, counts = connected_components(mask_pad)
+        labels = labels[:, :, :H, :W]
+        counts = counts[:, :, :H, :W]
+
+    return labels, counts

detect_tools/sam3/sam3/model/sam3_tracking_predictor.py

@@ -0,0 +1,1370 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates. All Rights Reserved
+
+import logging
+from collections import OrderedDict
+
+import torch
+
+from sam3.model.sam3_tracker_base import concat_points, NO_OBJ_SCORE, Sam3TrackerBase
+from sam3.model.sam3_tracker_utils import fill_holes_in_mask_scores
+from sam3.model.utils.sam2_utils import load_video_frames
+from tqdm.auto import tqdm
+
+
+class Sam3TrackerPredictor(Sam3TrackerBase):
+    """
+    The demo class that extends the `Sam3TrackerBase` to handle user interactions
+    and manage inference states, with support for multi-object tracking.
+    """
+
+    def __init__(
+        self,
+        # whether to clear non-conditioning memory of the surrounding frames (which may contain outdated information) after adding correction clicks;
+        # note that this would only apply to *single-object tracking* unless `clear_non_cond_mem_for_multi_obj` is also set to True)
+        clear_non_cond_mem_around_input=False,
+        # whether to also clear non-conditioning memory of the surrounding frames (only effective when `clear_non_cond_mem_around_input` is True).
+        clear_non_cond_mem_for_multi_obj=False,
+        # if fill_hole_area > 0, we fill small holes in the final masks up to this area (after resizing them to the original video resolution)
+        fill_hole_area=0,
+        # if always_start_from_first_ann_frame is True, we always start tracking from the frame where we receive the first annotation (clicks or mask)
+        # and ignore the `start_frame_idx` passed to `propagate_in_video`
+        always_start_from_first_ann_frame=False,
+        # the maximum number of points to be used in the prompt encoder, which reduce the domain gap between training (that only has 8 points)
+        # - if it's set to a positive integer, we only take the `max_point_num_in_prompt_enc//2` points and
+        #   the last `(max_point_num_in_prompt_enc - max_point_num_in_prompt_enc//2)` points in the prompt encoder
+        # - if it's set to 0 or negative, this option is turned off and we use all points in the prompt encoder
+        max_point_num_in_prompt_enc=16,
+        non_overlap_masks_for_output=True,
+        # checkpoint_file=None,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+        self.clear_non_cond_mem_around_input = clear_non_cond_mem_around_input
+        self.clear_non_cond_mem_for_multi_obj = clear_non_cond_mem_for_multi_obj
+        self.fill_hole_area = fill_hole_area
+        self.always_start_from_first_ann_frame = always_start_from_first_ann_frame
+        self.max_point_num_in_prompt_enc = max_point_num_in_prompt_enc
+        self.non_overlap_masks_for_output = non_overlap_masks_for_output
+
+        self.bf16_context = torch.autocast(device_type="cuda", dtype=torch.bfloat16)
+        self.bf16_context.__enter__()  # keep using for the entire model process
+
+        self.iter_use_prev_mask_pred = True
+        self.add_all_frames_to_correct_as_cond = True
+
+    @torch.inference_mode()
+    def init_state(
+        self,
+        video_height=None,
+        video_width=None,
+        num_frames=None,
+        video_path=None,
+        cached_features=None,
+        offload_video_to_cpu=False,
+        offload_state_to_cpu=False,
+        async_loading_frames=False,
+    ):
+        """Initialize a inference state."""
+        inference_state = {}
+        # whether to offload the video frames to CPU memory
+        # turning on this option saves the GPU memory with only a very small overhead
+        inference_state["offload_video_to_cpu"] = offload_video_to_cpu
+        # whether to offload the inference state to CPU memory
+        # turning on this option saves the GPU memory at the cost of a lower tracking fps
+        # (e.g. in a test case of 768x768 model, fps dropped from 27 to 24 when tracking one object
+        # and from 24 to 21 when tracking two objects)
+        inference_state["offload_state_to_cpu"] = offload_state_to_cpu
+        inference_state["device"] = self.device
+        if offload_state_to_cpu:
+            inference_state["storage_device"] = torch.device("cpu")
+        else:
+            inference_state["storage_device"] = torch.device("cuda")
+
+        if video_path is not None:
+            images, video_height, video_width = load_video_frames(
+                video_path=video_path,
+                image_size=self.image_size,
+                offload_video_to_cpu=offload_video_to_cpu,
+                async_loading_frames=async_loading_frames,
+                compute_device=inference_state["storage_device"],
+            )
+            inference_state["images"] = images
+            inference_state["num_frames"] = len(images)
+            inference_state["video_height"] = video_height
+            inference_state["video_width"] = video_width
+        else:
+            # the original video height and width, used for resizing final output scores
+            inference_state["video_height"] = video_height
+            inference_state["video_width"] = video_width
+            inference_state["num_frames"] = num_frames
+        # inputs on each frame
+        inference_state["point_inputs_per_obj"] = {}
+        inference_state["mask_inputs_per_obj"] = {}
+        # visual features on a small number of recently visited frames for quick interactions
+        inference_state["cached_features"] = (
+            {} if cached_features is None else cached_features
+        )
+        # values that don't change across frames (so we only need to hold one copy of them)
+        inference_state["constants"] = {}
+        # mapping between client-side object id and model-side object index
+        inference_state["obj_id_to_idx"] = OrderedDict()
+        inference_state["obj_idx_to_id"] = OrderedDict()
+        inference_state["obj_ids"] = []
+        # A storage to hold the model's tracking results and states on each frame
+        inference_state["output_dict"] = {
+            "cond_frame_outputs": {},  # dict containing {frame_idx: <out>}
+            "non_cond_frame_outputs": {},  # dict containing {frame_idx: <out>}
+        }
+        # The index of the frame that received the first annotation
+        inference_state["first_ann_frame_idx"] = None
+        # Slice (view) of each object tracking results, sharing the same memory with "output_dict"
+        inference_state["output_dict_per_obj"] = {}
+        # A temporary storage to hold new outputs when user interact with a frame
+        # to add clicks or mask (it's merged into "output_dict" before propagation starts)
+        inference_state["temp_output_dict_per_obj"] = {}
+        # Frames that already holds consolidated outputs from click or mask inputs
+        # (we directly use their consolidated outputs during tracking)
+        inference_state["consolidated_frame_inds"] = {
+            "cond_frame_outputs": set(),  # set containing frame indices
+            "non_cond_frame_outputs": set(),  # set containing frame indices
+        }
+        # metadata for each tracking frame (e.g. which direction it's tracked)
+        inference_state["tracking_has_started"] = False
+        inference_state["frames_already_tracked"] = {}
+        self.clear_all_points_in_video(inference_state)
+        return inference_state
+
+    def _obj_id_to_idx(self, inference_state, obj_id):
+        """Map client-side object id to model-side object index."""
+        obj_idx = inference_state["obj_id_to_idx"].get(obj_id, None)
+        if obj_idx is not None:
+            return obj_idx
+
+        # This is a new object id not sent to the server before. We only allow adding
+        # new objects *before* the tracking starts.
+        allow_new_object = not inference_state["tracking_has_started"]
+        if allow_new_object:
+            # get the next object slot
+            obj_idx = len(inference_state["obj_id_to_idx"])
+            inference_state["obj_id_to_idx"][obj_id] = obj_idx
+            inference_state["obj_idx_to_id"][obj_idx] = obj_id
+            inference_state["obj_ids"] = list(inference_state["obj_id_to_idx"])
+            # set up input and output structures for this object
+            inference_state["point_inputs_per_obj"][obj_idx] = {}
+            inference_state["mask_inputs_per_obj"][obj_idx] = {}
+            inference_state["output_dict_per_obj"][obj_idx] = {
+                "cond_frame_outputs": {},  # dict containing {frame_idx: <out>}
+                "non_cond_frame_outputs": {},  # dict containing {frame_idx: <out>}
+            }
+            inference_state["temp_output_dict_per_obj"][obj_idx] = {
+                "cond_frame_outputs": {},  # dict containing {frame_idx: <out>}
+                "non_cond_frame_outputs": {},  # dict containing {frame_idx: <out>}
+            }
+            return obj_idx
+        else:
+            raise RuntimeError(
+                f"Cannot add new object id {obj_id} after tracking starts. "
+                f"All existing object ids: {inference_state['obj_ids']}."
+            )
+
+    def _obj_idx_to_id(self, inference_state, obj_idx):
+        """Map model-side object index to client-side object id."""
+        return inference_state["obj_idx_to_id"][obj_idx]
+
+    def _get_obj_num(self, inference_state):
+        """Get the total number of unique object ids received so far in this session."""
+        return len(inference_state["obj_idx_to_id"])
+
+    @torch.inference_mode()
+    def add_new_points_or_box(
+        self,
+        inference_state,
+        frame_idx,
+        obj_id,
+        points=None,
+        labels=None,
+        clear_old_points=True,
+        rel_coordinates=True,
+        use_prev_mem_frame=False,
+        normalize_coords=True,
+        box=None,
+    ):
+        """Add new points to a frame."""
+        obj_idx = self._obj_id_to_idx(inference_state, obj_id)
+        point_inputs_per_frame = inference_state["point_inputs_per_obj"][obj_idx]
+        mask_inputs_per_frame = inference_state["mask_inputs_per_obj"][obj_idx]
+
+        if (points is not None) != (labels is not None):
+            raise ValueError("points and labels must be provided together")
+        if points is None and box is None:
+            raise ValueError("at least one of points or box must be provided as input")
+
+        if points is None:
+            points = torch.zeros(0, 2, dtype=torch.float32)
+        elif not isinstance(points, torch.Tensor):
+            points = torch.tensor(points, dtype=torch.float32)
+        if labels is None:
+            labels = torch.zeros(0, dtype=torch.int32)
+        elif not isinstance(labels, torch.Tensor):
+            labels = torch.tensor(labels, dtype=torch.int32)
+        if points.dim() == 2:
+            points = points.unsqueeze(0)  # add batch dimension
+        if labels.dim() == 1:
+            labels = labels.unsqueeze(0)  # add batch dimension
+
+        if rel_coordinates:
+            # convert the points from relative coordinates to absolute coordinates
+            if points is not None:
+                points = points * self.image_size
+            if box is not None:
+                box = box * self.image_size
+
+        # If `box` is provided, we add it as the first two points with labels 2 and 3
+        # along with the user-provided points (consistent with how SAM 2 is trained).
+        if box is not None:
+            if not clear_old_points:
+                raise ValueError(
+                    "cannot add box without clearing old points, since "
+                    "box prompt must be provided before any point prompt "
+                    "(please use clear_old_points=True instead)"
+                )
+            if not isinstance(box, torch.Tensor):
+                box = torch.tensor(box, dtype=torch.float32, device=points.device)
+            box_coords = box.reshape(1, 2, 2)
+            box_labels = torch.tensor([2, 3], dtype=torch.int32, device=labels.device)
+            box_labels = box_labels.reshape(1, 2)
+            points = torch.cat([box_coords, points], dim=1)
+            labels = torch.cat([box_labels, labels], dim=1)
+
+        points = points.to(inference_state["device"])
+        labels = labels.to(inference_state["device"])
+
+        if not clear_old_points:
+            point_inputs = point_inputs_per_frame.get(frame_idx, None)
+        else:
+            point_inputs = None
+        point_inputs = concat_points(point_inputs, points, labels)
+
+        point_inputs_per_frame[frame_idx] = point_inputs
+        mask_inputs_per_frame.pop(frame_idx, None)
+        # If this frame hasn't been tracked before, we treat it as an initial conditioning
+        # frame, meaning that the inputs points are to generate segments on this frame without
+        # using any memory from other frames, like in SAM. Otherwise (if it has been tracked),
+        # the input points will be used to correct the already tracked masks.
+        is_init_cond_frame = frame_idx not in inference_state["frames_already_tracked"]
+        # whether to track in reverse time order
+        if is_init_cond_frame:
+            reverse = False
+        else:
+            reverse = inference_state["frames_already_tracked"][frame_idx]["reverse"]
+        obj_output_dict = inference_state["output_dict_per_obj"][obj_idx]
+        obj_temp_output_dict = inference_state["temp_output_dict_per_obj"][obj_idx]
+        # Add a frame to conditioning output if it's an initial conditioning frame or
+        # if the model sees all frames receiving clicks/mask as conditioning frames.
+        is_cond = is_init_cond_frame or self.add_all_frames_to_correct_as_cond
+        storage_key = "cond_frame_outputs" if is_cond else "non_cond_frame_outputs"
+
+        # Limit to a maximum number of input points to the prompt encoder (to reduce domain gap)
+        num_points = point_inputs["point_coords"].size(1)
+        if num_points > self.max_point_num_in_prompt_enc > 0:
+            num_first = self.max_point_num_in_prompt_enc // 2
+            num_last = self.max_point_num_in_prompt_enc - num_first
+            point_inputs["point_coords"] = torch.cat(
+                [
+                    point_inputs["point_coords"][:, :num_first],
+                    point_inputs["point_coords"][:, -num_last:],
+                ],
+                dim=1,
+            )
+            point_inputs["point_labels"] = torch.cat(
+                [
+                    point_inputs["point_labels"][:, :num_first],
+                    point_inputs["point_labels"][:, -num_last:],
+                ],
+                dim=1,
+            )
+            logging.warning(
+                f"Too many points ({num_points}) are provided on frame {frame_idx}. Only "
+                f"the first {num_first} points and the last {num_last} points will be used."
+            )
+        # Get any previously predicted mask logits on this object and feed it along with
+        # the new clicks into the SAM mask decoder when `self.iter_use_prev_mask_pred=True`.
+        prev_sam_mask_logits = None
+        if self.iter_use_prev_mask_pred:
+            # lookup temporary output dict first, which contains the most recent output
+            # (if not found, then lookup conditioning and non-conditioning frame output)
+            prev_out = obj_temp_output_dict[storage_key].get(frame_idx)
+            if prev_out is None:
+                prev_out = obj_output_dict["cond_frame_outputs"].get(frame_idx)
+                if prev_out is None:
+                    prev_out = obj_output_dict["non_cond_frame_outputs"].get(frame_idx)
+
+            if prev_out is not None and prev_out["pred_masks"] is not None:
+                prev_sam_mask_logits = prev_out["pred_masks"].cuda(non_blocking=True)
+                # Clamp the scale of prev_sam_mask_logits to avoid rare numerical issues.
+                prev_sam_mask_logits = torch.clamp(prev_sam_mask_logits, -32.0, 32.0)
+        current_out, _ = self._run_single_frame_inference(
+            inference_state=inference_state,
+            output_dict=obj_output_dict,  # run on the slice of a single object
+            frame_idx=frame_idx,
+            batch_size=1,  # run on the slice of a single object
+            is_init_cond_frame=is_init_cond_frame,
+            point_inputs=point_inputs,
+            mask_inputs=None,
+            reverse=reverse,
+            # Skip the memory encoder when adding clicks or mask. We execute the memory encoder
+            # at the beginning of `propagate_in_video` (after user finalize their clicks). This
+            # allows us to enforce non-overlapping constraints on all objects before encoding
+            # them into memory.
+            run_mem_encoder=False,
+            prev_sam_mask_logits=prev_sam_mask_logits,
+            use_prev_mem_frame=use_prev_mem_frame,
+        )
+        # Add the output to the output dict (to be used as future memory)
+        obj_temp_output_dict[storage_key][frame_idx] = current_out
+
+        # Resize the output mask to the original video resolution
+        obj_ids = inference_state["obj_ids"]
+        consolidated_out = self._consolidate_temp_output_across_obj(
+            inference_state,
+            frame_idx,
+            is_cond=is_cond,
+            run_mem_encoder=False,
+            consolidate_at_video_res=True,
+        )
+        _, video_res_masks = self._get_orig_video_res_output(
+            inference_state, consolidated_out["pred_masks_video_res"]
+        )
+        low_res_masks = None  # not needed by the demo
+        return frame_idx, obj_ids, low_res_masks, video_res_masks
+
+    @torch.inference_mode()
+    def add_new_mask(
+        self,
+        inference_state,
+        frame_idx,
+        obj_id,
+        mask,
+        add_mask_to_memory=False,
+    ):
+        """Add new mask to a frame."""
+        obj_idx = self._obj_id_to_idx(inference_state, obj_id)
+        point_inputs_per_frame = inference_state["point_inputs_per_obj"][obj_idx]
+        mask_inputs_per_frame = inference_state["mask_inputs_per_obj"][obj_idx]
+
+        assert mask.dim() == 2
+        mask_H, mask_W = mask.shape
+        mask_inputs_orig = mask[None, None]  # add batch and channel dimension
+        mask_inputs_orig = mask_inputs_orig.float().to(inference_state["device"])
+
+        # resize the mask if it doesn't match the model's input mask size
+        if mask_H != self.input_mask_size or mask_W != self.input_mask_size:
+            mask_inputs = torch.nn.functional.interpolate(
+                mask_inputs_orig,
+                size=(self.input_mask_size, self.input_mask_size),
+                align_corners=False,
+                mode="bilinear",
+                antialias=True,  # use antialias for downsampling
+            )
+        else:
+            mask_inputs = mask_inputs_orig
+
+        # also get the mask at the original video resolution (for outputting)
+        video_H = inference_state["video_height"]
+        video_W = inference_state["video_width"]
+        if mask_H != video_H or mask_W != video_W:
+            mask_inputs_video_res = torch.nn.functional.interpolate(
+                mask_inputs_orig,
+                size=(video_H, video_W),
+                align_corners=False,
+                mode="bilinear",
+                antialias=True,  # use antialias for potential downsampling
+            )
+        else:
+            mask_inputs_video_res = mask_inputs_orig
+        # convert mask_inputs_video_res to binary (threshold at 0.5 as it is in range 0~1)
+        mask_inputs_video_res = mask_inputs_video_res > 0.5
+
+        mask_inputs_per_frame[frame_idx] = mask_inputs_video_res
+        point_inputs_per_frame.pop(frame_idx, None)
+        # If this frame hasn't been tracked before, we treat it as an initial conditioning
+        # frame, meaning that the inputs points are to generate segments on this frame without
+        # using any memory from other frames, like in SAM. Otherwise (if it has been tracked),
+        # the input points will be used to correct the already tracked masks.
+        is_init_cond_frame = frame_idx not in inference_state["frames_already_tracked"]
+        # whether to track in reverse time order
+        if is_init_cond_frame:
+            reverse = False
+        else:
+            reverse = inference_state["frames_already_tracked"][frame_idx]["reverse"]
+        obj_output_dict = inference_state["output_dict_per_obj"][obj_idx]
+        obj_temp_output_dict = inference_state["temp_output_dict_per_obj"][obj_idx]
+        # Add a frame to conditioning output if it's an initial conditioning frame or
+        # if the model sees all frames receiving clicks/mask as conditioning frames.
+        is_cond = is_init_cond_frame or self.add_all_frames_to_correct_as_cond
+        storage_key = "cond_frame_outputs" if is_cond else "non_cond_frame_outputs"
+
+        current_out, _ = self._run_single_frame_inference(
+            inference_state=inference_state,
+            output_dict=obj_output_dict,  # run on the slice of a single object
+            frame_idx=frame_idx,
+            batch_size=1,  # run on the slice of a single object
+            is_init_cond_frame=is_init_cond_frame,
+            point_inputs=None,
+            mask_inputs=mask_inputs,
+            reverse=reverse,
+            # Skip the memory encoder when adding clicks or mask. We execute the memory encoder
+            # at the beginning of `propagate_in_video` (after user finalize their clicks). This
+            # allows us to enforce non-overlapping constraints on all objects before encoding
+            # them into memory.
+            run_mem_encoder=False,
+        )
+        # We directly use the input mask at video resolution as the output mask for a better
+        # video editing experience (so that the masks don't change after each brushing).
+        # Here NO_OBJ_SCORE is a large negative value to represent the background and
+        # similarly -NO_OBJ_SCORE is a large positive value to represent the foreground.
+        current_out["pred_masks"] = None
+        current_out["pred_masks_video_res"] = torch.where(
+            mask_inputs_video_res, -NO_OBJ_SCORE, NO_OBJ_SCORE
+        )
+        # Add the output to the output dict (to be used as future memory)
+        obj_temp_output_dict[storage_key][frame_idx] = current_out
+        # Remove the overlapping proportion of other objects' input masks on this frame
+        temp_output_dict_per_obj = inference_state["temp_output_dict_per_obj"]
+        for obj_idx2, obj_temp_output_dict2 in temp_output_dict_per_obj.items():
+            if obj_idx2 == obj_idx:
+                continue
+            current_out2 = obj_temp_output_dict2[storage_key].get(frame_idx, None)
+            if current_out2 is not None and "pred_masks_video_res" in current_out2:
+                current_out2["pred_masks_video_res"] = torch.where(
+                    mask_inputs_video_res,
+                    NO_OBJ_SCORE,
+                    current_out2["pred_masks_video_res"],
+                )
+
+        # Resize the output mask to the original video resolution
+        obj_ids = inference_state["obj_ids"]
+        consolidated_out = self._consolidate_temp_output_across_obj(
+            inference_state,
+            frame_idx,
+            is_cond=is_cond,
+            run_mem_encoder=False,
+            consolidate_at_video_res=True,
+        )
+        _, video_res_masks = self._get_orig_video_res_output(
+            inference_state, consolidated_out["pred_masks_video_res"]
+        )
+        low_res_masks = None  # not needed by the demo
+        return frame_idx, obj_ids, low_res_masks, video_res_masks
+
+    def add_new_points(self, *args, **kwargs):
+        """Deprecated method. Please use `add_new_points_or_box` instead."""
+        return self.add_new_points_or_box(*args, **kwargs)
+
+    def _get_orig_video_res_output(self, inference_state, any_res_masks):
+        """
+        Resize the object scores to the original video resolution (video_res_masks)
+        and apply non-overlapping constraints for final output.
+        """
+        device = inference_state["device"]
+        video_H = inference_state["video_height"]
+        video_W = inference_state["video_width"]
+        any_res_masks = any_res_masks.to(device, non_blocking=True)
+        if any_res_masks.shape[-2:] == (video_H, video_W):
+            video_res_masks = any_res_masks
+        else:
+            video_res_masks = torch.nn.functional.interpolate(
+                any_res_masks,
+                size=(video_H, video_W),
+                mode="bilinear",
+                align_corners=False,
+            )
+        if self.non_overlap_masks_for_output:
+            video_res_masks = self._apply_non_overlapping_constraints(video_res_masks)
+        # potentially fill holes in the predicted masks
+        if self.fill_hole_area > 0:
+            video_res_masks = fill_holes_in_mask_scores(
+                video_res_masks, self.fill_hole_area
+            )
+        return any_res_masks, video_res_masks
+
+    def _consolidate_temp_output_across_obj(
+        self,
+        inference_state,
+        frame_idx,
+        is_cond,
+        run_mem_encoder,
+        consolidate_at_video_res=False,
+    ):
+        """
+        Consolidate the per-object temporary outputs in `temp_output_dict_per_obj` on
+        a frame into a single output for all objects, including
+        1) fill any missing objects either from `output_dict_per_obj` (if they exist in
+           `output_dict_per_obj` for this frame) or leave them as placeholder values
+           (if they don't exist in `output_dict_per_obj` for this frame);
+        2) if specified, rerun memory encoder after apply non-overlapping constraints
+           on the object scores.
+        """
+        batch_size = self._get_obj_num(inference_state)
+        storage_key = "cond_frame_outputs" if is_cond else "non_cond_frame_outputs"
+        # Optionally, we allow consolidating the temporary outputs at the original
+        # video resolution (to provide a better editing experience for mask prompts).
+        if consolidate_at_video_res:
+            assert not run_mem_encoder, "memory encoder cannot run at video resolution"
+            consolidated_H = inference_state["video_height"]
+            consolidated_W = inference_state["video_width"]
+            consolidated_mask_key = "pred_masks_video_res"
+        else:
+            consolidated_H = consolidated_W = self.low_res_mask_size
+            consolidated_mask_key = "pred_masks"
+
+        # Initialize `consolidated_out`. Its "maskmem_features" and "maskmem_pos_enc"
+        # will be added when rerunning the memory encoder after applying non-overlapping
+        # constraints to object scores. Its "pred_masks" are prefilled with a large
+        # negative value (NO_OBJ_SCORE) to represent missing objects.
+        consolidated_out = {
+            "maskmem_features": None,
+            "maskmem_pos_enc": None,
+            consolidated_mask_key: torch.full(
+                size=(batch_size, 1, consolidated_H, consolidated_W),
+                fill_value=NO_OBJ_SCORE,
+                dtype=torch.float32,
+                device=inference_state["storage_device"],
+            ),
+            "obj_ptr": torch.full(
+                size=(batch_size, self.hidden_dim),
+                fill_value=NO_OBJ_SCORE,
+                dtype=torch.float32,
+                device=inference_state["device"],
+            ),
+            "object_score_logits": torch.full(
+                size=(batch_size, 1),
+                # default to 10.0 for object_score_logits, i.e. assuming the object is
+                # present as sigmoid(10)=1, same as in `predict_masks` of `MaskDecoder`
+                fill_value=10.0,
+                dtype=torch.float32,
+                device=inference_state["device"],
+            ),
+        }
+        if self.use_memory_selection:
+            consolidated_out["iou_score"] = torch.full(
+                size=(batch_size, 1),
+                fill_value=0.0,
+                dtype=torch.float32,
+                device=inference_state["device"],
+            )
+        empty_mask_ptr = None
+        for obj_idx in range(batch_size):
+            obj_temp_output_dict = inference_state["temp_output_dict_per_obj"][obj_idx]
+            obj_output_dict = inference_state["output_dict_per_obj"][obj_idx]
+            out = obj_temp_output_dict[storage_key].get(frame_idx, None)
+            # If the object doesn't appear in "temp_output_dict_per_obj" on this frame,
+            # we fall back and look up its previous output in "output_dict_per_obj".
+            # We look up both "cond_frame_outputs" and "non_cond_frame_outputs" in
+            # "output_dict_per_obj" to find a previous output for this object.
+            if out is None:
+                out = obj_output_dict["cond_frame_outputs"].get(frame_idx, None)
+            if out is None:
+                out = obj_output_dict["non_cond_frame_outputs"].get(frame_idx, None)
+            # If the object doesn't appear in "output_dict_per_obj" either, we skip it
+            # and leave its mask scores to the default scores (i.e. the NO_OBJ_SCORE
+            # placeholder above) and set its object pointer to be a dummy pointer.
+            if out is None:
+                # Fill in dummy object pointers for those objects without any inputs or
+                # tracking outcomes on this frame (only do it under `run_mem_encoder=True`,
+                # i.e. when we need to build the memory for tracking).
+                if run_mem_encoder:
+                    if empty_mask_ptr is None:
+                        empty_mask_ptr = self._get_empty_mask_ptr(
+                            inference_state, frame_idx
+                        )
+                    # fill object pointer with a dummy pointer (based on an empty mask)
+                    consolidated_out["obj_ptr"][obj_idx : obj_idx + 1] = empty_mask_ptr
+                continue
+            # Add the temporary object output mask to consolidated output mask
+            # (use "pred_masks_video_res" if it's available)
+            obj_mask = out.get("pred_masks_video_res", out["pred_masks"])
+            consolidated_pred_masks = consolidated_out[consolidated_mask_key]
+            if obj_mask.shape[-2:] == consolidated_pred_masks.shape[-2:]:
+                consolidated_pred_masks[obj_idx : obj_idx + 1] = obj_mask
+            else:
+                # Resize first if temporary object mask has a different resolution
+                is_downsampling = "pred_masks_video_res" in out
+                resized_obj_mask = torch.nn.functional.interpolate(
+                    obj_mask,
+                    size=consolidated_pred_masks.shape[-2:],
+                    mode="bilinear",
+                    align_corners=False,
+                    antialias=is_downsampling,  # use antialias for downsampling
+                )
+                consolidated_pred_masks[obj_idx : obj_idx + 1] = resized_obj_mask
+            consolidated_out["obj_ptr"][obj_idx : obj_idx + 1] = out["obj_ptr"]
+            consolidated_out["object_score_logits"][obj_idx : obj_idx + 1] = out[
+                "object_score_logits"
+            ]
+            if self.use_memory_selection:
+                consolidated_out["iou_score"][obj_idx : obj_idx + 1] = out["iou_score"]
+        # Optionally, apply non-overlapping constraints on the consolidated scores
+        # and rerun the memory encoder
+        if run_mem_encoder:
+            device = inference_state["device"]
+            high_res_masks = torch.nn.functional.interpolate(
+                consolidated_out["pred_masks"].to(device, non_blocking=True),
+                size=(self.image_size, self.image_size),
+                mode="bilinear",
+                align_corners=False,
+            )
+            high_res_masks = self._apply_non_overlapping_constraints(high_res_masks)
+            maskmem_features, maskmem_pos_enc = self._run_memory_encoder(
+                inference_state=inference_state,
+                frame_idx=frame_idx,
+                batch_size=batch_size,
+                high_res_masks=high_res_masks,
+                object_score_logits=consolidated_out["object_score_logits"],
+                is_mask_from_pts=True,  # these frames are what the user interacted with
+            )
+            consolidated_out["maskmem_features"] = maskmem_features
+            consolidated_out["maskmem_pos_enc"] = maskmem_pos_enc
+
+        return consolidated_out
+
+    def _get_empty_mask_ptr(self, inference_state, frame_idx):
+        """Get a dummy object pointer based on an empty mask on the current frame."""
+        # A dummy (empty) mask with a single object
+        batch_size = 1
+        mask_inputs = torch.zeros(
+            (batch_size, 1, self.image_size, self.image_size),
+            dtype=torch.float32,
+            device=inference_state["device"],
+        )
+
+        # Retrieve correct image features
+        (
+            image,
+            _,
+            current_vision_feats,
+            current_vision_pos_embeds,
+            feat_sizes,
+        ) = self._get_image_feature(inference_state, frame_idx, batch_size)
+
+        # Feed the empty mask and image feature above to get a dummy object pointer
+        current_out = self.track_step(
+            frame_idx=frame_idx,
+            is_init_cond_frame=True,
+            current_vision_feats=current_vision_feats,
+            current_vision_pos_embeds=current_vision_pos_embeds,
+            feat_sizes=feat_sizes,
+            image=image,
+            point_inputs=None,
+            mask_inputs=mask_inputs,
+            gt_masks=None,
+            frames_to_add_correction_pt=[],
+            output_dict={
+                "cond_frame_outputs": {},
+                "non_cond_frame_outputs": {},
+            },
+            num_frames=inference_state["num_frames"],
+            track_in_reverse=False,
+            run_mem_encoder=False,
+            prev_sam_mask_logits=None,
+        )
+        return current_out["obj_ptr"]
+
+    @torch.inference_mode()
+    def propagate_in_video_preflight(self, inference_state, run_mem_encoder=True):
+        """Prepare inference_state and consolidate temporary outputs before tracking."""
+        # Tracking has started and we don't allow adding new objects until session is reset.
+        inference_state["tracking_has_started"] = True
+        batch_size = self._get_obj_num(inference_state)
+
+        # Consolidate per-object temporary outputs in "temp_output_dict_per_obj" and
+        # add them into "output_dict".
+        temp_output_dict_per_obj = inference_state["temp_output_dict_per_obj"]
+        output_dict = inference_state["output_dict"]
+        # "consolidated_frame_inds" contains indices of those frames where consolidated
+        # temporary outputs have been added (either in this call or any previous calls
+        # to `propagate_in_video_preflight`).
+        consolidated_frame_inds = inference_state["consolidated_frame_inds"]
+        for is_cond in [False, True]:
+            # Separately consolidate conditioning and non-conditioning temp outptus
+            storage_key = "cond_frame_outputs" if is_cond else "non_cond_frame_outputs"
+            # Find all the frames that contain temporary outputs for any objects
+            # (these should be the frames that have just received clicks for mask inputs
+            # via `add_new_points` or `add_new_mask`)
+            temp_frame_inds = set()
+            for obj_temp_output_dict in temp_output_dict_per_obj.values():
+                temp_frame_inds.update(obj_temp_output_dict[storage_key].keys())
+            consolidated_frame_inds[storage_key].update(temp_frame_inds)
+            # consolidate the temprary output across all objects on this frame
+            for frame_idx in temp_frame_inds:
+                consolidated_out = self._consolidate_temp_output_across_obj(
+                    inference_state,
+                    frame_idx,
+                    is_cond=is_cond,
+                    run_mem_encoder=run_mem_encoder,
+                )
+                # merge them into "output_dict" and also create per-object slices
+                output_dict[storage_key][frame_idx] = consolidated_out
+                self._add_output_per_object(
+                    inference_state, frame_idx, consolidated_out, storage_key
+                )
+                clear_non_cond_mem = self.clear_non_cond_mem_around_input and (
+                    self.clear_non_cond_mem_for_multi_obj or batch_size <= 1
+                )
+                if clear_non_cond_mem:
+                    # clear non-conditioning memory of the surrounding frames
+                    self._clear_non_cond_mem_around_input(inference_state, frame_idx)
+
+            # clear temporary outputs in `temp_output_dict_per_obj`
+            for obj_temp_output_dict in temp_output_dict_per_obj.values():
+                obj_temp_output_dict[storage_key].clear()
+
+        # edge case: if an output is added to "cond_frame_outputs", we remove any prior
+        # output on the same frame in "non_cond_frame_outputs"
+        for frame_idx in output_dict["cond_frame_outputs"]:
+            output_dict["non_cond_frame_outputs"].pop(frame_idx, None)
+        for obj_output_dict in inference_state["output_dict_per_obj"].values():
+            for frame_idx in obj_output_dict["cond_frame_outputs"]:
+                obj_output_dict["non_cond_frame_outputs"].pop(frame_idx, None)
+        for frame_idx in consolidated_frame_inds["cond_frame_outputs"]:
+            assert frame_idx in output_dict["cond_frame_outputs"]
+            consolidated_frame_inds["non_cond_frame_outputs"].discard(frame_idx)
+
+        # Make sure that the frame indices in "consolidated_frame_inds" are exactly those frames
+        # with either points or mask inputs (which should be true under a correct demo workflow).
+        all_consolidated_frame_inds = (
+            consolidated_frame_inds["cond_frame_outputs"]
+            | consolidated_frame_inds["non_cond_frame_outputs"]
+        )
+        input_frames_inds = set()
+        for point_inputs_per_frame in inference_state["point_inputs_per_obj"].values():
+            input_frames_inds.update(point_inputs_per_frame.keys())
+        for mask_inputs_per_frame in inference_state["mask_inputs_per_obj"].values():
+            input_frames_inds.update(mask_inputs_per_frame.keys())
+        assert all_consolidated_frame_inds == input_frames_inds
+        # Record the first interacted frame index (for tracking start)
+        if inference_state["first_ann_frame_idx"] is None:
+            inference_state["first_ann_frame_idx"] = min(
+                input_frames_inds, default=None
+            )
+        # In case `first_ann_frame_idx` is not in the conditioning frames (e.g. because
+        # we cleared the input points on that frame), pick the first conditioning frame
+        if (
+            inference_state["first_ann_frame_idx"]
+            not in output_dict["cond_frame_outputs"]
+        ):
+            inference_state["first_ann_frame_idx"] = min(
+                output_dict["cond_frame_outputs"], default=None
+            )
+
+    def _get_processing_order(
+        self, inference_state, start_frame_idx, max_frame_num_to_track, reverse
+    ):
+        num_frames = inference_state["num_frames"]
+        # set start index, end index, and processing order
+        if self.always_start_from_first_ann_frame:
+            # in this case, we always start tracking from the frame where we receive
+            # the initial annotation and ignore the provided start_frame_idx
+            start_frame_idx = inference_state["first_ann_frame_idx"]
+        if start_frame_idx is None:
+            # default: start from the earliest frame with input points
+            start_frame_idx = min(inference_state["output_dict"]["cond_frame_outputs"])
+        if max_frame_num_to_track is None:
+            # default: track all the frames in the video
+            max_frame_num_to_track = num_frames
+        if reverse:
+            end_frame_idx = max(start_frame_idx - max_frame_num_to_track, 0)
+            if start_frame_idx > 0:
+                processing_order = range(start_frame_idx, end_frame_idx - 1, -1)
+            else:
+                # this is the edge case where we start from frame 0 and track in reverse order;
+                # in this case, we track a single frame (frame 0)
+                processing_order = [0]
+        else:
+            end_frame_idx = min(
+                start_frame_idx + max_frame_num_to_track, num_frames - 1
+            )
+            processing_order = range(start_frame_idx, end_frame_idx + 1)
+        return processing_order
+
+    @torch.inference_mode()
+    def propagate_in_video(
+        self,
+        inference_state,
+        start_frame_idx,
+        max_frame_num_to_track,
+        reverse,
+        tqdm_disable=False,
+        obj_ids=None,
+        run_mem_encoder=True,
+        propagate_preflight=False,
+    ):
+        """Propagate the input points across frames to track in the entire video."""
+        if propagate_preflight:
+            self.propagate_in_video_preflight(inference_state)
+        # NOTE: This is a copy from the parent class, except that we return object scores as well.
+        output_dict = inference_state["output_dict"]
+        consolidated_frame_inds = inference_state["consolidated_frame_inds"]
+        if obj_ids is not None:
+            raise NotImplementedError(
+                "Per-object tracking yet for batched inference if not implemented."
+            )
+        obj_ids = inference_state["obj_ids"]
+        batch_size = self._get_obj_num(inference_state)
+        if len(output_dict["cond_frame_outputs"]) == 0:
+            raise RuntimeError("No points are provided; please add points first")
+        clear_non_cond_mem = self.clear_non_cond_mem_around_input and (
+            self.clear_non_cond_mem_for_multi_obj or batch_size <= 1
+        )
+
+        processing_order = self._get_processing_order(
+            inference_state,
+            start_frame_idx,
+            max_frame_num_to_track,
+            reverse,
+        )
+
+        for frame_idx in tqdm(
+            processing_order, desc="propagate in video", disable=tqdm_disable
+        ):
+            # We skip those frames already in consolidated outputs (these are frames
+            # that received input clicks or mask). Note that we cannot directly run
+            # batched forward on them via `_run_single_frame_inference` because the
+            # number of clicks on each object might be different.
+            if frame_idx in consolidated_frame_inds["cond_frame_outputs"]:
+                storage_key = "cond_frame_outputs"
+                current_out = output_dict[storage_key][frame_idx]
+                pred_masks = current_out["pred_masks"]
+                obj_scores = current_out["object_score_logits"]
+                if clear_non_cond_mem:
+                    # clear non-conditioning memory of the surrounding frames
+                    self._clear_non_cond_mem_around_input(inference_state, frame_idx)
+            elif frame_idx in consolidated_frame_inds["non_cond_frame_outputs"]:
+                storage_key = "non_cond_frame_outputs"
+                current_out = output_dict[storage_key][frame_idx]
+                pred_masks = current_out["pred_masks"]
+                obj_scores = current_out["object_score_logits"]
+            else:
+                storage_key = "non_cond_frame_outputs"
+                current_out, pred_masks = self._run_single_frame_inference(
+                    inference_state=inference_state,
+                    output_dict=output_dict,
+                    frame_idx=frame_idx,
+                    batch_size=batch_size,
+                    is_init_cond_frame=False,
+                    point_inputs=None,
+                    mask_inputs=None,
+                    reverse=reverse,
+                    run_mem_encoder=run_mem_encoder,
+                )
+                obj_scores = current_out["object_score_logits"]
+                output_dict[storage_key][frame_idx] = current_out
+            # Create slices of per-object outputs for subsequent interaction with each
+            # individual object after tracking.
+            self._add_output_per_object(
+                inference_state, frame_idx, current_out, storage_key
+            )
+            inference_state["frames_already_tracked"][frame_idx] = {"reverse": reverse}
+
+            # Resize the output mask to the original video resolution (we directly use
+            # the mask scores on GPU for output to avoid any CPU conversion in between)
+            low_res_masks, video_res_masks = self._get_orig_video_res_output(
+                inference_state, pred_masks
+            )
+            yield frame_idx, obj_ids, low_res_masks, video_res_masks, obj_scores
+
+    def _add_output_per_object(
+        self, inference_state, frame_idx, current_out, storage_key
+    ):
+        """
+        Split a multi-object output into per-object output slices and add them into
+        `output_dict_per_obj`. The resulting slices share the same tensor storage.
+        """
+        maskmem_features = current_out["maskmem_features"]
+        assert maskmem_features is None or isinstance(maskmem_features, torch.Tensor)
+
+        maskmem_pos_enc = current_out["maskmem_pos_enc"]
+        assert maskmem_pos_enc is None or isinstance(maskmem_pos_enc, list)
+
+        output_dict_per_obj = inference_state["output_dict_per_obj"]
+        for obj_idx, obj_output_dict in output_dict_per_obj.items():
+            obj_slice = slice(obj_idx, obj_idx + 1)
+            obj_out = {
+                "maskmem_features": None,
+                "maskmem_pos_enc": None,
+                "pred_masks": current_out["pred_masks"][obj_slice],
+                "obj_ptr": current_out["obj_ptr"][obj_slice],
+                "object_score_logits": current_out["object_score_logits"][obj_slice],
+            }
+            if self.use_memory_selection:
+                obj_out["iou_score"] = current_out["iou_score"][obj_slice]
+            if maskmem_features is not None:
+                obj_out["maskmem_features"] = maskmem_features[obj_slice]
+            if maskmem_pos_enc is not None:
+                obj_out["maskmem_pos_enc"] = [x[obj_slice] for x in maskmem_pos_enc]
+            obj_output_dict[storage_key][frame_idx] = obj_out
+
+    @torch.inference_mode()
+    def clear_all_points_in_frame(
+        self, inference_state, frame_idx, obj_id, need_output=True
+    ):
+        """Remove all input points or mask in a specific frame for a given object."""
+        obj_idx = self._obj_id_to_idx(inference_state, obj_id)
+
+        # Clear the conditioning information on the given frame
+        inference_state["point_inputs_per_obj"][obj_idx].pop(frame_idx, None)
+        inference_state["mask_inputs_per_obj"][obj_idx].pop(frame_idx, None)
+
+        temp_output_dict_per_obj = inference_state["temp_output_dict_per_obj"]
+        temp_output_dict_per_obj[obj_idx]["cond_frame_outputs"].pop(frame_idx, None)
+        temp_output_dict_per_obj[obj_idx]["non_cond_frame_outputs"].pop(frame_idx, None)
+
+        # Check and see if there are still any inputs left on this frame
+        batch_size = self._get_obj_num(inference_state)
+        frame_has_input = False
+        for obj_idx2 in range(batch_size):
+            if frame_idx in inference_state["point_inputs_per_obj"][obj_idx2]:
+                frame_has_input = True
+                break
+            if frame_idx in inference_state["mask_inputs_per_obj"][obj_idx2]:
+                frame_has_input = True
+                break
+
+        # If this frame has no remaining inputs for any objects, we further clear its
+        # conditioning frame status
+        if not frame_has_input:
+            output_dict = inference_state["output_dict"]
+            consolidated_frame_inds = inference_state["consolidated_frame_inds"]
+            consolidated_frame_inds["cond_frame_outputs"].discard(frame_idx)
+            consolidated_frame_inds["non_cond_frame_outputs"].discard(frame_idx)
+            # Remove the frame's conditioning output (possibly downgrading it to non-conditioning)
+            out = output_dict["cond_frame_outputs"].pop(frame_idx, None)
+            if out is not None:
+                # The frame is not a conditioning frame anymore since it's not receiving inputs,
+                # so we "downgrade" its output (if exists) to a non-conditioning frame output.
+                output_dict["non_cond_frame_outputs"][frame_idx] = out
+                inference_state["frames_already_tracked"].pop(frame_idx, None)
+            # Similarly, do it for the sliced output on each object.
+            for obj_idx2 in range(batch_size):
+                obj_output_dict = inference_state["output_dict_per_obj"][obj_idx2]
+                obj_out = obj_output_dict["cond_frame_outputs"].pop(frame_idx, None)
+                if obj_out is not None:
+                    obj_output_dict["non_cond_frame_outputs"][frame_idx] = obj_out
+
+            # If all the conditioning frames have been removed, we also clear the tracking outputs
+            if len(output_dict["cond_frame_outputs"]) == 0:
+                self._reset_tracking_results(inference_state)
+
+        if not need_output:
+            return
+        # Finally, output updated masks per object (after removing the inputs above)
+        obj_ids = inference_state["obj_ids"]
+        is_cond = any(
+            frame_idx in obj_temp_output_dict["cond_frame_outputs"]
+            for obj_temp_output_dict in temp_output_dict_per_obj.values()
+        )
+        consolidated_out = self._consolidate_temp_output_across_obj(
+            inference_state,
+            frame_idx,
+            is_cond=is_cond,
+            run_mem_encoder=False,
+            consolidate_at_video_res=True,
+        )
+        _, video_res_masks = self._get_orig_video_res_output(
+            inference_state, consolidated_out["pred_masks_video_res"]
+        )
+        low_res_masks = None  # not needed by the demo
+        return frame_idx, obj_ids, low_res_masks, video_res_masks
+
+    @torch.inference_mode()
+    def clear_all_points_in_video(self, inference_state):
+        """Remove all input points or mask in all frames throughout the video."""
+        self._reset_tracking_results(inference_state)
+        # Remove all object ids
+        inference_state["obj_id_to_idx"].clear()
+        inference_state["obj_idx_to_id"].clear()
+        inference_state["obj_ids"].clear()
+        inference_state["point_inputs_per_obj"].clear()
+        inference_state["mask_inputs_per_obj"].clear()
+        inference_state["output_dict_per_obj"].clear()
+        inference_state["temp_output_dict_per_obj"].clear()
+
+    def _reset_tracking_results(self, inference_state):
+        """Reset all tracking inputs and results across the videos."""
+        for v in inference_state["point_inputs_per_obj"].values():
+            v.clear()
+        for v in inference_state["mask_inputs_per_obj"].values():
+            v.clear()
+        for v in inference_state["output_dict_per_obj"].values():
+            v["cond_frame_outputs"].clear()
+            v["non_cond_frame_outputs"].clear()
+        for v in inference_state["temp_output_dict_per_obj"].values():
+            v["cond_frame_outputs"].clear()
+            v["non_cond_frame_outputs"].clear()
+        inference_state["output_dict"]["cond_frame_outputs"].clear()
+        inference_state["output_dict"]["non_cond_frame_outputs"].clear()
+        inference_state["consolidated_frame_inds"]["cond_frame_outputs"].clear()
+        inference_state["consolidated_frame_inds"]["non_cond_frame_outputs"].clear()
+        inference_state["tracking_has_started"] = False
+        inference_state["frames_already_tracked"].clear()
+        inference_state["first_ann_frame_idx"] = None
+
+    def _get_image_feature(self, inference_state, frame_idx, batch_size):
+        """Compute the image features on a given frame."""
+        # Look up in the cache
+        image, backbone_out = inference_state["cached_features"].get(
+            frame_idx, (None, None)
+        )
+        if backbone_out is None:
+            if self.backbone is None:
+                raise RuntimeError(
+                    f"Image features for frame {frame_idx} are not cached. "
+                    "Please run inference on this frame first."
+                )
+            else:
+                # Cache miss -- we will run inference on a single image
+                image = inference_state["images"][frame_idx].cuda().float().unsqueeze(0)
+                backbone_out = self.forward_image(image)
+                # Cache the most recent frame's feature (for repeated interactions with
+                # a frame; we can use an LRU cache for more frames in the future).
+                inference_state["cached_features"] = {frame_idx: (image, backbone_out)}
+        if "tracker_backbone_out" in backbone_out:
+            backbone_out = backbone_out["tracker_backbone_out"]  # get backbone output
+
+        # expand the features to have the same dimension as the number of objects
+        expanded_image = image.expand(batch_size, -1, -1, -1)
+        expanded_backbone_out = {
+            "backbone_fpn": backbone_out["backbone_fpn"].copy(),
+            "vision_pos_enc": backbone_out["vision_pos_enc"].copy(),
+        }
+        for i, feat in enumerate(expanded_backbone_out["backbone_fpn"]):
+            feat = feat.expand(batch_size, -1, -1, -1)
+            expanded_backbone_out["backbone_fpn"][i] = feat
+        for i, pos in enumerate(expanded_backbone_out["vision_pos_enc"]):
+            pos = pos.expand(batch_size, -1, -1, -1)
+            expanded_backbone_out["vision_pos_enc"][i] = pos
+
+        features = self._prepare_backbone_features(expanded_backbone_out)
+        features = (expanded_image,) + features
+        return features
+
+    def _run_single_frame_inference(
+        self,
+        inference_state,
+        output_dict,
+        frame_idx,
+        batch_size,
+        is_init_cond_frame,
+        point_inputs,
+        mask_inputs,
+        reverse,
+        run_mem_encoder,
+        prev_sam_mask_logits=None,
+        use_prev_mem_frame=True,
+    ):
+        """Run tracking on a single frame based on current inputs and previous memory."""
+        # Retrieve correct image features
+        (
+            image,
+            _,
+            current_vision_feats,
+            current_vision_pos_embeds,
+            feat_sizes,
+        ) = self._get_image_feature(inference_state, frame_idx, batch_size)
+
+        # point and mask should not appear as input simultaneously on the same frame
+        assert point_inputs is None or mask_inputs is None
+        current_out = self.track_step(
+            frame_idx=frame_idx,
+            is_init_cond_frame=is_init_cond_frame,
+            current_vision_feats=current_vision_feats,
+            current_vision_pos_embeds=current_vision_pos_embeds,
+            feat_sizes=feat_sizes,
+            image=image,
+            point_inputs=point_inputs,
+            mask_inputs=mask_inputs,
+            output_dict=output_dict,
+            num_frames=inference_state["num_frames"],
+            track_in_reverse=reverse,
+            run_mem_encoder=run_mem_encoder,
+            prev_sam_mask_logits=prev_sam_mask_logits,
+            use_prev_mem_frame=use_prev_mem_frame,
+        )
+
+        # optionally offload the output to CPU memory to save GPU space
+        storage_device = inference_state["storage_device"]
+        maskmem_features = current_out["maskmem_features"]
+        if maskmem_features is not None:
+            maskmem_features = maskmem_features.to(torch.bfloat16)
+            maskmem_features = maskmem_features.to(storage_device, non_blocking=True)
+        pred_masks_gpu = current_out["pred_masks"]
+        pred_masks = pred_masks_gpu.to(storage_device, non_blocking=True)
+        # "maskmem_pos_enc" is the same across frames, so we only need to store one copy of it
+        maskmem_pos_enc = self._get_maskmem_pos_enc(inference_state, current_out)
+        # object pointer is a small tensor, so we always keep it on GPU memory for fast access
+        obj_ptr = current_out["obj_ptr"]
+        object_score_logits = current_out["object_score_logits"]
+        # make a compact version of this frame's output to reduce the state size
+        compact_current_out = {
+            "maskmem_features": maskmem_features,
+            "maskmem_pos_enc": maskmem_pos_enc,
+            "pred_masks": pred_masks,
+            "obj_ptr": obj_ptr,
+            "object_score_logits": object_score_logits,
+        }
+        if self.use_memory_selection:
+            compact_current_out["iou_score"] = current_out["iou_score"]
+            compact_current_out["eff_iou_score"] = current_out["eff_iou_score"]
+        return compact_current_out, pred_masks_gpu
+
+    def _run_memory_encoder(
+        self,
+        inference_state,
+        frame_idx,
+        batch_size,
+        high_res_masks,
+        object_score_logits,
+        is_mask_from_pts,
+    ):
+        """
+        Run the memory encoder on `high_res_masks`. This is usually after applying
+        non-overlapping constraints to object scores. Since their scores changed, their
+        memory also need to be computed again with the memory encoder.
+        """
+        # Retrieve correct image features
+        image, _, current_vision_feats, _, feat_sizes = self._get_image_feature(
+            inference_state, frame_idx, batch_size
+        )
+        maskmem_features, maskmem_pos_enc = self._encode_new_memory(
+            image=image,
+            current_vision_feats=current_vision_feats,
+            feat_sizes=feat_sizes,
+            pred_masks_high_res=high_res_masks,
+            object_score_logits=object_score_logits,
+            is_mask_from_pts=is_mask_from_pts,
+        )
+
+        # optionally offload the output to CPU memory to save GPU space
+        storage_device = inference_state["storage_device"]
+        maskmem_features = maskmem_features.to(torch.bfloat16)
+        maskmem_features = maskmem_features.to(storage_device, non_blocking=True)
+        # "maskmem_pos_enc" is the same across frames, so we only need to store one copy of it
+        maskmem_pos_enc = self._get_maskmem_pos_enc(
+            inference_state, {"maskmem_pos_enc": maskmem_pos_enc}
+        )
+        return maskmem_features, maskmem_pos_enc
+
+    def _get_maskmem_pos_enc(self, inference_state, current_out):
+        """
+        `maskmem_pos_enc` is the same across frames and objects, so we cache it as
+        a constant in the inference session to reduce session storage size.
+        """
+        model_constants = inference_state["constants"]
+        # "out_maskmem_pos_enc" should be either a list of tensors or None
+        out_maskmem_pos_enc = current_out["maskmem_pos_enc"]
+        if out_maskmem_pos_enc is not None:
+            if "maskmem_pos_enc" not in model_constants:
+                assert isinstance(out_maskmem_pos_enc, list)
+                # only take the slice for one object, since it's same across objects
+                maskmem_pos_enc = [x[0:1].clone() for x in out_maskmem_pos_enc]
+                model_constants["maskmem_pos_enc"] = maskmem_pos_enc
+            else:
+                maskmem_pos_enc = model_constants["maskmem_pos_enc"]
+            # expand the cached maskmem_pos_enc to the actual batch size
+            batch_size = out_maskmem_pos_enc[0].size(0)
+            expanded_maskmem_pos_enc = [
+                x.expand(batch_size, -1, -1, -1) for x in maskmem_pos_enc
+            ]
+        else:
+            expanded_maskmem_pos_enc = None
+        return expanded_maskmem_pos_enc
+
+    @torch.inference_mode()
+    def remove_object(self, inference_state, obj_id, strict=False, need_output=True):
+        """
+        Remove an object id from the tracking state. If strict is True, we check whether
+        the object id actually exists and raise an error if it doesn't exist.
+        """
+        old_obj_idx_to_rm = inference_state["obj_id_to_idx"].get(obj_id, None)
+        updated_frames = []
+        # Check whether this object_id to remove actually exists and possibly raise an error.
+        if old_obj_idx_to_rm is None:
+            if not strict:
+                return inference_state["obj_ids"], updated_frames
+            raise RuntimeError(
+                f"Cannot remove object id {obj_id} as it doesn't exist. "
+                f"All existing object ids: {inference_state['obj_ids']}."
+            )
+
+        # If this is the only remaining object id, we simply reset the state.
+        if len(inference_state["obj_id_to_idx"]) == 1:
+            self.clear_all_points_in_video(inference_state)
+            return inference_state["obj_ids"], updated_frames
+
+        # There are still remaining objects after removing this object id. In this case,
+        # we need to delete the object storage from inference state tensors.
+        # Step 0: clear the input on those frames where this object id has point or mask input
+        # (note that this step is required as it might downgrade conditioning frames to
+        # non-conditioning ones)
+        obj_input_frames_inds = set()
+        obj_input_frames_inds.update(
+            inference_state["point_inputs_per_obj"][old_obj_idx_to_rm]
+        )
+        obj_input_frames_inds.update(
+            inference_state["mask_inputs_per_obj"][old_obj_idx_to_rm]
+        )
+        for frame_idx in obj_input_frames_inds:
+            self.clear_all_points_in_frame(
+                inference_state, frame_idx, obj_id, need_output=False
+            )
+
+        # Step 1: Update the object id mapping (note that it must be done after Step 0,
+        # since Step 0 still requires the old object id mappings in inference_state)
+        old_obj_ids = inference_state["obj_ids"]
+        old_obj_inds = list(range(len(old_obj_ids)))
+        remain_old_obj_inds = old_obj_inds.copy()
+        remain_old_obj_inds.remove(old_obj_idx_to_rm)
+        new_obj_ids = [old_obj_ids[old_idx] for old_idx in remain_old_obj_inds]
+        new_obj_inds = list(range(len(new_obj_ids)))
+        # build new mappings
+        old_idx_to_new_idx = dict(zip(remain_old_obj_inds, new_obj_inds))
+        inference_state["obj_id_to_idx"] = dict(zip(new_obj_ids, new_obj_inds))
+        inference_state["obj_idx_to_id"] = dict(zip(new_obj_inds, new_obj_ids))
+        inference_state["obj_ids"] = new_obj_ids
+
+        # Step 2: For per-object tensor storage, we shift their obj_idx in the dict keys.
+        # (note that "consolidated_frame_inds" doesn't need to be updated in this step as
+        # it's already handled in Step 0)
+        def _map_keys(container):
+            new_kvs = []
+            for k in old_obj_inds:
+                v = container.pop(k)
+                if k in old_idx_to_new_idx:
+                    new_kvs.append((old_idx_to_new_idx[k], v))
+            container.update(new_kvs)
+
+        _map_keys(inference_state["point_inputs_per_obj"])
+        _map_keys(inference_state["mask_inputs_per_obj"])
+        _map_keys(inference_state["output_dict_per_obj"])
+        _map_keys(inference_state["temp_output_dict_per_obj"])
+
+        # Step 3: For packed tensor storage, we index the remaining ids and rebuild the per-object slices.
+        def _slice_state(output_dict, storage_key):
+            for frame_idx, out in output_dict[storage_key].items():
+                out["maskmem_features"] = out["maskmem_features"][remain_old_obj_inds]
+                out["maskmem_pos_enc"] = [
+                    x[remain_old_obj_inds] for x in out["maskmem_pos_enc"]
+                ]
+                # "maskmem_pos_enc" is the same across frames, so we only need to store one copy of it
+                out["maskmem_pos_enc"] = self._get_maskmem_pos_enc(inference_state, out)
+                out["pred_masks"] = out["pred_masks"][remain_old_obj_inds]
+                out["obj_ptr"] = out["obj_ptr"][remain_old_obj_inds]
+                out["object_score_logits"] = out["object_score_logits"][
+                    remain_old_obj_inds
+                ]
+                if self.use_memory_selection:
+                    out["iou_score"] = out["iou_score"][remain_old_obj_inds]
+                    out["eff_iou_score"] = self.cal_mem_score(
+                        out["object_score_logits"], out["iou_score"]
+                    )  # recalculate the memory frame score
+                # also update the per-object slices
+                self._add_output_per_object(
+                    inference_state, frame_idx, out, storage_key
+                )
+
+        _slice_state(inference_state["output_dict"], "cond_frame_outputs")
+        _slice_state(inference_state["output_dict"], "non_cond_frame_outputs")
+
+        # Step 4: Further collect the outputs on those frames in `obj_input_frames_inds`, which
+        # could show an updated mask for objects previously occluded by the object being removed
+        if need_output:
+            temp_output_dict_per_obj = inference_state["temp_output_dict_per_obj"]
+            for frame_idx in obj_input_frames_inds:
+                is_cond = any(
+                    frame_idx in obj_temp_output_dict["cond_frame_outputs"]
+                    for obj_temp_output_dict in temp_output_dict_per_obj.values()
+                )
+                consolidated_out = self._consolidate_temp_output_across_obj(
+                    inference_state,
+                    frame_idx,
+                    is_cond=is_cond,
+                    run_mem_encoder=False,
+                    consolidate_at_video_res=True,
+                )
+                _, video_res_masks = self._get_orig_video_res_output(
+                    inference_state, consolidated_out["pred_masks_video_res"]
+                )
+                updated_frames.append((frame_idx, video_res_masks))
+
+        return inference_state["obj_ids"], updated_frames
+
+    def _clear_non_cond_mem_around_input(self, inference_state, frame_idx):
+        """
+        Remove the non-conditioning memory around the input frame. When users provide
+        correction clicks, the surrounding frames' non-conditioning memories can still
+        contain outdated object appearance information and could confuse the model.
+
+        This method clears those non-conditioning memories surrounding the interacted
+        frame to avoid giving the model both old and new information about the object.
+        """
+        r = self.memory_temporal_stride_for_eval
+        frame_idx_begin = frame_idx - r * self.num_maskmem
+        frame_idx_end = frame_idx + r * self.num_maskmem
+        batch_size = self._get_obj_num(inference_state)
+        for obj_idx in range(batch_size):
+            obj_output_dict = inference_state["output_dict_per_obj"][obj_idx]
+            non_cond_frame_outputs = obj_output_dict["non_cond_frame_outputs"]
+            for t in range(frame_idx_begin, frame_idx_end + 1):
+                non_cond_frame_outputs.pop(t, None)
+
+    def _suppress_shrinked_masks(
+        self, pred_masks, new_pred_masks, shrink_threshold=0.3
+    ):
+        area_before = (pred_masks > 0).sum(dim=(-1, -2))
+        area_after = (new_pred_masks > 0).sum(dim=(-1, -2))
+        area_before = torch.clamp(area_before, min=1.0)
+        area_ratio = area_after / area_before
+        keep = area_ratio >= shrink_threshold
+        keep_mask = keep[..., None, None].expand_as(pred_masks)
+        pred_masks_after = torch.where(
+            keep_mask, pred_masks, torch.clamp(pred_masks, max=-10.0)
+        )
+        return pred_masks_after
+
+    def _suppress_object_pw_area_shrinkage(self, pred_masks):
+        """
+        This function suppresses masks that shrink in area after applying pixelwise non-overlapping constriants.
+        Note that the final output can still be overlapping.
+        """
+        # Apply pixel-wise non-overlapping constraint based on mask scores
+        pixel_level_non_overlapping_masks = super()._apply_non_overlapping_constraints(
+            pred_masks
+        )
+        # Fully suppress masks with high shrinkage (probably noisy) based on the pixel wise non-overlapping constraints
+        # NOTE: The output of this function can be a no op if none of the masks shrinked by a large factor.
+        pred_masks = self._suppress_shrinked_masks(
+            pred_masks, pixel_level_non_overlapping_masks
+        )
+        return pred_masks
+
+    def _apply_object_wise_non_overlapping_constraints(
+        self, pred_masks, obj_scores, background_value=-10.0
+    ):
+        """
+        Applies non-overlapping constraints object wise (i.e. only one object can claim the overlapping region)
+        """
+        # Replace pixel scores with object scores
+        pred_masks_single_score = torch.where(
+            pred_masks > 0, obj_scores[..., None, None], background_value
+        )
+        # Apply pixel-wise non-overlapping constraint based on mask scores
+        pixel_level_non_overlapping_masks = super()._apply_non_overlapping_constraints(
+            pred_masks_single_score
+        )
+        # Replace object scores with pixel scores. Note, that now only one object can claim the overlapping region
+        pred_masks = torch.where(
+            pixel_level_non_overlapping_masks > 0,
+            pred_masks,
+            torch.clamp(pred_masks, max=background_value),
+        )
+        return pred_masks

detect_tools/sam3/sam3/model/sam3_video_base.py

@@ -0,0 +1,1767 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates. All Rights Reserved
+
+import datetime
+import logging
+import math
+import os
+from collections import defaultdict
+from copy import deepcopy
+from enum import Enum
+from typing import Any, Dict, List, Set
+
+import numpy as np
+import numpy.typing as npt
+import torch
+import torch.distributed as dist
+import torch.nn.functional as F
+
+from sam3 import perflib
+from sam3.logger import get_logger
+from sam3.model.box_ops import fast_diag_box_iou
+from sam3.model.data_misc import BatchedDatapoint
+from sam3.model.sam3_tracker_utils import fill_holes_in_mask_scores, mask_to_box
+from sam3.perflib.masks_ops import mask_iou
+from sam3.train.masks_ops import rle_encode
+from torch import nn, Tensor
+
+logger = get_logger(__name__)
+
+
+class MaskletConfirmationStatus(Enum):
+    UNCONFIRMED = 1  # newly added masklet, not confirmed by any detection yet
+    CONFIRMED = 2  # confirmed by at least one detection
+
+
+class Sam3VideoBase(nn.Module):
+    def __init__(
+        self,
+        detector: nn.Module,
+        tracker: nn.Module,
+        # prob threshold for detection outputs -- only keep detections above this threshold
+        # enters NMS and det-to-track matching
+        score_threshold_detection=0.5,
+        # IoU threshold for detection NMS
+        det_nms_thresh=0.0,
+        # IoU threshold for det-to-track matching -- a detection is considered "matched" to a tracklet it
+        # overlaps with a tracklet above this threshold -- it is often a loose threshold like 0.1
+        assoc_iou_thresh=0.5,
+        # IoU threshold for det-to-track matching, which is used to determine whether a masklet is "unmatched"
+        # by any detections -- it is often a stricter threshold like 0.5
+        trk_assoc_iou_thresh=0.5,
+        # prob threshold for a detection to be added as a new object
+        new_det_thresh=0.0,
+        # hotstart parameters: we hold off the outputs for `hotstart_delay` frames and
+        # 1) remove those tracklets unmatched by any detections based on `hotstart_unmatch_thresh`
+        # 2) remove those tracklets overlapping with one another based on `hotstart_dup_thresh`
+        hotstart_delay=0,
+        hotstart_unmatch_thresh=3,
+        hotstart_dup_thresh=3,
+        # Whether to suppress masks only within hotstart. If False, we can suppress masks even if they start before hotstart period.
+        suppress_unmatched_only_within_hotstart=True,
+        init_trk_keep_alive=0,
+        max_trk_keep_alive=8,
+        min_trk_keep_alive=-4,
+        # Threshold for suppressing overlapping objects based on recent occlusion
+        suppress_overlapping_based_on_recent_occlusion_threshold=0.0,
+        decrease_trk_keep_alive_for_empty_masklets=False,
+        o2o_matching_masklets_enable=False,  # Enable hungarian matching to match existing masklets
+        suppress_det_close_to_boundary=False,
+        fill_hole_area=16,
+        # The maximum number of objects (masklets) to track across all GPUs (for no limit, set it to -1)
+        max_num_objects=-1,
+        recondition_every_nth_frame=-1,
+        # masket confirmation status (to suppress unconfirmed masklets)
+        masklet_confirmation_enable=False,
+        # a masklet is confirmed after being consecutively detected and matched for
+        # `masklet_confirmation_consecutive_det_thresh`
+        masklet_confirmation_consecutive_det_thresh=3,
+        # bbox heuristic parameters
+        reconstruction_bbox_iou_thresh=0.0,
+        reconstruction_bbox_det_score=0.0,
+    ):
+        super().__init__()
+        self.detector = detector
+        self.tracker = tracker
+        self.score_threshold_detection = score_threshold_detection
+        self.det_nms_thresh = det_nms_thresh
+        self.assoc_iou_thresh = assoc_iou_thresh
+        self.trk_assoc_iou_thresh = trk_assoc_iou_thresh
+        self.new_det_thresh = new_det_thresh
+
+        # hotstart parameters
+        if hotstart_delay > 0:
+            assert hotstart_unmatch_thresh <= hotstart_delay
+            assert hotstart_dup_thresh <= hotstart_delay
+        self.hotstart_delay = hotstart_delay
+        self.hotstart_unmatch_thresh = hotstart_unmatch_thresh
+        self.hotstart_dup_thresh = hotstart_dup_thresh
+        self.suppress_unmatched_only_within_hotstart = (
+            suppress_unmatched_only_within_hotstart
+        )
+        self.init_trk_keep_alive = init_trk_keep_alive
+        self.max_trk_keep_alive = max_trk_keep_alive
+        self.min_trk_keep_alive = min_trk_keep_alive
+        self.suppress_overlapping_based_on_recent_occlusion_threshold = (
+            suppress_overlapping_based_on_recent_occlusion_threshold
+        )
+        self.suppress_det_close_to_boundary = suppress_det_close_to_boundary
+        self.decrease_trk_keep_alive_for_empty_masklets = (
+            decrease_trk_keep_alive_for_empty_masklets
+        )
+        self.o2o_matching_masklets_enable = o2o_matching_masklets_enable
+        self.fill_hole_area = fill_hole_area
+        self.eval()
+        self.rank = int(os.getenv("RANK", "0"))
+        self.world_size = int(os.getenv("WORLD_SIZE", "1"))
+        self._dist_pg_cpu = None  # CPU process group (lazy-initialized on first use)
+
+        # the maximum object number
+        if max_num_objects > 0:
+            num_obj_for_compile = math.ceil(max_num_objects / self.world_size)
+        else:
+            max_num_objects = 10000  # no limit
+            num_obj_for_compile = 16
+        logger.info(f"setting {max_num_objects=} and {num_obj_for_compile=}")
+        self.max_num_objects = max_num_objects
+        self.num_obj_for_compile = num_obj_for_compile
+        self.recondition_every_nth_frame = recondition_every_nth_frame
+        self.masklet_confirmation_enable = masklet_confirmation_enable
+        self.masklet_confirmation_consecutive_det_thresh = (
+            masklet_confirmation_consecutive_det_thresh
+        )
+        self.reconstruction_bbox_iou_thresh = reconstruction_bbox_iou_thresh
+        self.reconstruction_bbox_det_score = reconstruction_bbox_det_score
+
+    @property
+    def device(self):
+        self._device = getattr(self, "_device", None) or next(self.parameters()).device
+        return self._device
+
+    def _init_dist_pg_cpu(self):
+        # a short 3-min timeout to quickly detect any synchronization failures
+        timeout_sec = int(os.getenv("SAM3_COLLECTIVE_OP_TIMEOUT_SEC", "180"))
+        timeout = datetime.timedelta(seconds=timeout_sec)
+        self._dist_pg_cpu = dist.new_group(backend="gloo", timeout=timeout)
+
+    def broadcast_python_obj_cpu(self, python_obj_list, src):
+        if self._dist_pg_cpu is None:
+            self._init_dist_pg_cpu()
+        dist.broadcast_object_list(python_obj_list, src=src, group=self._dist_pg_cpu)
+
+    def _det_track_one_frame(
+        self,
+        frame_idx: int,
+        num_frames: int,
+        reverse: bool,
+        input_batch: BatchedDatapoint,
+        geometric_prompt: Any,
+        tracker_states_local: List[Any],
+        tracker_metadata_prev: Dict[str, Any],
+        feature_cache: Dict,
+        orig_vid_height: int,
+        orig_vid_width: int,
+        is_image_only: bool = False,
+        allow_new_detections: bool = True,
+    ):
+        """
+        This function handles one-step inference for the DenseTracking model in an SPMD manner.
+        At a high-level, all GPUs execute the same function calls as if it's done on a single GPU,
+        while under the hood, some function calls involve distributed computation based on sharded
+        SAM2 states.
+
+        - `input_batch` contains image and other inputs on the entire video; it should be identical across GPUs
+        - `tracker_states_local` holds the local masklet information in this GPU shard
+        - `tracker_metadata_prev` manages the metadata for SAM2 objects, such as which masklet is hold on which GPUs
+          it contains both global and local masklet information
+        """
+
+        # Step 1: run backbone and detector in a distributed manner -- this is done via Sam3ImageOnVideoMultiGPU,
+        # a MultiGPU model (assigned to `self.detector`) that shards frames in a round-robin manner.
+        # It returns a "det_out" dict for `frame_idx` and fills SAM2 backbone features for `frame_idx`
+        # into `feature_cache`. Despite its distributed inference under the hood, the results would be
+        # the same as if it is running backbone and detector for every frame on a single GPU.
+        det_out = self.run_backbone_and_detection(
+            frame_idx=frame_idx,
+            num_frames=num_frames,
+            reverse=reverse,
+            input_batch=input_batch,
+            geometric_prompt=geometric_prompt,
+            feature_cache=feature_cache,
+            allow_new_detections=allow_new_detections,
+        )
+
+        # Step 2: each GPU propagates its local SAM2 states to get the SAM2 prediction masks.
+        # the returned `tracker_low_res_masks_global` contains the concatenated masklet predictions
+        # gathered from all GPUs (as if they are propagated on a single GPU). Note that this step only
+        # runs the SAM2 propagation step, but doesn't encode new memory for the predicted masks;
+        # we defer memory encoding to `run_tracker_update_execution_phase` after resolving all heuristics.
+        if tracker_metadata_prev == {}:
+            # initialize masklet metadata if it's uninitialized (empty dict)
+            tracker_metadata_prev.update(self._initialize_metadata())
+        tracker_low_res_masks_global, tracker_obj_scores_global = (
+            self.run_tracker_propagation(
+                frame_idx=frame_idx,
+                num_frames=num_frames,
+                reverse=reverse,
+                tracker_states_local=tracker_states_local,
+                tracker_metadata_prev=tracker_metadata_prev,
+            )
+        )
+
+        # Step 3: based on detection outputs and the propagated SAM2 prediction masks, we make plans
+        # for SAM2 masklet updates (i.e. which objects to add and remove, how to load-balance them, etc).
+        # We also run SAM2 memory encoder globally in this step to resolve non-overlapping constraints.
+        # **This step should involve all the heuristics needed for any updates.** Most of the update
+        # planning will be done on the master rank (GPU 0) and the resulting plan `tracker_update_plan` is
+        # broadcasted to other GPUs (to be executed in a distributed manner). This step also generates the
+        # new masklet metadata `tracker_metadata_new` (based on its previous version `tracker_metadata_prev`).
+        tracker_update_plan, tracker_metadata_new = (
+            self.run_tracker_update_planning_phase(
+                frame_idx=frame_idx,
+                num_frames=num_frames,
+                reverse=reverse,
+                det_out=det_out,
+                tracker_low_res_masks_global=tracker_low_res_masks_global,
+                tracker_obj_scores_global=tracker_obj_scores_global,
+                tracker_metadata_prev=tracker_metadata_prev,
+                tracker_states_local=tracker_states_local,
+                is_image_only=is_image_only,
+            )
+        )
+
+        # Get reconditioning info from the update plan
+        reconditioned_obj_ids = tracker_update_plan.get("reconditioned_obj_ids", set())
+        det_to_matched_trk_obj_ids = tracker_update_plan.get(
+            "det_to_matched_trk_obj_ids", {}
+        )
+
+        # Step 4: based on `tracker_update_plan`, each GPU executes the update w.r.t. its local SAM2 inference states
+        tracker_states_local_new = self.run_tracker_update_execution_phase(
+            frame_idx=frame_idx,
+            num_frames=num_frames,
+            reverse=reverse,
+            det_out=det_out,
+            tracker_states_local=tracker_states_local,
+            tracker_update_plan=tracker_update_plan,
+            orig_vid_height=orig_vid_height,
+            orig_vid_width=orig_vid_width,
+            feature_cache=feature_cache,
+        )
+
+        # Step 5: finally, build the outputs for this frame (it only needs to be done on GPU 0 since
+        # only GPU 0 will send outputs to the server).
+        if self.rank == 0:
+            obj_id_to_mask = self.build_outputs(
+                frame_idx=frame_idx,
+                num_frames=num_frames,
+                reverse=reverse,
+                det_out=det_out,
+                tracker_low_res_masks_global=tracker_low_res_masks_global,
+                tracker_obj_scores_global=tracker_obj_scores_global,
+                tracker_metadata_prev=tracker_metadata_prev,
+                tracker_update_plan=tracker_update_plan,
+                orig_vid_height=orig_vid_height,
+                orig_vid_width=orig_vid_width,
+                reconditioned_obj_ids=reconditioned_obj_ids,
+                det_to_matched_trk_obj_ids=det_to_matched_trk_obj_ids,
+            )
+            obj_id_to_score = tracker_metadata_new["obj_id_to_score"]
+        else:
+            obj_id_to_mask, obj_id_to_score = {}, {}  # dummy outputs on other GPUs
+        # a few statistics for the current frame as a part of the output
+        frame_stats = {
+            "num_obj_tracked": np.sum(tracker_metadata_new["num_obj_per_gpu"]),
+            "num_obj_dropped": tracker_update_plan["num_obj_dropped_due_to_limit"],
+        }
+        # add tracker scores to metadata, it should be fired for frames except the first frame
+        if tracker_obj_scores_global.shape[0] > 0:
+            # Convert tracker_obj_scores_global to sigmoid scores before updating
+            tracker_obj_scores_global = tracker_obj_scores_global.sigmoid().tolist()
+            tracker_obj_ids = tracker_metadata_prev["obj_ids_all_gpu"]
+            tracker_metadata_new["obj_id_to_tracker_score_frame_wise"][
+                frame_idx
+            ].update(dict(zip(tracker_obj_ids, tracker_obj_scores_global)))
+        return (
+            obj_id_to_mask,  # a dict: obj_id --> output mask
+            obj_id_to_score,  # a dict: obj_id --> output score (prob)
+            tracker_states_local_new,
+            tracker_metadata_new,
+            frame_stats,
+            tracker_obj_scores_global,  # a dict: obj_id --> tracker frame-level scores
+        )
+
+    def _suppress_detections_close_to_boundary(self, boxes, margin=0.025):
+        """
+        Suppress detections too close to image edges (for normalized boxes).
+
+        boxes: (N, 4) in xyxy format, normalized [0,1]
+        margin: fraction of image
+        """
+        x_min, y_min, x_max, y_max = boxes.unbind(-1)
+        x_c = (x_min + x_max) / 2
+        y_c = (y_min + y_max) / 2
+        keep = (
+            (x_c > margin)
+            & (x_c < 1.0 - margin)
+            & (y_c > margin)
+            & (y_c < 1.0 - margin)
+        )
+
+        return keep
+
+    def run_backbone_and_detection(
+        self,
+        frame_idx: int,
+        num_frames: int,
+        input_batch: BatchedDatapoint,
+        geometric_prompt: Any,
+        feature_cache: Dict,
+        reverse: bool,
+        allow_new_detections: bool,
+    ):
+        # Step 1: if text feature is not cached in `feature_cache`, compute and cache it
+        text_batch_key = tuple(input_batch.find_text_batch)
+        if "text" not in feature_cache or text_batch_key not in feature_cache["text"]:
+            text_outputs = self.detector.backbone.forward_text(
+                input_batch.find_text_batch, device=self.device
+            )
+            # note: we only cache the text feature of the most recent prompt
+            feature_cache["text"] = {text_batch_key: text_outputs}
+        else:
+            text_outputs = feature_cache["text"][text_batch_key]
+
+        # Step 2: run backbone, detector, and post-processing with NMS
+        if "multigpu_buffer" not in feature_cache:
+            # "multigpu_buffer" is a buffer cache used by `self.detector` and it needs
+            # to be passed to `forward_video_grounding_multigpu` for every call
+            feature_cache["multigpu_buffer"] = {}
+
+        # Extract max_frame_num_to_track from feature_cache if available
+        tracking_bounds = feature_cache.get("tracking_bounds", {})
+        max_frame_num_to_track = tracking_bounds.get("max_frame_num_to_track")
+        start_frame_idx = tracking_bounds.get("propagate_in_video_start_frame_idx")
+
+        sam3_image_out, _ = self.detector.forward_video_grounding_multigpu(
+            backbone_out={
+                "img_batch_all_stages": input_batch.img_batch,
+                **text_outputs,
+            },
+            find_inputs=input_batch.find_inputs,
+            geometric_prompt=geometric_prompt,
+            frame_idx=frame_idx,
+            num_frames=num_frames,
+            multigpu_buffer=feature_cache["multigpu_buffer"],
+            track_in_reverse=reverse,
+            # also get the SAM2 backbone features
+            return_tracker_backbone_feats=True,
+            # run NMS as a part of distributed computation
+            run_nms=self.det_nms_thresh > 0.0,
+            nms_prob_thresh=self.score_threshold_detection,
+            nms_iou_thresh=self.det_nms_thresh,
+            # pass max_frame_num_to_track to respect tracking limits
+            max_frame_num_to_track=max_frame_num_to_track,
+            propagate_in_video_start_frame_idx=start_frame_idx,
+        )
+        # note: detections in `sam3_image_out` has already gone through NMS
+        pred_probs = sam3_image_out["pred_logits"].squeeze(-1).sigmoid()
+        if not allow_new_detections:
+            pred_probs = pred_probs - 1e8  # make sure no detections are kept
+        pred_boxes_xyxy = sam3_image_out["pred_boxes_xyxy"]
+        pred_masks = sam3_image_out["pred_masks"]
+        # get the positive detection outputs above threshold
+        pos_pred_idx = torch.where(pred_probs > self.score_threshold_detection)
+        det_out = {
+            "bbox": pred_boxes_xyxy[pos_pred_idx[0], pos_pred_idx[1]],
+            "mask": pred_masks[pos_pred_idx[0], pos_pred_idx[1]],
+            "scores": pred_probs[pos_pred_idx[0], pos_pred_idx[1]],
+        }
+
+        # Step 3: build SAM2 backbone features and store them in `feature_cache`
+        backbone_cache = {}
+        sam_mask_decoder = self.tracker.sam_mask_decoder
+        tracker_backbone_fpn = [
+            sam_mask_decoder.conv_s0(sam3_image_out["tracker_backbone_fpn_0"]),
+            sam_mask_decoder.conv_s1(sam3_image_out["tracker_backbone_fpn_1"]),
+            sam3_image_out["tracker_backbone_fpn_2"],  # fpn_2 doesn't need conv
+        ]
+        tracker_backbone_out = {
+            "vision_features": tracker_backbone_fpn[-1],  # top-level feature
+            "vision_pos_enc": sam3_image_out["tracker_backbone_pos_enc"],
+            "backbone_fpn": tracker_backbone_fpn,
+        }
+        backbone_cache["tracker_backbone_out"] = tracker_backbone_out
+        feature_cache[frame_idx] = (
+            input_batch.img_batch[frame_idx],
+            backbone_cache,
+        )
+        # remove from `feature_cache` old features to save GPU memory
+        feature_cache.pop(frame_idx - 1 if not reverse else frame_idx + 1, None)
+        return det_out
+
+    def run_tracker_propagation(
+        self,
+        frame_idx: int,
+        num_frames: int,
+        reverse: bool,
+        tracker_states_local: List[Any],
+        tracker_metadata_prev: Dict[str, npt.NDArray],
+    ):
+        # Step 1: propagate the local SAM2 states to get the current frame's prediction
+        # `low_res_masks_local` of the existing masklets on this GPU
+        # - obj_ids_local: List[int] -- list of object IDs
+        # - low_res_masks_local: Tensor -- (num_local_obj, H_mask, W_mask)
+        obj_ids_local, low_res_masks_local, obj_scores_local = (
+            self._propogate_tracker_one_frame_local_gpu(
+                tracker_states_local, frame_idx=frame_idx, reverse=reverse
+            )
+        )
+
+        assert np.all(
+            obj_ids_local == tracker_metadata_prev["obj_ids_per_gpu"][self.rank]
+        ), "{} != {}".format(
+            obj_ids_local, tracker_metadata_prev["obj_ids_per_gpu"][self.rank]
+        )
+
+        # Step 2: all-gather `low_res_masks_local` into `low_res_masks_global`
+        # - low_res_masks_global: Tensor -- (num_global_obj, H_mask, W_mask)
+        _, H_mask, W_mask = low_res_masks_local.shape
+        if self.world_size > 1:
+            # `low_res_masks_local` and `obj_scores_local` need to be contiguous and float32
+            # (they could be non-contiguous due to slicing and/or bfloat16 due to autocast)
+            low_res_masks_local = low_res_masks_local.float().contiguous()
+            obj_scores_local = obj_scores_local.float().contiguous()
+            num_obj_this_gpu = tracker_metadata_prev["num_obj_per_gpu"][self.rank]
+            assert low_res_masks_local.size(0) == num_obj_this_gpu
+            assert obj_scores_local.size(0) == num_obj_this_gpu
+            low_res_masks_peers = [
+                low_res_masks_local.new_empty(num_obj, H_mask, W_mask)
+                for num_obj in tracker_metadata_prev["num_obj_per_gpu"]
+            ]
+            obj_scores_peers = [
+                obj_scores_local.new_empty(num_obj)
+                for num_obj in tracker_metadata_prev["num_obj_per_gpu"]
+            ]
+            dist.all_gather(low_res_masks_peers, low_res_masks_local)
+            dist.all_gather(obj_scores_peers, obj_scores_local)
+            low_res_masks_global = torch.cat(low_res_masks_peers, dim=0)
+            obj_scores_global = torch.cat(obj_scores_peers, dim=0)
+        else:
+            low_res_masks_global = low_res_masks_local
+            obj_scores_global = obj_scores_local
+        return low_res_masks_global, obj_scores_global
+
+    def _recondition_masklets(
+        self,
+        frame_idx,
+        det_out: Dict[str, Tensor],
+        trk_id_to_max_iou_high_conf_det: List[int],
+        tracker_states_local: List[Any],
+        tracker_metadata: Dict[str, npt.NDArray],
+        tracker_obj_scores_global: Tensor,
+    ):
+        # Recondition the masklets based on the new detections
+        for trk_obj_id, det_idx in trk_id_to_max_iou_high_conf_det.items():
+            new_mask = det_out["mask"][det_idx : det_idx + 1]
+            input_mask_res = self.tracker.input_mask_size
+            new_mask_binary = (
+                F.interpolate(
+                    new_mask.unsqueeze(1),
+                    size=(input_mask_res, input_mask_res),
+                    mode="bilinear",
+                    align_corners=False,
+                ).squeeze(1)[0]
+                > 0
+            )
+            HIGH_CONF_THRESH = 0.8
+            reconditioned_states_idx = set()
+            obj_idx = np.where(tracker_metadata["obj_ids_all_gpu"] == trk_obj_id)[
+                0
+            ].item()
+            obj_score = tracker_obj_scores_global[obj_idx]
+            for state_idx, inference_state in enumerate(tracker_states_local):
+                if (
+                    trk_obj_id in inference_state["obj_ids"]
+                    # NOTE: Goal of this condition is to avoid reconditioning masks that are occluded/low qualiy.
+                    # Unfortunately, these can get reconditioned anyway due to batching. We should consider removing these heuristics.
+                    and obj_score > HIGH_CONF_THRESH
+                ):
+                    logger.debug(
+                        f"Adding new mask for track {trk_obj_id} at frame {frame_idx}. Objects {inference_state['obj_ids']} are all reconditioned."
+                    )
+                    self.tracker.add_new_mask(
+                        inference_state=inference_state,
+                        frame_idx=frame_idx,
+                        obj_id=trk_obj_id,
+                        mask=new_mask_binary,
+                    )
+                    reconditioned_states_idx.add(state_idx)
+
+            for idx in reconditioned_states_idx:
+                self.tracker.propagate_in_video_preflight(
+                    tracker_states_local[idx], run_mem_encoder=True
+                )
+        return tracker_states_local
+
+    def run_tracker_update_planning_phase(
+        self,
+        frame_idx: int,
+        num_frames: int,
+        reverse: bool,
+        det_out: Dict[str, Tensor],
+        tracker_low_res_masks_global: Tensor,
+        tracker_obj_scores_global: Tensor,
+        tracker_metadata_prev: Dict[str, npt.NDArray],
+        tracker_states_local: List[Any],
+        is_image_only: bool = False,
+    ):
+        # initialize new metadata from previous metadata (its values will be updated later)
+        tracker_metadata_new = {
+            "obj_ids_per_gpu": deepcopy(tracker_metadata_prev["obj_ids_per_gpu"]),
+            "obj_ids_all_gpu": None,  # will be filled later
+            "num_obj_per_gpu": deepcopy(tracker_metadata_prev["num_obj_per_gpu"]),
+            "obj_id_to_score": deepcopy(tracker_metadata_prev["obj_id_to_score"]),
+            "obj_id_to_tracker_score_frame_wise": deepcopy(
+                tracker_metadata_prev["obj_id_to_tracker_score_frame_wise"]
+            ),
+            "obj_id_to_last_occluded": {},  # will be filled later
+            "max_obj_id": deepcopy(tracker_metadata_prev["max_obj_id"]),
+        }
+
+        # Initialize reconditioned_obj_ids early to avoid UnboundLocalError
+        reconditioned_obj_ids = set()
+
+        # Step 1: make the update plan and resolve heuristics on GPU 0
+        det_mask_preds: Tensor = det_out["mask"]  # low-res mask logits
+        det_scores_np: npt.NDArray = det_out["scores"].float().cpu().numpy()
+        det_bbox_xyxy: Tensor = det_out["bbox"]
+        if self.rank == 0:
+            # a) match detector and tracker masks and find new objects
+            (
+                new_det_fa_inds,
+                unmatched_trk_obj_ids,
+                det_to_matched_trk_obj_ids,
+                trk_id_to_max_iou_high_conf_det,
+                empty_trk_obj_ids,
+            ) = self._associate_det_trk(
+                det_masks=det_mask_preds,
+                det_scores_np=det_scores_np,
+                trk_masks=tracker_low_res_masks_global,
+                trk_obj_ids=tracker_metadata_prev["obj_ids_all_gpu"],
+            )
+            if self.suppress_det_close_to_boundary:
+                keep = self._suppress_detections_close_to_boundary(
+                    det_bbox_xyxy[new_det_fa_inds]
+                )
+                new_det_fa_inds = new_det_fa_inds[keep.cpu().numpy()]
+
+            # check whether we've hit the maximum number of objects we can track (and if so, drop some detections)
+            prev_obj_num = np.sum(tracker_metadata_prev["num_obj_per_gpu"])
+            new_det_num = len(new_det_fa_inds)
+            num_obj_dropped_due_to_limit = 0
+            if not is_image_only and prev_obj_num + new_det_num > self.max_num_objects:
+                logger.warning(
+                    f"hitting {self.max_num_objects=} with {new_det_num=} and {prev_obj_num=}"
+                )
+                new_det_num_to_keep = self.max_num_objects - prev_obj_num
+                num_obj_dropped_due_to_limit = new_det_num - new_det_num_to_keep
+                new_det_fa_inds = self._drop_new_det_with_obj_limit(
+                    new_det_fa_inds, det_scores_np, new_det_num_to_keep
+                )
+                assert len(new_det_fa_inds) == new_det_num_to_keep
+                new_det_num = len(new_det_fa_inds)
+
+            # assign object IDs to new detections and decide which GPU to place them
+            new_det_start_obj_id = tracker_metadata_prev["max_obj_id"] + 1
+            new_det_obj_ids = new_det_start_obj_id + np.arange(new_det_num)
+            prev_workload_per_gpu = tracker_metadata_prev["num_obj_per_gpu"]
+            new_det_gpu_ids = self._assign_new_det_to_gpus(
+                new_det_num=new_det_num,
+                prev_workload_per_gpu=prev_workload_per_gpu,
+            )
+
+            # b) handle hotstart heuristics to remove objects
+            # here `rank0_metadata` contains metadata stored on (and only accessible to) GPU 0;
+            # we avoid broadcasting them to other GPUs to save communication cost, assuming
+            # that `rank0_metadata` is not needed by other GPUs
+            rank0_metadata_new = deepcopy(tracker_metadata_prev["rank0_metadata"])
+            if not hasattr(self, "_warm_up_complete") or self._warm_up_complete:
+                obj_ids_newly_removed, rank0_metadata_new = self._process_hotstart(
+                    frame_idx=frame_idx,
+                    num_frames=num_frames,
+                    reverse=reverse,
+                    det_to_matched_trk_obj_ids=det_to_matched_trk_obj_ids,
+                    new_det_obj_ids=new_det_obj_ids,
+                    empty_trk_obj_ids=empty_trk_obj_ids,
+                    unmatched_trk_obj_ids=unmatched_trk_obj_ids,
+                    rank0_metadata=rank0_metadata_new,
+                    tracker_metadata=tracker_metadata_prev,
+                )
+            else:
+                # if warm-up is not complete, we don't remove any objects
+                obj_ids_newly_removed = set()
+            tracker_metadata_new["rank0_metadata"] = rank0_metadata_new
+
+        # Step 2: broadcast the update plan to other GPUs
+        NUM_BROADCAST_ITEMS = 9
+        if self.rank == 0 and self.world_size > 1:
+            # `num_obj_per_gpu_on_rank0` is used for metadata consistency check on other GPUs
+            # (it's a small array with length==self.world_size, so broadcasting it is cheap)
+            num_obj_per_gpu_on_rank0 = tracker_metadata_prev["num_obj_per_gpu"]
+            update_plan = [
+                new_det_fa_inds,
+                new_det_obj_ids,
+                new_det_gpu_ids,
+                num_obj_per_gpu_on_rank0,
+                unmatched_trk_obj_ids,
+                det_to_matched_trk_obj_ids,
+                obj_ids_newly_removed,
+                num_obj_dropped_due_to_limit,
+                trk_id_to_max_iou_high_conf_det,
+            ]
+            assert (
+                len(update_plan) == NUM_BROADCAST_ITEMS
+            ), f"Manually update NUM_BROADCAST_ITEMS to be: {len(update_plan)}"
+            self.broadcast_python_obj_cpu(update_plan, src=0)
+        elif self.rank > 0 and self.world_size > 1:
+            update_plan = [
+                None
+            ] * NUM_BROADCAST_ITEMS  # other ranks receive the plan from rank 0
+            self.broadcast_python_obj_cpu(update_plan, src=0)
+            (
+                new_det_fa_inds,
+                new_det_obj_ids,
+                new_det_gpu_ids,
+                num_obj_per_gpu_on_rank0,
+                unmatched_trk_obj_ids,
+                det_to_matched_trk_obj_ids,
+                obj_ids_newly_removed,
+                num_obj_dropped_due_to_limit,
+                trk_id_to_max_iou_high_conf_det,
+            ) = update_plan
+            # metadata consistency check: verify that the received `num_obj_per_gpu_on_rank0` is consistent with the local metadata
+            # it's critical that all GPUs agree on the previous number of objects (otherwise the inference might hang or fail silently)
+            if not np.all(
+                num_obj_per_gpu_on_rank0 == tracker_metadata_prev["num_obj_per_gpu"]
+            ):
+                raise RuntimeError(
+                    f"{self.rank=} received {num_obj_per_gpu_on_rank0=}, which is inconsistent with local record "
+                    f"{tracker_metadata_prev['num_obj_per_gpu']=}. There's likely a bug in update planning or execution."
+                )
+
+        # `tracker_update_plan` should be identical on all GPUs after broadcasting
+        tracker_update_plan = {
+            "new_det_fa_inds": new_det_fa_inds,  # npt.NDArray
+            "new_det_obj_ids": new_det_obj_ids,  # npt.NDArray
+            "new_det_gpu_ids": new_det_gpu_ids,  # npt.NDArray
+            "unmatched_trk_obj_ids": unmatched_trk_obj_ids,  # npt.NDArray
+            "det_to_matched_trk_obj_ids": det_to_matched_trk_obj_ids,  # dict
+            "obj_ids_newly_removed": obj_ids_newly_removed,  # set
+            "num_obj_dropped_due_to_limit": num_obj_dropped_due_to_limit,  # int
+            "trk_id_to_max_iou_high_conf_det": trk_id_to_max_iou_high_conf_det,  # dict
+            "reconditioned_obj_ids": reconditioned_obj_ids,  # set
+        }
+
+        # Step 3 (optional): recondition masklets based on high-confidence detections before memory encoding
+        # NOTE: Running this in execution phase (after memory encoding) can lead to suboptimal results
+        should_recondition_iou = False
+
+        # Evaluate tracklets for reconditioning based on bbox IoU mismatch with detections
+        if (
+            self.reconstruction_bbox_iou_thresh > 0
+            and len(trk_id_to_max_iou_high_conf_det) > 0
+        ):
+            for trk_obj_id, det_idx in trk_id_to_max_iou_high_conf_det.items():
+                det_box = det_out["bbox"][det_idx]
+                det_score = det_out["scores"][det_idx]
+
+                try:
+                    trk_idx = list(tracker_metadata_prev["obj_ids_all_gpu"]).index(
+                        trk_obj_id
+                    )
+                except ValueError:
+                    continue  # Skip if tracklet not found
+
+                tracker_mask = tracker_low_res_masks_global[trk_idx]
+                mask_binary = tracker_mask > 0
+                mask_area = mask_binary.sum().item()
+
+                if mask_area == 0:
+                    continue  # Skip tracklets with zero mask area
+
+                # Get bounding box from SAM2 mask and convert to normalized coordinates
+                tracker_box_pixels = (
+                    mask_to_box(mask_binary.unsqueeze(0).unsqueeze(0))
+                    .squeeze(0)
+                    .squeeze(0)
+                )
+                mask_height, mask_width = tracker_mask.shape[-2:]
+                tracker_box_normalized = torch.tensor(
+                    [
+                        tracker_box_pixels[0] / mask_width,
+                        tracker_box_pixels[1] / mask_height,
+                        tracker_box_pixels[2] / mask_width,
+                        tracker_box_pixels[3] / mask_height,
+                    ],
+                    device=tracker_box_pixels.device,
+                )
+
+                # Compute IoU between detection and SAM2 tracklet bounding boxes
+                det_box_batch = det_box.unsqueeze(0)
+                tracker_box_batch = tracker_box_normalized.unsqueeze(0)
+                iou = fast_diag_box_iou(det_box_batch, tracker_box_batch)[0]
+
+                if (
+                    iou < self.reconstruction_bbox_iou_thresh
+                    and det_score >= self.reconstruction_bbox_det_score
+                ):
+                    should_recondition_iou = True
+                    reconditioned_obj_ids.add(trk_obj_id)
+
+        should_recondition_periodic = (
+            self.recondition_every_nth_frame > 0
+            and frame_idx % self.recondition_every_nth_frame == 0
+            and len(trk_id_to_max_iou_high_conf_det) > 0
+        )
+
+        # Recondition if periodic or IoU condition met
+        if should_recondition_periodic or should_recondition_iou:
+            self._recondition_masklets(
+                frame_idx,
+                det_out,
+                trk_id_to_max_iou_high_conf_det,
+                tracker_states_local,
+                tracker_metadata_prev,
+                tracker_obj_scores_global,
+            )
+
+        # Step 4: Run SAM2 memory encoder on the current frame's prediction masks
+        # This is done on all GPUs
+        batch_size = tracker_low_res_masks_global.size(0)
+        if batch_size > 0:
+            if not hasattr(self, "_warm_up_complete") or self._warm_up_complete:
+                if self.suppress_overlapping_based_on_recent_occlusion_threshold > 0.0:
+                    # NOTE: tracker_low_res_masks_global is updated in-place then returned
+                    tracker_low_res_masks_global = (
+                        self._suppress_overlapping_based_on_recent_occlusion(
+                            frame_idx,
+                            tracker_low_res_masks_global,
+                            tracker_metadata_prev,
+                            tracker_metadata_new,
+                            obj_ids_newly_removed,
+                            reverse,
+                        )
+                    )
+
+            self._tracker_update_memories(
+                tracker_states_local,
+                frame_idx,
+                tracker_metadata=tracker_metadata_prev,
+                low_res_masks=tracker_low_res_masks_global,
+            )
+
+        # Step 4: update the SAM2 metadata based on the update plan
+        # note: except for "rank0_metadata" (that is only available on GPU 0),
+        # the updated `tracker_metadata_new` should be identical on all GPUs
+        for rank in range(self.world_size):
+            new_det_obj_ids_this_gpu = new_det_obj_ids[new_det_gpu_ids == rank]
+            updated_obj_ids_this_gpu = tracker_metadata_new["obj_ids_per_gpu"][rank]
+            if len(new_det_obj_ids_this_gpu) > 0:
+                updated_obj_ids_this_gpu = np.concatenate(
+                    [updated_obj_ids_this_gpu, new_det_obj_ids_this_gpu]
+                )
+            if len(obj_ids_newly_removed) > 0:
+                is_removed = np.isin(
+                    updated_obj_ids_this_gpu, list(obj_ids_newly_removed)
+                )
+                updated_obj_ids_this_gpu = updated_obj_ids_this_gpu[~is_removed]
+            tracker_metadata_new["obj_ids_per_gpu"][rank] = updated_obj_ids_this_gpu
+            tracker_metadata_new["num_obj_per_gpu"][rank] = len(
+                updated_obj_ids_this_gpu
+            )
+        tracker_metadata_new["obj_ids_all_gpu"] = np.concatenate(
+            tracker_metadata_new["obj_ids_per_gpu"]
+        )
+        # update object scores and the maximum object ID assigned so far
+        if len(new_det_obj_ids) > 0:
+            tracker_metadata_new["obj_id_to_score"].update(
+                zip(new_det_obj_ids, det_scores_np[new_det_fa_inds])
+            )
+            # tracker scores are not available for new objects, use det score instead.
+            tracker_metadata_new["obj_id_to_tracker_score_frame_wise"][
+                frame_idx
+            ].update(zip(new_det_obj_ids, det_scores_np[new_det_fa_inds]))
+            tracker_metadata_new["max_obj_id"] = max(
+                tracker_metadata_new["max_obj_id"],
+                np.max(new_det_obj_ids),
+            )
+        # for removed objects, we set their scores to a very low value (-1e4) but still
+        # keep them in "obj_id_to_score" (it's easier to handle outputs this way)
+        for obj_id in obj_ids_newly_removed:
+            tracker_metadata_new["obj_id_to_score"][obj_id] = -1e4
+            tracker_metadata_new["obj_id_to_tracker_score_frame_wise"][frame_idx][
+                obj_id
+            ] = -1e4
+            tracker_metadata_new["obj_id_to_last_occluded"].pop(obj_id, None)
+        # check that "rank0_metadata" is in tracker_metadata_new if and only if it's GPU 0
+        assert ("rank0_metadata" in tracker_metadata_new) == (self.rank == 0)
+        if self.rank == 0 and self.masklet_confirmation_enable:
+            rank0_metadata = self.update_masklet_confirmation_status(
+                rank0_metadata=tracker_metadata_new["rank0_metadata"],
+                obj_ids_all_gpu_prev=tracker_metadata_prev["obj_ids_all_gpu"],
+                obj_ids_all_gpu_updated=tracker_metadata_new["obj_ids_all_gpu"],
+                det_to_matched_trk_obj_ids=det_to_matched_trk_obj_ids,
+                new_det_obj_ids=new_det_obj_ids,
+            )
+            tracker_metadata_new["rank0_metadata"] = rank0_metadata
+
+        return tracker_update_plan, tracker_metadata_new
+
+    def _suppress_overlapping_based_on_recent_occlusion(
+        self,
+        frame_idx: int,
+        tracker_low_res_masks_global: Tensor,
+        tracker_metadata_prev: Dict[str, Any],
+        tracker_metadata_new: Dict[str, Any],
+        obj_ids_newly_removed: Set[int],
+        reverse: bool = False,
+    ):
+        """
+        Suppress overlapping masks based on the most recent occlusion information. If an object is removed by hotstart, we always suppress it if it overlaps with any other object.
+        Args:
+            frame_idx (int): The current frame index.
+            tracker_low_res_masks_global (Tensor): The low-resolution masks for the current frame.
+            tracker_metadata_prev (Dict[str, Any]): The metadata from the previous frame.
+            tracker_metadata_new (Dict[str, Any]): The metadata for the current frame.
+            obj_ids_newly_removed (Set[int]): The object IDs that have been removed.
+        Return:
+            Tensor: The updated low-resolution masks with some objects suppressed.
+        """
+        obj_ids_global = tracker_metadata_prev["obj_ids_all_gpu"]
+        binary_tracker_low_res_masks_global = tracker_low_res_masks_global > 0
+        batch_size = tracker_low_res_masks_global.size(0)
+        if batch_size > 0:
+            assert (
+                len(obj_ids_global) == batch_size
+            ), f"Mismatch in number of objects: {len(obj_ids_global)} vs {batch_size}"
+            NEVER_OCCLUDED = -1
+            ALWAYS_OCCLUDED = 100000  # This value should be larger than any possible frame index, indicates that the object was removed by hotstart logic
+            last_occluded_prev = torch.cat(
+                [
+                    tracker_metadata_prev["obj_id_to_last_occluded"].get(
+                        obj_id,
+                        torch.full(
+                            (1,),
+                            fill_value=(
+                                NEVER_OCCLUDED
+                                if obj_id not in obj_ids_newly_removed
+                                else ALWAYS_OCCLUDED
+                            ),
+                            device=binary_tracker_low_res_masks_global.device,
+                            dtype=torch.long,
+                        ),
+                    )
+                    for obj_id in obj_ids_global
+                ],
+                dim=0,
+            )
+            to_suppress = self._get_objects_to_suppress_based_on_most_recently_occluded(
+                binary_tracker_low_res_masks_global,
+                last_occluded_prev,
+                obj_ids_global,
+                frame_idx,
+                reverse,
+            )
+
+            # Update metadata with occlusion information
+            is_obj_occluded = ~(binary_tracker_low_res_masks_global.any(dim=(-1, -2)))
+            is_obj_occluded_or_suppressed = is_obj_occluded | to_suppress
+            last_occluded_new = last_occluded_prev.clone()
+            last_occluded_new[is_obj_occluded_or_suppressed] = frame_idx
+            # Slice out the last occluded frame for each object
+            tracker_metadata_new["obj_id_to_last_occluded"] = {
+                obj_id: last_occluded_new[obj_idx : obj_idx + 1]
+                for obj_idx, obj_id in enumerate(obj_ids_global)
+            }
+
+            # Zero out suppressed masks before memory encoding
+            NO_OBJ_LOGIT = -10
+            tracker_low_res_masks_global[to_suppress] = NO_OBJ_LOGIT
+
+        return tracker_low_res_masks_global
+
+    def run_tracker_update_execution_phase(
+        self,
+        frame_idx: int,
+        num_frames: int,
+        reverse: bool,
+        det_out: Dict[str, Tensor],
+        tracker_states_local: List[Any],
+        tracker_update_plan: Dict[str, npt.NDArray],
+        orig_vid_height: int,
+        orig_vid_width: int,
+        feature_cache: Dict,
+    ):
+        # initialize tracking scores with detection scores
+        new_det_fa_inds: npt.NDArray = tracker_update_plan["new_det_fa_inds"]
+        new_det_obj_ids: npt.NDArray = tracker_update_plan["new_det_obj_ids"]
+        new_det_gpu_ids: npt.NDArray = tracker_update_plan["new_det_gpu_ids"]
+        is_on_this_gpu: npt.NDArray = new_det_gpu_ids == self.rank
+        new_det_obj_ids_local: npt.NDArray = new_det_obj_ids[is_on_this_gpu]
+        new_det_fa_inds_local: npt.NDArray = new_det_fa_inds[is_on_this_gpu]
+        obj_ids_newly_removed: Set[int] = tracker_update_plan["obj_ids_newly_removed"]
+
+        # Step 1: add new objects from the detector to SAM2 inference states
+        if len(new_det_fa_inds_local) > 0:
+            new_det_fa_inds_local_t = torch.from_numpy(new_det_fa_inds_local)
+            new_det_masks: Tensor = det_out["mask"][new_det_fa_inds_local_t]
+            # initialize SAM2 with new object masks
+            tracker_states_local = self._tracker_add_new_objects(
+                frame_idx=frame_idx,
+                num_frames=num_frames,
+                new_obj_ids=new_det_obj_ids_local,
+                new_obj_masks=new_det_masks,
+                tracker_states_local=tracker_states_local,
+                orig_vid_height=orig_vid_height,
+                orig_vid_width=orig_vid_width,
+                feature_cache=feature_cache,
+            )
+
+        # Step 2: remove from SAM2 inference states those objects removed by heuristics
+        if len(obj_ids_newly_removed) > 0:
+            self._tracker_remove_objects(tracker_states_local, obj_ids_newly_removed)
+
+        return tracker_states_local
+
+    def build_outputs(
+        self,
+        frame_idx: int,
+        num_frames: int,
+        reverse: bool,
+        det_out: Dict[str, Tensor],
+        tracker_low_res_masks_global: Tensor,
+        tracker_obj_scores_global: Tensor,
+        tracker_metadata_prev: Dict[str, npt.NDArray],
+        tracker_update_plan: Dict[str, npt.NDArray],
+        orig_vid_height: int,
+        orig_vid_width: int,
+        reconditioned_obj_ids: set = None,
+        det_to_matched_trk_obj_ids: dict = None,
+    ):
+        new_det_fa_inds: npt.NDArray = tracker_update_plan["new_det_fa_inds"]
+        new_det_obj_ids: npt.NDArray = tracker_update_plan["new_det_obj_ids"]
+        obj_id_to_mask = {}  # obj_id --> output mask tensor
+
+        # Part 1: masks from previous SAM2 propagation
+        existing_masklet_obj_ids = tracker_metadata_prev["obj_ids_all_gpu"]
+        existing_masklet_video_res_masks = F.interpolate(
+            tracker_low_res_masks_global.unsqueeze(1),
+            size=(orig_vid_height, orig_vid_width),
+            mode="bilinear",
+            align_corners=False,
+        )  # (num_obj, 1, H_video, W_video)
+        existing_masklet_binary = existing_masklet_video_res_masks > 0
+        assert len(existing_masklet_obj_ids) == len(existing_masklet_binary)
+        for obj_id, mask in zip(existing_masklet_obj_ids, existing_masklet_binary):
+            obj_id_to_mask[obj_id] = mask  # (1, H_video, W_video)
+
+        # Part 2: masks from new detections
+        new_det_fa_inds_t = torch.from_numpy(new_det_fa_inds)
+        new_det_low_res_masks = det_out["mask"][new_det_fa_inds_t].unsqueeze(1)
+        new_det_low_res_masks = fill_holes_in_mask_scores(
+            new_det_low_res_masks,
+            max_area=self.fill_hole_area,
+            fill_holes=True,
+            remove_sprinkles=True,
+        )
+        new_masklet_video_res_masks = F.interpolate(
+            new_det_low_res_masks,
+            size=(orig_vid_height, orig_vid_width),
+            mode="bilinear",
+            align_corners=False,
+        )  # (num_obj, 1, H_video, W_video)
+
+        new_masklet_binary = new_masklet_video_res_masks > 0
+        assert len(new_det_obj_ids) == len(new_masklet_video_res_masks)
+        for obj_id, mask in zip(new_det_obj_ids, new_masklet_binary):
+            obj_id_to_mask[obj_id] = mask  # (1, H_video, W_video)
+
+        # Part 3: Override masks for reconditioned objects using detection masks
+        if reconditioned_obj_ids is not None and len(reconditioned_obj_ids) > 0:
+            trk_id_to_max_iou_high_conf_det = tracker_update_plan.get(
+                "trk_id_to_max_iou_high_conf_det", {}
+            )
+
+            for obj_id in reconditioned_obj_ids:
+                det_idx = trk_id_to_max_iou_high_conf_det.get(obj_id)
+
+                if det_idx is not None:
+                    det_mask = det_out["mask"][det_idx]
+                    det_mask = det_mask.unsqueeze(0).unsqueeze(0)
+                    det_mask_resized = (
+                        F.interpolate(
+                            det_mask.float(),
+                            size=(orig_vid_height, orig_vid_width),
+                            mode="bilinear",
+                            align_corners=False,
+                        )
+                        > 0
+                    )
+
+                    det_mask_final = det_mask_resized.squeeze(0)
+                    obj_id_to_mask[obj_id] = det_mask_final
+
+        return obj_id_to_mask
+
+    def _get_objects_to_suppress_based_on_most_recently_occluded(
+        self,
+        binary_low_res_masks: Tensor,
+        last_occluded: List[int],
+        obj_ids: List[int],
+        frame_idx: int = None,
+        reverse: bool = False,
+    ):
+        # Suppress overlapping masks for objects that were most recently occluded
+        assert (
+            binary_low_res_masks.dtype == torch.bool
+        ), f"Expected boolean tensor, got {binary_low_res_masks.dtype}"
+        to_suppress = torch.zeros(
+            binary_low_res_masks.size(0),
+            device=binary_low_res_masks.device,
+            dtype=torch.bool,
+        )
+        if len(obj_ids) <= 1:
+            return to_suppress
+
+        iou = mask_iou(binary_low_res_masks, binary_low_res_masks)  # [N,N]
+
+        # Create masks for upper triangular matrix (i < j) and IoU threshold
+        mask_iou_thresh = (
+            iou >= self.suppress_overlapping_based_on_recent_occlusion_threshold
+        )
+        overlapping_pairs = torch.triu(mask_iou_thresh, diagonal=1)  # [N,N]
+
+        last_occ_expanded_i = last_occluded.unsqueeze(1)  # (N, 1)
+        last_occ_expanded_j = last_occluded.unsqueeze(0)  # (1, N)
+        # Suppress most recently occluded
+        cmp_op = torch.gt if not reverse else torch.lt
+        suppress_i_mask = (
+            overlapping_pairs
+            & cmp_op(
+                last_occ_expanded_i, last_occ_expanded_j
+            )  # (last_occ_expanded_i > last_occ_expanded_j)
+            & (
+                last_occ_expanded_j > -1
+            )  # j can suppress i only if i was previously occluded
+        )
+        suppress_j_mask = (
+            overlapping_pairs
+            & cmp_op(last_occ_expanded_j, last_occ_expanded_i)
+            & (
+                last_occ_expanded_i > -1
+            )  # i can suppress j only if j was previously occluded
+        )
+        # Apply suppression
+        to_suppress = suppress_i_mask.any(dim=1) | suppress_j_mask.any(dim=0)
+
+        # Log for debugging
+        if (
+            self.rank == 0
+            and logger.isEnabledFor(logging.DEBUG)
+            and frame_idx is not None
+        ):
+            suppress_i_mask = suppress_i_mask.cpu().numpy()
+            suppress_j_mask = suppress_j_mask.cpu().numpy()
+            last_occluded = last_occluded.cpu().numpy()
+
+            # Find all suppression pairs without using torch.where
+            batch_size = suppress_i_mask.shape[0]
+
+            # Log i-suppression cases (where i gets suppressed in favor of j)
+            for i in range(batch_size):
+                for j in range(batch_size):
+                    if suppress_i_mask[i, j]:
+                        logger.debug(
+                            f"{frame_idx=}: Suppressing obj {obj_ids[i]} last occluded {last_occluded[i]} in favor of {obj_ids[j]} last occluded {last_occluded[j]}"
+                        )
+
+            # Log j-suppression cases (where j gets suppressed in favor of i)
+            for i in range(batch_size):
+                for j in range(batch_size):
+                    if suppress_j_mask[i, j]:
+                        logger.debug(
+                            f"{frame_idx=}: Suppressing obj {obj_ids[j]} last occluded {last_occluded[j]} in favor of {obj_ids[i]} last occluded {last_occluded[i]}"
+                        )
+
+        return to_suppress
+
+    def _propogate_tracker_one_frame_local_gpu(
+        self,
+        inference_states: List[Any],
+        frame_idx: int,
+        reverse: bool,
+        # by default, we disable memory encoding until we gather all outputs
+        run_mem_encoder: bool = False,
+    ):
+        """
+        inference_states: List of inference states, each state corresponds to a different set of objects.
+        """
+        obj_ids_local = []
+        low_res_masks_list = []
+        obj_scores_list = []
+        for inference_state in inference_states:
+            if len(inference_state["obj_ids"]) == 0:
+                continue  # skip propagation on empty inference states
+
+            # propagate one frame
+            num_frames_propagated = 0
+            for out in self.tracker.propagate_in_video(
+                inference_state,
+                start_frame_idx=frame_idx,
+                # end_frame_idx = start_frame_idx + max_frame_num_to_track
+                # (i.e. propagating 1 frame since end_frame_idx is inclusive)
+                max_frame_num_to_track=0,
+                reverse=reverse,
+                tqdm_disable=True,
+                run_mem_encoder=run_mem_encoder,
+            ):
+                out_frame_idx, out_obj_ids, out_low_res_masks, _, out_obj_scores = out
+                num_frames_propagated += 1
+
+            # only 1 frames should be propagated
+            assert (
+                num_frames_propagated == 1 and out_frame_idx == frame_idx
+            ), f"num_frames_propagated: {num_frames_propagated}, out_frame_idx: {out_frame_idx}, frame_idx: {frame_idx}"
+            assert isinstance(out_obj_ids, list)
+            obj_ids_local.extend(out_obj_ids)
+            low_res_masks_list.append(out_low_res_masks.squeeze(1))
+            obj_scores_list.append(out_obj_scores.squeeze(1))
+
+        # concatenate the output masklets from all local inference states
+        H_mask = W_mask = self.tracker.low_res_mask_size
+        if len(low_res_masks_list) > 0:
+            low_res_masks_local = torch.cat(low_res_masks_list, dim=0)
+            obj_scores_local = torch.cat(obj_scores_list, dim=0)
+            assert low_res_masks_local.shape[1:] == (H_mask, W_mask)
+
+            # Apply hole filling to the masks
+            low_res_masks_local = fill_holes_in_mask_scores(
+                low_res_masks_local.unsqueeze(1),
+                max_area=self.fill_hole_area,
+                fill_holes=True,
+                remove_sprinkles=True,
+            )
+            low_res_masks_local = low_res_masks_local.squeeze(1)
+        else:
+            low_res_masks_local = torch.zeros(0, H_mask, W_mask, device=self.device)
+            obj_scores_local = torch.zeros(0, device=self.device)
+
+        return obj_ids_local, low_res_masks_local, obj_scores_local
+
+    def _associate_det_trk(
+        self,
+        det_masks: Tensor,
+        det_scores_np: npt.NDArray,
+        trk_masks: Tensor,
+        trk_obj_ids: npt.NDArray,
+    ):
+        """
+        Match detections on the current frame with the existing masklets.
+
+        Args:
+          - det_masks: (N, H, W) tensor of predicted masks
+          - det_scores_np: (N,) array of detection scores
+          - trk_masks: (M, H, W) tensor of track masks
+          - trk_obj_ids: (M,) array of object IDs corresponding to trk_masks
+
+        Returns:
+          - new_det_fa_inds: array of new object indices.
+          - unmatched_trk_obj_ids: array of existing masklet object IDs that are not matched
+            to any detections on this frame (for unmatched, we only count masklets with >0 area)
+          - det_to_matched_trk_obj_ids: dict[int, npt.NDArray]: mapping from detector's detection indices
+            to the list of matched tracklet object IDs
+          - empty_trk_obj_ids: array of existing masklet object IDs with zero area in SAM2 prediction
+        """
+        iou_threshold = self.assoc_iou_thresh
+        iou_threshold_trk = self.trk_assoc_iou_thresh
+        new_det_thresh = self.new_det_thresh
+
+        assert det_masks.is_floating_point(), "float tensor expected (do not binarize)"
+        assert trk_masks.is_floating_point(), "float tensor expected (do not binarize)"
+        assert (
+            trk_masks.size(0) == len(trk_obj_ids)
+        ), f"trk_masks and trk_obj_ids should have the same length, {trk_masks.size(0)} vs {len(trk_obj_ids)}"
+        if trk_masks.size(0) == 0:
+            # all detections are new
+            new_det_fa_inds = np.arange(det_masks.size(0))
+            unmatched_trk_obj_ids = np.array([], np.int64)
+            empty_trk_obj_ids = np.array([], np.int64)
+            det_to_matched_trk_obj_ids = {}
+            trk_id_to_max_iou_high_conf_det = {}
+            return (
+                new_det_fa_inds,
+                unmatched_trk_obj_ids,
+                det_to_matched_trk_obj_ids,
+                trk_id_to_max_iou_high_conf_det,
+                empty_trk_obj_ids,
+            )
+        elif det_masks.size(0) == 0:
+            # all previous tracklets are unmatched if they have a non-zero area
+            new_det_fa_inds = np.array([], np.int64)
+            trk_is_nonempty = (trk_masks > 0).any(dim=(1, 2)).cpu().numpy()
+            unmatched_trk_obj_ids = trk_obj_ids[trk_is_nonempty]
+            empty_trk_obj_ids = trk_obj_ids[~trk_is_nonempty]
+            det_to_matched_trk_obj_ids = {}
+            trk_id_to_max_iou_high_conf_det = {}
+            return (
+                new_det_fa_inds,
+                unmatched_trk_obj_ids,
+                det_to_matched_trk_obj_ids,
+                trk_id_to_max_iou_high_conf_det,
+                empty_trk_obj_ids,
+            )
+
+        if det_masks.shape[-2:] != trk_masks.shape[-2:]:
+            # resize to the smaller size to save GPU memory
+            if np.prod(det_masks.shape[-2:]) < np.prod(trk_masks.shape[-2:]):
+                trk_masks = F.interpolate(
+                    trk_masks.unsqueeze(1),
+                    size=det_masks.shape[-2:],
+                    mode="bilinear",
+                    align_corners=False,
+                ).squeeze(1)
+            else:
+                # resize detections to track size
+                det_masks = F.interpolate(
+                    det_masks.unsqueeze(1),
+                    size=trk_masks.shape[-2:],
+                    mode="bilinear",
+                    align_corners=False,
+                ).squeeze(1)
+
+        det_masks_binary = det_masks > 0
+        trk_masks_binary = trk_masks > 0
+        ious = mask_iou(det_masks_binary, trk_masks_binary)  # (N, M)
+
+        ious_np = ious.cpu().numpy()
+        if self.o2o_matching_masklets_enable:
+            from scipy.optimize import linear_sum_assignment
+
+            # Hungarian matching for tracks (one-to-one: each track matches at most one detection)
+            cost_matrix = 1 - ious_np  # Hungarian solves for minimum cost
+            row_ind, col_ind = linear_sum_assignment(cost_matrix)
+            trk_is_matched = np.zeros(trk_masks.size(0), dtype=bool)
+            for d, t in zip(row_ind, col_ind):
+                if ious_np[d, t] >= iou_threshold_trk:
+                    trk_is_matched[t] = True
+        else:
+            trk_is_matched = (ious_np >= iou_threshold_trk).any(axis=0)
+        # Non-empty tracks not matched by Hungarian assignment above threshold are unmatched
+        trk_is_nonempty = trk_masks_binary.any(dim=(1, 2)).cpu().numpy()
+        trk_is_unmatched = np.logical_and(trk_is_nonempty, ~trk_is_matched)
+        unmatched_trk_obj_ids = trk_obj_ids[trk_is_unmatched]
+        # also record masklets that have zero area in SAM 2 prediction
+        empty_trk_obj_ids = trk_obj_ids[~trk_is_nonempty]
+
+        # For detections: allow many tracks to match to the same detection (many-to-one)
+        # So, a detection is 'new' if it does not match any track above threshold
+        is_new_det = np.logical_and(
+            det_scores_np >= new_det_thresh,
+            np.logical_not(np.any(ious_np >= iou_threshold, axis=1)),
+        )
+        new_det_fa_inds = np.nonzero(is_new_det)[0]
+
+        # for each detection, which tracks it matched to (above threshold)
+        det_to_matched_trk_obj_ids = {}
+        trk_id_to_max_iou_high_conf_det = {}  # trk id --> exactly one detection idx
+        HIGH_CONF_THRESH = 0.8
+        HIGH_IOU_THRESH = 0.8
+        det_to_max_iou_trk_idx = np.argmax(ious_np, axis=1)
+        det_is_high_conf = (det_scores_np >= HIGH_CONF_THRESH) & ~is_new_det
+        det_is_high_iou = np.max(ious_np, axis=1) >= HIGH_IOU_THRESH
+        det_is_high_conf_and_iou = set(
+            np.nonzero(det_is_high_conf & det_is_high_iou)[0]
+        )
+        for d in range(det_masks.size(0)):
+            det_to_matched_trk_obj_ids[d] = trk_obj_ids[ious_np[d, :] >= iou_threshold]
+            if d in det_is_high_conf_and_iou:
+                trk_obj_id = trk_obj_ids[det_to_max_iou_trk_idx[d]].item()
+                trk_id_to_max_iou_high_conf_det[trk_obj_id] = d
+
+        return (
+            new_det_fa_inds,
+            unmatched_trk_obj_ids,
+            det_to_matched_trk_obj_ids,
+            trk_id_to_max_iou_high_conf_det,
+            empty_trk_obj_ids,
+        )
+
+    def _assign_new_det_to_gpus(self, new_det_num, prev_workload_per_gpu):
+        """Distribute the new objects to the GPUs with the least workload."""
+        workload_per_gpu: npt.NDArray = prev_workload_per_gpu.copy()
+        new_det_gpu_ids = np.zeros(new_det_num, np.int64)
+
+        # assign the objects one by one
+        for i in range(len(new_det_gpu_ids)):
+            # find the GPU with the least workload
+            min_gpu = np.argmin(workload_per_gpu)
+            new_det_gpu_ids[i] = min_gpu
+            workload_per_gpu[min_gpu] += 1
+        return new_det_gpu_ids
+
+    def _process_hotstart(
+        self,
+        frame_idx: int,
+        num_frames: int,
+        reverse: bool,
+        det_to_matched_trk_obj_ids: Dict[int, npt.NDArray],
+        new_det_obj_ids: npt.NDArray,
+        empty_trk_obj_ids: npt.NDArray,
+        unmatched_trk_obj_ids: npt.NDArray,
+        rank0_metadata: Dict[str, Any],
+        tracker_metadata: Dict[str, Any],
+    ):
+        """Handle hotstart heuristics to remove unmatched or duplicated objects."""
+        # obj_id --> first frame index where the object was detected
+        obj_first_frame_idx = rank0_metadata["obj_first_frame_idx"]
+        # obj_id --> [mismatched frame indices]
+        unmatched_frame_inds = rank0_metadata["unmatched_frame_inds"]
+        trk_keep_alive = rank0_metadata["trk_keep_alive"]
+        # (first_appear_obj_id, obj_id) --> [overlap frame indices]
+        overlap_pair_to_frame_inds = rank0_metadata["overlap_pair_to_frame_inds"]
+        # removed_obj_ids: object IDs that are suppressed via hot-start
+        removed_obj_ids = rank0_metadata["removed_obj_ids"]
+        suppressed_obj_ids = rank0_metadata["suppressed_obj_ids"][frame_idx]
+
+        obj_ids_newly_removed = set()  # object IDs to be newly removed on this frame
+        hotstart_diff = (
+            frame_idx - self.hotstart_delay
+            if not reverse
+            else frame_idx + self.hotstart_delay
+        )
+
+        # Step 1: log the frame index where each object ID first appears
+        for obj_id in new_det_obj_ids:
+            if obj_id not in obj_first_frame_idx:
+                obj_first_frame_idx[obj_id] = frame_idx
+            assert obj_id not in trk_keep_alive
+            trk_keep_alive[obj_id] = self.init_trk_keep_alive
+
+        matched_trks = set()
+        # We use the det-->tracks list to check for matched objects. Otherwise, we need to compute areas to decide whether they're occluded
+        for matched_trks_per_det in det_to_matched_trk_obj_ids.values():
+            matched_trks.update(matched_trks_per_det)
+        for obj_id in matched_trks:
+            # NOTE: To minimize number of configurable params, we use the hotstart_unmatch_thresh to set the max value of trk_keep_alive
+            trk_keep_alive[obj_id] = min(
+                self.max_trk_keep_alive, trk_keep_alive[obj_id] + 1
+            )
+        for obj_id in unmatched_trk_obj_ids:
+            unmatched_frame_inds[obj_id].append(frame_idx)
+            # NOTE: To minimize number of configurable params, we use the hotstart_unmatch_thresh to set the min value of trk_keep_alive
+            # The max keep alive is 2x the min, means the model prefers to keep the prediction rather than suppress it if it was matched long enough.
+            trk_keep_alive[obj_id] = max(
+                self.min_trk_keep_alive, trk_keep_alive[obj_id] - 1
+            )
+        if self.decrease_trk_keep_alive_for_empty_masklets:
+            for obj_id in empty_trk_obj_ids:
+                # NOTE: To minimize number of configurable params, we use the hotstart_unmatch_thresh to set the min value of trk_keep_alive
+                trk_keep_alive[obj_id] = max(
+                    self.min_trk_keep_alive, trk_keep_alive[obj_id] - 1
+                )
+
+        # Step 2: removed tracks that has not matched with detections for `hotstart_unmatch_thresh` frames with hotstart period
+        # a) add unmatched frame indices for each existing object ID
+        # note that `unmatched_trk_obj_ids` contains those frames where the SAM2 output mask
+        # doesn't match any detection; it excludes those frames where SAM2 gives an empty mask
+        # b) remove a masklet if it first appears after `hotstart_diff` and is unmatched for more
+        # than `self.hotstart_unmatch_thresh` frames
+        for obj_id, frame_indices in unmatched_frame_inds.items():
+            if obj_id in removed_obj_ids or obj_id in obj_ids_newly_removed:
+                continue  # skip if the object is already removed
+            if len(frame_indices) >= self.hotstart_unmatch_thresh:
+                is_within_hotstart = (
+                    obj_first_frame_idx[obj_id] > hotstart_diff and not reverse
+                ) or (obj_first_frame_idx[obj_id] < hotstart_diff and reverse)
+                if is_within_hotstart:
+                    obj_ids_newly_removed.add(obj_id)
+                    logger.debug(
+                        f"Removing object {obj_id} at frame {frame_idx} "
+                        f"since it is unmatched for frames: {frame_indices}"
+                    )
+            if (
+                trk_keep_alive[obj_id] <= 0  # Object has not been matched for too long
+                and not self.suppress_unmatched_only_within_hotstart
+                and obj_id not in removed_obj_ids
+                and obj_id not in obj_ids_newly_removed
+            ):
+                logger.debug(
+                    f"Suppressing object {obj_id} at frame {frame_idx}, due to being unmatched"
+                )
+                suppressed_obj_ids.add(obj_id)
+
+        # Step 3: removed tracks that overlaps with another track for `hotstart_dup_thresh` frames
+        # a) find overlaps tracks -- we consider overlap if they match to the same detection
+        for _, matched_trk_obj_ids in det_to_matched_trk_obj_ids.items():
+            if len(matched_trk_obj_ids) < 2:
+                continue  # only count detections that are matched to multiple (>=2) masklets
+            # if there are multiple matched track ids, we need to find the one that appeared first;
+            # these later appearing ids may be removed since they may be considered as duplicates
+            first_appear_obj_id = (
+                min(matched_trk_obj_ids, key=lambda x: obj_first_frame_idx[x])
+                if not reverse
+                else max(matched_trk_obj_ids, key=lambda x: obj_first_frame_idx[x])
+            )
+            for obj_id in matched_trk_obj_ids:
+                if obj_id != first_appear_obj_id:
+                    key = (first_appear_obj_id, obj_id)
+                    overlap_pair_to_frame_inds[key].append(frame_idx)
+
+        # b) remove a masklet if it first appears after `hotstart_diff` and it overlaps with another
+        # masklet (that appears earlier) for more than `self.hotstart_dup_thresh` frames
+        for (first_obj_id, obj_id), frame_indices in overlap_pair_to_frame_inds.items():
+            if obj_id in removed_obj_ids or obj_id in obj_ids_newly_removed:
+                continue  # skip if the object is already removed
+            if (obj_first_frame_idx[obj_id] > hotstart_diff and not reverse) or (
+                obj_first_frame_idx[obj_id] < hotstart_diff and reverse
+            ):
+                if len(frame_indices) >= self.hotstart_dup_thresh:
+                    obj_ids_newly_removed.add(obj_id)
+                    logger.debug(
+                        f"Removing object {obj_id} at frame {frame_idx} "
+                        f"since it overlaps with another track {first_obj_id} at frames: {frame_indices}"
+                    )
+
+        removed_obj_ids.update(obj_ids_newly_removed)
+        return obj_ids_newly_removed, rank0_metadata
+
+    def _tracker_update_memories(
+        self,
+        tracker_inference_states: List[Any],
+        frame_idx: int,
+        tracker_metadata: Dict[str, Any],
+        low_res_masks: Tensor,
+    ):
+        """
+        Run Sam2 memory encoder, enforcing non-overlapping constraints globally.
+        """
+        if len(tracker_inference_states) == 0:
+            return
+        # Avoid an extra interpolation step by directly interpolating to `interpol_size`
+        high_res_H, high_res_W = (
+            self.tracker.maskmem_backbone.mask_downsampler.interpol_size
+        )
+        # NOTE: inspect this part if we observe OOMs in the demo
+        high_res_masks = F.interpolate(
+            low_res_masks.unsqueeze(1),
+            size=(high_res_H, high_res_W),
+            mode="bilinear",
+            align_corners=False,
+        )
+        # We first apply non-overlapping constraints before memory encoding. This may include some suppression heuristics.
+        if not hasattr(self, "_warm_up_complete") or self._warm_up_complete:
+            high_res_masks = self.tracker._suppress_object_pw_area_shrinkage(
+                high_res_masks
+            )
+        # Instead of gathering the predicted object scores, we use mask areas as a proxy.
+        object_score_logits = torch.where(
+            (high_res_masks > 0).any(dim=(-1, -2)), 10.0, -10.0
+        )
+
+        # Run the memory encoder on local slices for each GPU
+        start_idx_gpu = sum(tracker_metadata["num_obj_per_gpu"][: self.rank])
+        start_idx_state = start_idx_gpu
+        for tracker_state in tracker_inference_states:
+            num_obj_per_state = len(tracker_state["obj_ids"])
+            if num_obj_per_state == 0:
+                continue
+            # Get the local high-res masks and object score logits for this inference state
+            end_idx_state = start_idx_state + num_obj_per_state
+            local_high_res_masks = high_res_masks[start_idx_state:end_idx_state]
+            local_object_score_logits = object_score_logits[
+                start_idx_state:end_idx_state
+            ]
+            local_batch_size = local_high_res_masks.size(0)
+            # Run Sam2 memory encoder. Note that we do not re-enforce the non-overlapping constraint as it is turned off by default
+
+            encoded_mem = self.tracker._run_memory_encoder(
+                tracker_state,
+                frame_idx,
+                local_batch_size,
+                local_high_res_masks,
+                local_object_score_logits,
+                is_mask_from_pts=False,
+            )
+            local_maskmem_features, local_maskmem_pos_enc = encoded_mem
+            # Store encoded memories in the local inference state
+            output_dict = tracker_state["output_dict"]
+            for storage_key in ["cond_frame_outputs", "non_cond_frame_outputs"]:
+                if frame_idx not in output_dict[storage_key]:
+                    continue
+                output_dict[storage_key][frame_idx]["maskmem_features"] = (
+                    local_maskmem_features
+                )
+                output_dict[storage_key][frame_idx]["maskmem_pos_enc"] = [
+                    pos for pos in local_maskmem_pos_enc
+                ]
+                # for batched inference state, we also need to add per-object
+                # memory slides to support instance interactivity
+                self.tracker._add_output_per_object(
+                    inference_state=tracker_state,
+                    frame_idx=frame_idx,
+                    current_out=output_dict[storage_key][frame_idx],
+                    storage_key=storage_key,
+                )
+            start_idx_state += num_obj_per_state
+
+    def _tracker_add_new_objects(
+        self,
+        frame_idx: int,
+        num_frames: int,
+        new_obj_ids: List[int],
+        new_obj_masks: Tensor,
+        tracker_states_local: List[Any],
+        orig_vid_height: int,
+        orig_vid_width: int,
+        feature_cache: Dict,
+    ):
+        """Add a new object to SAM2 inference states."""
+        prev_tracker_state = (
+            tracker_states_local[0] if len(tracker_states_local) > 0 else None
+        )
+
+        # prepare inference_state
+        # batch objects that first appear on the same frame together
+        # Clear inference state. Keep the cached image features if available.
+        new_tracker_state = self.tracker.init_state(
+            cached_features=feature_cache,
+            video_height=orig_vid_height,
+            video_width=orig_vid_width,
+            num_frames=num_frames,
+        )
+        new_tracker_state["backbone_out"] = (
+            prev_tracker_state.get("backbone_out", None)
+            if prev_tracker_state is not None
+            else None
+        )
+
+        assert len(new_obj_ids) == new_obj_masks.size(0)
+        assert new_obj_masks.is_floating_point()
+        input_mask_res = self.tracker.input_mask_size
+        new_obj_masks = F.interpolate(
+            new_obj_masks.unsqueeze(1),
+            size=(input_mask_res, input_mask_res),
+            mode="bilinear",
+            align_corners=False,
+        ).squeeze(1)
+        new_obj_masks = new_obj_masks > 0
+
+        # add object one by one
+        for new_obj_id, new_mask in zip(new_obj_ids, new_obj_masks):
+            self.tracker.add_new_mask(
+                inference_state=new_tracker_state,
+                frame_idx=frame_idx,
+                obj_id=new_obj_id,
+                mask=new_mask,
+                add_mask_to_memory=True,
+            )
+        # NOTE: we skip enforcing the non-overlapping constraint **globally** when adding new objects.
+        self.tracker.propagate_in_video_preflight(
+            new_tracker_state, run_mem_encoder=True
+        )
+        tracker_states_local.append(new_tracker_state)
+        return tracker_states_local
+
+    def _tracker_remove_object(self, tracker_states_local: List[Any], obj_id: int):
+        """
+        Remove an object from SAM2 inference states. This would remove the object from
+        all frames in the video.
+        """
+        tracker_states_local_before_removal = tracker_states_local.copy()
+        tracker_states_local.clear()
+        for tracker_inference_state in tracker_states_local_before_removal:
+            # we try to remove `obj_id` on every inference state with `strict=False`
+            # it will not do anything if an inference state doesn't contain `obj_id`
+            new_obj_ids, _ = self.tracker.remove_object(
+                tracker_inference_state, obj_id, strict=False, need_output=False
+            )
+            # only keep an inference state if it's non-empty after object removal
+            if len(new_obj_ids) > 0:
+                tracker_states_local.append(tracker_inference_state)
+
+    def _tracker_remove_objects(
+        self, tracker_states_local: List[Any], obj_ids: list[int]
+    ):
+        """
+        Remove an object from SAM2 inference states. This would remove the object from
+        all frames in the video.
+        """
+        for obj_id in obj_ids:
+            self._tracker_remove_object(tracker_states_local, obj_id)
+
+    def _initialize_metadata(self):
+        """Initialize metadata for the masklets."""
+        tracker_metadata = {
+            "obj_ids_per_gpu": [np.array([], np.int64) for _ in range(self.world_size)],
+            "obj_ids_all_gpu": np.array([], np.int64),
+            "num_obj_per_gpu": np.zeros(self.world_size, np.int64),
+            "max_obj_id": -1,
+            "obj_id_to_score": {},
+            "obj_id_to_tracker_score_frame_wise": defaultdict(dict),
+            "obj_id_to_last_occluded": {},
+        }
+        if self.rank == 0:
+            # "rank0_metadata" contains metadata that is only stored on (and accessible to) GPU 0
+            # - obj_first_frame_idx: obj_id --> first frame index where the object was detected
+            # - unmatched_frame_inds: obj_id --> [mismatched frame indices]
+            # - overlap_pair_to_frame_inds: (first_appear_obj_id, obj_id) --> [overlap frame indices]
+            # - removed_obj_ids: object IDs that are suppressed via hot-start
+            rank0_metadata = {
+                "obj_first_frame_idx": {},
+                "unmatched_frame_inds": defaultdict(list),
+                "trk_keep_alive": defaultdict(
+                    int
+                ),  # This is used only for object suppression not for removal
+                "overlap_pair_to_frame_inds": defaultdict(list),
+                "removed_obj_ids": set(),
+                "suppressed_obj_ids": defaultdict(
+                    set
+                ),  # frame_idx --> set of objects with suppressed outputs, but still continue to be tracked
+            }
+            if self.masklet_confirmation_enable:
+                # all the following are npt.NDArray with the same shape as `obj_ids_all_gpu`
+                rank0_metadata["masklet_confirmation"] = {
+                    # "status" is the confirmation status of each masklet (in `MaskletConfirmationStatus`)
+                    "status": np.array([], np.int64),
+                    # "consecutive_det_num" is the number of consecutive frames where the masklet is
+                    # detected by the detector (with a matched detection)
+                    "consecutive_det_num": np.array([], np.int64),
+                }
+            tracker_metadata["rank0_metadata"] = rank0_metadata
+
+        return tracker_metadata
+
+    def update_masklet_confirmation_status(
+        self,
+        rank0_metadata: Dict[str, Any],
+        obj_ids_all_gpu_prev: npt.NDArray,
+        obj_ids_all_gpu_updated: npt.NDArray,
+        det_to_matched_trk_obj_ids: Dict[int, npt.NDArray],
+        new_det_obj_ids: npt.NDArray,
+    ):
+        confirmation_data = rank0_metadata["masklet_confirmation"]
+
+        # a) first, expand "confirmation_data" to include new masklets added in this frame
+        status_prev = confirmation_data["status"]
+        consecutive_det_num_prev = confirmation_data["consecutive_det_num"]
+        assert (
+            status_prev.shape == obj_ids_all_gpu_prev.shape
+        ), f"Got {status_prev.shape} vs {obj_ids_all_gpu_prev.shape}"
+
+        obj_id_to_updated_idx = {
+            obj_id: idx for idx, obj_id in enumerate(obj_ids_all_gpu_updated)
+        }
+        prev_elem_is_in_updated = np.isin(obj_ids_all_gpu_prev, obj_ids_all_gpu_updated)
+        prev_elem_obj_ids_in_updated = obj_ids_all_gpu_prev[prev_elem_is_in_updated]
+        prev_elem_inds_in_updated = np.array(
+            [obj_id_to_updated_idx[obj_id] for obj_id in prev_elem_obj_ids_in_updated],
+            dtype=np.int64,
+        )
+        # newly added masklets are initialized to "UNCONFIRMED" status
+        unconfirmed_val = MaskletConfirmationStatus.UNCONFIRMED.value
+        status = np.full_like(obj_ids_all_gpu_updated, fill_value=unconfirmed_val)
+        status[prev_elem_inds_in_updated] = status_prev[prev_elem_is_in_updated]
+        consecutive_det_num = np.zeros_like(obj_ids_all_gpu_updated)
+        consecutive_det_num[prev_elem_inds_in_updated] = consecutive_det_num_prev[
+            prev_elem_is_in_updated
+        ]
+
+        # b) update the confirmation status of all masklets based on the current frame
+        # b.1) update "consecutive_det_num"
+        # "is_matched": whether a masklet is matched to a detection on this frame
+        is_matched = np.isin(obj_ids_all_gpu_updated, new_det_obj_ids)
+        for matched_trk_obj_ids in det_to_matched_trk_obj_ids.values():
+            is_matched |= np.isin(obj_ids_all_gpu_updated, matched_trk_obj_ids)
+        consecutive_det_num = np.where(is_matched, consecutive_det_num + 1, 0)
+
+        # b.2) update "status"
+        change_to_confirmed = (
+            consecutive_det_num >= self.masklet_confirmation_consecutive_det_thresh
+        )
+        status[change_to_confirmed] = MaskletConfirmationStatus.CONFIRMED.value
+
+        confirmation_data["status"] = status
+        confirmation_data["consecutive_det_num"] = consecutive_det_num
+        return rank0_metadata
+
+    def forward(self, input: BatchedDatapoint, is_inference: bool = False):
+        raise NotImplementedError("Evaluation outside demo is not implemented yet")
+
+    def _load_checkpoint(self, ckpt_path: str, strict: bool = True):
+        sd = torch.load(ckpt_path, map_location="cpu", weights_only=True)["model"]
+        missing_keys, unexpected_keys = self.load_state_dict(sd, strict=strict)
+        if len(missing_keys) > 0 or len(unexpected_keys) > 0:
+            logger.warning(f"Loaded ckpt with {missing_keys=}, {unexpected_keys=}")
+        else:
+            logger.info("Loaded ckpt successfully without missing or unexpected keys")
+
+    def prep_for_evaluator(self, video_frames, tracking_res, scores_labels):
+        """This method is only used for benchmark eval (not used in the demo)."""
+        num_frames = len(video_frames)
+        w, h = video_frames[0].size
+        zero_mask = torch.zeros((1, h, w), dtype=torch.bool)
+        object_ids = list(scores_labels.keys())
+        preds = {"scores": [], "labels": [], "boxes": [], "masks_rle": []}
+        for oid in object_ids:
+            o_masks = []
+            o_score = scores_labels[oid][0].item()
+            o_label = scores_labels[oid][1]
+            for frame_idx in range(num_frames):
+                if frame_idx not in tracking_res:
+                    o_masks.append(zero_mask)
+                else:
+                    o_masks.append(tracking_res[frame_idx].get(oid, zero_mask))
+
+            o_masks = torch.cat(o_masks, dim=0)  # (n_frames, H, W)
+            preds["scores"].append(o_score)
+            preds["labels"].append(o_label)
+            preds["boxes"].append(mask_to_box(o_masks.unsqueeze(1)).squeeze())
+            preds["masks_rle"].append(rle_encode(o_masks, return_areas=True))
+
+        preds["boxes"] = (
+            torch.stack(preds["boxes"], dim=0)
+            if len(preds["boxes"]) > 0
+            else torch.empty(
+                (0, num_frames, 4), dtype=torch.float32, device=self.device
+            )
+        )
+        preds["scores"] = (
+            torch.tensor(preds["scores"], device=self.device)
+            if len(preds["scores"]) > 0
+            else torch.empty((0,), device=self.device)
+        )
+        preds["per_frame_scores"] = preds["scores"]
+        preds["labels"] = (
+            torch.tensor(preds["labels"], device=self.device)
+            if len(preds["labels"]) > 0
+            else torch.empty((0,), device=self.device)
+        )
+        return preds
+
+    def _encode_prompt(self, **kwargs):
+        return self.detector._encode_prompt(**kwargs)
+
+    def _drop_new_det_with_obj_limit(self, new_det_fa_inds, det_scores_np, num_to_keep):
+        """
+        Drop a few new detections based on the maximum number of objects. We drop new objects based
+        on their detection scores, keeping the high-scoring ones and dropping the low-scoring ones.
+        """
+        assert 0 <= num_to_keep <= len(new_det_fa_inds)
+        if num_to_keep == 0:
+            return np.array([], np.int64)  # keep none
+        if num_to_keep == len(new_det_fa_inds):
+            return new_det_fa_inds  # keep all
+
+        # keep the top-scoring detections
+        score_order = np.argsort(det_scores_np[new_det_fa_inds])[::-1]
+        new_det_fa_inds = new_det_fa_inds[score_order[:num_to_keep]]
+        return new_det_fa_inds

detect_tools/sam3/sam3/model/sam3_video_inference.py

@@ -0,0 +1,1709 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates. All Rights Reserved
+
+import logging
+from collections import defaultdict
+
+import numpy as np
+import torch
+import torch.distributed as dist
+import torch.nn.functional as F
+
+from sam3 import perflib
+from sam3.logger import get_logger
+from sam3.model.act_ckpt_utils import clone_output_wrapper
+from sam3.model.box_ops import box_xywh_to_cxcywh, box_xyxy_to_xywh
+from sam3.model.data_misc import BatchedDatapoint, convert_my_tensors, FindStage
+from sam3.model.geometry_encoders import Prompt
+from sam3.model.io_utils import IMAGE_EXTS, load_resource_as_video_frames
+from sam3.model.sam3_tracker_utils import fill_holes_in_mask_scores
+from sam3.model.sam3_video_base import MaskletConfirmationStatus, Sam3VideoBase
+from sam3.model.utils.misc import copy_data_to_device
+from sam3.perflib.compile import compile_wrapper, shape_logging_wrapper
+from sam3.perflib.masks_ops import masks_to_boxes as perf_masks_to_boxes
+from torchvision.ops import masks_to_boxes
+from tqdm.auto import tqdm
+
+logger = get_logger(__name__)
+
+
+class Sam3VideoInference(Sam3VideoBase):
+    TEXT_ID_FOR_TEXT = 0
+    TEXT_ID_FOR_VISUAL = 1
+
+    def __init__(
+        self,
+        image_size=1008,
+        image_mean=(0.5, 0.5, 0.5),
+        image_std=(0.5, 0.5, 0.5),
+        compile_model=False,
+        **kwargs,
+    ):
+        """
+        hotstart_delay: int, the delay (in #frames) before the model starts to yield output, 0 to disable hotstart delay.
+        hotstart_unmatch_thresh: int, remove the object if it has this many unmatched frames within its hotstart_delay period.
+            If `hotstart_delay` is set to 0, this parameter is ignored.
+        hotstart_dup_thresh: int, remove the object if it has overlapped with another object this many frames within its hotstart_delay period.
+        """
+        super().__init__(**kwargs)
+        self.image_size = image_size
+        self.image_mean = image_mean
+        self.image_std = image_std
+        self.compile_model = compile_model
+
+    @torch.inference_mode()
+    def init_state(
+        self,
+        resource_path,
+        offload_video_to_cpu=False,
+        async_loading_frames=False,
+        video_loader_type="cv2",
+    ):
+        """Initialize an inference state from `resource_path` (an image or a video)."""
+        images, orig_height, orig_width = load_resource_as_video_frames(
+            resource_path=resource_path,
+            image_size=self.image_size,
+            offload_video_to_cpu=offload_video_to_cpu,
+            img_mean=self.image_mean,
+            img_std=self.image_std,
+            async_loading_frames=async_loading_frames,
+            video_loader_type=video_loader_type,
+        )
+        inference_state = {}
+        inference_state["image_size"] = self.image_size
+        inference_state["num_frames"] = len(images)
+        # the original video height and width, used for resizing final output scores
+        inference_state["orig_height"] = orig_height
+        inference_state["orig_width"] = orig_width
+        # values that don't change across frames (so we only need to hold one copy of them)
+        inference_state["constants"] = {}
+        # inputs on each frame
+        self._construct_initial_input_batch(inference_state, images)
+        # initialize extra states
+        inference_state["tracker_inference_states"] = []
+        inference_state["tracker_metadata"] = {}
+        inference_state["feature_cache"] = {}
+        inference_state["cached_frame_outputs"] = {}
+        inference_state["action_history"] = []  # for logging user actions
+        inference_state["is_image_only"] = is_image_type(resource_path)
+        return inference_state
+
+    @torch.inference_mode()
+    def reset_state(self, inference_state):
+        """Revert `inference_state` to what it was right after initialization."""
+        inference_state["input_batch"].find_text_batch[0] = "<text placeholder>"
+        inference_state["text_prompt"] = None
+        for t in range(inference_state["num_frames"]):
+            inference_state["input_batch"].find_inputs[t].text_ids[...] = 0
+            # constructing an output list in inference state (we start with an empty list)
+            inference_state["previous_stages_out"][t] = None
+            inference_state["per_frame_raw_point_input"][t] = None
+            inference_state["per_frame_raw_box_input"][t] = None
+            inference_state["per_frame_visual_prompt"][t] = None
+            inference_state["per_frame_geometric_prompt"][t] = None
+            inference_state["per_frame_cur_step"][t] = 0
+
+        inference_state["visual_prompt_embed"] = None
+        inference_state["visual_prompt_mask"] = None
+        inference_state["tracker_inference_states"].clear()
+        inference_state["tracker_metadata"].clear()
+        inference_state["feature_cache"].clear()
+        inference_state["cached_frame_outputs"].clear()
+        inference_state["action_history"].clear()  # for logging user actions
+
+    def _construct_initial_input_batch(self, inference_state, images):
+        """Construct an initial `BatchedDatapoint` instance as input."""
+        # 1) img_batch
+        num_frames = len(images)
+        device = self.device
+
+        # 2) find_text_batch
+        # "<text placeholder>" will be replaced by the actual text prompt when adding prompts
+        find_text_batch = ["<text placeholder>", "visual"]
+
+        # 3) find_inputs
+        input_box_embedding_dim = 258  # historical default
+        input_points_embedding_dim = 257  # historical default
+        stages = [
+            FindStage(
+                img_ids=[stage_id],
+                text_ids=[0],
+                input_boxes=[torch.zeros(input_box_embedding_dim)],
+                input_boxes_mask=[torch.empty(0, dtype=torch.bool)],
+                input_boxes_label=[torch.empty(0, dtype=torch.long)],
+                input_points=[torch.empty(0, input_points_embedding_dim)],
+                input_points_mask=[torch.empty(0)],
+                object_ids=[],
+            )
+            for stage_id in range(num_frames)
+        ]
+        for i in range(len(stages)):
+            stages[i] = convert_my_tensors(stages[i])
+
+        # construct the final `BatchedDatapoint` and cast to GPU
+        input_batch = BatchedDatapoint(
+            img_batch=images,
+            find_text_batch=find_text_batch,
+            find_inputs=stages,
+            find_targets=[None] * num_frames,
+            find_metadatas=[None] * num_frames,
+        )
+        input_batch = copy_data_to_device(input_batch, device, non_blocking=True)
+        inference_state["input_batch"] = input_batch
+
+        # construct the placeholder interactive prompts and tracking queries
+        bs = 1
+        inference_state["constants"]["empty_geometric_prompt"] = Prompt(
+            box_embeddings=torch.zeros(0, bs, 4, device=device),
+            box_mask=torch.zeros(bs, 0, device=device, dtype=torch.bool),
+            box_labels=torch.zeros(0, bs, device=device, dtype=torch.long),
+            point_embeddings=torch.zeros(0, bs, 2, device=device),
+            point_mask=torch.zeros(bs, 0, device=device, dtype=torch.bool),
+            point_labels=torch.zeros(0, bs, device=device, dtype=torch.long),
+        )
+
+        # constructing an output list in inference state (we start with an empty list)
+        inference_state["previous_stages_out"] = [None] * num_frames
+        inference_state["text_prompt"] = None
+        inference_state["per_frame_raw_point_input"] = [None] * num_frames
+        inference_state["per_frame_raw_box_input"] = [None] * num_frames
+        inference_state["per_frame_visual_prompt"] = [None] * num_frames
+        inference_state["per_frame_geometric_prompt"] = [None] * num_frames
+        inference_state["per_frame_cur_step"] = [0] * num_frames
+
+        # placeholders for cached outputs
+        # (note: currently, a single visual prompt embedding is shared for all frames)
+        inference_state["visual_prompt_embed"] = None
+        inference_state["visual_prompt_mask"] = None
+
+    def _get_visual_prompt(self, inference_state, frame_idx, boxes_cxcywh, box_labels):
+        """
+        Handle the case of visual prompt. Currently, in the inference API we do not
+        explicitly distinguish between initial box as visual prompt vs subsequent boxes
+        or boxes after inference for refinement.
+        """
+        # If the frame hasn't had any inference results before (prompting or propagation),
+        # we treat the first added box prompt as a visual prompt; otherwise, we treat
+        # the first box just as a refinement prompt.
+        is_new_visual_prompt = (
+            inference_state["per_frame_visual_prompt"][frame_idx] is None
+            and inference_state["previous_stages_out"][frame_idx] is None
+        )
+        if is_new_visual_prompt:
+            if boxes_cxcywh.size(0) != 1:
+                raise RuntimeError(
+                    "visual prompts (box as an initial prompt) should only have one box, "
+                    f"but got {boxes_cxcywh.shape=}"
+                )
+            if not box_labels.item():
+                logging.warning("A negative box is added as a visual prompt.")
+            # take the first box prompt as a visual prompt
+            device = self.device
+            new_visual_prompt = Prompt(
+                box_embeddings=boxes_cxcywh[None, 0:1, :].to(device),  # (seq, bs, 4)
+                box_mask=None,
+                box_labels=box_labels[None, 0:1].to(device),  # (seq, bs)
+                point_embeddings=None,
+                point_mask=None,
+                point_labels=None,
+            )
+            inference_state["per_frame_visual_prompt"][frame_idx] = new_visual_prompt
+        else:
+            new_visual_prompt = None
+
+        # `boxes_cxcywh` and `box_labels` contains all the raw box inputs added so far
+        # strip any visual prompt from the input boxes (for geometric prompt encoding)
+        if inference_state["per_frame_visual_prompt"][frame_idx] is not None:
+            boxes_cxcywh = boxes_cxcywh[1:]
+            box_labels = box_labels[1:]
+
+        return boxes_cxcywh, box_labels, new_visual_prompt
+
+    def _get_processing_order(
+        self, inference_state, start_frame_idx, max_frame_num_to_track, reverse
+    ):
+        num_frames = inference_state["num_frames"]
+        previous_stages_out = inference_state["previous_stages_out"]
+        if all(out is None for out in previous_stages_out) and start_frame_idx is None:
+            raise RuntimeError(
+                "No prompts are received on any frames. Please add prompt on at least one frame before propagation."
+            )
+        # set start index, end index, and processing order
+        if start_frame_idx is None:
+            # default: start from the earliest frame with input points
+            start_frame_idx = min(
+                t for t, out in enumerate(previous_stages_out) if out is not None
+            )
+        if max_frame_num_to_track is None:
+            # default: track all the frames in the video
+            max_frame_num_to_track = num_frames
+        if reverse:
+            end_frame_idx = start_frame_idx - max_frame_num_to_track
+            end_frame_idx = max(end_frame_idx, 0)
+            processing_order = range(start_frame_idx - 1, end_frame_idx - 1, -1)
+        else:
+            end_frame_idx = start_frame_idx + max_frame_num_to_track
+            end_frame_idx = min(end_frame_idx, num_frames - 1)
+            processing_order = range(start_frame_idx, end_frame_idx + 1)
+        return processing_order, end_frame_idx
+
+    @torch.inference_mode()
+    def propagate_in_video(
+        self,
+        inference_state,
+        start_frame_idx=None,
+        max_frame_num_to_track=None,
+        reverse=False,
+    ):
+        """
+        Propagate the prompts to get grounding results for the entire video. This method
+        is a generator and yields inference outputs for all frames in the range specified
+        by `start_frame_idx`, `max_frame_num_to_track`, and `reverse`.
+        """
+        # compile the model (it's a no-op if the model is already compiled)
+        # note that it's intentionally added to `self.propagate_in_video`, so that the first
+        # `self.add_prompt` call will be done in eager mode to fill in the decoder buffers
+        # such as positional encoding cache)
+        self._compile_model()
+
+        processing_order, end_frame_idx = self._get_processing_order(
+            inference_state,
+            start_frame_idx,
+            max_frame_num_to_track,
+            reverse=reverse,
+        )
+
+        # Store max_frame_num_to_track in feature_cache for downstream methods
+        inference_state["feature_cache"]["tracking_bounds"] = {
+            "max_frame_num_to_track": max_frame_num_to_track,
+            "propagate_in_video_start_frame_idx": start_frame_idx,
+        }
+
+        hotstart_buffer = []
+        hotstart_removed_obj_ids = set()
+        # when deciding whether to output a masklet on `yield_frame_idx`, we check whether the object is confirmed
+        # in a future frame (`unconfirmed_frame_delay` frames after the current frame). For example, if we require
+        # an object to be detected in 3 consecutive frames to be confirmed, then we look 2 frames in the future --
+        # e.g., we output an object on frame 4 only if it becomes confirmed on frame 6.
+        unconfirmed_status_delay = self.masklet_confirmation_consecutive_det_thresh - 1
+        unconfirmed_obj_ids_per_frame = {}  # frame_idx -> hidden_obj_ids
+        for frame_idx in tqdm(
+            processing_order, desc="propagate_in_video", disable=self.rank > 0
+        ):
+            out = self._run_single_frame_inference(inference_state, frame_idx, reverse)
+
+            if self.hotstart_delay > 0:
+                # accumulate the outputs for the first `hotstart_delay` frames
+                hotstart_buffer.append([frame_idx, out])
+                # update the object IDs removed by hotstart so that we don't output them
+                if self.rank == 0:
+                    hotstart_removed_obj_ids.update(out["removed_obj_ids"])
+                    unconfirmed_obj_ids = out.get("unconfirmed_obj_ids", None)
+                    if unconfirmed_obj_ids is not None:
+                        unconfirmed_obj_ids_per_frame[frame_idx] = unconfirmed_obj_ids
+
+                if frame_idx == end_frame_idx:
+                    # we reached the end of propagation -- yield all frames in the buffer
+                    yield_list = hotstart_buffer
+                    hotstart_buffer = []
+                elif len(hotstart_buffer) >= self.hotstart_delay:
+                    # we have enough frames -- yield and remove the first (oldest) frame from the buffer
+                    yield_list = hotstart_buffer[:1]
+                    hotstart_buffer = hotstart_buffer[1:]
+                else:
+                    # not enough frames yet -- skip yielding
+                    yield_list = []
+            else:
+                yield_list = [(frame_idx, out)]  # output the current frame
+
+            for yield_frame_idx, yield_out in yield_list:
+                # post-process the output and yield it
+                if self.rank == 0:
+                    suppressed_obj_ids = yield_out["suppressed_obj_ids"]
+                    unconfirmed_status_frame_idx = (
+                        yield_frame_idx + unconfirmed_status_delay
+                        if not reverse
+                        else yield_frame_idx - unconfirmed_status_delay
+                    )
+
+                    # Clamp the frame index to stay within video bounds
+                    num_frames = inference_state["num_frames"]
+                    unconfirmed_status_frame_idx = max(
+                        0, min(unconfirmed_status_frame_idx, num_frames - 1)
+                    )
+
+                    unconfirmed_obj_ids = unconfirmed_obj_ids_per_frame.get(
+                        unconfirmed_status_frame_idx, None
+                    )
+                    postprocessed_out = self._postprocess_output(
+                        inference_state,
+                        yield_out,
+                        hotstart_removed_obj_ids,
+                        suppressed_obj_ids,
+                        unconfirmed_obj_ids,
+                    )
+
+                    self._cache_frame_outputs(
+                        inference_state,
+                        yield_frame_idx,
+                        yield_out["obj_id_to_mask"],
+                        suppressed_obj_ids=suppressed_obj_ids,
+                        removed_obj_ids=hotstart_removed_obj_ids,
+                        unconfirmed_obj_ids=unconfirmed_obj_ids,
+                    )
+                else:
+                    postprocessed_out = None  # no output on other GPUs
+                yield yield_frame_idx, postprocessed_out
+
+    def _run_single_frame_inference(self, inference_state, frame_idx, reverse):
+        """
+        Perform inference on a single frame and get its inference results. This would
+        also update `inference_state`.
+        """
+        # prepare inputs
+        input_batch = inference_state["input_batch"]
+        tracker_states_local = inference_state["tracker_inference_states"]
+        has_text_prompt = inference_state["text_prompt"] is not None
+        has_geometric_prompt = (
+            inference_state["per_frame_geometric_prompt"][frame_idx] is not None
+        )
+        # run inference for the current frame
+        (
+            obj_id_to_mask,
+            obj_id_to_score,
+            tracker_states_local_new,
+            tracker_metadata_new,
+            frame_stats,
+            _,
+        ) = self._det_track_one_frame(
+            frame_idx=frame_idx,
+            num_frames=inference_state["num_frames"],
+            reverse=reverse,
+            input_batch=input_batch,
+            geometric_prompt=(
+                inference_state["constants"]["empty_geometric_prompt"]
+                if not has_geometric_prompt
+                else inference_state["per_frame_geometric_prompt"][frame_idx]
+            ),
+            tracker_states_local=tracker_states_local,
+            tracker_metadata_prev=inference_state["tracker_metadata"],
+            feature_cache=inference_state["feature_cache"],
+            orig_vid_height=inference_state["orig_height"],
+            orig_vid_width=inference_state["orig_width"],
+            is_image_only=inference_state["is_image_only"],
+            allow_new_detections=has_text_prompt or has_geometric_prompt,
+        )
+        # update inference state
+        inference_state["tracker_inference_states"] = tracker_states_local_new
+        inference_state["tracker_metadata"] = tracker_metadata_new
+        # use a dummy string in "previous_stages_out" to indicate this frame has outputs
+        inference_state["previous_stages_out"][frame_idx] = "_THIS_FRAME_HAS_OUTPUTS_"
+
+        if self.rank == 0:
+            self._cache_frame_outputs(inference_state, frame_idx, obj_id_to_mask)
+
+        out = {
+            "obj_id_to_mask": obj_id_to_mask,
+            "obj_id_to_score": obj_id_to_score,  # first frame detection score
+            "obj_id_to_tracker_score": tracker_metadata_new[
+                "obj_id_to_tracker_score_frame_wise"
+            ][frame_idx],
+        }
+        # removed_obj_ids is only needed on rank 0 to handle hotstart delay buffer
+        if self.rank == 0:
+            rank0_metadata = tracker_metadata_new["rank0_metadata"]
+            removed_obj_ids = rank0_metadata["removed_obj_ids"]
+            out["removed_obj_ids"] = removed_obj_ids
+            out["suppressed_obj_ids"] = rank0_metadata["suppressed_obj_ids"][frame_idx]
+            out["frame_stats"] = frame_stats
+            if self.masklet_confirmation_enable:
+                status = rank0_metadata["masklet_confirmation"]["status"]
+                is_unconfirmed = status == MaskletConfirmationStatus.UNCONFIRMED.value
+                out["unconfirmed_obj_ids"] = tracker_metadata_new["obj_ids_all_gpu"][
+                    is_unconfirmed
+                ].tolist()
+            else:
+                out["unconfirmed_obj_ids"] = []
+
+        return out
+
+    def _postprocess_output(
+        self,
+        inference_state,
+        out,
+        removed_obj_ids=None,
+        suppressed_obj_ids=None,
+        unconfirmed_obj_ids=None,
+    ):
+        obj_id_to_mask = out["obj_id_to_mask"]  # low res masks
+        curr_obj_ids = sorted(obj_id_to_mask.keys())
+        H_video, W_video = inference_state["orig_height"], inference_state["orig_width"]
+        if len(curr_obj_ids) == 0:
+            out_obj_ids = torch.zeros(0, dtype=torch.int64)
+            out_probs = torch.zeros(0, dtype=torch.float32)
+            out_binary_masks = torch.zeros(0, H_video, W_video, dtype=torch.bool)
+            out_boxes_xywh = torch.zeros(0, 4, dtype=torch.float32)
+        else:
+            out_obj_ids = torch.tensor(curr_obj_ids, dtype=torch.int64)
+            out_probs = torch.tensor(
+                [out["obj_id_to_score"][obj_id] for obj_id in curr_obj_ids]
+            )
+            out_tracker_probs = torch.tensor(
+                [
+                    (
+                        out["obj_id_to_tracker_score"][obj_id]
+                        if obj_id in out["obj_id_to_tracker_score"]
+                        else 0.0
+                    )
+                    for obj_id in curr_obj_ids
+                ]
+            )
+            out_binary_masks = torch.cat(
+                [obj_id_to_mask[obj_id] for obj_id in curr_obj_ids], dim=0
+            )
+
+            assert out_binary_masks.dtype == torch.bool
+            keep = out_binary_masks.any(dim=(1, 2)).cpu()  # remove masks with 0 areas
+            # hide outputs for those object IDs in `obj_ids_to_hide`
+            obj_ids_to_hide = []
+            if suppressed_obj_ids is not None:
+                obj_ids_to_hide.extend(suppressed_obj_ids)
+            if removed_obj_ids is not None:
+                obj_ids_to_hide.extend(removed_obj_ids)
+            if unconfirmed_obj_ids is not None:
+                obj_ids_to_hide.extend(unconfirmed_obj_ids)
+            if len(obj_ids_to_hide) > 0:
+                obj_ids_to_hide_t = torch.tensor(obj_ids_to_hide, dtype=torch.int64)
+                keep &= ~torch.isin(out_obj_ids, obj_ids_to_hide_t)
+
+            # slice those valid entries from the original outputs
+            keep_idx = torch.nonzero(keep, as_tuple=True)[0]
+            keep_idx_gpu = keep_idx.pin_memory().to(
+                device=out_binary_masks.device, non_blocking=True
+            )
+
+            out_obj_ids = torch.index_select(out_obj_ids, 0, keep_idx)
+            out_probs = torch.index_select(out_probs, 0, keep_idx)
+            out_tracker_probs = torch.index_select(out_tracker_probs, 0, keep_idx)
+            out_binary_masks = torch.index_select(out_binary_masks, 0, keep_idx_gpu)
+
+            if perflib.is_enabled:
+                out_boxes_xyxy = perf_masks_to_boxes(
+                    out_binary_masks, out_obj_ids.tolist()
+                )
+            else:
+                out_boxes_xyxy = masks_to_boxes(out_binary_masks)
+
+            out_boxes_xywh = box_xyxy_to_xywh(out_boxes_xyxy)  # convert to xywh format
+            # normalize boxes
+            out_boxes_xywh[..., 0] /= W_video
+            out_boxes_xywh[..., 1] /= H_video
+            out_boxes_xywh[..., 2] /= W_video
+            out_boxes_xywh[..., 3] /= H_video
+
+        # apply non-overlapping constraints on the existing masklets
+        if out_binary_masks.shape[0] > 1:
+            assert len(out_binary_masks) == len(out_tracker_probs)
+            out_binary_masks = (
+                self.tracker._apply_object_wise_non_overlapping_constraints(
+                    out_binary_masks.unsqueeze(1),
+                    out_tracker_probs.unsqueeze(1).to(out_binary_masks.device),
+                    background_value=0,
+                ).squeeze(1)
+            ) > 0
+
+        outputs = {
+            "out_obj_ids": out_obj_ids.cpu().numpy(),
+            "out_probs": out_probs.cpu().numpy(),
+            "out_boxes_xywh": out_boxes_xywh.cpu().numpy(),
+            "out_binary_masks": out_binary_masks.cpu().numpy(),
+            "frame_stats": out.get("frame_stats", None),
+        }
+        return outputs
+
+    def _cache_frame_outputs(
+        self,
+        inference_state,
+        frame_idx,
+        obj_id_to_mask,
+        suppressed_obj_ids=None,
+        removed_obj_ids=None,
+        unconfirmed_obj_ids=None,
+    ):
+        # Filter out suppressed, removed, and unconfirmed objects from the cache
+        filtered_obj_id_to_mask = obj_id_to_mask.copy()
+
+        objects_to_exclude = set()
+        if suppressed_obj_ids is not None:
+            objects_to_exclude.update(suppressed_obj_ids)
+        if removed_obj_ids is not None:
+            objects_to_exclude.update(removed_obj_ids)
+        if unconfirmed_obj_ids is not None:
+            objects_to_exclude.update(unconfirmed_obj_ids)
+
+        if objects_to_exclude:
+            for obj_id in objects_to_exclude:
+                if obj_id in filtered_obj_id_to_mask:
+                    del filtered_obj_id_to_mask[obj_id]
+
+        inference_state["cached_frame_outputs"][frame_idx] = filtered_obj_id_to_mask
+
+    def _build_tracker_output(
+        self, inference_state, frame_idx, refined_obj_id_to_mask=None
+    ):
+        assert (
+            "cached_frame_outputs" in inference_state
+            and frame_idx in inference_state["cached_frame_outputs"]
+        ), "No cached outputs found. Ensure normal propagation has run first to populate the cache."
+        cached_outputs = inference_state["cached_frame_outputs"][frame_idx]
+
+        obj_id_to_mask = cached_outputs.copy()
+
+        # Update with refined masks if provided
+        if refined_obj_id_to_mask is not None:
+            for obj_id, refined_mask in refined_obj_id_to_mask.items():
+                assert (
+                    refined_mask is not None
+                ), f"Refined mask data must be provided for obj_id {obj_id}"
+                obj_id_to_mask[obj_id] = refined_mask
+
+        return obj_id_to_mask
+
+    def _compile_model(self):
+        """Compile the SAM model with torch.compile for speedup."""
+        is_compiled = getattr(self, "_model_is_compiled", False)
+        if is_compiled or not self.compile_model:
+            return
+
+        import torch._dynamo
+
+        # a larger cache size to hold varying number of shapes for torch.compile
+        # see https://github.com/pytorch/pytorch/blob/v2.5.1/torch/_dynamo/config.py#L42-L49
+        torch._dynamo.config.cache_size_limit = 128
+        torch._dynamo.config.accumulated_cache_size_limit = 2048
+        torch._dynamo.config.capture_scalar_outputs = True
+        torch._dynamo.config.suppress_errors = True
+
+        # Compile module components
+        # skip compilation of `_encode_prompt` since it sometimes tiggger SymInt errors
+        # self._encode_prompt = clone_output_wrapper(
+        #     torch.compile(self._encode_prompt, fullgraph=True, mode="max-autotune")
+        # )
+
+        ## Compile SAM3 model components
+        self.detector.backbone.vision_backbone.forward = clone_output_wrapper(
+            torch.compile(
+                self.detector.backbone.vision_backbone.forward,
+                fullgraph=True,
+                mode="max-autotune",
+            )
+        )
+        self.detector.transformer.encoder.forward = clone_output_wrapper(
+            torch.compile(
+                self.detector.transformer.encoder.forward,
+                fullgraph=True,
+                mode="max-autotune",
+            )
+        )
+        self.detector.transformer.decoder.forward = clone_output_wrapper(
+            torch.compile(
+                self.detector.transformer.decoder.forward,
+                fullgraph=True,
+                mode="max-autotune",
+                dynamic=False,
+            )
+        )
+
+        self.detector.segmentation_head.forward = clone_output_wrapper(
+            torch.compile(
+                self.detector.segmentation_head.forward,
+                fullgraph=True,
+                mode="max-autotune",
+            )
+        )
+
+        ## Compile Tracker model components
+        self.tracker.maskmem_backbone.forward = compile_wrapper(
+            self.tracker.maskmem_backbone.forward,
+            mode="max-autotune",
+            fullgraph=True,
+            dynamic=False,
+        )
+
+        self.tracker.transformer.encoder.forward = shape_logging_wrapper(
+            compile_wrapper(
+                self.tracker.transformer.encoder.forward,
+                mode="max-autotune-no-cudagraphs",
+                fullgraph=True,
+                dynamic=True,
+            ),
+            keep_kwargs=["src", "src_pos", "prompt", "prompt_pos"],
+        )
+
+        self.tracker.sam_mask_decoder.forward = compile_wrapper(
+            self.tracker.sam_mask_decoder.forward,
+            mode="max-autotune",
+            fullgraph=True,
+            dynamic=False,  # Accuracy regression on True
+        )
+
+        self._model_is_compiled = True
+
+    def _warm_up_vg_propagation(self, inference_state, start_frame_idx=0):
+        # use different tracking score thresholds for each round to simulate different number of output objects
+        num_objects_list = range(self.num_obj_for_compile + 1)
+        new_det_score_thresh_list = [0.3, 0.5, 0.7]
+        num_rounds = len(new_det_score_thresh_list)
+        orig_new_det_thresh = self.new_det_thresh
+
+        for i, thresh in enumerate(new_det_score_thresh_list):
+            self.new_det_thresh = thresh
+            for num_objects in num_objects_list:
+                logger.info(f"{i+1}/{num_rounds} warming up model compilation")
+                self.add_prompt(
+                    inference_state, frame_idx=start_frame_idx, text_str="cat"
+                )
+                logger.info(
+                    f"{i+1}/{num_rounds} warming up model compilation -- simulating {num_objects}/{self.num_obj_for_compile} objects"
+                )
+                inference_state = self.add_fake_objects_to_inference_state(
+                    inference_state, num_objects, frame_idx=start_frame_idx
+                )
+                inference_state["tracker_metadata"]["rank0_metadata"].update(
+                    {
+                        "masklet_confirmation": {
+                            "status": np.zeros(num_objects, dtype=np.int64),
+                            "consecutive_det_num": np.zeros(
+                                num_objects, dtype=np.int64
+                            ),
+                        }
+                    }
+                )
+                for _ in self.propagate_in_video(
+                    inference_state, start_frame_idx, reverse=False
+                ):
+                    pass
+                for _ in self.propagate_in_video(
+                    inference_state, start_frame_idx, reverse=True
+                ):
+                    pass
+                self.reset_state(inference_state)
+                logger.info(
+                    f"{i+1}/{num_rounds} warming up model compilation -- completed round {i+1} out of {num_rounds}"
+                )
+
+        # Warm up Tracker memory encoder with varying input shapes
+        num_iters = 3
+        feat_size = self.tracker.sam_image_embedding_size**2  # 72 * 72 = 5184
+        hidden_dim = self.tracker.hidden_dim  # 256
+        mem_dim = self.tracker.mem_dim  # 64
+        for _ in tqdm(range(num_iters)):
+            for b in range(1, self.num_obj_for_compile + 1):
+                for i in range(
+                    1,
+                    self.tracker.max_cond_frames_in_attn + self.tracker.num_maskmem,
+                ):
+                    for j in range(
+                        self.tracker.max_cond_frames_in_attn
+                        + self.tracker.max_obj_ptrs_in_encoder
+                    ):
+                        num_obj_ptr_tokens = (hidden_dim // mem_dim) * j
+                        src = torch.randn(feat_size, b, hidden_dim, device=self.device)
+                        src_pos = torch.randn(
+                            feat_size, b, hidden_dim, device=self.device
+                        )
+                        prompt = torch.randn(
+                            feat_size * i + num_obj_ptr_tokens,
+                            b,
+                            mem_dim,
+                            device=self.device,
+                        )
+                        prompt_pos = torch.randn(
+                            feat_size * i + num_obj_ptr_tokens,
+                            b,
+                            mem_dim,
+                            device=self.device,
+                        )
+
+                        self.tracker.transformer.encoder.forward(
+                            src=src,
+                            src_pos=src_pos,
+                            prompt=prompt,
+                            prompt_pos=prompt_pos,
+                            num_obj_ptr_tokens=num_obj_ptr_tokens,
+                        )
+
+        self.new_det_thresh = orig_new_det_thresh
+        return inference_state
+
+    def add_fake_objects_to_inference_state(
+        self, inference_state, num_objects, frame_idx
+    ):
+        new_det_obj_ids_local = np.arange(num_objects)
+        high_res_H, high_res_W = (
+            self.tracker.maskmem_backbone.mask_downsampler.interpol_size
+        )
+        new_det_masks = torch.ones(
+            len(new_det_obj_ids_local), high_res_H, high_res_W
+        ).to(self.device)
+
+        inference_state["tracker_inference_states"] = self._tracker_add_new_objects(
+            frame_idx=frame_idx,
+            num_frames=inference_state["num_frames"],
+            new_obj_ids=new_det_obj_ids_local,
+            new_obj_masks=new_det_masks,
+            tracker_states_local=inference_state["tracker_inference_states"],
+            orig_vid_height=inference_state["orig_height"],
+            orig_vid_width=inference_state["orig_width"],
+            feature_cache=inference_state["feature_cache"],
+        )
+
+        # Synthesize obj_id_to_mask data for cached_frame_outputs to support _build_tracker_output during warmup
+        obj_id_to_mask = {}
+        if num_objects > 0:
+            H_video = inference_state["orig_height"]
+            W_video = inference_state["orig_width"]
+
+            video_res_masks = F.interpolate(
+                new_det_masks.unsqueeze(1),  # Add channel dimension for interpolation
+                size=(H_video, W_video),
+                mode="bilinear",
+                align_corners=False,
+            )  # (num_objects, 1, H_video, W_video)
+            for i, obj_id in enumerate(new_det_obj_ids_local):
+                obj_id_to_mask[obj_id] = (video_res_masks[i] > 0.0).to(torch.bool)
+        if self.rank == 0:
+            for fidx in range(inference_state["num_frames"]):
+                self._cache_frame_outputs(inference_state, fidx, obj_id_to_mask)
+
+        inference_state["tracker_metadata"].update(
+            {
+                "obj_ids_per_gpu": [np.arange(num_objects)],
+                "obj_ids_all_gpu": np.arange(num_objects),  # Same as 1 GPU
+                "num_obj_per_gpu": [num_objects],
+                "obj_id_to_score": {i: 1.0 for i in range(num_objects)},
+                "max_obj_id": num_objects,
+                "rank0_metadata": {
+                    "masklet_confirmation": {
+                        "status": np.zeros(num_objects, dtype=np.int64),
+                        "consecutive_det_num": np.zeros(num_objects, dtype=np.int64),
+                    },
+                    "removed_obj_ids": set(),
+                    "suppressed_obj_ids": defaultdict(set),
+                },
+            }
+        )
+        return inference_state
+
+    @torch.inference_mode()
+    @torch.autocast(device_type="cuda", dtype=torch.bfloat16)
+    def warm_up_compilation(self):
+        """
+        Warm up the model by running a dummy inference to compile the model. This is
+        useful to avoid the compilation overhead in the first inference call.
+        """
+        if not self.compile_model:
+            return
+        self._warm_up_complete = False
+        if self.device.type != "cuda":
+            raise RuntimeError(
+                f"The model must be on CUDA for warm-up compilation, got {self.device=}."
+            )
+
+        # temporally set to single GPU temporarily for warm-up compilation
+        orig_rank = self.rank
+        orig_world_size = self.world_size
+        self.rank = self.detector.rank = 0
+        self.world_size = self.detector.world_size = 1
+        orig_recondition_every_nth_frame = self.recondition_every_nth_frame
+        # self.recondition_every_nth_frame = 2
+
+        # Get a random video
+        inference_state = self.init_state(resource_path="<load-dummy-video-30>")
+        start_frame_idx = 0
+
+        # Run basic propagation warm-up
+        inference_state = self._warm_up_vg_propagation(inference_state, start_frame_idx)
+
+        logger.info("Warm-up compilation completed.")
+
+        # revert to the original GPU and rank
+        self.rank = self.detector.rank = orig_rank
+        self.world_size = self.detector.world_size = orig_world_size
+        self.recondition_every_nth_frame = orig_recondition_every_nth_frame
+        self._warm_up_complete = True
+        self.tracker.transformer.encoder.forward.set_logging(True)
+
+    @torch.inference_mode()
+    def add_prompt(
+        self,
+        inference_state,
+        frame_idx,
+        text_str=None,
+        boxes_xywh=None,
+        box_labels=None,
+    ):
+        """
+        Add text, point or box prompts on a single frame. This method returns the inference
+        outputs only on the prompted frame.
+
+        Note that text prompts are NOT associated with a particular frame (i.e. they apply
+        to all frames). However, we only run inference on the frame specified in `frame_idx`.
+        """
+        logger.debug("Running add_prompt on frame %d", frame_idx)
+
+        num_frames = inference_state["num_frames"]
+        assert (
+            text_str is not None or boxes_xywh is not None
+        ), "at least one type of prompt (text, boxes) must be provided"
+        assert (
+            0 <= frame_idx < num_frames
+        ), f"{frame_idx=} is out of range for a total of {num_frames} frames"
+
+        # since it's a semantic prompt, we start over
+        self.reset_state(inference_state)
+
+        # 1) add text prompt
+        if text_str is not None and text_str != "visual":
+            inference_state["text_prompt"] = text_str
+            inference_state["input_batch"].find_text_batch[0] = text_str
+            text_id = self.TEXT_ID_FOR_TEXT
+        else:
+            inference_state["text_prompt"] = None
+            inference_state["input_batch"].find_text_batch[0] = "<text placeholder>"
+            text_id = self.TEXT_ID_FOR_VISUAL
+        for t in range(inference_state["num_frames"]):
+            inference_state["input_batch"].find_inputs[t].text_ids[...] = text_id
+
+        # 2) handle box prompt
+        assert (boxes_xywh is not None) == (box_labels is not None)
+        if boxes_xywh is not None:
+            boxes_xywh = torch.as_tensor(boxes_xywh, dtype=torch.float32)
+            box_labels = torch.as_tensor(box_labels, dtype=torch.long)
+            # input boxes are expected to be [xmin, ymin, width, height] format
+            # in normalized coordinates of range 0~1, similar to FA
+            assert boxes_xywh.dim() == 2
+            assert boxes_xywh.size(0) > 0 and boxes_xywh.size(-1) == 4
+            assert box_labels.dim() == 1 and box_labels.size(0) == boxes_xywh.size(0)
+            boxes_cxcywh = box_xywh_to_cxcywh(boxes_xywh)
+            assert (boxes_xywh >= 0).all().item() and (boxes_xywh <= 1).all().item()
+            assert (boxes_cxcywh >= 0).all().item() and (boxes_cxcywh <= 1).all().item()
+
+            new_box_input = boxes_cxcywh, box_labels
+            inference_state["per_frame_raw_box_input"][frame_idx] = new_box_input
+
+            # handle the case of visual prompt (also added as an input box from the UI)
+            boxes_cxcywh, box_labels, geometric_prompt = self._get_visual_prompt(
+                inference_state, frame_idx, boxes_cxcywh, box_labels
+            )
+
+            inference_state["per_frame_geometric_prompt"][frame_idx] = geometric_prompt
+
+        out = self._run_single_frame_inference(
+            inference_state, frame_idx, reverse=False
+        )
+        return frame_idx, self._postprocess_output(inference_state, out)
+
+    @torch.autocast(device_type="cuda", dtype=torch.bfloat16)
+    def forward(self, input: BatchedDatapoint, is_inference: bool = False):
+        """This method is only used for benchmark eval (not used in the demo)."""
+        # set the model to single GPU for benchmark evaluation (to be compatible with trainer)
+        orig_rank = self.rank
+        orig_world_size = self.world_size
+        self.rank = self.detector.rank = 0
+        self.world_size = self.detector.world_size = 1
+
+        # get data
+        text_prompt_ids = input.find_metadatas[0].original_category_id
+        text_prompt_list = input.find_text_batch
+
+        # loop over txt prompts
+        tracking_res = defaultdict(dict)  # frame_idx --> {obj_id: mask}
+        scores_labels = defaultdict(tuple)  # obj_id --> (score, text_prompt_id)
+        inference_state = self.init_state(resource_path=input.raw_images)
+        for prompt_id, prompt in zip(text_prompt_ids, text_prompt_list):
+            self.add_prompt(inference_state, frame_idx=0, text_str=prompt)
+            start_obj_id = max(scores_labels.keys(), default=-1) + 1  # prev max + 1
+
+            # propagate the prompts
+            obj_ids_this_prompt = set()
+            for frame_idx, out in self.propagate_in_video(
+                inference_state,
+                start_frame_idx=0,
+                max_frame_num_to_track=inference_state["num_frames"],
+                reverse=False,
+            ):
+                current_frame_res = tracking_res[frame_idx]
+                for obj_id, mask in zip(out["out_obj_ids"], out["out_binary_masks"]):
+                    mask_tensor = torch.tensor(mask[None], dtype=torch.bool)
+                    current_frame_res[obj_id + start_obj_id] = mask_tensor
+                obj_ids_this_prompt.update(current_frame_res.keys())
+
+            obj_id_to_score = inference_state["tracker_metadata"]["obj_id_to_score"]
+            for obj_id, score in obj_id_to_score.items():
+                if obj_id + start_obj_id in obj_ids_this_prompt:
+                    score_tensor = torch.tensor(score, dtype=torch.float32)
+                    scores_labels[obj_id + start_obj_id] = (score_tensor, prompt_id)
+
+            self.reset_state(inference_state)
+
+        video_id = input.find_metadatas[0].original_image_id[0].cpu().item()
+        preds = self.prep_for_evaluator(input.raw_images, tracking_res, scores_labels)
+
+        # revert the model to the original GPU and rank
+        self.rank = self.detector.rank = orig_rank
+        self.world_size = self.detector.world_size = orig_world_size
+        return {video_id: preds}
+
+    def back_convert(self, targets):
+        # Needed for retraining compatibility with trainer
+        return targets
+
+
+class Sam3VideoInferenceWithInstanceInteractivity(Sam3VideoInference):
+    def __init__(
+        self,
+        use_prev_mem_frame=False,
+        use_stateless_refinement=False,
+        refinement_detector_cond_frame_removal_window=16,
+        **kwargs,
+    ):
+        """
+        use_prev_mem_frame: bool, whether to condition on previous memory frames for adding points
+        use_stateless_refinement: bool, whether to enable stateless refinement behavior
+        refinement_detector_cond_frame_removal_window: int, we remove a detector conditioning frame if it
+            is within this many frames of a user refined frame. Set to a large value (e.g. 10000) to
+            always remove detector conditioning frames if there is any user refinement in the video.
+        """
+        super().__init__(**kwargs)
+        self.use_prev_mem_frame = use_prev_mem_frame
+        self.use_stateless_refinement = use_stateless_refinement
+        self.refinement_detector_cond_frame_removal_window = (
+            refinement_detector_cond_frame_removal_window
+        )
+
+    def _init_new_tracker_state(self, inference_state):
+        return self.tracker.init_state(
+            cached_features=inference_state["feature_cache"],
+            video_height=inference_state["orig_height"],
+            video_width=inference_state["orig_width"],
+            num_frames=inference_state["num_frames"],
+        )
+
+    @torch.inference_mode()
+    def propagate_in_video(
+        self,
+        inference_state,
+        start_frame_idx=None,
+        max_frame_num_to_track=None,
+        reverse=False,
+    ):
+        # step 1: check which type of propagation to run, should be the same for all GPUs.
+        propagation_type, obj_ids = self.parse_action_history_for_propagation(
+            inference_state
+        )
+        self.add_action_history(
+            inference_state,
+            action_type=propagation_type,
+            obj_ids=obj_ids,
+            frame_idx=start_frame_idx,
+        )
+
+        # step 2: run full VG propagation
+        if propagation_type == "propagation_full":
+            logger.debug(f"Running full VG propagation (reverse={reverse}).")
+            yield from super().propagate_in_video(
+                inference_state,
+                start_frame_idx=start_frame_idx,
+                max_frame_num_to_track=max_frame_num_to_track,
+                reverse=reverse,
+            )
+            return
+
+        # step 3: run Tracker partial propagation or direct fetch existing predictions
+        assert propagation_type in ["propagation_partial", "propagation_fetch"]
+        logger.debug(
+            f"Running Tracker propagation for objects {obj_ids} and merging it with existing VG predictions (reverse={reverse})."
+            if propagation_type == "propagation_partial"
+            else f"Fetching existing VG predictions without running any propagation (reverse={reverse})."
+        )
+        processing_order, _ = self._get_processing_order(
+            inference_state,
+            start_frame_idx=start_frame_idx,
+            max_frame_num_to_track=max_frame_num_to_track,
+            reverse=reverse,
+        )
+
+        tracker_metadata = inference_state["tracker_metadata"]
+
+        # if fetch just return from output
+        if propagation_type == "propagation_fetch":
+            for frame_idx in tqdm(processing_order):
+                if self.rank == 0:
+                    obj_id_to_mask = inference_state["cached_frame_outputs"].get(
+                        frame_idx, {}
+                    )
+                    # post processing - remove suppressed obj_ids
+                    obj_id_to_score = tracker_metadata["obj_id_to_score"]
+                    suppressed_obj_ids = tracker_metadata["rank0_metadata"][
+                        "suppressed_obj_ids"
+                    ][frame_idx]
+                    obj_id_to_tracker_score = tracker_metadata[
+                        "obj_id_to_tracker_score_frame_wise"
+                    ][frame_idx]
+
+                    out = {
+                        "obj_id_to_mask": obj_id_to_mask,
+                        "obj_id_to_score": obj_id_to_score,
+                        "obj_id_to_tracker_score": obj_id_to_tracker_score,
+                    }
+                    yield (
+                        frame_idx,
+                        self._postprocess_output(
+                            inference_state, out, suppressed_obj_ids=suppressed_obj_ids
+                        ),
+                    )
+                else:
+                    yield frame_idx, None
+
+            return
+
+        # get Tracker inference states containing selected obj_ids
+        if propagation_type == "propagation_partial":
+            # can be empty for GPUs where objects are not in their inference states
+            tracker_states_local = self._get_tracker_inference_states_by_obj_ids(
+                inference_state, obj_ids
+            )
+            for tracker_state in tracker_states_local:
+                self.tracker.propagate_in_video_preflight(
+                    tracker_state, run_mem_encoder=True
+                )
+
+        for frame_idx in tqdm(processing_order):
+            # run Tracker propagation
+            if propagation_type == "propagation_partial":
+                self._prepare_backbone_feats(inference_state, frame_idx, reverse)
+                obj_ids_local, low_res_masks_local, tracker_scores_local = (
+                    self._propogate_tracker_one_frame_local_gpu(
+                        tracker_states_local,
+                        frame_idx=frame_idx,
+                        reverse=reverse,
+                        run_mem_encoder=True,
+                    )
+                )
+
+                # broadcast refined object tracker scores and masks to all GPUs
+                # handle multiple objects that can be located on different GPUs
+                refined_obj_data = {}  # obj_id -> (score, mask_video_res)
+
+                # Collect data for objects on this GPU
+                local_obj_data = {}
+                for obj_id in obj_ids:
+                    obj_rank = self._get_gpu_id_by_obj_id(inference_state, obj_id)
+                    if self.rank == obj_rank and obj_id in obj_ids_local:
+                        refined_obj_idx = obj_ids_local.index(obj_id)
+                        refined_mask_low_res = low_res_masks_local[
+                            refined_obj_idx
+                        ]  # (H_low_res, W_low_res)
+                        refined_score = tracker_scores_local[refined_obj_idx]
+
+                        # Keep low resolution for broadcasting to reduce communication cost
+                        local_obj_data[obj_id] = (refined_score, refined_mask_low_res)
+
+                # Broadcast data from each GPU that has refined objects
+                if self.world_size > 1:
+                    for obj_id in obj_ids:
+                        obj_rank = self._get_gpu_id_by_obj_id(inference_state, obj_id)
+                        if self.rank == obj_rank:
+                            # This GPU has the object, broadcast its data
+                            data_to_broadcast = local_obj_data.get(obj_id, None)
+                            data_list = [
+                                (data_to_broadcast[0].cpu(), data_to_broadcast[1].cpu())
+                            ]
+                            self.broadcast_python_obj_cpu(data_list, src=obj_rank)
+                            if data_to_broadcast is not None:
+                                refined_obj_data[obj_id] = data_to_broadcast
+                        elif self.rank != obj_rank:
+                            # This GPU doesn't have the object, receive data
+                            data_list = [None]
+                            self.broadcast_python_obj_cpu(data_list, src=obj_rank)
+                            refined_obj_data[obj_id] = (
+                                data_list[0][0].to(self.device),
+                                data_list[0][1].to(self.device),
+                            )
+                else:
+                    # Single GPU case
+                    refined_obj_data = local_obj_data
+
+                # Update Tracker scores for all refined objects
+                for obj_id, (refined_score, _) in refined_obj_data.items():
+                    tracker_metadata["obj_id_to_tracker_score_frame_wise"][
+                        frame_idx
+                    ].update({obj_id: refined_score.item()})
+
+                if self.rank == 0:
+                    # get predictions from Tracker inference states, it includes the original
+                    # VG predictions and the refined predictions from interactivity.
+
+                    # Prepare refined masks dictionary - upscale to video resolution after broadcast
+                    refined_obj_id_to_mask = {}
+                    for obj_id, (_, refined_mask_low_res) in refined_obj_data.items():
+                        refined_mask_video_res = (
+                            self._convert_low_res_mask_to_video_res(
+                                refined_mask_low_res, inference_state
+                            )
+                        )  # (1, H_video, W_video) bool
+                        refined_obj_id_to_mask[obj_id] = refined_mask_video_res
+
+                    obj_id_to_mask = self._build_tracker_output(
+                        inference_state, frame_idx, refined_obj_id_to_mask
+                    )
+                    out = {
+                        "obj_id_to_mask": obj_id_to_mask,
+                        "obj_id_to_score": tracker_metadata["obj_id_to_score"],
+                        "obj_id_to_tracker_score": tracker_metadata[
+                            "obj_id_to_tracker_score_frame_wise"
+                        ][frame_idx],
+                    }
+                    suppressed_obj_ids = tracker_metadata["rank0_metadata"][
+                        "suppressed_obj_ids"
+                    ][frame_idx]
+                    self._cache_frame_outputs(
+                        inference_state,
+                        frame_idx,
+                        obj_id_to_mask,
+                        suppressed_obj_ids=suppressed_obj_ids,
+                    )
+                    suppressed_obj_ids = tracker_metadata["rank0_metadata"][
+                        "suppressed_obj_ids"
+                    ][frame_idx]
+                    yield (
+                        frame_idx,
+                        self._postprocess_output(
+                            inference_state, out, suppressed_obj_ids=suppressed_obj_ids
+                        ),
+                    )
+                else:
+                    yield frame_idx, None
+
+    def add_action_history(
+        self, inference_state, action_type, frame_idx=None, obj_ids=None
+    ):
+        """
+        action_history is used to automatically decide what to do during propagation.
+        action_type: one of ["add", "remove", "refine"] + ["propagation_full", "propagation_partial", "propagation_fetch"]
+        """
+        instance_actions = ["add", "remove", "refine"]
+        propagation_actions = [
+            "propagation_full",
+            "propagation_partial",
+            "propagation_fetch",
+        ]
+        assert (
+            action_type in instance_actions + propagation_actions
+        ), f"Invalid action type: {action_type}, must be one of {instance_actions + propagation_actions}"
+        action = {
+            "type": action_type,
+            "frame_idx": frame_idx,
+            "obj_ids": obj_ids,
+        }
+        inference_state["action_history"].append(action)
+
+    def _has_object_been_refined(self, inference_state, obj_id):
+        action_history = inference_state["action_history"]
+        for action in action_history:
+            if action["type"] in ["add", "refine"] and action.get("obj_ids"):
+                if obj_id in action["obj_ids"]:
+                    return True
+        return False
+
+    def parse_action_history_for_propagation(self, inference_state):
+        """
+        Parse the actions in history before the last propagation and prepare for the next propagation.
+        We support multiple actions (add/remove/refine) between two propagations. If we had an action
+        history similar to this ["propagate", "add", "refine", "remove", "add"], the next propagation
+        would remove the removed object, and also propagate the two added/refined objects.
+
+        Returns:
+            propagation_type: one of ["propagation_full", "propagation_partial", "propagation_fetch"]
+                - "propagation_full": run VG propagation for all objects
+                - "propagation_partial": run Tracker propagation for selected objects, useful for add/refine actions
+                - "propagation_fetch": fetch existing VG predictions without running any propagation
+            obj_ids: list of object ids to run Tracker propagation on if propagation_type is "propagation_partial".
+        """
+        action_history = inference_state["action_history"]
+        if len(action_history) == 0:
+            # we run propagation for the first time
+            return "propagation_full", None
+
+        if "propagation" in action_history[-1]["type"]:
+            if action_history[-1]["type"] in ["propagation_fetch"]:
+                # last propagation is direct fetch, we fetch existing predictions
+                return "propagation_fetch", None
+            elif action_history[-1]["type"] in [
+                "propagation_partial",
+                "propagation_full",
+            ]:
+                # we do fetch prediction if we have already run propagation twice or we have run
+                # propagation once and it is from the first frame or last frame.
+                if (
+                    len(action_history) > 1
+                    and action_history[-2]["type"]
+                    in ["propagation_partial", "propagation_full"]
+                ) or action_history[-1]["frame_idx"] in [
+                    0,
+                    inference_state["num_frames"] - 1,
+                ]:
+                    # we have run both forward and backward partial/full propagation
+                    return "propagation_fetch", None
+                else:
+                    # we have run partial/full forward or backward propagation once, need run it for the rest of the frames
+                    return action_history[-1]["type"], action_history[-1]["obj_ids"]
+
+        # parse actions since last propagation
+        obj_ids = []
+        for action in action_history[::-1]:
+            if "propagation" in action["type"]:
+                # we reached the last propagation action, stop parsing
+                break
+            if action["type"] in ["add", "refine"]:
+                obj_ids.extend(action["obj_ids"])
+            # else action["type"] == "remove": noop
+        obj_ids = list(set(obj_ids)) if len(obj_ids) > 0 else None
+        propagation_type = (
+            "propagation_partial" if obj_ids is not None else "propagation_fetch"
+        )
+        return propagation_type, obj_ids
+
+    def remove_object(self, inference_state, obj_id, is_user_action=False):
+        """
+        We try to remove object from tracker states on every GPU, it will do nothing
+        for states without this object.
+        """
+        obj_rank = self._get_gpu_id_by_obj_id(inference_state, obj_id)
+        assert obj_rank is not None, f"Object {obj_id} not found in any GPU."
+
+        tracker_states_local = inference_state["tracker_inference_states"]
+        if self.rank == obj_rank:
+            self._tracker_remove_object(tracker_states_local, obj_id)
+
+        if is_user_action:
+            self.add_action_history(
+                inference_state, action_type="remove", obj_ids=[obj_id]
+            )
+
+        # update metadata
+        tracker_metadata = inference_state["tracker_metadata"]
+        _obj_ids = tracker_metadata["obj_ids_per_gpu"][obj_rank]
+        tracker_metadata["obj_ids_per_gpu"][obj_rank] = _obj_ids[_obj_ids != obj_id]
+        tracker_metadata["num_obj_per_gpu"][obj_rank] = len(
+            tracker_metadata["obj_ids_per_gpu"][obj_rank]
+        )
+        tracker_metadata["obj_ids_all_gpu"] = np.concatenate(
+            tracker_metadata["obj_ids_per_gpu"]
+        )
+        tracker_metadata["obj_id_to_score"].pop(obj_id, None)
+        # tracker_metadata["max_obj_id"] # we do not reuse the object id, so we do not update it here
+
+        # Clean up cached frame outputs to remove references to the deleted object
+        if "cached_frame_outputs" in inference_state:
+            for frame_idx in inference_state["cached_frame_outputs"]:
+                frame_cache = inference_state["cached_frame_outputs"][frame_idx]
+                if obj_id in frame_cache:
+                    del frame_cache[obj_id]
+
+    def _get_gpu_id_by_obj_id(self, inference_state, obj_id):
+        """
+        Locate GPU ID for a given object.
+        """
+        obj_ids_per_gpu = inference_state["tracker_metadata"]["obj_ids_per_gpu"]
+        for rank, obj_ids in enumerate(obj_ids_per_gpu):
+            if obj_id in obj_ids:
+                return rank
+        return None  # object not found in any GPU
+
+    def _get_tracker_inference_states_by_obj_ids(self, inference_state, obj_ids):
+        """
+        Get the Tracker inference states that contain the given object ids.
+        This is used to run partial Tracker propagation on a single object/bucket.
+        Possibly multiple or zero states can be returned.
+        """
+        states = [
+            state
+            for state in inference_state["tracker_inference_states"]
+            if set(obj_ids) & set(state["obj_ids"])
+        ]
+        return states
+
+    def _prepare_backbone_feats(self, inference_state, frame_idx, reverse):
+        input_batch = inference_state["input_batch"]
+        feature_cache = inference_state["feature_cache"]
+        num_frames = inference_state["num_frames"]
+        geometric_prompt = (
+            inference_state["constants"]["empty_geometric_prompt"]
+            if inference_state["per_frame_geometric_prompt"][frame_idx] is None
+            else inference_state["per_frame_geometric_prompt"][frame_idx]
+        )
+        _ = self.run_backbone_and_detection(
+            frame_idx=frame_idx,
+            num_frames=num_frames,
+            input_batch=input_batch,
+            geometric_prompt=geometric_prompt,
+            feature_cache=feature_cache,
+            reverse=reverse,
+            allow_new_detections=True,
+        )
+
+    @torch.inference_mode()
+    def add_prompt(
+        self,
+        inference_state,
+        frame_idx,
+        text_str=None,
+        boxes_xywh=None,
+        box_labels=None,
+        points=None,
+        point_labels=None,
+        obj_id=None,
+        rel_coordinates=True,
+    ):
+        if points is not None:
+            # Tracker instance prompts
+            assert (
+                text_str is None and boxes_xywh is None
+            ), "When points are provided, text_str and boxes_xywh must be None."
+            assert (
+                obj_id is not None
+            ), "When points are provided, obj_id must be provided."
+            return self.add_tracker_new_points(
+                inference_state,
+                frame_idx,
+                obj_id=obj_id,
+                points=points,
+                labels=point_labels,
+                rel_coordinates=rel_coordinates,
+                use_prev_mem_frame=self.use_prev_mem_frame,
+            )
+        else:
+            # SAM3 prompts
+            return super().add_prompt(
+                inference_state,
+                frame_idx,
+                text_str=text_str,
+                boxes_xywh=boxes_xywh,
+                box_labels=box_labels,
+            )
+
+    @torch.inference_mode()
+    def add_tracker_new_points(
+        self,
+        inference_state,
+        frame_idx,
+        obj_id,
+        points,
+        labels,
+        rel_coordinates=True,
+        use_prev_mem_frame=False,
+    ):
+        """Add a new point prompt to Tracker. Suppporting instance refinement to existing
+        objects by passing existing obj_id or adding a new object by passing a new obj_id.
+        use_prev_mem_frame=False to disable cross attention to previous memory frames.
+        Every GPU returns the same results, and results should contain all masks including
+        these masks not refined or not added by the current user points.
+        """
+        assert obj_id is not None, "obj_id must be provided to add new points"
+        tracker_metadata = inference_state["tracker_metadata"]
+        if tracker_metadata == {}:
+            # initialize masklet metadata if it's uninitialized (empty dict)
+            tracker_metadata.update(self._initialize_metadata())
+
+        obj_rank = self._get_gpu_id_by_obj_id(inference_state, obj_id)
+
+        # prepare feature
+        self._prepare_backbone_feats(inference_state, frame_idx, reverse=False)
+
+        object_has_been_refined = self._has_object_been_refined(inference_state, obj_id)
+        if (
+            obj_rank is not None
+            and self.use_stateless_refinement
+            and not object_has_been_refined
+        ):
+            # The first time we start refinement on the object, we remove it.
+            logger.debug(
+                f"[rank={self.rank}] Removing object {obj_id} before refinement."
+            )
+            self.remove_object(inference_state, obj_id, is_user_action=False)
+            obj_rank = None
+
+        if obj_rank is None:
+            # new object, we assign it a GPU and create a new inference state if limit allows
+            num_prev_obj = np.sum(tracker_metadata["num_obj_per_gpu"])
+            if num_prev_obj >= self.max_num_objects:
+                logger.warning(
+                    f"add_tracker_new_points: cannot add a new object as we are already tracking {num_prev_obj=} "
+                    f"masklets (under {self.max_num_objects=})"
+                )
+                obj_ids = []
+                H_low_res = W_low_res = self.tracker.low_res_mask_size
+                H_video_res = inference_state["orig_height"]
+                W_video_res = inference_state["orig_width"]
+                low_res_masks = torch.zeros(0, 1, H_low_res, W_low_res)
+                video_res_masks = torch.zeros(0, 1, H_video_res, W_video_res)
+                return frame_idx, obj_ids, low_res_masks, video_res_masks
+
+            new_det_gpu_ids = self._assign_new_det_to_gpus(
+                new_det_num=1,
+                prev_workload_per_gpu=tracker_metadata["num_obj_per_gpu"],
+            )
+            obj_rank = new_det_gpu_ids[0]
+
+            # get tracker inference state for the new object
+            if self.rank == obj_rank:
+                # for batched inference, we create a new inference state
+                tracker_state = self._init_new_tracker_state(inference_state)
+                inference_state["tracker_inference_states"].append(tracker_state)
+
+            # update metadata
+            tracker_metadata["obj_ids_per_gpu"][obj_rank] = np.concatenate(
+                [
+                    tracker_metadata["obj_ids_per_gpu"][obj_rank],
+                    np.array([obj_id], dtype=np.int64),
+                ]
+            )
+            tracker_metadata["num_obj_per_gpu"][obj_rank] = len(
+                tracker_metadata["obj_ids_per_gpu"][obj_rank]
+            )
+            tracker_metadata["obj_ids_all_gpu"] = np.concatenate(
+                tracker_metadata["obj_ids_per_gpu"]
+            )
+            tracker_metadata["max_obj_id"] = max(tracker_metadata["max_obj_id"], obj_id)
+
+            logger.debug(
+                f"[rank={self.rank}] Adding new object with id {obj_id} at frame {frame_idx}."
+            )
+            self.add_action_history(
+                inference_state, "add", frame_idx=frame_idx, obj_ids=[obj_id]
+            )
+        else:
+            # existing object, for refinement
+            if self.rank == obj_rank:
+                tracker_states = self._get_tracker_inference_states_by_obj_ids(
+                    inference_state, [obj_id]
+                )
+                assert (
+                    len(tracker_states) == 1
+                ), f"[rank={self.rank}] Multiple Tracker inference states found for the same object id."
+                tracker_state = tracker_states[0]
+
+            # log
+            logger.debug(
+                f"[rank={self.rank}] Refining existing object with id {obj_id} at frame {frame_idx}."
+            )
+            self.add_action_history(
+                inference_state, "refine", frame_idx=frame_idx, obj_ids=[obj_id]
+            )
+
+        # assign higher score to added/refined object
+        tracker_metadata["obj_id_to_score"][obj_id] = 1.0
+        tracker_metadata["obj_id_to_tracker_score_frame_wise"][frame_idx][obj_id] = 1.0
+
+        if self.rank == 0:
+            rank0_metadata = tracker_metadata.get("rank0_metadata", {})
+
+            if "removed_obj_ids" in rank0_metadata:
+                rank0_metadata["removed_obj_ids"].discard(obj_id)
+
+            if "suppressed_obj_ids" in rank0_metadata:
+                for frame_id in rank0_metadata["suppressed_obj_ids"]:
+                    rank0_metadata["suppressed_obj_ids"][frame_id].discard(obj_id)
+
+            if "masklet_confirmation" in rank0_metadata:
+                obj_ids_all_gpu = tracker_metadata["obj_ids_all_gpu"]
+                obj_indices = np.where(obj_ids_all_gpu == obj_id)[0]
+                if len(obj_indices) > 0:
+                    obj_idx = obj_indices[0]
+                    if obj_idx < len(rank0_metadata["masklet_confirmation"]["status"]):
+                        rank0_metadata["masklet_confirmation"]["status"][obj_idx] = 1
+                        rank0_metadata["masklet_confirmation"]["consecutive_det_num"][
+                            obj_idx
+                        ] = self.masklet_confirmation_consecutive_det_thresh
+
+        if self.rank == obj_rank:
+            frame_idx, obj_ids, low_res_masks, video_res_masks = (
+                self.tracker.add_new_points(
+                    inference_state=tracker_state,
+                    frame_idx=frame_idx,
+                    obj_id=obj_id,
+                    points=points,
+                    labels=labels,
+                    clear_old_points=True,
+                    rel_coordinates=rel_coordinates,
+                    use_prev_mem_frame=use_prev_mem_frame,
+                )
+            )
+
+            if video_res_masks is not None and len(video_res_masks) > 0:
+                video_res_masks = fill_holes_in_mask_scores(
+                    video_res_masks,  # shape (N, 1, H_video, W_video)
+                    max_area=self.fill_hole_area,
+                    fill_holes=True,
+                    remove_sprinkles=True,
+                )
+
+            # Since the mem encoder has already run for the current input points?
+            self.tracker.propagate_in_video_preflight(
+                tracker_state, run_mem_encoder=True
+            )
+            # Clear detector conditioning frames when user clicks are received to allow
+            # model updating masks on these frames. It is a noop if user is refining on the
+            # detector conditioning frames or adding new objects.
+            self.clear_detector_added_cond_frame_in_tracker(
+                tracker_state, obj_id, frame_idx
+            )
+
+        # fetch results from states and gather across GPUs
+        # Use optimized caching approach to avoid reprocessing unmodified objects
+        if self.rank == obj_rank and len(obj_ids) > 0:
+            new_mask_data = (video_res_masks[obj_ids.index(obj_id)] > 0.0).to(
+                torch.bool
+            )
+        else:
+            new_mask_data = None
+        # Broadcast the new mask data across all ranks for consistency
+        if self.world_size > 1:
+            data_list = [new_mask_data.cpu() if new_mask_data is not None else None]
+            self.broadcast_python_obj_cpu(data_list, src=obj_rank)
+            new_mask_data = data_list[0].to(self.device)
+
+        if self.rank == 0:
+            obj_id_to_mask = self._build_tracker_output(
+                inference_state,
+                frame_idx,
+                {obj_id: new_mask_data} if new_mask_data is not None else None,
+            )
+            # post processing - remove suppressed obj_ids
+            obj_id_to_score = tracker_metadata["obj_id_to_score"]
+            suppressed_obj_ids = tracker_metadata["rank0_metadata"][
+                "suppressed_obj_ids"
+            ][frame_idx]
+            obj_id_to_tracker_score = tracker_metadata[
+                "obj_id_to_tracker_score_frame_wise"
+            ][frame_idx]
+
+            out = {
+                "obj_id_to_mask": obj_id_to_mask,
+                "obj_id_to_score": obj_id_to_score,
+                "obj_id_to_tracker_score": obj_id_to_tracker_score,
+            }
+            self._cache_frame_outputs(
+                inference_state,
+                frame_idx,
+                obj_id_to_mask,
+                suppressed_obj_ids=suppressed_obj_ids,
+            )
+            return frame_idx, self._postprocess_output(
+                inference_state, out, suppressed_obj_ids=suppressed_obj_ids
+            )
+        else:
+            return frame_idx, None  # no output on other GPUs
+
+    def _gather_obj_id_to_mask_across_gpus(self, inference_state, obj_id_to_mask_local):
+        """Gather obj_id_to_mask from all GPUs. Optionally resize the masks to the video resolution."""
+        tracker_metadata = inference_state["tracker_metadata"]
+
+        # concatenate the output masklets from all local inference states
+        H_mask = W_mask = self.tracker.low_res_mask_size
+        obj_ids_local = tracker_metadata["obj_ids_per_gpu"][self.rank]
+        low_res_masks_local = []
+        for obj_id in obj_ids_local:
+            if obj_id in obj_id_to_mask_local:
+                low_res_masks_local.append(obj_id_to_mask_local[obj_id])
+            else:
+                low_res_masks_local.append(
+                    torch.full((H_mask, W_mask), -1024.0, device=self.device)
+                )
+        if len(low_res_masks_local) > 0:
+            low_res_masks_local = torch.stack(low_res_masks_local, dim=0)  # (N, H, W)
+            assert low_res_masks_local.shape[1:] == (H_mask, W_mask)
+        else:
+            low_res_masks_local = torch.zeros(0, H_mask, W_mask, device=self.device)
+
+        # all-gather `low_res_masks_local` into `low_res_masks_global`
+        # - low_res_masks_global: Tensor -- (num_global_obj, H_mask, W_mask)
+        if self.world_size > 1:
+            low_res_masks_local = low_res_masks_local.float().contiguous()
+            low_res_masks_peers = [
+                low_res_masks_local.new_empty(num_obj, H_mask, W_mask)
+                for num_obj in tracker_metadata["num_obj_per_gpu"]
+            ]
+            dist.all_gather(low_res_masks_peers, low_res_masks_local)
+            low_res_masks_global = torch.cat(low_res_masks_peers, dim=0)
+        else:
+            low_res_masks_global = low_res_masks_local
+        return low_res_masks_global
+
+    def _convert_low_res_mask_to_video_res(self, low_res_mask, inference_state):
+        """
+        Convert a low-res mask to video resolution, matching the format expected by _build_tracker_output.
+
+        Args:
+            low_res_mask: Tensor of shape (H_low_res, W_low_res)
+            inference_state: Contains video dimensions
+
+        Returns:
+            video_res_mask: Tensor of shape (1, H_video, W_video) bool
+        """
+        if low_res_mask is None:
+            return None
+
+        # Convert to 3D for interpolation: (H_low_res, W_low_res) -> (1, H_low_res, W_low_res)
+        low_res_mask_3d = low_res_mask.unsqueeze(0).unsqueeze(0)
+
+        # Get video dimensions
+        H_video = inference_state["orig_height"]
+        W_video = inference_state["orig_width"]
+
+        video_res_mask = F.interpolate(
+            low_res_mask_3d.float(),
+            size=(H_video, W_video),
+            mode="bilinear",
+            align_corners=False,
+        )  # (1, H_video, W_video)
+
+        # Convert to boolean - already in the right shape!
+        return (video_res_mask.squeeze(0) > 0.0).to(torch.bool)
+
+    def clear_detector_added_cond_frame_in_tracker(
+        self, tracker_state, obj_id, refined_frame_idx
+    ):
+        """Clear detector added conditioning frame if it is within a predefined window
+        of the refined frame. This allow model to update masks on these frames."""
+        obj_idx = self.tracker._obj_id_to_idx(tracker_state, obj_id)
+
+        mask_only_cond_frame_indices = []
+        window = self.refinement_detector_cond_frame_removal_window
+        for frame_idx in tracker_state["mask_inputs_per_obj"][obj_idx]:
+            if frame_idx not in tracker_state["point_inputs_per_obj"][obj_idx]:
+                # clear conditioning frames within a window of the refined frame
+                if abs(frame_idx - refined_frame_idx) <= window:
+                    mask_only_cond_frame_indices.append(frame_idx)
+
+        # clear
+        if len(mask_only_cond_frame_indices) > 0:
+            for frame_idx in mask_only_cond_frame_indices:
+                # obj_ids_on_this_frame is essentially all obj_ids in the state
+                # since they are bucket batched
+                obj_ids_on_this_frame = tracker_state["obj_id_to_idx"].keys()
+                for obj_id2 in obj_ids_on_this_frame:
+                    self.tracker.clear_all_points_in_frame(
+                        tracker_state, frame_idx, obj_id2, need_output=False
+                    )
+            logger.debug(
+                f"Cleared detector mask only conditioning frames ({mask_only_cond_frame_indices}) in Tracker."
+            )
+        return
+
+
+def is_image_type(resource_path: str) -> bool:
+    if isinstance(resource_path, list):
+        return len(resource_path) == 1
+    return resource_path.lower().endswith(tuple(IMAGE_EXTS))

detect_tools/sam3/sam3/model/sam3_video_predictor.py

@@ -0,0 +1,521 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates. All Rights Reserved
+
+import datetime
+import gc
+import multiprocessing as mp
+import os
+import queue
+import socket
+import sys
+import time
+import uuid
+from contextlib import closing
+from typing import List, Optional
+
+import psutil
+import torch
+
+from sam3.logger import get_logger
+
+logger = get_logger(__name__)
+
+
+class Sam3VideoPredictor:
+    # a global dictionary that holds all inference states for this model (key is session_id)
+    _ALL_INFERENCE_STATES = {}
+
+    def __init__(
+        self,
+        checkpoint_path=None,
+        bpe_path=None,
+        has_presence_token=True,
+        geo_encoder_use_img_cross_attn=True,
+        strict_state_dict_loading=True,
+        async_loading_frames=False,
+        video_loader_type="cv2",
+        apply_temporal_disambiguation: bool = True,
+    ):
+        self.async_loading_frames = async_loading_frames
+        self.video_loader_type = video_loader_type
+        from sam3.model_builder import build_sam3_video_model
+
+        self.model = (
+            build_sam3_video_model(
+                checkpoint_path=checkpoint_path,
+                bpe_path=bpe_path,
+                has_presence_token=has_presence_token,
+                geo_encoder_use_img_cross_attn=geo_encoder_use_img_cross_attn,
+                strict_state_dict_loading=strict_state_dict_loading,
+                apply_temporal_disambiguation=apply_temporal_disambiguation,
+            )
+            .cuda()
+            .eval()
+        )
+
+    @torch.inference_mode()
+    def handle_request(self, request):
+        """Dispatch a request based on its type."""
+        request_type = request["type"]
+        if request_type == "start_session":
+            return self.start_session(
+                resource_path=request["resource_path"],
+                session_id=request.get("session_id", None),
+            )
+        elif request_type == "add_prompt":
+            return self.add_prompt(
+                session_id=request["session_id"],
+                frame_idx=request["frame_index"],
+                text=request.get("text", None),
+                points=request.get("points", None),
+                point_labels=request.get("point_labels", None),
+                bounding_boxes=request.get("bounding_boxes", None),
+                bounding_box_labels=request.get("bounding_box_labels", None),
+                obj_id=request.get("obj_id", None),
+            )
+        elif request_type == "remove_object":
+            return self.remove_object(
+                session_id=request["session_id"],
+                obj_id=request["obj_id"],
+                is_user_action=request.get("is_user_action", True),
+            )
+        elif request_type == "reset_session":
+            return self.reset_session(session_id=request["session_id"])
+        elif request_type == "close_session":
+            return self.close_session(session_id=request["session_id"])
+        else:
+            raise RuntimeError(f"invalid request type: {request_type}")
+
+    @torch.inference_mode()
+    def handle_stream_request(self, request):
+        """Dispatch a stream request based on its type."""
+        request_type = request["type"]
+        if request_type == "propagate_in_video":
+            yield from self.propagate_in_video(
+                session_id=request["session_id"],
+                propagation_direction=request.get("propagation_direction", "both"),
+                start_frame_idx=request.get("start_frame_index", None),
+                max_frame_num_to_track=request.get("max_frame_num_to_track", None),
+            )
+        else:
+            raise RuntimeError(f"invalid request type: {request_type}")
+
+    def start_session(self, resource_path, session_id=None):
+        """
+        Start a new inference session on an image or a video. Here `resource_path`
+        can be either a path to an image file (for image inference) or an MP4 file
+        or directory with JPEG video frames (for video inference).
+
+        If `session_id` is defined, it will be used as identifier for the
+        session. If it is not defined, the start_session function will create
+        a session id and return it.
+        """
+        # get an initial inference_state from the model
+        inference_state = self.model.init_state(
+            resource_path=resource_path,
+            async_loading_frames=self.async_loading_frames,
+            video_loader_type=self.video_loader_type,
+        )
+        if not session_id:
+            session_id = str(uuid.uuid4())
+        self._ALL_INFERENCE_STATES[session_id] = {
+            "state": inference_state,
+            "session_id": session_id,
+            "start_time": time.time(),
+        }
+        logger.debug(
+            f"started new session {session_id}; {self._get_session_stats()}; "
+            f"{self._get_torch_and_gpu_properties()}"
+        )
+        return {"session_id": session_id}
+
+    def add_prompt(
+        self,
+        session_id: str,
+        frame_idx: int,
+        text: Optional[str] = None,
+        points: Optional[List[List[float]]] = None,
+        point_labels: Optional[List[int]] = None,
+        bounding_boxes: Optional[List[List[float]]] = None,
+        bounding_box_labels: Optional[List[int]] = None,
+        obj_id: Optional[int] = None,
+    ):
+        """Add text, box and/or point prompt on a specific video frame."""
+        logger.debug(
+            f"add prompt on frame {frame_idx} in session {session_id}: "
+            f"{text=}, {points=}, {point_labels=}, "
+            f"{bounding_boxes=}, {bounding_box_labels=}"
+        )
+        session = self._get_session(session_id)
+        inference_state = session["state"]
+
+        frame_idx, outputs = self.model.add_prompt(
+            inference_state=inference_state,
+            frame_idx=frame_idx,
+            text_str=text,
+            points=points,
+            point_labels=point_labels,
+            boxes_xywh=bounding_boxes,
+            box_labels=bounding_box_labels,
+            obj_id=obj_id,
+        )
+        return {"frame_index": frame_idx, "outputs": outputs}
+
+    def remove_object(
+        self,
+        session_id: str,
+        obj_id: int,
+        is_user_action: bool = True,
+    ):
+        """Remove an object from tracking."""
+        logger.debug(
+            f"remove object {obj_id} in session {session_id}: " f"{is_user_action=}"
+        )
+        session = self._get_session(session_id)
+        inference_state = session["state"]
+
+        self.model.remove_object(
+            inference_state=inference_state,
+            obj_id=obj_id,
+            is_user_action=is_user_action,
+        )
+        return {"is_success": True}
+
+    def propagate_in_video(
+        self,
+        session_id,
+        propagation_direction,
+        start_frame_idx,
+        max_frame_num_to_track,
+    ):
+        """Propagate the added prompts to get grounding results on all video frames."""
+        logger.debug(
+            f"propagate in video in session {session_id}: "
+            f"{propagation_direction=}, {start_frame_idx=}, {max_frame_num_to_track=}"
+        )
+        try:
+            session = self._get_session(session_id)
+            inference_state = session["state"]
+            if propagation_direction not in ["both", "forward", "backward"]:
+                raise ValueError(
+                    f"invalid propagation direction: {propagation_direction}"
+                )
+
+            # First doing the forward propagation
+            if propagation_direction in ["both", "forward"]:
+                for frame_idx, outputs in self.model.propagate_in_video(
+                    inference_state=inference_state,
+                    start_frame_idx=start_frame_idx,
+                    max_frame_num_to_track=max_frame_num_to_track,
+                    reverse=False,
+                ):
+                    yield {"frame_index": frame_idx, "outputs": outputs}
+            # Then doing the backward propagation (reverse in time)
+            if propagation_direction in ["both", "backward"]:
+                for frame_idx, outputs in self.model.propagate_in_video(
+                    inference_state=inference_state,
+                    start_frame_idx=start_frame_idx,
+                    max_frame_num_to_track=max_frame_num_to_track,
+                    reverse=True,
+                ):
+                    yield {"frame_index": frame_idx, "outputs": outputs}
+        finally:
+            # Log upon completion (so that e.g. we can see if two propagations happen in parallel).
+            # Using `finally` here to log even when the tracking is aborted with GeneratorExit.
+            logger.debug(
+                f"propagation ended in session {session_id}; {self._get_session_stats()}"
+            )
+
+    def reset_session(self, session_id):
+        """Reset the session to its initial state (as when it's initial opened)."""
+        logger.debug(f"reset session {session_id}")
+        session = self._get_session(session_id)
+        inference_state = session["state"]
+        self.model.reset_state(inference_state)
+        return {"is_success": True}
+
+    def close_session(self, session_id):
+        """
+        Close a session. This method is idempotent and can be called multiple
+        times on the same "session_id".
+        """
+        session = self._ALL_INFERENCE_STATES.pop(session_id, None)
+        if session is None:
+            logger.warning(
+                f"cannot close session {session_id} as it does not exist (it might have expired); "
+                f"{self._get_session_stats()}"
+            )
+        else:
+            del session
+            gc.collect()
+            logger.info(f"removed session {session_id}; {self._get_session_stats()}")
+        return {"is_success": True}
+
+    def _get_session(self, session_id):
+        session = self._ALL_INFERENCE_STATES.get(session_id, None)
+        if session is None:
+            raise RuntimeError(
+                f"Cannot find session {session_id}; it might have expired"
+            )
+        return session
+
+    def _get_session_stats(self):
+        """Get a statistics string for live sessions and their GPU usage."""
+        # print both the session ids and their video frame numbers
+        live_session_strs = [
+            f"'{session_id}' ({session['state']['num_frames']} frames)"
+            for session_id, session in self._ALL_INFERENCE_STATES.items()
+        ]
+        session_stats_str = (
+            f"live sessions: [{', '.join(live_session_strs)}], GPU memory: "
+            f"{torch.cuda.memory_allocated() // 1024**2} MiB used and "
+            f"{torch.cuda.memory_reserved() // 1024**2} MiB reserved"
+            f" (max over time: {torch.cuda.max_memory_allocated() // 1024**2} MiB used "
+            f"and {torch.cuda.max_memory_reserved() // 1024**2} MiB reserved)"
+        )
+        return session_stats_str
+
+    def _get_torch_and_gpu_properties(self):
+        """Get a string for PyTorch and GPU properties (for logging and debugging)."""
+        torch_and_gpu_str = (
+            f"torch: {torch.__version__} with CUDA arch {torch.cuda.get_arch_list()}, "
+            f"GPU device: {torch.cuda.get_device_properties(torch.cuda.current_device())}"
+        )
+        return torch_and_gpu_str
+
+    def shutdown(self):
+        """Shutdown the predictor and clear all sessions."""
+        self._ALL_INFERENCE_STATES.clear()
+
+
+class Sam3VideoPredictorMultiGPU(Sam3VideoPredictor):
+    def __init__(self, *model_args, gpus_to_use=None, **model_kwargs):
+        if gpus_to_use is None:
+            # if not specified, use only the current GPU by default
+            gpus_to_use = [torch.cuda.current_device()]
+
+        IS_MAIN_PROCESS = os.getenv("IS_MAIN_PROCESS", "1") == "1"
+        if IS_MAIN_PROCESS:
+            gpus_to_use = sorted(set(gpus_to_use))
+            logger.info(f"using the following GPU IDs: {gpus_to_use}")
+            assert len(gpus_to_use) > 0 and all(isinstance(i, int) for i in gpus_to_use)
+            assert all(0 <= i < torch.cuda.device_count() for i in gpus_to_use)
+            os.environ["MASTER_ADDR"] = "localhost"
+            os.environ["MASTER_PORT"] = f"{self._find_free_port()}"
+            os.environ["RANK"] = "0"
+            os.environ["WORLD_SIZE"] = f"{len(gpus_to_use)}"
+
+        self.gpus_to_use = gpus_to_use
+        self.rank = int(os.environ["RANK"])
+        self.world_size = int(os.environ["WORLD_SIZE"])
+        self.rank_str = f"rank={self.rank} with world_size={self.world_size}"
+        self.device = torch.device(f"cuda:{self.gpus_to_use[self.rank]}")
+        torch.cuda.set_device(self.device)
+        self.has_shutdown = False
+        if self.rank == 0:
+            logger.info("\n\n\n\t*** START loading model on all ranks ***\n\n")
+
+        logger.info(f"loading model on {self.rank_str} -- this could take a while ...")
+        super().__init__(*model_args, **model_kwargs)
+        logger.info(f"loading model on {self.rank_str} -- DONE locally")
+
+        if self.world_size > 1 and self.rank == 0:
+            # start the worker processes *after* the model is loaded in the main process
+            # so that the main process can run torch.compile and fill the cache first
+            self._start_worker_processes(*model_args, **model_kwargs)
+            for rank in range(1, self.world_size):
+                self.command_queues[rank].put(("start_nccl_process_group", None))
+            self._start_nccl_process_group()
+
+        if self.rank == 0:
+            logger.info("\n\n\n\t*** DONE loading model on all ranks ***\n\n")
+
+    @torch.inference_mode()
+    def handle_request(self, request):
+        """Dispatch a request based on its type."""
+        if self.has_shutdown:
+            raise RuntimeError(
+                "cannot handle request after the predictor has shutdown; please create a new predictor"
+            )
+
+        # when starting a session, we need to create a session id before dispatching
+        # the request to the workers
+        if request["type"] == "start_session" and request.get("session_id") is None:
+            request["session_id"] = str(uuid.uuid4())
+        # dispatch the request to all worker processes
+        if self.world_size > 1 and self.rank == 0:
+            for rank in range(1, self.world_size):
+                self.command_queues[rank].put((request, False))
+
+        response = super().handle_request(request)
+
+        if self.world_size > 1:
+            torch.distributed.barrier()  # wait for all ranks to finish
+        return response
+
+    @torch.inference_mode()
+    def handle_stream_request(self, request):
+        """Dispatch a stream request based on its type."""
+        if self.has_shutdown:
+            raise RuntimeError(
+                "cannot handle request after the predictor has shutdown; please create a new predictor"
+            )
+
+        # dispatch the request to all worker processes
+        if self.world_size > 1 and self.rank == 0:
+            for rank in range(1, self.world_size):
+                self.command_queues[rank].put((request, True))
+
+        yield from super().handle_stream_request(request)
+
+        if self.world_size > 1:
+            torch.distributed.barrier()  # wait for all ranks to finish
+
+    def _start_worker_processes(self, *model_args, **model_kwargs):
+        """Start worker processes for handling model inference."""
+        world_size = self.world_size
+        logger.info(f"spawning {world_size - 1} worker processes")
+        # Use "spawn" (instead of "fork") for different PyTorch or CUDA context
+        mp_ctx = mp.get_context("spawn")
+        self.command_queues = {rank: mp_ctx.Queue() for rank in range(1, world_size)}
+        self.result_queues = {rank: mp_ctx.Queue() for rank in range(1, world_size)}
+        parent_pid = os.getpid()
+        for rank in range(1, world_size):
+            # set the environment variables for each worker process
+            os.environ["IS_MAIN_PROCESS"] = "0"  # mark this as a worker process
+            os.environ["RANK"] = f"{rank}"
+            worker_process = mp_ctx.Process(
+                target=Sam3VideoPredictorMultiGPU._worker_process_command_loop,
+                args=(
+                    rank,
+                    world_size,
+                    self.command_queues[rank],
+                    self.result_queues[rank],
+                    model_args,
+                    model_kwargs,
+                    self.gpus_to_use,
+                    parent_pid,
+                ),
+                daemon=True,
+            )
+            worker_process.start()
+        # revert the environment variables for the main process
+        os.environ["IS_MAIN_PROCESS"] = "1"
+        os.environ["RANK"] = "0"
+        # wait for all the worker processes to load the model and collect their PIDs
+        self.worker_pids = {}
+        for rank in range(1, self.world_size):
+            # a large timeout to cover potentially long model loading time due to compilation
+            _, worker_pid = self.result_queues[rank].get(timeout=7200)
+            self.worker_pids[rank] = worker_pid
+        logger.info(f"spawned {world_size - 1} worker processes")
+
+    def _start_nccl_process_group(self):
+        rank = int(os.environ["RANK"])
+        world_size = int(os.environ["WORLD_SIZE"])
+        if world_size == 1:
+            return
+
+        logger.debug(f"starting NCCL process group on {rank=} with {world_size=}")
+        assert not torch.distributed.is_initialized()
+        # use the "env://" init method with environment variables set in start_worker_processes
+        # a short 3-min timeout to quickly detect any synchronization failures
+        timeout_sec = int(os.getenv("SAM3_COLLECTIVE_OP_TIMEOUT_SEC", "180"))
+        timeout = datetime.timedelta(seconds=timeout_sec)
+        torch.distributed.init_process_group(
+            backend="nccl",
+            init_method="env://",
+            timeout=timeout,
+            device_id=self.device,
+        )
+        # warm-up the NCCL process group by running a dummy all-reduce
+        tensor = torch.ones(1024, 1024).cuda()
+        torch.distributed.all_reduce(tensor)
+        logger.debug(f"started NCCL process group on {rank=} with {world_size=}")
+
+    def _find_free_port(self) -> int:
+        """
+        Find a free port (a random free port from 1024 to 65535 will be selected)
+        https://stackoverflow.com/questions/1365265/on-localhost-how-do-i-pick-a-free-port-number)
+        """
+        with closing(socket.socket(socket.AF_INET, socket.SOCK_STREAM)) as s:
+            s.bind(("", 0))
+            s.setsockopt(socket.SOL_SOCKET, socket.SO_REUSEADDR, 1)
+            return s.getsockname()[1]
+
+    @staticmethod
+    def _worker_process_command_loop(
+        rank,
+        world_size,
+        command_queue,
+        result_queue,
+        model_args,
+        model_kwargs,
+        gpus_to_use,
+        parent_pid,
+    ):
+        """
+        The command loop for each worker process. It listens to commands from the main process
+        and executes them using the model.
+        """
+        logger.info(f"starting worker process {rank=} with {world_size=}")
+        # verify that the environment variables are set correctly
+        assert int(os.environ["IS_MAIN_PROCESS"]) == 0
+        assert int(os.environ["RANK"]) == rank
+        assert int(os.environ["WORLD_SIZE"]) == world_size
+        # load the model in this worker process
+        predictor = Sam3VideoPredictorMultiGPU(
+            *model_args, gpus_to_use=gpus_to_use, **model_kwargs
+        )
+        logger.info(f"started worker {rank=} with {world_size=}")
+        # return the worker process id to the main process for bookkeeping
+        worker_pid = os.getpid()
+        result_queue.put(("load_model", worker_pid))
+
+        # wait for the command to start the NCCL process group
+        request_type, _ = command_queue.get(timeout=7200)
+        assert request_type == "start_nccl_process_group"
+        predictor._start_nccl_process_group()
+
+        # keep listening to commands from the main process
+        while True:
+            try:
+                request, is_stream_request = command_queue.get(timeout=5.0)
+                if request == "shutdown":
+                    logger.info(f"worker {rank=} shutting down")
+                    torch.distributed.destroy_process_group()
+                    result_queue.put(("shutdown", True))  # acknowledge the shutdown
+                    sys.exit(0)
+
+                logger.debug(f"worker {rank=} received request {request['type']=}")
+                if is_stream_request:
+                    for _ in predictor.handle_stream_request(request):
+                        pass  # handle stream requests in a generator fashion
+                else:
+                    predictor.handle_request(request)
+            except queue.Empty:
+                # Usually Python's multiprocessing module will shutdown all the daemon worker
+                # processes when the main process exits gracefully. However, the user may kill
+                # the main process using SIGKILL and thereby leaving no chance for the main process
+                # to clean up its daemon child processes. So here we manually check whether the
+                # parent process still exists (every 5 sec as in `command_queue.get` timeout).
+                if not psutil.pid_exists(parent_pid):
+                    logger.info(
+                        f"stopping worker {rank=} as its parent process has exited"
+                    )
+                    sys.exit(1)
+            except Exception as e:
+                logger.error(f"worker {rank=} exception: {e}", exc_info=True)
+
+    def shutdown(self):
+        """Shutdown all worker processes."""
+        if self.rank == 0 and self.world_size > 1:
+            logger.info(f"shutting down {self.world_size - 1} worker processes")
+            for rank in range(1, self.world_size):
+                self.command_queues[rank].put(("shutdown", False))
+            torch.distributed.destroy_process_group()
+            for rank in range(1, self.world_size):
+                self.result_queues[rank].get()  # wait for the worker to acknowledge
+            logger.info(f"shut down {self.world_size - 1} worker processes")
+        self.has_shutdown = True
+
+        super().shutdown()

detect_tools/sam3/sam3/model/text_encoder_ve.py

@@ -0,0 +1,328 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates. All Rights Reserved
+
+from collections import OrderedDict
+from typing import Callable, List, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+from torch.utils.checkpoint import checkpoint
+
+from .model_misc import LayerScale
+
+
+class ResidualAttentionBlock(nn.Module):
+    def __init__(
+        self,
+        d_model: int,
+        n_head: int,
+        mlp_ratio: float = 4.0,
+        ls_init_value: Optional[float] = None,
+        act_layer: Callable[[], nn.Module] = nn.GELU,
+        norm_layer: Callable[[int], nn.Module] = nn.LayerNorm,
+    ):
+        super().__init__()
+        # Attention
+        self.attn = nn.MultiheadAttention(d_model, n_head, batch_first=True)
+
+        # LayerNorm, LayerScale
+        self.ln_1 = norm_layer(d_model)
+        self.ln_2 = norm_layer(d_model)
+
+        self.ls_1 = (
+            LayerScale(d_model, ls_init_value)
+            if ls_init_value is not None
+            else nn.Identity()
+        )
+        self.ls_2 = (
+            LayerScale(d_model, ls_init_value)
+            if ls_init_value is not None
+            else nn.Identity()
+        )
+
+        # MLP
+        mlp_width = int(d_model * mlp_ratio)
+        self.mlp = nn.Sequential(
+            OrderedDict(
+                [
+                    ("c_fc", nn.Linear(d_model, mlp_width)),
+                    ("gelu", act_layer()),
+                    ("c_proj", nn.Linear(mlp_width, d_model)),
+                ]
+            )
+        )
+
+    def attention(
+        self,
+        q_x: torch.Tensor,
+        k_x: Optional[torch.Tensor] = None,
+        v_x: Optional[torch.Tensor] = None,
+        attn_mask: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        k_x = k_x if k_x is not None else q_x
+        v_x = v_x if v_x is not None else q_x
+        if attn_mask is not None:
+            # Leave boolean masks as is
+            if not attn_mask.dtype == torch.bool:
+                attn_mask = attn_mask.to(q_x.dtype)
+
+        return self.attn(q_x, k_x, v_x, need_weights=False, attn_mask=attn_mask)[0]
+
+    def forward(
+        self,
+        q_x: torch.Tensor,
+        k_x: Optional[torch.Tensor] = None,
+        v_x: Optional[torch.Tensor] = None,
+        attn_mask: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        k_x = (
+            self.ln_1_kv(k_x) if hasattr(self, "ln_1_kv") and k_x is not None else None
+        )
+        v_x = (
+            self.ln_1_kv(v_x) if hasattr(self, "ln_1_kv") and v_x is not None else None
+        )
+        x = q_x + self.ls_1(
+            self.attention(q_x=self.ln_1(q_x), k_x=k_x, v_x=v_x, attn_mask=attn_mask)
+        )
+        x = x + self.ls_2(self.mlp(self.ln_2(x)))
+        return x
+
+
+class Transformer(nn.Module):
+    def __init__(
+        self,
+        width: int,
+        layers: int,
+        heads: int,
+        mlp_ratio: float = 4.0,
+        ls_init_value: Optional[float] = None,
+        act_layer: Callable[[], nn.Module] = nn.GELU,
+        norm_layer: Callable[[int], nn.Module] = nn.LayerNorm,
+        compile_mode: Optional[str] = None,
+        use_act_checkpoint: bool = False,
+    ):
+        super().__init__()
+        self.width = width
+        self.layers = layers
+        self.grad_checkpointing = use_act_checkpoint
+        self.resblocks = nn.ModuleList(
+            [
+                ResidualAttentionBlock(
+                    width,
+                    heads,
+                    mlp_ratio,
+                    ls_init_value=ls_init_value,
+                    act_layer=act_layer,
+                    norm_layer=norm_layer,
+                )
+                for _ in range(layers)
+            ]
+        )
+
+        if compile_mode is not None:
+            self.forward = torch.compile(
+                self.forward, mode=compile_mode, fullgraph=True
+            )
+            if self.grad_checkpointing:
+                torch._dynamo.config.optimize_ddp = False
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        attn_mask: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        for _, r in enumerate(self.resblocks):
+            if (
+                self.grad_checkpointing
+                and not torch.jit.is_scripting()
+                and self.training
+            ):
+                x = checkpoint(r, x, None, None, attn_mask, use_reentrant=False)
+            else:
+                x = r(
+                    x,
+                    attn_mask=attn_mask,
+                )
+        return x
+
+
+def text_global_pool(
+    x: torch.Tensor, text: Optional[torch.Tensor] = None, pool_type: str = "argmax"
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    if pool_type == "first":
+        pooled, tokens = x[:, 0], x[:, 1:]
+    elif pool_type == "last":
+        pooled, tokens = x[:, -1], x[:, :-1]
+    elif pool_type == "argmax":
+        # take features from the eot embedding (eot_token is the highest number in each sequence)
+        assert text is not None
+        pooled, tokens = x[torch.arange(x.shape[0]), text.argmax(dim=-1)], x
+    else:
+        pooled = tokens = x
+    return pooled, tokens
+
+
+class TextTransformer(nn.Module):
+    def __init__(
+        self,
+        context_length: int = 77,
+        vocab_size: int = 49408,
+        width: int = 512,
+        heads: int = 8,
+        layers: int = 12,
+        mlp_ratio: float = 4.0,
+        ls_init_value: Optional[float] = None,
+        output_dim: int = 512,
+        no_causal_mask: bool = False,
+        pool_type: str = "none",  # no pooling
+        proj_bias: bool = False,
+        act_layer: Callable = nn.GELU,
+        norm_layer: Callable = nn.LayerNorm,
+        output_tokens: bool = False,
+        use_ln_post: bool = True,
+        compile_mode: Optional[str] = None,
+        use_act_checkpoint: bool = False,
+    ):
+        super().__init__()
+        assert pool_type in ("first", "last", "argmax", "none")
+        self.output_tokens = output_tokens
+        self.num_pos = self.context_length = context_length
+        self.vocab_size = vocab_size
+        self.width = width
+        self.output_dim = output_dim
+        self.heads = heads
+        self.pool_type = pool_type
+
+        self.token_embedding = nn.Embedding(self.vocab_size, width)
+        self.positional_embedding = nn.Parameter(torch.empty(self.num_pos, width))
+        self.transformer = Transformer(
+            width=width,
+            layers=layers,
+            heads=heads,
+            mlp_ratio=mlp_ratio,
+            ls_init_value=ls_init_value,
+            act_layer=act_layer,
+            norm_layer=norm_layer,
+            compile_mode=compile_mode,
+            use_act_checkpoint=use_act_checkpoint,
+        )
+        self.ln_final = norm_layer(width) if use_ln_post else nn.Identity()
+        if no_causal_mask:
+            self.attn_mask = None
+        else:
+            self.register_buffer(
+                "attn_mask", self.build_causal_mask(), persistent=False
+            )
+        if proj_bias:
+            self.text_projection = nn.Linear(width, output_dim)
+        else:
+            self.text_projection = nn.Parameter(torch.empty(width, output_dim))
+
+    def build_causal_mask(self) -> torch.Tensor:
+        # lazily create causal attention mask, with full attention between the tokens
+        # pytorch uses additive attention mask; fill with -inf
+        mask = torch.empty(self.num_pos, self.num_pos)
+        mask.fill_(float("-inf"))
+        mask.triu_(1)  # zero out the lower diagonal
+        return mask
+
+    def forward(
+        self, text: torch.Tensor
+    ) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
+        seq_len = text.shape[1]
+        x = self.token_embedding(text)  # [batch_size, n_ctx, d_model]
+
+        attn_mask = self.attn_mask
+        if attn_mask is not None:
+            attn_mask = attn_mask[:seq_len, :seq_len]
+
+        x = x + self.positional_embedding[:seq_len]
+        x = self.transformer(x, attn_mask=attn_mask)
+
+        x = self.ln_final(x)
+        pooled, tokens = text_global_pool(x, text, pool_type=self.pool_type)
+        if self.text_projection is not None:
+            if isinstance(self.text_projection, nn.Linear):
+                pooled = self.text_projection(pooled)
+            else:
+                pooled = pooled @ self.text_projection
+        if self.output_tokens:
+            return pooled, tokens
+        return pooled
+
+
+class VETextEncoder(nn.Module):
+    def __init__(
+        self,
+        d_model: int,
+        tokenizer: Callable,
+        width: int = 1024,
+        heads: int = 16,
+        layers: int = 24,
+        context_length: int = 32,
+        vocab_size: int = 49408,
+        use_ln_post: bool = True,
+        compile_mode: Optional[str] = None,
+        use_act_checkpoint: bool = True,
+    ):
+        super().__init__()
+        self.context_length = context_length
+        self.use_ln_post = use_ln_post
+        self.tokenizer = tokenizer
+
+        self.encoder = TextTransformer(
+            context_length=self.context_length,
+            vocab_size=vocab_size,
+            width=width,
+            heads=heads,
+            layers=layers,
+            # we want the tokens, not just the pooled output
+            output_tokens=True,
+            use_ln_post=use_ln_post,
+            compile_mode=compile_mode,
+            use_act_checkpoint=use_act_checkpoint,
+        )
+        self.resizer = nn.Linear(self.encoder.width, d_model)
+
+    def forward(
+        self,
+        text: Union[List[str], Tuple[torch.Tensor, torch.Tensor, dict]],
+        input_boxes: Optional[List] = None,
+        device: torch.device = None,
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        if isinstance(text[0], str):
+            # no use case for this
+            assert input_boxes is None or len(input_boxes) == 0, "not supported"
+
+            # Encode the text
+            tokenized = self.tokenizer(text, context_length=self.context_length).to(
+                device
+            )  # [b, seq_len]
+            text_attention_mask = (tokenized != 0).bool()
+
+            # manually embed the tokens
+            inputs_embeds = self.encoder.token_embedding(
+                tokenized
+            )  # [b, seq_len, d=1024]
+            _, text_memory = self.encoder(tokenized)  # [b, seq_len, d=1024]
+
+            assert text_memory.shape[1] == inputs_embeds.shape[1]
+            # Invert attention mask because its the opposite in pytorch transformer
+            text_attention_mask = text_attention_mask.ne(1)
+            # Transpose memory because pytorch's attention expects sequence first
+            text_memory = text_memory.transpose(0, 1)
+            # Resize the encoder hidden states to be of the same d_model as the decoder
+            text_memory_resized = self.resizer(text_memory)
+        else:
+            # The text is already encoded, use as is.
+            text_attention_mask, text_memory_resized, tokenized = text
+            inputs_embeds = tokenized["inputs_embeds"]
+            assert (
+                input_boxes is None or len(input_boxes) == 0
+            ), "Can't replace boxes in text if it's already encoded"
+
+        # Note that the input_embeds are returned in pytorch's convention (sequence first)
+        return (
+            text_attention_mask,
+            text_memory_resized,
+            inputs_embeds.transpose(0, 1),
+        )

detect_tools/sam3/sam3/model/tokenizer_ve.py

@@ -0,0 +1,253 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates. All Rights Reserved
+
+"""
+Text Tokenizer.
+
+Copied and lightly adapted from VE repo, which in turn copied
+from open_clip and openAI CLIP.
+"""
+
+import gzip
+import html
+import io
+import os
+import string
+from functools import lru_cache
+from typing import List, Optional, Union
+
+import ftfy
+import regex as re
+import torch
+from iopath.common.file_io import g_pathmgr
+
+
+# https://stackoverflow.com/q/62691279
+os.environ["TOKENIZERS_PARALLELISM"] = "false"
+DEFAULT_CONTEXT_LENGTH = 77
+
+
+@lru_cache()
+def bytes_to_unicode():
+    """
+    Returns list of utf-8 byte and a corresponding list of unicode strings.
+    The reversible bpe codes work on unicode strings.
+    This means you need a large # of unicode characters in your vocab if you want to avoid UNKs.
+    When you're at something like a 10B token dataset you end up needing around 5K for decent coverage.
+    This is a significant percentage of your normal, say, 32K bpe vocab.
+    To avoid that, we want lookup tables between utf-8 bytes and unicode strings.
+    And avoids mapping to whitespace/control characters the bpe code barfs on.
+    """
+    bs = (
+        list(range(ord("!"), ord("~") + 1))
+        + list(range(ord("¡"), ord("¬") + 1))
+        + list(range(ord("®"), ord("ÿ") + 1))
+    )
+    cs = bs[:]
+    n = 0
+    for b in range(2**8):
+        if b not in bs:
+            bs.append(b)
+            cs.append(2**8 + n)
+            n += 1
+    cs = [chr(n) for n in cs]
+    return dict(zip(bs, cs))
+
+
+def get_pairs(word):
+    """Return set of symbol pairs in a word.
+    Word is represented as tuple of symbols (symbols being variable-length strings).
+    """
+    pairs = set()
+    prev_char = word[0]
+    for char in word[1:]:
+        pairs.add((prev_char, char))
+        prev_char = char
+    return pairs
+
+
+def basic_clean(text):
+    text = ftfy.fix_text(text)
+    text = html.unescape(html.unescape(text))
+    return text.strip()
+
+
+def whitespace_clean(text):
+    text = re.sub(r"\s+", " ", text)
+    text = text.strip()
+    return text
+
+
+def _clean_canonicalize(x):
+    # basic, remove whitespace, remove punctuation, lower case
+    return canonicalize_text(basic_clean(x))
+
+
+def _clean_lower(x):
+    # basic, remove whitespace, lower case
+    return whitespace_clean(basic_clean(x)).lower()
+
+
+def _clean_whitespace(x):
+    # basic, remove whitespace
+    return whitespace_clean(basic_clean(x))
+
+
+def get_clean_fn(type: str):
+    if type == "canonicalize":
+        return _clean_canonicalize
+    elif type == "lower":
+        return _clean_lower
+    elif type == "whitespace":
+        return _clean_whitespace
+    else:
+        assert False, f"Invalid clean function ({type})."
+
+
+def canonicalize_text(text, *, keep_punctuation_exact_string=None):
+    """Returns canonicalized `text` (lowercase and punctuation removed).
+    From: https://github.com/google-research/big_vision/blob/53f18caf27a9419231bbf08d3388b07671616d3d/big_vision/evaluators/proj/image_text/prompt_engineering.py#L94
+    Args:
+      text: string to be canonicalized.
+      keep_punctuation_exact_string: If provided, then this exact string kept.
+        For example providing '{}' will keep any occurrences of '{}' (but will
+        still remove '{' and '}' that appear separately).
+    """
+    text = text.replace("_", " ")
+    if keep_punctuation_exact_string:
+        text = keep_punctuation_exact_string.join(
+            part.translate(str.maketrans("", "", string.punctuation))
+            for part in text.split(keep_punctuation_exact_string)
+        )
+    else:
+        text = text.translate(str.maketrans("", "", string.punctuation))
+    text = text.lower()
+    text = re.sub(r"\s+", " ", text)
+    return text.strip()
+
+
+class SimpleTokenizer(object):
+    def __init__(
+        self,
+        bpe_path: Union[str, os.PathLike],
+        additional_special_tokens: Optional[List[str]] = None,
+        context_length: Optional[int] = DEFAULT_CONTEXT_LENGTH,
+        clean: str = "lower",
+    ):
+        self.byte_encoder = bytes_to_unicode()
+        self.byte_decoder = {v: k for k, v in self.byte_encoder.items()}
+        with g_pathmgr.open(bpe_path, "rb") as fh:
+            bpe_bytes = io.BytesIO(fh.read())
+            merges = gzip.open(bpe_bytes).read().decode("utf-8").split("\n")
+        # merges = gzip.open(bpe_path).read().decode("utf-8").split("\n")
+        merges = merges[1 : 49152 - 256 - 2 + 1]
+        merges = [tuple(merge.split()) for merge in merges]
+        vocab = list(bytes_to_unicode().values())
+        vocab = vocab + [v + "</w>" for v in vocab]
+        for merge in merges:
+            vocab.append("".join(merge))
+        special_tokens = ["<start_of_text>", "<end_of_text>"]
+        if additional_special_tokens:
+            special_tokens += additional_special_tokens
+        vocab.extend(special_tokens)
+        self.encoder = dict(zip(vocab, range(len(vocab))))
+        self.decoder = {v: k for k, v in self.encoder.items()}
+        self.bpe_ranks = dict(zip(merges, range(len(merges))))
+        self.cache = {t: t for t in special_tokens}
+        special = "|".join(special_tokens)
+        self.pat = re.compile(
+            special + r"""|'s|'t|'re|'ve|'m|'ll|'d|[\p{L}]+|[\p{N}]|[^\s\p{L}\p{N}]+""",
+            re.IGNORECASE,
+        )
+        self.vocab_size = len(self.encoder)
+        self.all_special_ids = [self.encoder[t] for t in special_tokens]
+        self.sot_token_id = self.all_special_ids[0]
+        self.eot_token_id = self.all_special_ids[1]
+        self.context_length = context_length
+        self.clean_fn = get_clean_fn(clean)
+
+    def bpe(self, token):
+        if token in self.cache:
+            return self.cache[token]
+        word = tuple(token[:-1]) + (token[-1] + "</w>",)
+        pairs = get_pairs(word)
+        if not pairs:
+            return token + "</w>"
+        while True:
+            bigram = min(pairs, key=lambda pair: self.bpe_ranks.get(pair, float("inf")))
+            if bigram not in self.bpe_ranks:
+                break
+            first, second = bigram
+            new_word = []
+            i = 0
+            while i < len(word):
+                try:
+                    j = word.index(first, i)
+                    new_word.extend(word[i:j])
+                    i = j
+                except:
+                    new_word.extend(word[i:])
+                    break
+                if word[i] == first and i < len(word) - 1 and word[i + 1] == second:
+                    new_word.append(first + second)
+                    i += 2
+                else:
+                    new_word.append(word[i])
+                    i += 1
+            new_word = tuple(new_word)
+            word = new_word
+            if len(word) == 1:
+                break
+            else:
+                pairs = get_pairs(word)
+        word = " ".join(word)
+        self.cache[token] = word
+        return word
+
+    def encode(self, text):
+        bpe_tokens = []
+        text = self.clean_fn(text)
+        for token in re.findall(self.pat, text):
+            token = "".join(self.byte_encoder[b] for b in token.encode("utf-8"))
+            bpe_tokens.extend(
+                self.encoder[bpe_token] for bpe_token in self.bpe(token).split(" ")
+            )
+        return bpe_tokens
+
+    def decode(self, tokens):
+        text = "".join([self.decoder[token] for token in tokens])
+        text = (
+            bytearray([self.byte_decoder[c] for c in text])
+            .decode("utf-8", errors="replace")
+            .replace("</w>", " ")
+        )
+        return text
+
+    def __call__(
+        self, texts: Union[str, List[str]], context_length: Optional[int] = None
+    ) -> torch.LongTensor:
+        """Returns the tokenized representation of given input string(s)
+        Parameters
+        ----------
+        texts : Union[str, List[str]]
+            An input string or a list of input strings to tokenize
+        context_length : int
+            The context length to use; all CLIP models use 77 as the context length
+        Returns
+        -------
+        A two-dimensional tensor containing the resulting tokens, shape = [number of input strings, context_length]
+        """
+        if isinstance(texts, str):
+            texts = [texts]
+        context_length = context_length or self.context_length
+        assert context_length, "Please set a valid context length"
+        all_tokens = [
+            [self.sot_token_id] + self.encode(text) + [self.eot_token_id]
+            for text in texts
+        ]
+        result = torch.zeros(len(all_tokens), context_length, dtype=torch.long)
+        for i, tokens in enumerate(all_tokens):
+            if len(tokens) > context_length:
+                tokens = tokens[:context_length]  # Truncate
+                tokens[-1] = self.eot_token_id
+            result[i, : len(tokens)] = torch.tensor(tokens)
+        return result

detect_tools/sam3/sam3/model/utils/__init__.py

@@ -0,0 +1,5 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates. All Rights Reserved
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.

detect_tools/sam3/sam3/model/utils/misc.py

@@ -0,0 +1,77 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates. All Rights Reserved
+
+from collections import defaultdict
+from dataclasses import fields, is_dataclass
+from typing import Any, Mapping, Protocol, runtime_checkable
+
+import torch
+
+
+def _is_named_tuple(x) -> bool:
+    return isinstance(x, tuple) and hasattr(x, "_asdict") and hasattr(x, "_fields")
+
+
+@runtime_checkable
+class _CopyableData(Protocol):
+    def to(self, device: torch.device, *args: Any, **kwargs: Any):
+        """Copy data to the specified device"""
+        ...
+
+
+def copy_data_to_device(data, device: torch.device, *args: Any, **kwargs: Any):
+    """Function that recursively copies data to a torch.device.
+
+    Args:
+        data: The data to copy to device
+        device: The device to which the data should be copied
+        args: positional arguments that will be passed to the `to` call
+        kwargs: keyword arguments that will be passed to the `to` call
+
+    Returns:
+        The data on the correct device
+    """
+
+    if _is_named_tuple(data):
+        return type(data)(
+            **copy_data_to_device(data._asdict(), device, *args, **kwargs)
+        )
+    elif isinstance(data, (list, tuple)):
+        return type(data)(copy_data_to_device(e, device, *args, **kwargs) for e in data)
+    elif isinstance(data, defaultdict):
+        return type(data)(
+            data.default_factory,
+            {
+                k: copy_data_to_device(v, device, *args, **kwargs)
+                for k, v in data.items()
+            },
+        )
+    elif isinstance(data, Mapping):
+        return type(data)(
+            {
+                k: copy_data_to_device(v, device, *args, **kwargs)
+                for k, v in data.items()
+            }
+        )
+    elif is_dataclass(data) and not isinstance(data, type):
+        new_data_class = type(data)(
+            **{
+                field.name: copy_data_to_device(
+                    getattr(data, field.name), device, *args, **kwargs
+                )
+                for field in fields(data)
+                if field.init
+            }
+        )
+        for field in fields(data):
+            if not field.init:
+                setattr(
+                    new_data_class,
+                    field.name,
+                    copy_data_to_device(
+                        getattr(data, field.name), device, *args, **kwargs
+                    ),
+                )
+        return new_data_class
+    elif isinstance(data, _CopyableData):
+        return data.to(device, *args, **kwargs)
+    return data

detect_tools/sam3/sam3/model/utils/sam1_utils.py

@@ -0,0 +1,119 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates. All Rights Reserved
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import warnings
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torchvision.transforms import Normalize, Resize, ToTensor
+
+
+# Adapted from https://github.com/facebookresearch/sam2/blob/main/sam2/utils/transforms.py
+class SAM2Transforms(nn.Module):
+    def __init__(
+        self, resolution, mask_threshold, max_hole_area=0.0, max_sprinkle_area=0.0
+    ):
+        """
+        Transforms for SAM2.
+        """
+        super().__init__()
+        self.resolution = resolution
+        self.mask_threshold = mask_threshold
+        self.max_hole_area = max_hole_area
+        self.max_sprinkle_area = max_sprinkle_area
+        self.mean = [0.5, 0.5, 0.5]
+        self.std = [0.5, 0.5, 0.5]
+        self.to_tensor = ToTensor()
+        self.transforms = torch.jit.script(
+            nn.Sequential(
+                Resize((self.resolution, self.resolution)),
+                Normalize(self.mean, self.std),
+            )
+        )
+
+    def __call__(self, x):
+        x = self.to_tensor(x)
+        return self.transforms(x)
+
+    def forward_batch(self, img_list):
+        img_batch = [self.transforms(self.to_tensor(img)) for img in img_list]
+        img_batch = torch.stack(img_batch, dim=0)
+        return img_batch
+
+    def transform_coords(
+        self, coords: torch.Tensor, normalize=False, orig_hw=None
+    ) -> torch.Tensor:
+        """
+        Expects a torch tensor with length 2 in the last dimension. The coordinates can be in absolute image or normalized coordinates,
+        If the coords are in absolute image coordinates, normalize should be set to True and original image size is required.
+
+        Returns
+            Un-normalized coordinates in the range of [0, 1] which is expected by the SAM2 model.
+        """
+        if normalize:
+            assert orig_hw is not None
+            h, w = orig_hw
+            coords = coords.clone()
+            coords[..., 0] = coords[..., 0] / w
+            coords[..., 1] = coords[..., 1] / h
+
+        coords = coords * self.resolution  # unnormalize coords
+        return coords
+
+    def transform_boxes(
+        self, boxes: torch.Tensor, normalize=False, orig_hw=None
+    ) -> torch.Tensor:
+        """
+        Expects a tensor of shape Bx4. The coordinates can be in absolute image or normalized coordinates,
+        if the coords are in absolute image coordinates, normalize should be set to True and original image size is required.
+        """
+        boxes = self.transform_coords(boxes.reshape(-1, 2, 2), normalize, orig_hw)
+        return boxes
+
+    def postprocess_masks(self, masks: torch.Tensor, orig_hw) -> torch.Tensor:
+        """
+        Perform PostProcessing on output masks.
+        """
+        masks = masks.float()
+        input_masks = masks
+        mask_flat = masks.flatten(0, 1).unsqueeze(1)  # flatten as 1-channel image
+        try:
+            from sam3.perflib.connected_components import connected_components
+
+            if self.max_hole_area > 0:
+                # Holes are those connected components in background with area <= self.fill_hole_area
+                # (background regions are those with mask scores <= self.mask_threshold)
+                labels, areas = connected_components(
+                    (mask_flat <= self.mask_threshold).to(torch.uint8)
+                )
+                is_hole = (labels > 0) & (areas <= self.max_hole_area)
+                is_hole = is_hole.reshape_as(masks)
+                # We fill holes with a small positive mask score (10.0) to change them to foreground.
+                masks = torch.where(is_hole, self.mask_threshold + 10.0, masks)
+
+            if self.max_sprinkle_area > 0:
+                labels, areas = connected_components(
+                    (mask_flat > self.mask_threshold).to(torch.uint8)
+                )
+                is_hole = (labels > 0) & (areas <= self.max_sprinkle_area)
+                is_hole = is_hole.reshape_as(masks)
+                # We fill holes with negative mask score (-10.0) to change them to background.
+                masks = torch.where(is_hole, self.mask_threshold - 10.0, masks)
+        except Exception as e:
+            # Skip the post-processing step if the CUDA kernel fails
+            warnings.warn(
+                f"{e}\n\nSkipping the post-processing step due to the error above. You can "
+                "still use SAM 3 and it's OK to ignore the error above, although some post-processing "
+                "functionality may be limited (which doesn't affect the results in most cases; see "
+                "https://github.com/facebookresearch/sam3/blob/main/INSTALL.md).",
+                category=UserWarning,
+                stacklevel=2,
+            )
+            masks = input_masks
+
+        masks = F.interpolate(masks, orig_hw, mode="bilinear", align_corners=False)
+        return masks

detect_tools/sam3/sam3/model/utils/sam2_utils.py

@@ -0,0 +1,233 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates. All Rights Reserved
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import os
+from threading import Thread
+
+import numpy as np
+import torch
+from PIL import Image
+from tqdm import tqdm
+
+
+def _load_img_as_tensor(img_path, image_size):
+    img_pil = Image.open(img_path)
+    img_np = np.array(img_pil.convert("RGB").resize((image_size, image_size)))
+    if img_np.dtype == np.uint8:  # np.uint8 is expected for JPEG images
+        img_np = img_np / 255.0
+    else:
+        raise RuntimeError(f"Unknown image dtype: {img_np.dtype} on {img_path}")
+    img = torch.from_numpy(img_np).permute(2, 0, 1)
+    video_width, video_height = img_pil.size  # the original video size
+    return img, video_height, video_width
+
+
+class AsyncVideoFrameLoader:
+    """
+    A list of video frames to be load asynchronously without blocking session start.
+    """
+
+    def __init__(
+        self,
+        img_paths,
+        image_size,
+        offload_video_to_cpu,
+        img_mean,
+        img_std,
+        compute_device,
+    ):
+        self.img_paths = img_paths
+        self.image_size = image_size
+        self.offload_video_to_cpu = offload_video_to_cpu
+        self.img_mean = img_mean
+        self.img_std = img_std
+        # items in `self.images` will be loaded asynchronously
+        self.images = [None] * len(img_paths)
+        # catch and raise any exceptions in the async loading thread
+        self.exception = None
+        # video_height and video_width be filled when loading the first image
+        self.video_height = None
+        self.video_width = None
+        self.compute_device = compute_device
+
+        # load the first frame to fill video_height and video_width and also
+        # to cache it (since it's most likely where the user will click)
+        self.__getitem__(0)
+
+        # load the rest of frames asynchronously without blocking the session start
+        def _load_frames():
+            try:
+                for n in tqdm(range(len(self.images)), desc="frame loading (JPEG)"):
+                    self.__getitem__(n)
+            except Exception as e:
+                self.exception = e
+
+        self.thread = Thread(target=_load_frames, daemon=True)
+        self.thread.start()
+
+    def __getitem__(self, index):
+        if self.exception is not None:
+            raise RuntimeError("Failure in frame loading thread") from self.exception
+
+        img = self.images[index]
+        if img is not None:
+            return img
+
+        img, video_height, video_width = _load_img_as_tensor(
+            self.img_paths[index], self.image_size
+        )
+        self.video_height = video_height
+        self.video_width = video_width
+        # normalize by mean and std
+        img -= self.img_mean
+        img /= self.img_std
+        if not self.offload_video_to_cpu:
+            img = img.to(self.compute_device, non_blocking=True)
+        self.images[index] = img
+        return img
+
+    def __len__(self):
+        return len(self.images)
+
+
+def load_video_frames(
+    video_path,
+    image_size,
+    offload_video_to_cpu,
+    img_mean=(0.485, 0.456, 0.406),
+    img_std=(0.229, 0.224, 0.225),
+    async_loading_frames=False,
+    compute_device=torch.device("cuda"),
+):
+    """
+    Load the video frames from video_path. The frames are resized to image_size as in
+    the model and are loaded to GPU if offload_video_to_cpu=False. This is used by the demo.
+    """
+    is_bytes = isinstance(video_path, bytes)
+    is_str = isinstance(video_path, str)
+    is_mp4_path = is_str and os.path.splitext(video_path)[-1] in [".mp4", ".MP4"]
+    if is_bytes or is_mp4_path:
+        return load_video_frames_from_video_file(
+            video_path=video_path,
+            image_size=image_size,
+            offload_video_to_cpu=offload_video_to_cpu,
+            img_mean=img_mean,
+            img_std=img_std,
+            compute_device=compute_device,
+        )
+    elif is_str and os.path.isdir(video_path):
+        return load_video_frames_from_jpg_images(
+            video_path=video_path,
+            image_size=image_size,
+            offload_video_to_cpu=offload_video_to_cpu,
+            img_mean=img_mean,
+            img_std=img_std,
+            async_loading_frames=async_loading_frames,
+            compute_device=compute_device,
+        )
+    else:
+        raise NotImplementedError(
+            "Only MP4 video and JPEG folder are supported at this moment"
+        )
+
+
+def load_video_frames_from_jpg_images(
+    video_path,
+    image_size,
+    offload_video_to_cpu,
+    img_mean=(0.485, 0.456, 0.406),
+    img_std=(0.229, 0.224, 0.225),
+    async_loading_frames=False,
+    compute_device=torch.device("cuda"),
+):
+    """
+    Load the video frames from a directory of JPEG files ("<frame_index>.jpg" format).
+
+    The frames are resized to image_size x image_size and are loaded to GPU if
+    `offload_video_to_cpu` is `False` and to CPU if `offload_video_to_cpu` is `True`.
+
+    You can load a frame asynchronously by setting `async_loading_frames` to `True`.
+    """
+    if isinstance(video_path, str) and os.path.isdir(video_path):
+        jpg_folder = video_path
+    else:
+        raise NotImplementedError(
+            "Only JPEG frames are supported at this moment. For video files, you may use "
+            "ffmpeg (https://ffmpeg.org/) to extract frames into a folder of JPEG files, such as \n"
+            "```\n"
+            "ffmpeg -i <your_video>.mp4 -q:v 2 -start_number 0 <output_dir>/'%05d.jpg'\n"
+            "```\n"
+            "where `-q:v` generates high-quality JPEG frames and `-start_number 0` asks "
+            "ffmpeg to start the JPEG file from 00000.jpg."
+        )
+
+    frame_names = [
+        p
+        for p in os.listdir(jpg_folder)
+        if os.path.splitext(p)[-1] in [".jpg", ".jpeg", ".JPG", ".JPEG"]
+    ]
+    frame_names.sort(key=lambda p: int(os.path.splitext(p)[0]))
+    num_frames = len(frame_names)
+    if num_frames == 0:
+        raise RuntimeError(f"no images found in {jpg_folder}")
+    img_paths = [os.path.join(jpg_folder, frame_name) for frame_name in frame_names]
+    img_mean = torch.tensor(img_mean, dtype=torch.float32)[:, None, None]
+    img_std = torch.tensor(img_std, dtype=torch.float32)[:, None, None]
+
+    if async_loading_frames:
+        lazy_images = AsyncVideoFrameLoader(
+            img_paths,
+            image_size,
+            offload_video_to_cpu,
+            img_mean,
+            img_std,
+            compute_device,
+        )
+        return lazy_images, lazy_images.video_height, lazy_images.video_width
+
+    images = torch.zeros(num_frames, 3, image_size, image_size, dtype=torch.float32)
+    for n, img_path in enumerate(tqdm(img_paths, desc="frame loading (JPEG)")):
+        images[n], video_height, video_width = _load_img_as_tensor(img_path, image_size)
+    if not offload_video_to_cpu:
+        images = images.to(compute_device)
+        img_mean = img_mean.to(compute_device)
+        img_std = img_std.to(compute_device)
+    # normalize by mean and std
+    images -= img_mean
+    images /= img_std
+    return images, video_height, video_width
+
+
+def load_video_frames_from_video_file(
+    video_path,
+    image_size,
+    offload_video_to_cpu,
+    img_mean=(0.485, 0.456, 0.406),
+    img_std=(0.229, 0.224, 0.225),
+    compute_device=torch.device("cuda"),
+):
+    """Load the video frames from a video file."""
+    import decord
+
+    img_mean = torch.tensor(img_mean, dtype=torch.float32)[:, None, None]
+    img_std = torch.tensor(img_std, dtype=torch.float32)[:, None, None]
+    # Get the original video height and width
+    decord.bridge.set_bridge("torch")
+    video_height, video_width, _ = decord.VideoReader(video_path).next().shape
+    # Iterate over all frames in the video
+    images = []
+    for frame in decord.VideoReader(video_path, width=image_size, height=image_size):
+        images.append(frame.permute(2, 0, 1))
+
+    images = torch.stack(images, dim=0).float() / 255.0
+    if not offload_video_to_cpu:
+        images = images.to(compute_device)
+        img_mean = img_mean.to(compute_device)
+        img_std = img_std.to(compute_device)
+    # normalize by mean and std
+    images -= img_mean
+    images /= img_std
+    return images, video_height, video_width

detect_tools/sam3/sam3/model/vitdet.py

@@ -0,0 +1,879 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates. All Rights Reserved
+
+"""
+ViTDet backbone adapted from Detectron2.
+This module implements Vision Transformer (ViT) backbone for object detection.
+
+Rope embedding code adopted from:
+1. https://github.com/meta-llama/codellama/blob/main/llama/model.py
+2. https://github.com/naver-ai/rope-vit
+3. https://github.com/lucidrains/rotary-embedding-torch
+"""
+
+import math
+from functools import partial
+from typing import Callable, List, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import torch.utils.checkpoint as checkpoint
+
+try:
+    from timm.layers import DropPath, Mlp, trunc_normal_
+except ModuleNotFoundError:
+    # compatibility for older timm versions
+    from timm.models.layers import DropPath, Mlp, trunc_normal_
+from torch import Tensor
+
+from .model_misc import LayerScale
+
+
+def init_t_xy(
+    end_x: int, end_y: int, scale: float = 1.0, offset: int = 0
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    t = torch.arange(end_x * end_y, dtype=torch.float32)
+    t_x = (t % end_x).float()
+    t_y = torch.div(t, end_x, rounding_mode="floor").float()
+    return t_x * scale + offset, t_y * scale + offset
+
+
+def compute_axial_cis(
+    dim: int,
+    end_x: int,
+    end_y: int,
+    theta: float = 10000.0,
+    scale_pos: float = 1.0,
+    offset: int = 0,
+) -> torch.Tensor:
+    freqs_x = 1.0 / (theta ** (torch.arange(0, dim, 4)[: (dim // 4)].float() / dim))
+    freqs_y = 1.0 / (theta ** (torch.arange(0, dim, 4)[: (dim // 4)].float() / dim))
+
+    t_x, t_y = init_t_xy(end_x, end_y, scale_pos, offset)
+    freqs_x = torch.outer(t_x, freqs_x)
+    freqs_y = torch.outer(t_y, freqs_y)
+    freqs_cis_x = torch.polar(torch.ones_like(freqs_x), freqs_x)
+    freqs_cis_y = torch.polar(torch.ones_like(freqs_y), freqs_y)
+    return torch.cat([freqs_cis_x, freqs_cis_y], dim=-1)
+
+
+def reshape_for_broadcast(freqs_cis: torch.Tensor, x: torch.Tensor) -> torch.Tensor:
+    ndim = x.ndim
+    assert 0 <= 1 < ndim
+    assert freqs_cis.shape == (x.shape[-2], x.shape[-1])
+    shape = [d if i >= ndim - 2 else 1 for i, d in enumerate(x.shape)]
+    return freqs_cis.view(*shape)
+
+
+def apply_rotary_enc(
+    xq: torch.Tensor,
+    xk: torch.Tensor,
+    freqs_cis: torch.Tensor,
+    repeat_freqs_k: bool = False,
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    xq_ = torch.view_as_complex(xq.float().reshape(*xq.shape[:-1], -1, 2))
+    xk_ = (
+        torch.view_as_complex(xk.float().reshape(*xk.shape[:-1], -1, 2))
+        if xk.shape[-2] != 0
+        else None
+    )
+    freqs_cis = reshape_for_broadcast(freqs_cis, xq_)
+    xq_out = torch.view_as_real(xq_ * freqs_cis).flatten(3)
+    if xk_ is None:
+        # no keys to rotate, due to dropout
+        return xq_out.type_as(xq).to(xq.device), xk
+    # repeat freqs along seq_len dim to match k seq_len
+    if repeat_freqs_k:
+        r = xk_.shape[-2] // xq_.shape[-2]
+        freqs_cis = freqs_cis.repeat(*([1] * (freqs_cis.ndim - 2)), r, 1)
+    xk_out = torch.view_as_real(xk_ * freqs_cis).flatten(3)
+    return xq_out.type_as(xq).to(xq.device), xk_out.type_as(xk).to(xk.device)
+
+
+def window_partition(x: Tensor, window_size: int) -> Tuple[Tensor, Tuple[int, int]]:
+    """
+    Partition into non-overlapping windows with padding if needed.
+    Args:
+        x (tensor): input tokens with [B, H, W, C].
+        window_size (int): window size.
+    Returns:
+        windows: windows after partition with [B * num_windows, window_size, window_size, C].
+        (Hp, Wp): padded height and width before partition
+    """
+    B, H, W, C = x.shape
+
+    pad_h = (window_size - H % window_size) % window_size
+    pad_w = (window_size - W % window_size) % window_size
+    if pad_h > 0 or pad_w > 0:
+        x = F.pad(x, (0, 0, 0, pad_w, 0, pad_h))
+    Hp, Wp = H + pad_h, W + pad_w
+
+    x = x.view(B, Hp // window_size, window_size, Wp // window_size, window_size, C)
+    windows = x.permute(0, 1, 3, 2, 4, 5).reshape(-1, window_size, window_size, C)
+    return windows, (Hp, Wp)
+
+
+def window_unpartition(
+    windows: Tensor, window_size: int, pad_hw: Tuple[int, int], hw: Tuple[int, int]
+) -> Tensor:
+    """
+    Window unpartition into original sequences and removing padding.
+    Args:
+        x (tensor): input tokens with [B * num_windows, window_size, window_size, C].
+        window_size (int): window size.
+        pad_hw (Tuple): padded height and width (Hp, Wp).
+        hw (Tuple): original height and width (H, W) before padding.
+    Returns:
+        x: unpartitioned sequences with [B, H, W, C].
+    """
+    Hp, Wp = pad_hw
+    H, W = hw
+    B = windows.shape[0] // (Hp * Wp // window_size // window_size)
+    x = windows.reshape(
+        B, Hp // window_size, Wp // window_size, window_size, window_size, -1
+    )
+    x = x.permute(0, 1, 3, 2, 4, 5).reshape(B, Hp, Wp, -1)
+
+    if Hp > H or Wp > W:
+        x = x[:, :H, :W, :]
+    return x
+
+
+def get_rel_pos(q_size: int, k_size: int, rel_pos: Tensor) -> Tensor:
+    """
+    Get relative positional embeddings according to the relative positions of
+        query and key sizes.
+    Args:
+        q_size (int): size of query q.
+        k_size (int): size of key k.
+        rel_pos (Tensor): relative position embeddings (L, C).
+    Returns:
+        Extracted positional embeddings according to relative positions.
+    """
+    max_rel_dist = int(2 * max(q_size, k_size) - 1)
+    # Interpolate rel pos if needed.
+    if rel_pos.shape[0] != max_rel_dist:
+        # Interpolate rel pos.
+        rel_pos_resized = F.interpolate(
+            rel_pos.reshape(1, rel_pos.shape[0], -1).permute(0, 2, 1),
+            size=max_rel_dist,
+            mode="linear",
+            align_corners=False,
+        )
+        rel_pos_resized = rel_pos_resized.reshape(-1, max_rel_dist).permute(1, 0)
+    else:
+        rel_pos_resized = rel_pos
+
+    # Scale the coords with short length if shapes for q and k are different.
+    q_coords = torch.arange(q_size)[:, None] * max(k_size / q_size, 1.0)
+    k_coords = torch.arange(k_size)[None, :] * max(q_size / k_size, 1.0)
+    relative_coords = (q_coords - k_coords) + (k_size - 1) * max(q_size / k_size, 1.0)
+
+    return rel_pos_resized[relative_coords.long()]
+
+
+def get_abs_pos(
+    abs_pos: Tensor,
+    has_cls_token: bool,
+    hw: Tuple[int, int],
+    retain_cls_token: bool = False,
+    tiling: bool = False,
+) -> Tensor:
+    """
+    Calculate absolute positional embeddings. If needed, resize embeddings and remove cls_token
+        dimension for the original embeddings.
+    Args:
+        abs_pos (Tensor): absolute positional embeddings with (1, num_position, C).
+        has_cls_token (bool): If true, has 1 embedding in abs_pos for cls token.
+        hw (Tuple): size of input image tokens.
+        retain_cls_token: whether to retain the cls_token
+        tiling: whether to tile the embeddings, *instead* of interpolation (a la abs_win)
+    Returns:
+        Absolute positional embeddings after processing with shape (1, H, W, C),
+        if retain_cls_token is False, otherwise (1, 1+H*W, C)
+    """
+    if retain_cls_token:
+        assert has_cls_token
+
+    h, w = hw
+    if has_cls_token:
+        cls_pos = abs_pos[:, :1]
+        abs_pos = abs_pos[:, 1:]
+
+    xy_num = abs_pos.shape[1]
+    size = int(math.sqrt(xy_num))
+    assert size * size == xy_num
+
+    if size != h or size != w:
+        new_abs_pos = abs_pos.reshape(1, size, size, -1).permute(0, 3, 1, 2)
+        if tiling:
+            new_abs_pos = new_abs_pos.tile(
+                [1, 1] + [x // y + 1 for x, y in zip((h, w), new_abs_pos.shape[2:])]
+            )[:, :, :h, :w]
+        else:
+            new_abs_pos = F.interpolate(
+                new_abs_pos,
+                size=(h, w),
+                mode="bicubic",
+                align_corners=False,
+            )
+
+        if not retain_cls_token:
+            return new_abs_pos.permute(0, 2, 3, 1)
+        else:
+            # add cls_token back, flatten spatial dims
+            assert has_cls_token
+            return torch.cat(
+                [cls_pos, new_abs_pos.permute(0, 2, 3, 1).reshape(1, h * w, -1)],
+                dim=1,
+            )
+
+    else:
+        if not retain_cls_token:
+            return abs_pos.reshape(1, h, w, -1)
+        else:
+            assert has_cls_token
+            return torch.cat([cls_pos, abs_pos], dim=1)
+
+
+def concat_rel_pos(
+    q: Tensor,
+    k: Tensor,
+    q_hw: Tuple[int, int],
+    k_hw: Tuple[int, int],
+    rel_pos_h: Tensor,
+    rel_pos_w: Tensor,
+    rescale: bool = False,
+    relative_coords: Optional[Tensor] = None,
+) -> Tuple[Tensor, Tensor]:
+    """
+    Concatenate rel pos coeffs to the q & k tensors, so that qk^T is now
+    effectively including rel pos biases.
+    Args:
+        q (Tensor): q tensor with shape (B, L_q, C).
+        k (Tensor): k tensor with shape (B, L_k, C).
+        q_hw, k_hw: These are spatial size of q & k tensors.
+        rel_pos_h, rel_pos_w: These are relative pos embeddings/params of height, width.
+        rescale (bool): whether to rescale. e.g. for use when using sdpa, pytorch will
+            scale by the wrong factor due to the concat.
+    Returns:
+        q, k: But, padded so that qk^T accounts for rel pos biases
+    """
+    q_h, q_w = q_hw
+    k_h, k_w = k_hw
+
+    assert (q_h == q_w) and (k_h == k_w), "only square inputs supported"
+
+    if relative_coords is not None:
+        Rh = rel_pos_h[relative_coords]
+        Rw = rel_pos_w[relative_coords]
+    else:
+        Rh = get_rel_pos(q_h, k_h, rel_pos_h)
+        Rw = get_rel_pos(q_w, k_w, rel_pos_w)
+
+    B, _, dim = q.shape
+    r_q = q.reshape(B, q_h, q_w, dim)
+
+    old_scale = dim**0.5
+    new_scale = (dim + k_h + k_w) ** 0.5 if rescale else old_scale  # for sdpa
+    # attn will be divided by new_scale, but we want to divide q by old_scale
+    scale_ratio = new_scale / old_scale
+
+    rel_h = torch.einsum("bhwc,hkc->bhwk", r_q, Rh) * new_scale  # (B, q_h, q_w, k_h)
+    rel_w = torch.einsum("bhwc,wkc->bhwk", r_q, Rw) * new_scale  # (B, q_h, q_w, k_w)
+
+    eye_h = torch.eye(k_h, dtype=q.dtype, device=q.device)
+    eye_w = torch.eye(k_w, dtype=q.dtype, device=q.device)
+
+    eye_h = eye_h.view(1, k_h, 1, k_h).expand([B, k_h, k_w, k_h])
+    eye_w = eye_w.view(1, 1, k_w, k_w).expand([B, k_h, k_w, k_w])
+
+    q = torch.cat([r_q * scale_ratio, rel_h, rel_w], dim=-1).view(B, q_h * q_w, -1)
+    k = torch.cat([k.view(B, k_h, k_w, -1), eye_h, eye_w], dim=-1).view(
+        B, k_h * k_w, -1
+    )
+
+    return q, k
+
+
+class PatchEmbed(nn.Module):
+    """
+    Image to Patch Embedding.
+    """
+
+    def __init__(
+        self,
+        kernel_size: Tuple[int, int] = (16, 16),
+        stride: Tuple[int, int] = (16, 16),
+        padding: Tuple[int, int] = (0, 0),
+        in_chans: int = 3,
+        embed_dim: int = 768,
+        bias: bool = True,
+    ):
+        """
+        Args:
+            kernel_size (Tuple): kernel size of the projection layer.
+            stride (Tuple): stride of the projection layer.
+            padding (Tuple): padding size of the projection layer.
+            in_chans (int): Number of input image channels.
+            embed_dim (int):  embed_dim (int): Patch embedding dimension.
+        """
+        super().__init__()
+
+        self.proj = nn.Conv2d(
+            in_chans,
+            embed_dim,
+            kernel_size=kernel_size,
+            stride=stride,
+            padding=padding,
+            bias=bias,
+        )
+
+    def forward(self, x: Tensor) -> Tensor:
+        x = self.proj(x)
+        # B C H W -> B H W C
+        x = x.permute(0, 2, 3, 1)
+        return x
+
+
+class Attention(nn.Module):
+    """Multi-head Attention block with relative position embeddings and 2d-rope."""
+
+    def __init__(
+        self,
+        dim: int,
+        num_heads: int = 8,
+        qkv_bias: bool = True,
+        use_rel_pos: bool = False,
+        rel_pos_zero_init: bool = True,
+        input_size: Optional[Tuple[int, int]] = None,
+        cls_token: bool = False,
+        use_rope: bool = False,
+        rope_theta: float = 10000.0,
+        rope_pt_size: Optional[Tuple[int, int]] = None,
+        rope_interp: bool = False,
+    ):
+        """
+        Args:
+            dim (int): Number of input channels.
+            num_heads (int): Number of attention heads.
+            qkv_bias (bool:  If True, add a learnable bias to query, key, value.
+            rel_pos (bool): If True, add relative positional embeddings to the attention map.
+            rel_pos_zero_init (bool): If True, zero initialize relative positional parameters.
+            input_size (int or None): Input resolution for calculating the relative positional
+                parameter size or rope size.
+            attn_type: Type of attention operation, e.g. "vanilla", "vanilla-xformer".
+            cls_token: whether a cls_token is present.
+            use_rope: whether to use rope 2d (indep of use_rel_pos, as it can be used together)
+            rope_theta: control frequencies of rope
+            rope_pt_size: size of rope in previous stage of training, needed for interpolation or tiling
+            rope_interp: whether to interpolate (or extrapolate) rope to match input size
+        """
+        super().__init__()
+        self.num_heads = num_heads
+        self.head_dim = dim // num_heads
+        self.scale = self.head_dim**-0.5
+        self.cls_token = cls_token
+
+        self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
+        self.proj = nn.Linear(dim, dim)
+
+        # rel_pos embeddings and rope
+        self.use_rel_pos = use_rel_pos
+        self.input_size = input_size
+
+        self.use_rope = use_rope
+        self.rope_theta = rope_theta
+        self.rope_pt_size = rope_pt_size
+        self.rope_interp = rope_interp
+
+        # init rel_pos embeddings and rope
+        self._setup_rel_pos(rel_pos_zero_init)
+        self._setup_rope_freqs()
+
+    def _setup_rel_pos(self, rel_pos_zero_init: bool = True) -> None:
+        if not self.use_rel_pos:
+            self.rel_pos_h = None
+            self.rel_pos_w = None
+            return
+
+        assert self.input_size is not None
+        assert self.cls_token is False, "not supported"
+        # initialize relative positional embeddings
+        self.rel_pos_h = nn.Parameter(
+            torch.zeros(2 * self.input_size[0] - 1, self.head_dim)
+        )
+        self.rel_pos_w = nn.Parameter(
+            torch.zeros(2 * self.input_size[1] - 1, self.head_dim)
+        )
+
+        if not rel_pos_zero_init:
+            trunc_normal_(self.rel_pos_h, std=0.02)
+            trunc_normal_(self.rel_pos_w, std=0.02)
+
+        # Precompute the relative coords
+        H, W = self.input_size
+        q_coords = torch.arange(H)[:, None]
+        k_coords = torch.arange(W)[None, :]
+        relative_coords = (q_coords - k_coords) + (H - 1)
+        self.register_buffer("relative_coords", relative_coords.long())
+
+    def _setup_rope_freqs(self) -> None:
+        if not self.use_rope:
+            self.freqs_cis = None
+            return
+
+        assert self.input_size is not None
+        # determine rope input size
+        if self.rope_pt_size is None:
+            self.rope_pt_size = self.input_size
+
+        # initialize 2d rope freqs
+        self.compute_cis = partial(
+            compute_axial_cis,
+            dim=self.head_dim,
+            theta=self.rope_theta,
+        )
+
+        # interpolate rope
+        scale_pos = 1.0
+        if self.rope_interp:
+            scale_pos = self.rope_pt_size[0] / self.input_size[0]
+        # get scaled freqs_cis
+        freqs_cis = self.compute_cis(
+            end_x=self.input_size[0],
+            end_y=self.input_size[1],
+            scale_pos=scale_pos,
+        )
+        if self.cls_token:
+            t = torch.zeros(
+                self.head_dim // 2,
+                dtype=torch.float32,
+                device=freqs_cis.device,
+            )
+            cls_freqs_cis = torch.polar(torch.ones_like(t), t)[None, :]
+            freqs_cis = torch.cat([cls_freqs_cis, freqs_cis], dim=0)
+
+        self.register_buffer("freqs_cis", freqs_cis)
+
+    def _apply_rope(self, q, k) -> Tuple[Tensor, Tensor]:
+        if not self.use_rope:
+            return q, k
+
+        assert self.freqs_cis is not None
+        return apply_rotary_enc(q, k, freqs_cis=self.freqs_cis)
+
+    def forward(self, x: Tensor) -> Tensor:
+        s = 1 if self.cls_token else 0  # used to exclude cls_token
+        if x.ndim == 4:
+            B, H, W, _ = x.shape
+            assert s == 0  # no cls_token
+            L = H * W
+            ndim = 4
+        else:
+            assert x.ndim == 3
+            B, L, _ = x.shape
+            ndim = 3
+            H = W = math.sqrt(L - s)
+
+        # qkv with shape (3, B, nHead, L, C)
+        qkv = self.qkv(x).reshape(B, L, 3, self.num_heads, -1)
+        # q, k, v with shape (B, nHead, L, C)
+        q, k, v = qkv.permute(2, 0, 3, 1, 4).unbind(0)
+
+        # handle rope and rel pos embeddings
+        q, k = self._apply_rope(q, k)
+        if self.use_rel_pos:
+            q, k = concat_rel_pos(
+                q.flatten(0, 1),
+                k.flatten(0, 1),
+                (H, W),
+                x.shape[1:3],
+                self.rel_pos_h,
+                self.rel_pos_w,
+                rescale=True,
+                relative_coords=self.relative_coords,
+            )
+
+            # sdpa expects [B, nheads, H*W, C] so we transpose back
+            q = q.reshape(B, self.num_heads, H * W, -1)
+            k = k.reshape(B, self.num_heads, H * W, -1)
+
+        x = F.scaled_dot_product_attention(q, k, v)
+
+        if ndim == 4:
+            x = (
+                x.view(B, self.num_heads, H, W, -1)
+                .permute(0, 2, 3, 1, 4)
+                .reshape(B, H, W, -1)
+            )
+        else:
+            x = x.view(B, self.num_heads, L, -1).permute(0, 2, 1, 3).reshape(B, L, -1)
+
+        x = self.proj(x)
+
+        return x
+
+
+class Block(nn.Module):
+    """Transformer blocks with support of window attention"""
+
+    def __init__(
+        self,
+        dim: int,
+        num_heads: int,
+        mlp_ratio: float = 4.0,
+        qkv_bias: bool = True,
+        drop_path: float = 0.0,
+        norm_layer: Callable[..., nn.Module] = nn.LayerNorm,
+        act_layer: Callable[..., nn.Module] = nn.GELU,
+        use_rel_pos: bool = False,
+        rel_pos_zero_init: bool = True,
+        window_size: int = 0,
+        input_size: Optional[Tuple[int, int]] = None,
+        use_rope: bool = False,
+        rope_pt_size: Optional[Tuple[int, int]] = None,
+        rope_tiled: bool = False,
+        rope_interp: bool = False,
+        use_ve_rope: bool = False,
+        cls_token: bool = False,
+        dropout: float = 0.0,
+        init_values: Optional[float] = None,
+    ):
+        """
+        Args:
+            dim (int): Number of input channels.
+            num_heads (int): Number of attention heads in each ViT block.
+            mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
+            qkv_bias (bool): If True, add a learnable bias to query, key, value.
+            drop_path (float): Stochastic depth rate.
+            norm_layer (nn.Module): Normalization layer.
+            act_layer (nn.Module): Activation layer.
+            use_rel_pos (bool): If True, add relative positional embeddings to the attention map.
+            rel_pos_zero_init (bool): If True, zero initialize relative positional parameters.
+            window_size (int): Window size for window attention blocks. If it equals 0, then not
+                use window attention.
+            input_size (int or None): Input resolution for calculating the relative positional
+                parameter size.
+            dropout (float): Dropout rate.
+            cls_token: whether a cls_token is present.
+            use_rope: whether to use rope 2d (indep of use_rel_pos, as it can be used together)
+            rope_pt_size: size of rope in previous stage of training, needed for interpolation or tiling
+            rope_interp: whether to interpolate (or extrapolate) rope to match target input size,
+                expected to specify source size as rope_pt_size.
+        """
+        super().__init__()
+        self.norm1 = norm_layer(dim)
+        self.attn = Attention(
+            dim,
+            num_heads=num_heads,
+            qkv_bias=qkv_bias,
+            use_rel_pos=use_rel_pos,
+            rel_pos_zero_init=rel_pos_zero_init,
+            input_size=input_size if window_size == 0 else (window_size, window_size),
+            use_rope=use_rope,
+            rope_pt_size=rope_pt_size,
+            rope_interp=rope_interp,
+            cls_token=cls_token,
+        )
+        self.ls1 = (
+            LayerScale(dim, init_values=init_values) if init_values else nn.Identity()
+        )
+        self.drop_path = DropPath(drop_path) if drop_path > 0.0 else nn.Identity()
+
+        self.norm2 = norm_layer(dim)
+        self.mlp = Mlp(
+            in_features=dim,
+            hidden_features=int(dim * mlp_ratio),
+            act_layer=act_layer,
+            drop=(dropout, 0.0),
+        )
+        self.ls2 = (
+            LayerScale(dim, init_values=init_values) if init_values else nn.Identity()
+        )
+        self.dropout = nn.Dropout(dropout)
+        self.window_size = window_size
+
+    def forward(self, x: Tensor) -> Tensor:
+        shortcut = x
+        x = self.norm1(x)
+        # Window partition
+        if self.window_size > 0:
+            H, W = x.shape[1], x.shape[2]
+            x, pad_hw = window_partition(x, self.window_size)
+
+        x = self.ls1(self.attn(x))
+        # Reverse window partition
+        if self.window_size > 0:
+            x = window_unpartition(x, self.window_size, pad_hw, (H, W))
+
+        x = shortcut + self.dropout(self.drop_path(x))
+        x = x + self.dropout(self.drop_path(self.ls2(self.mlp(self.norm2(x)))))
+
+        return x
+
+
+class ViT(nn.Module):
+    """
+    This module implements Vision Transformer (ViT) backbone in :paper:`vitdet`.
+    "Exploring Plain Vision Transformer Backbones for Object Detection",
+    https://arxiv.org/abs/2203.16527
+    """
+
+    def __init__(
+        self,
+        img_size: int = 1024,
+        patch_size: int = 16,
+        in_chans: int = 3,
+        embed_dim: int = 768,
+        depth: int = 12,
+        num_heads: int = 12,
+        mlp_ratio: float = 4.0,
+        qkv_bias: bool = True,
+        drop_path_rate: float = 0.0,
+        norm_layer: Union[Callable[..., nn.Module], str] = "LayerNorm",
+        act_layer: Callable[..., nn.Module] = nn.GELU,
+        use_abs_pos: bool = True,
+        tile_abs_pos: bool = True,
+        rel_pos_blocks: Union[Tuple[int, ...], bool] = (2, 5, 8, 11),
+        rel_pos_zero_init: bool = True,
+        window_size: int = 14,
+        global_att_blocks: Tuple[int, ...] = (2, 5, 8, 11),
+        use_rope: bool = False,
+        rope_pt_size: Optional[int] = None,
+        use_interp_rope: bool = False,
+        pretrain_img_size: int = 224,
+        pretrain_use_cls_token: bool = True,
+        retain_cls_token: bool = True,
+        dropout: float = 0.0,
+        return_interm_layers: bool = False,
+        init_values: Optional[float] = None,  # for layerscale
+        ln_pre: bool = False,
+        ln_post: bool = False,
+        bias_patch_embed: bool = True,
+        compile_mode: Optional[str] = None,
+        use_act_checkpoint: bool = True,
+    ):
+        """
+        Args:
+            img_size (int): Input image size. Only relevant for rel pos or rope.
+            patch_size (int): Patch size.
+            in_chans (int): Number of input image channels.
+            embed_dim (int): Patch embedding dimension.
+            depth (int): Depth of ViT.
+            num_heads (int): Number of attention heads in each ViT block.
+            mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
+            qkv_bias (bool): If True, add a learnable bias to query, key, value.
+            drop_path_rate (float): Stochastic depth rate.
+            norm_layer (nn.Module): Normalization layer.
+            act_layer (nn.Module): Activation layer.
+            use_abs_pos (bool): If True, use absolute positional embeddings.
+            tile_abs_pos (bool): If True, tile absolute positional embeddings instead of interpolation.
+            rel_pos_blocks (list): Blocks which have rel pos embeddings.
+            rel_pos_zero_init (bool): If True, zero initialize relative positional parameters.
+            window_size (int): Window size for window attention blocks.
+            global_att_blocks (list): Indexes for blocks using global attention (other blocks use window attention).
+            use_rope (bool): whether to use rope 2d (indep of rel_pos_blocks, as it can be used together).
+            rope_pt_size (int): size of rope in previous stage of training, needed for interpolation or tiling.
+            use_interp_rope: whether to interpolate (or extrapolate) rope to match target input size,
+                expected to specify source size as rope_pt_size.
+            use_act_checkpoint (bool): If True, use activation checkpointing.
+            pretrain_img_size (int): input image size for pretraining models.
+            pretrain_use_cls_token (bool): If True, pretraining models use class token.
+            retain_cls_token: whether cls_token should be retained.
+            dropout (float): Dropout rate. Applied in residual blocks of attn, mlp and inside the mlp.
+
+            return_interm_layers (bool): Whether to return intermediate layers (all global attention blocks).
+            init_values: layer scale init, None for no layer scale.
+
+            ln_pre (bool): If True, apply layer norm before transformer blocks.
+            ln_post (bool): If True, apply layer norm after transformer blocks.
+            bias_patch_embed (bool): bias in conv for patch embed?
+            compile_mode (str): mode to compile the forward
+        """
+        super().__init__()
+        self.pretrain_use_cls_token = pretrain_use_cls_token
+
+        window_block_indexes = [i for i in range(depth) if i not in global_att_blocks]
+        self.full_attn_ids = list(global_att_blocks)
+        self.rel_pos_blocks = [False] * depth
+        if isinstance(rel_pos_blocks, bool) and rel_pos_blocks:
+            self.rel_pos_blocks = [True] * depth
+        else:
+            for i in rel_pos_blocks:
+                self.rel_pos_blocks[i] = True
+
+        self.retain_cls_token = retain_cls_token
+        if self.retain_cls_token:
+            assert pretrain_use_cls_token
+            assert (
+                len(window_block_indexes) == 0
+            ), "windowing not supported with cls token"
+
+            assert sum(self.rel_pos_blocks) == 0, "rel pos not supported with cls token"
+
+            scale = embed_dim**-0.5
+            self.class_embedding = nn.Parameter(scale * torch.randn(1, 1, embed_dim))
+
+        if isinstance(norm_layer, str):
+            norm_layer = partial(getattr(nn, norm_layer), eps=1e-5)
+
+        self.patch_embed = PatchEmbed(
+            kernel_size=(patch_size, patch_size),
+            stride=(patch_size, patch_size),
+            in_chans=in_chans,
+            embed_dim=embed_dim,
+            bias=bias_patch_embed,
+        )
+
+        # Handle absolute positional embedding
+        self.tile_abs_pos = tile_abs_pos
+        self.use_abs_pos = use_abs_pos
+        if self.tile_abs_pos:
+            assert self.use_abs_pos
+
+        if self.use_abs_pos:
+            # Initialize absolute positional embedding with pretrain image size.
+            num_patches = (pretrain_img_size // patch_size) * (
+                pretrain_img_size // patch_size
+            )
+            num_positions = (num_patches + 1) if pretrain_use_cls_token else num_patches
+            self.pos_embed = nn.Parameter(torch.zeros(1, num_positions, embed_dim))
+        else:
+            self.pos_embed = None
+
+        # stochastic depth decay rule
+        dpr = [x.item() for x in torch.linspace(0, drop_path_rate, depth)]
+
+        self.blocks = nn.ModuleList()
+        cur_stage = 1
+        for i in range(depth):
+            block = Block(
+                dim=embed_dim,
+                num_heads=num_heads,
+                mlp_ratio=mlp_ratio,
+                qkv_bias=qkv_bias,
+                drop_path=dpr[i],
+                norm_layer=norm_layer,
+                act_layer=act_layer,
+                use_rel_pos=self.rel_pos_blocks[i],
+                rel_pos_zero_init=rel_pos_zero_init,
+                window_size=window_size if i in window_block_indexes else 0,
+                input_size=(img_size // patch_size, img_size // patch_size),
+                use_rope=use_rope,
+                rope_pt_size=(
+                    (window_size, window_size)
+                    if rope_pt_size is None
+                    else (rope_pt_size, rope_pt_size)
+                ),
+                rope_interp=use_interp_rope,
+                cls_token=self.retain_cls_token,
+                dropout=dropout,
+                init_values=init_values,
+            )
+
+            if i not in window_block_indexes:
+                cur_stage += 1
+
+            self.use_act_checkpoint = use_act_checkpoint
+
+            self.blocks.append(block)
+
+        self.return_interm_layers = return_interm_layers
+        self.channel_list = (
+            [embed_dim] * len(self.full_attn_ids)
+            if return_interm_layers
+            else [embed_dim]
+        )
+
+        if self.pos_embed is not None:
+            trunc_normal_(self.pos_embed, std=0.02)
+
+        self.ln_pre = norm_layer(embed_dim) if ln_pre else nn.Identity()
+        self.ln_post = norm_layer(embed_dim) if ln_post else nn.Identity()
+
+        self.apply(self._init_weights)
+
+        if compile_mode is not None:
+            self.forward = torch.compile(
+                self.forward, mode=compile_mode, fullgraph=True
+            )
+            if self.use_act_checkpoint and self.training:
+                torch._dynamo.config.optimize_ddp = False
+
+    def _init_weights(self, m: nn.Module) -> None:
+        if isinstance(m, nn.Linear):
+            trunc_normal_(m.weight, std=0.02)
+            if isinstance(m, nn.Linear) and m.bias is not None:
+                nn.init.constant_(m.bias, 0)
+        elif isinstance(m, nn.LayerNorm):
+            nn.init.constant_(m.bias, 0)
+            nn.init.constant_(m.weight, 1.0)
+
+    def forward(self, x: torch.Tensor) -> List[torch.Tensor]:
+        x = self.patch_embed(x)
+        h, w = x.shape[1], x.shape[2]
+
+        s = 0
+        if self.retain_cls_token:
+            # If cls_token is retained, we don't
+            # maintain spatial shape
+            x = torch.cat([self.class_embedding, x.flatten(1, 2)], dim=1)
+            s = 1
+
+        if self.pos_embed is not None:
+            x = x + get_abs_pos(
+                self.pos_embed,
+                self.pretrain_use_cls_token,
+                (h, w),
+                self.retain_cls_token,
+                tiling=self.tile_abs_pos,
+            )
+
+        x = self.ln_pre(x)
+
+        outputs = []
+        for i, blk in enumerate(self.blocks):
+            if self.use_act_checkpoint and self.training:
+                x = checkpoint.checkpoint(blk, x, use_reentrant=False)
+            else:
+                x = blk(x)
+            if (i == self.full_attn_ids[-1]) or (
+                self.return_interm_layers and i in self.full_attn_ids
+            ):
+                if i == self.full_attn_ids[-1]:
+                    x = self.ln_post(x)
+
+                feats = x[:, s:]
+                if feats.ndim == 4:
+                    feats = feats.permute(0, 3, 1, 2)
+                else:
+                    assert feats.ndim == 3
+                    h = w = math.sqrt(feats.shape[1])
+                    feats = feats.reshape(
+                        feats.shape[0], h, w, feats.shape[-1]
+                    ).permute(0, 3, 1, 2)
+
+                outputs.append(feats)
+
+        return outputs
+
+    def get_layer_id(self, layer_name: str) -> int:
+        # https://github.com/microsoft/unilm/blob/master/beit/optim_factory.py#L33
+        num_layers = self.get_num_layers()
+
+        if layer_name.find("rel_pos") != -1:
+            return num_layers + 1
+        elif layer_name.find("ln_pre") != -1:
+            return 0
+        elif layer_name.find("pos_embed") != -1 or layer_name.find("cls_token") != -1:
+            return 0
+        elif layer_name.find("patch_embed") != -1:
+            return 0
+        elif layer_name.find("blocks") != -1:
+            return int(layer_name.split("blocks")[1].split(".")[1]) + 1
+        else:
+            return num_layers + 1
+
+    def get_num_layers(self) -> int:
+        return len(self.blocks)

detect_tools/sam3/sam3/model/vl_combiner.py

@@ -0,0 +1,176 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates. All Rights Reserved
+
+"""Provides utility to combine a vision backbone with a language backbone."""
+
+from copy import copy
+from typing import List, Optional
+
+import torch
+import torch.nn as nn
+
+from torch.nn.attention import sdpa_kernel, SDPBackend
+
+from .act_ckpt_utils import activation_ckpt_wrapper
+from .necks import Sam3DualViTDetNeck
+
+
+class SAM3VLBackbone(nn.Module):
+    """This backbone combines a vision backbone and a language backbone without fusion.
+    As such it is more of a convenience wrapper to handle the two backbones together.
+
+    It adds support for activation checkpointing and compilation.
+    """
+
+    def __init__(
+        self,
+        visual: Sam3DualViTDetNeck,
+        text,
+        compile_visual: bool = False,
+        act_ckpt_whole_vision_backbone: bool = False,
+        act_ckpt_whole_language_backbone: bool = False,
+        scalp=0,
+    ):
+        """Initialize the backbone combiner.
+
+        :param visual: The vision backbone to use
+        :param text: The text encoder to use
+        """
+        super().__init__()
+        self.vision_backbone: Sam3DualViTDetNeck = (
+            torch.compile(visual) if compile_visual else visual
+        )
+        self.language_backbone = text
+        self.scalp = scalp
+        # allow running activation checkpointing on the entire vision and language backbones
+        self.act_ckpt_whole_vision_backbone = act_ckpt_whole_vision_backbone
+        self.act_ckpt_whole_language_backbone = act_ckpt_whole_language_backbone
+
+    def forward(
+        self,
+        samples: torch.Tensor,
+        captions: List[str],
+        input_boxes: Optional[torch.Tensor] = None,
+        additional_text: Optional[List[str]] = None,
+    ):
+        """Forward pass of the backbone combiner.
+
+        :param samples: The input images
+        :param captions: The input captions
+        :param input_boxes: If the text contains place-holders for boxes, this
+            parameter contains the tensor containing their spatial features
+        :param additional_text: This can be used to encode some additional text
+            (different from the captions) in the same forward of the backbone
+        :return: Output dictionary with the following keys:
+            - vision_features: The output of the vision backbone
+            - language_features: The output of the language backbone
+            - language_mask: The attention mask of the language backbone
+            - vision_pos_enc: The positional encoding of the vision backbone
+            - (optional) additional_text_features: The output of the language
+                backbone for the additional text
+            - (optional) additional_text_mask: The attention mask of the
+                language backbone for the additional text
+        """
+        output = self.forward_image(samples)
+        device = output["vision_features"].device
+        output.update(self.forward_text(captions, input_boxes, additional_text, device))
+        return output
+
+    def forward_image(self, samples: torch.Tensor):
+        return activation_ckpt_wrapper(self._forward_image_no_act_ckpt)(
+            samples=samples,
+            act_ckpt_enable=self.act_ckpt_whole_vision_backbone and self.training,
+        )
+
+    def _forward_image_no_act_ckpt(self, samples):
+        # Forward through backbone
+        sam3_features, sam3_pos, sam2_features, sam2_pos = self.vision_backbone.forward(
+            samples
+        )
+        if self.scalp > 0:
+            # Discard the lowest resolution features
+            sam3_features, sam3_pos = (
+                sam3_features[: -self.scalp],
+                sam3_pos[: -self.scalp],
+            )
+            if sam2_features is not None and sam2_pos is not None:
+                sam2_features, sam2_pos = (
+                    sam2_features[: -self.scalp],
+                    sam2_pos[: -self.scalp],
+                )
+
+        sam2_output = None
+
+        if sam2_features is not None and sam2_pos is not None:
+            sam2_src = sam2_features[-1]
+            sam2_output = {
+                "vision_features": sam2_src,
+                "vision_pos_enc": sam2_pos,
+                "backbone_fpn": sam2_features,
+            }
+
+        sam3_src = sam3_features[-1]
+        output = {
+            "vision_features": sam3_src,
+            "vision_pos_enc": sam3_pos,
+            "backbone_fpn": sam3_features,
+            "sam2_backbone_out": sam2_output,
+        }
+
+        return output
+
+    def forward_text(
+        self, captions, input_boxes=None, additional_text=None, device="cuda"
+    ):
+        return activation_ckpt_wrapper(self._forward_text_no_ack_ckpt)(
+            captions=captions,
+            input_boxes=input_boxes,
+            additional_text=additional_text,
+            device=device,
+            act_ckpt_enable=self.act_ckpt_whole_language_backbone and self.training,
+        )
+
+    def _forward_text_no_ack_ckpt(
+        self,
+        captions,
+        input_boxes=None,
+        additional_text=None,
+        device="cuda",
+    ):
+        output = {}
+
+        # Forward through text_encoder
+        text_to_encode = copy(captions)
+        if additional_text is not None:
+            # if there are additional_text, we piggy-back them into this forward.
+            # They'll be used later for output alignment
+            text_to_encode += additional_text
+
+        sdpa_context = sdpa_kernel(
+            [
+                SDPBackend.MATH,
+                SDPBackend.EFFICIENT_ATTENTION,
+                SDPBackend.FLASH_ATTENTION,
+            ]
+        )
+
+        with sdpa_context:
+            text_attention_mask, text_memory, text_embeds = self.language_backbone(
+                text_to_encode, input_boxes, device=device
+            )
+
+        if additional_text is not None:
+            output["additional_text_features"] = text_memory[:, -len(additional_text) :]
+            output["additional_text_mask"] = text_attention_mask[
+                -len(additional_text) :
+            ]
+
+        text_memory = text_memory[:, : len(captions)]
+        text_attention_mask = text_attention_mask[: len(captions)]
+        text_embeds = text_embeds[:, : len(captions)]
+        output["language_features"] = text_memory
+        output["language_mask"] = text_attention_mask
+        output["language_embeds"] = (
+            text_embeds  # Text embeddings before forward to the encoder
+        )
+
+        return output

detect_tools/sam3/sam3/model_builder.py

@@ -0,0 +1,793 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates. All Rights Reserved
+
+import os
+from typing import Optional
+
+import torch
+import torch.nn as nn
+from huggingface_hub import hf_hub_download
+from iopath.common.file_io import g_pathmgr
+from sam3.model.decoder import (
+    TransformerDecoder,
+    TransformerDecoderLayer,
+    TransformerDecoderLayerv2,
+    TransformerEncoderCrossAttention,
+)
+from sam3.model.encoder import TransformerEncoderFusion, TransformerEncoderLayer
+from sam3.model.geometry_encoders import SequenceGeometryEncoder
+from sam3.model.maskformer_segmentation import PixelDecoder, UniversalSegmentationHead
+from sam3.model.memory import (
+    CXBlock,
+    SimpleFuser,
+    SimpleMaskDownSampler,
+    SimpleMaskEncoder,
+)
+from sam3.model.model_misc import (
+    DotProductScoring,
+    MLP,
+    MultiheadAttentionWrapper as MultiheadAttention,
+    TransformerWrapper,
+)
+from sam3.model.necks import Sam3DualViTDetNeck
+from sam3.model.position_encoding import PositionEmbeddingSine
+from sam3.model.sam1_task_predictor import SAM3InteractiveImagePredictor
+from sam3.model.sam3_image import Sam3Image, Sam3ImageOnVideoMultiGPU
+from sam3.model.sam3_tracking_predictor import Sam3TrackerPredictor
+from sam3.model.sam3_video_inference import Sam3VideoInferenceWithInstanceInteractivity
+from sam3.model.sam3_video_predictor import Sam3VideoPredictorMultiGPU
+from sam3.model.text_encoder_ve import VETextEncoder
+from sam3.model.tokenizer_ve import SimpleTokenizer
+from sam3.model.vitdet import ViT
+from sam3.model.vl_combiner import SAM3VLBackbone
+from sam3.sam.transformer import RoPEAttention
+
+
+# Setup TensorFloat-32 for Ampere GPUs if available
+def _setup_tf32() -> None:
+    """Enable TensorFloat-32 for Ampere GPUs if available."""
+    if torch.cuda.is_available():
+        device_props = torch.cuda.get_device_properties(0)
+        if device_props.major >= 8:
+            torch.backends.cuda.matmul.allow_tf32 = True
+            torch.backends.cudnn.allow_tf32 = True
+
+
+_setup_tf32()
+
+
+def _create_position_encoding(precompute_resolution=None):
+    """Create position encoding for visual backbone."""
+    return PositionEmbeddingSine(
+        num_pos_feats=256,
+        normalize=True,
+        scale=None,
+        temperature=10000,
+        precompute_resolution=precompute_resolution,
+    )
+
+
+def _create_vit_backbone(compile_mode=None):
+    """Create ViT backbone for visual feature extraction."""
+    return ViT(
+        img_size=1008,
+        pretrain_img_size=336,
+        patch_size=14,
+        embed_dim=1024,
+        depth=32,
+        num_heads=16,
+        mlp_ratio=4.625,
+        norm_layer="LayerNorm",
+        drop_path_rate=0.1,
+        qkv_bias=True,
+        use_abs_pos=True,
+        tile_abs_pos=True,
+        global_att_blocks=(7, 15, 23, 31),
+        rel_pos_blocks=(),
+        use_rope=True,
+        use_interp_rope=True,
+        window_size=24,
+        pretrain_use_cls_token=True,
+        retain_cls_token=False,
+        ln_pre=True,
+        ln_post=False,
+        return_interm_layers=False,
+        bias_patch_embed=False,
+        compile_mode=compile_mode,
+    )
+
+
+def _create_vit_neck(position_encoding, vit_backbone, enable_inst_interactivity=False):
+    """Create ViT neck for feature pyramid."""
+    return Sam3DualViTDetNeck(
+        position_encoding=position_encoding,
+        d_model=256,
+        scale_factors=[4.0, 2.0, 1.0, 0.5],
+        trunk=vit_backbone,
+        add_sam2_neck=enable_inst_interactivity,
+    )
+
+
+def _create_vl_backbone(vit_neck, text_encoder):
+    """Create visual-language backbone."""
+    return SAM3VLBackbone(visual=vit_neck, text=text_encoder, scalp=1)
+
+
+def _create_transformer_encoder() -> TransformerEncoderFusion:
+    """Create transformer encoder with its layer."""
+    encoder_layer = TransformerEncoderLayer(
+        activation="relu",
+        d_model=256,
+        dim_feedforward=2048,
+        dropout=0.1,
+        pos_enc_at_attn=True,
+        pos_enc_at_cross_attn_keys=False,
+        pos_enc_at_cross_attn_queries=False,
+        pre_norm=True,
+        self_attention=MultiheadAttention(
+            num_heads=8,
+            dropout=0.1,
+            embed_dim=256,
+            batch_first=True,
+        ),
+        cross_attention=MultiheadAttention(
+            num_heads=8,
+            dropout=0.1,
+            embed_dim=256,
+            batch_first=True,
+        ),
+    )
+
+    encoder = TransformerEncoderFusion(
+        layer=encoder_layer,
+        num_layers=6,
+        d_model=256,
+        num_feature_levels=1,
+        frozen=False,
+        use_act_checkpoint=True,
+        add_pooled_text_to_img_feat=False,
+        pool_text_with_mask=True,
+    )
+    return encoder
+
+
+def _create_transformer_decoder() -> TransformerDecoder:
+    """Create transformer decoder with its layer."""
+    decoder_layer = TransformerDecoderLayer(
+        activation="relu",
+        d_model=256,
+        dim_feedforward=2048,
+        dropout=0.1,
+        cross_attention=MultiheadAttention(
+            num_heads=8,
+            dropout=0.1,
+            embed_dim=256,
+        ),
+        n_heads=8,
+        use_text_cross_attention=True,
+    )
+
+    decoder = TransformerDecoder(
+        layer=decoder_layer,
+        num_layers=6,
+        num_queries=200,
+        return_intermediate=True,
+        box_refine=True,
+        num_o2m_queries=0,
+        dac=True,
+        boxRPB="log",
+        d_model=256,
+        frozen=False,
+        interaction_layer=None,
+        dac_use_selfatt_ln=True,
+        resolution=1008,
+        stride=14,
+        use_act_checkpoint=True,
+        presence_token=True,
+    )
+    return decoder
+
+
+def _create_dot_product_scoring():
+    """Create dot product scoring module."""
+    prompt_mlp = MLP(
+        input_dim=256,
+        hidden_dim=2048,
+        output_dim=256,
+        num_layers=2,
+        dropout=0.1,
+        residual=True,
+        out_norm=nn.LayerNorm(256),
+    )
+    return DotProductScoring(d_model=256, d_proj=256, prompt_mlp=prompt_mlp)
+
+
+def _create_segmentation_head(compile_mode=None):
+    """Create segmentation head with pixel decoder."""
+    pixel_decoder = PixelDecoder(
+        num_upsampling_stages=3,
+        interpolation_mode="nearest",
+        hidden_dim=256,
+        compile_mode=compile_mode,
+    )
+
+    cross_attend_prompt = MultiheadAttention(
+        num_heads=8,
+        dropout=0,
+        embed_dim=256,
+    )
+
+    segmentation_head = UniversalSegmentationHead(
+        hidden_dim=256,
+        upsampling_stages=3,
+        aux_masks=False,
+        presence_head=False,
+        dot_product_scorer=None,
+        act_ckpt=True,
+        cross_attend_prompt=cross_attend_prompt,
+        pixel_decoder=pixel_decoder,
+    )
+    return segmentation_head
+
+
+def _create_geometry_encoder():
+    """Create geometry encoder with all its components."""
+    # Create position encoding for geometry encoder
+    geo_pos_enc = _create_position_encoding()
+    # Create CX block for fuser
+    cx_block = CXBlock(
+        dim=256,
+        kernel_size=7,
+        padding=3,
+        layer_scale_init_value=1.0e-06,
+        use_dwconv=True,
+    )
+    # Create geometry encoder layer
+    geo_layer = TransformerEncoderLayer(
+        activation="relu",
+        d_model=256,
+        dim_feedforward=2048,
+        dropout=0.1,
+        pos_enc_at_attn=False,
+        pre_norm=True,
+        self_attention=MultiheadAttention(
+            num_heads=8,
+            dropout=0.1,
+            embed_dim=256,
+            batch_first=False,
+        ),
+        pos_enc_at_cross_attn_queries=False,
+        pos_enc_at_cross_attn_keys=True,
+        cross_attention=MultiheadAttention(
+            num_heads=8,
+            dropout=0.1,
+            embed_dim=256,
+            batch_first=False,
+        ),
+    )
+
+    # Create geometry encoder
+    input_geometry_encoder = SequenceGeometryEncoder(
+        pos_enc=geo_pos_enc,
+        encode_boxes_as_points=False,
+        points_direct_project=True,
+        points_pool=True,
+        points_pos_enc=True,
+        boxes_direct_project=True,
+        boxes_pool=True,
+        boxes_pos_enc=True,
+        d_model=256,
+        num_layers=3,
+        layer=geo_layer,
+        use_act_ckpt=True,
+        add_cls=True,
+        add_post_encode_proj=True,
+    )
+    return input_geometry_encoder
+
+
+def _create_sam3_model(
+    backbone,
+    transformer,
+    input_geometry_encoder,
+    segmentation_head,
+    dot_prod_scoring,
+    inst_interactive_predictor,
+    eval_mode,
+):
+    """Create the SAM3 image model."""
+    common_params = {
+        "backbone": backbone,
+        "transformer": transformer,
+        "input_geometry_encoder": input_geometry_encoder,
+        "segmentation_head": segmentation_head,
+        "num_feature_levels": 1,
+        "o2m_mask_predict": True,
+        "dot_prod_scoring": dot_prod_scoring,
+        "use_instance_query": False,
+        "multimask_output": True,
+        "inst_interactive_predictor": inst_interactive_predictor,
+    }
+
+    matcher = None
+    if not eval_mode:
+        from sam3.train.matcher import BinaryHungarianMatcherV2
+
+        matcher = BinaryHungarianMatcherV2(
+            focal=True,
+            cost_class=2.0,
+            cost_bbox=5.0,
+            cost_giou=2.0,
+            alpha=0.25,
+            gamma=2,
+            stable=False,
+        )
+    common_params["matcher"] = matcher
+    model = Sam3Image(**common_params)
+
+    return model
+
+
+def _create_tracker_maskmem_backbone():
+    """Create the SAM3 Tracker memory encoder."""
+    # Position encoding for mask memory backbone
+    position_encoding = PositionEmbeddingSine(
+        num_pos_feats=64,
+        normalize=True,
+        scale=None,
+        temperature=10000,
+        precompute_resolution=1008,
+    )
+
+    # Mask processing components
+    mask_downsampler = SimpleMaskDownSampler(
+        kernel_size=3, stride=2, padding=1, interpol_size=[1152, 1152]
+    )
+
+    cx_block_layer = CXBlock(
+        dim=256,
+        kernel_size=7,
+        padding=3,
+        layer_scale_init_value=1.0e-06,
+        use_dwconv=True,
+    )
+
+    fuser = SimpleFuser(layer=cx_block_layer, num_layers=2)
+
+    maskmem_backbone = SimpleMaskEncoder(
+        out_dim=64,
+        position_encoding=position_encoding,
+        mask_downsampler=mask_downsampler,
+        fuser=fuser,
+    )
+
+    return maskmem_backbone
+
+
+def _create_tracker_transformer():
+    """Create the SAM3 Tracker transformer components."""
+    # Self attention
+    self_attention = RoPEAttention(
+        embedding_dim=256,
+        num_heads=1,
+        downsample_rate=1,
+        dropout=0.1,
+        rope_theta=10000.0,
+        feat_sizes=[72, 72],
+        use_fa3=False,
+        use_rope_real=False,
+    )
+
+    # Cross attention
+    cross_attention = RoPEAttention(
+        embedding_dim=256,
+        num_heads=1,
+        downsample_rate=1,
+        dropout=0.1,
+        kv_in_dim=64,
+        rope_theta=10000.0,
+        feat_sizes=[72, 72],
+        rope_k_repeat=True,
+        use_fa3=False,
+        use_rope_real=False,
+    )
+
+    # Encoder layer
+    encoder_layer = TransformerDecoderLayerv2(
+        cross_attention_first=False,
+        activation="relu",
+        dim_feedforward=2048,
+        dropout=0.1,
+        pos_enc_at_attn=False,
+        pre_norm=True,
+        self_attention=self_attention,
+        d_model=256,
+        pos_enc_at_cross_attn_keys=True,
+        pos_enc_at_cross_attn_queries=False,
+        cross_attention=cross_attention,
+    )
+
+    # Encoder
+    encoder = TransformerEncoderCrossAttention(
+        remove_cross_attention_layers=[],
+        batch_first=True,
+        d_model=256,
+        frozen=False,
+        pos_enc_at_input=True,
+        layer=encoder_layer,
+        num_layers=4,
+        use_act_checkpoint=False,
+    )
+
+    # Transformer wrapper
+    transformer = TransformerWrapper(
+        encoder=encoder,
+        decoder=None,
+        d_model=256,
+    )
+
+    return transformer
+
+
+def build_tracker(
+    apply_temporal_disambiguation: bool, with_backbone: bool = False, compile_mode=None
+) -> Sam3TrackerPredictor:
+    """
+    Build the SAM3 Tracker module for video tracking.
+
+    Returns:
+        Sam3TrackerPredictor: Wrapped SAM3 Tracker module
+    """
+
+    # Create model components
+    maskmem_backbone = _create_tracker_maskmem_backbone()
+    transformer = _create_tracker_transformer()
+    backbone = None
+    if with_backbone:
+        vision_backbone = _create_vision_backbone(compile_mode=compile_mode)
+        backbone = SAM3VLBackbone(scalp=1, visual=vision_backbone, text=None)
+    # Create the Tracker module
+    model = Sam3TrackerPredictor(
+        image_size=1008,
+        num_maskmem=7,
+        backbone=backbone,
+        backbone_stride=14,
+        transformer=transformer,
+        maskmem_backbone=maskmem_backbone,
+        # SAM parameters
+        multimask_output_in_sam=True,
+        # Evaluation
+        forward_backbone_per_frame_for_eval=True,
+        trim_past_non_cond_mem_for_eval=False,
+        # Multimask
+        multimask_output_for_tracking=True,
+        multimask_min_pt_num=0,
+        multimask_max_pt_num=1,
+        # Additional settings
+        always_start_from_first_ann_frame=False,
+        # Mask overlap
+        non_overlap_masks_for_mem_enc=False,
+        non_overlap_masks_for_output=False,
+        max_cond_frames_in_attn=4,
+        offload_output_to_cpu_for_eval=False,
+        # SAM decoder settings
+        sam_mask_decoder_extra_args={
+            "dynamic_multimask_via_stability": True,
+            "dynamic_multimask_stability_delta": 0.05,
+            "dynamic_multimask_stability_thresh": 0.98,
+        },
+        clear_non_cond_mem_around_input=True,
+        fill_hole_area=0,
+        use_memory_selection=apply_temporal_disambiguation,
+    )
+
+    return model
+
+
+def _create_text_encoder(bpe_path: str) -> VETextEncoder:
+    """Create SAM3 text encoder."""
+    tokenizer = SimpleTokenizer(bpe_path=bpe_path)
+    return VETextEncoder(
+        tokenizer=tokenizer,
+        d_model=256,
+        width=1024,
+        heads=16,
+        layers=24,
+    )
+
+
+def _create_vision_backbone(
+    compile_mode=None, enable_inst_interactivity=True
+) -> Sam3DualViTDetNeck:
+    """Create SAM3 visual backbone with ViT and neck."""
+    # Position encoding
+    position_encoding = _create_position_encoding(precompute_resolution=1008)
+    # ViT backbone
+    vit_backbone: ViT = _create_vit_backbone(compile_mode=compile_mode)
+    vit_neck: Sam3DualViTDetNeck = _create_vit_neck(
+        position_encoding,
+        vit_backbone,
+        enable_inst_interactivity=enable_inst_interactivity,
+    )
+    # Visual neck
+    return vit_neck
+
+
+def _create_sam3_transformer(has_presence_token: bool = True) -> TransformerWrapper:
+    """Create SAM3 transformer encoder and decoder."""
+    encoder: TransformerEncoderFusion = _create_transformer_encoder()
+    decoder: TransformerDecoder = _create_transformer_decoder()
+
+    return TransformerWrapper(encoder=encoder, decoder=decoder, d_model=256)
+
+
+def _load_checkpoint(model, checkpoint_path):
+    """Load model checkpoint from file."""
+    with g_pathmgr.open(checkpoint_path, "rb") as f:
+        ckpt = torch.load(f, map_location="cpu", weights_only=True)
+    if "model" in ckpt and isinstance(ckpt["model"], dict):
+        ckpt = ckpt["model"]
+    sam3_image_ckpt = {
+        k.replace("detector.", ""): v for k, v in ckpt.items() if "detector" in k
+    }
+    if model.inst_interactive_predictor is not None:
+        sam3_image_ckpt.update(
+            {
+                k.replace("tracker.", "inst_interactive_predictor.model."): v
+                for k, v in ckpt.items()
+                if "tracker" in k
+            }
+        )
+    missing_keys, _ = model.load_state_dict(sam3_image_ckpt, strict=False)
+    if len(missing_keys) > 0:
+        print(
+            f"loaded {checkpoint_path} and found "
+            f"missing and/or unexpected keys:\n{missing_keys=}"
+        )
+
+
+def _setup_device_and_mode(model, device, eval_mode):
+    """Setup model device and evaluation mode."""
+    if device == "cuda":
+        model = model.cuda()
+    if eval_mode:
+        model.eval()
+    return model
+
+
+def build_sam3_image_model(
+    bpe_path=None,
+    device="cuda" if torch.cuda.is_available() else "cpu",
+    eval_mode=True,
+    checkpoint_path=None,
+    load_from_HF=True,
+    enable_segmentation=True,
+    enable_inst_interactivity=False,
+    compile=False,
+):
+    """
+    Build SAM3 image model
+
+    Args:
+        bpe_path: Path to the BPE tokenizer vocabulary
+        device: Device to load the model on ('cuda' or 'cpu')
+        eval_mode: Whether to set the model to evaluation mode
+        checkpoint_path: Optional path to model checkpoint
+        enable_segmentation: Whether to enable segmentation head
+        enable_inst_interactivity: Whether to enable instance interactivity (SAM 1 task)
+        compile_mode: To enable compilation, set to "default"
+
+    Returns:
+        A SAM3 image model
+    """
+    if bpe_path is None:
+        bpe_path = os.path.join(
+            os.path.dirname(__file__), "..", "assets", "bpe_simple_vocab_16e6.txt.gz"
+        )
+    # Create visual components
+    compile_mode = "default" if compile else None
+    vision_encoder = _create_vision_backbone(
+        compile_mode=compile_mode, enable_inst_interactivity=enable_inst_interactivity
+    )
+
+    # Create text components
+    text_encoder = _create_text_encoder(bpe_path)
+
+    # Create visual-language backbone
+    backbone = _create_vl_backbone(vision_encoder, text_encoder)
+
+    # Create transformer components
+    transformer = _create_sam3_transformer()
+
+    # Create dot product scoring
+    dot_prod_scoring = _create_dot_product_scoring()
+
+    # Create segmentation head if enabled
+    segmentation_head = (
+        _create_segmentation_head(compile_mode=compile_mode)
+        if enable_segmentation
+        else None
+    )
+
+    # Create geometry encoder
+    input_geometry_encoder = _create_geometry_encoder()
+    if enable_inst_interactivity:
+        sam3_pvs_base = build_tracker(apply_temporal_disambiguation=False)
+        inst_predictor = SAM3InteractiveImagePredictor(sam3_pvs_base)
+    else:
+        inst_predictor = None
+    # Create the SAM3 model
+    model = _create_sam3_model(
+        backbone,
+        transformer,
+        input_geometry_encoder,
+        segmentation_head,
+        dot_prod_scoring,
+        inst_predictor,
+        eval_mode,
+    )
+    if load_from_HF and checkpoint_path is None:
+        checkpoint_path = download_ckpt_from_hf()
+    # Load checkpoint if provided
+    if checkpoint_path is not None:
+        _load_checkpoint(model, checkpoint_path)
+
+    # Setup device and mode
+    model = _setup_device_and_mode(model, device, eval_mode)
+
+    return model
+
+
+def download_ckpt_from_hf():
+    SAM3_MODEL_ID = "facebook/sam3"
+    SAM3_CKPT_NAME = "sam3.pt"
+    SAM3_CFG_NAME = "config.json"
+    _ = hf_hub_download(repo_id=SAM3_MODEL_ID, filename=SAM3_CFG_NAME)
+    checkpoint_path = hf_hub_download(repo_id=SAM3_MODEL_ID, filename=SAM3_CKPT_NAME)
+    return checkpoint_path
+
+
+def build_sam3_video_model(
+    checkpoint_path: Optional[str] = None,
+    load_from_HF=True,
+    bpe_path: Optional[str] = None,
+    has_presence_token: bool = True,
+    geo_encoder_use_img_cross_attn: bool = True,
+    strict_state_dict_loading: bool = True,
+    apply_temporal_disambiguation: bool = True,
+    device="cuda" if torch.cuda.is_available() else "cpu",
+    compile=False,
+) -> Sam3VideoInferenceWithInstanceInteractivity:
+    """
+    Build SAM3 dense tracking model.
+
+    Args:
+        checkpoint_path: Optional path to checkpoint file
+        bpe_path: Path to the BPE tokenizer file
+
+    Returns:
+        Sam3VideoInferenceWithInstanceInteractivity: The instantiated dense tracking model
+    """
+    if bpe_path is None:
+        bpe_path = os.path.join(
+            os.path.dirname(__file__), "..", "assets", "bpe_simple_vocab_16e6.txt.gz"
+        )
+
+    # Build Tracker module
+    tracker = build_tracker(apply_temporal_disambiguation=apply_temporal_disambiguation)
+
+    # Build Detector components
+    visual_neck = _create_vision_backbone()
+    text_encoder = _create_text_encoder(bpe_path)
+    backbone = SAM3VLBackbone(scalp=1, visual=visual_neck, text=text_encoder)
+    transformer = _create_sam3_transformer(has_presence_token=has_presence_token)
+    segmentation_head: UniversalSegmentationHead = _create_segmentation_head()
+    input_geometry_encoder = _create_geometry_encoder()
+
+    # Create main dot product scoring
+    main_dot_prod_mlp = MLP(
+        input_dim=256,
+        hidden_dim=2048,
+        output_dim=256,
+        num_layers=2,
+        dropout=0.1,
+        residual=True,
+        out_norm=nn.LayerNorm(256),
+    )
+    main_dot_prod_scoring = DotProductScoring(
+        d_model=256, d_proj=256, prompt_mlp=main_dot_prod_mlp
+    )
+
+    # Build Detector module
+    detector = Sam3ImageOnVideoMultiGPU(
+        num_feature_levels=1,
+        backbone=backbone,
+        transformer=transformer,
+        segmentation_head=segmentation_head,
+        semantic_segmentation_head=None,
+        input_geometry_encoder=input_geometry_encoder,
+        use_early_fusion=True,
+        use_dot_prod_scoring=True,
+        dot_prod_scoring=main_dot_prod_scoring,
+        supervise_joint_box_scores=has_presence_token,
+    )
+
+    # Build the main SAM3 video model
+    if apply_temporal_disambiguation:
+        model = Sam3VideoInferenceWithInstanceInteractivity(
+            detector=detector,
+            tracker=tracker,
+            score_threshold_detection=0.5,
+            assoc_iou_thresh=0.1,
+            det_nms_thresh=0.1,
+            new_det_thresh=0.7,
+            hotstart_delay=15,
+            hotstart_unmatch_thresh=8,
+            hotstart_dup_thresh=8,
+            suppress_unmatched_only_within_hotstart=True,
+            min_trk_keep_alive=-1,
+            max_trk_keep_alive=30,
+            init_trk_keep_alive=30,
+            suppress_overlapping_based_on_recent_occlusion_threshold=0.7,
+            suppress_det_close_to_boundary=False,
+            fill_hole_area=16,
+            recondition_every_nth_frame=16,
+            masklet_confirmation_enable=False,
+            decrease_trk_keep_alive_for_empty_masklets=False,
+            image_size=1008,
+            image_mean=(0.5, 0.5, 0.5),
+            image_std=(0.5, 0.5, 0.5),
+            compile_model=compile,
+        )
+    else:
+        # a version without any heuristics for ablation studies
+        model = Sam3VideoInferenceWithInstanceInteractivity(
+            detector=detector,
+            tracker=tracker,
+            score_threshold_detection=0.5,
+            assoc_iou_thresh=0.1,
+            det_nms_thresh=0.1,
+            new_det_thresh=0.7,
+            hotstart_delay=0,
+            hotstart_unmatch_thresh=0,
+            hotstart_dup_thresh=0,
+            suppress_unmatched_only_within_hotstart=True,
+            min_trk_keep_alive=-1,
+            max_trk_keep_alive=30,
+            init_trk_keep_alive=30,
+            suppress_overlapping_based_on_recent_occlusion_threshold=0.7,
+            suppress_det_close_to_boundary=False,
+            fill_hole_area=16,
+            recondition_every_nth_frame=0,
+            masklet_confirmation_enable=False,
+            decrease_trk_keep_alive_for_empty_masklets=False,
+            image_size=1008,
+            image_mean=(0.5, 0.5, 0.5),
+            image_std=(0.5, 0.5, 0.5),
+            compile_model=compile,
+        )
+
+    # Load checkpoint if provided
+    if load_from_HF and checkpoint_path is None:
+        checkpoint_path = download_ckpt_from_hf()
+    if checkpoint_path is not None:
+        with g_pathmgr.open(checkpoint_path, "rb") as f:
+            ckpt = torch.load(f, map_location="cpu", weights_only=True)
+        if "model" in ckpt and isinstance(ckpt["model"], dict):
+            ckpt = ckpt["model"]
+
+        missing_keys, unexpected_keys = model.load_state_dict(
+            ckpt, strict=strict_state_dict_loading
+        )
+        if missing_keys:
+            print(f"Missing keys: {missing_keys}")
+        if unexpected_keys:
+            print(f"Unexpected keys: {unexpected_keys}")
+
+    model.to(device=device)
+    return model
+
+
+def build_sam3_video_predictor(*model_args, gpus_to_use=None, **model_kwargs):
+    return Sam3VideoPredictorMultiGPU(
+        *model_args, gpus_to_use=gpus_to_use, **model_kwargs
+    )

detect_tools/sam3/sam3/perflib/__init__.py

@@ -0,0 +1,8 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates. All Rights Reserved
+
+import os
+
+is_enabled = False
+if os.getenv("USE_PERFLIB", "1") == "1":
+    # print("Enabled the use of perflib.\n", end="")
+    is_enabled = True

detect_tools/sam3/sam3/perflib/associate_det_trk.py

@@ -0,0 +1,137 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates. All Rights Reserved
+
+from collections import defaultdict
+
+import torch
+import torch.nn.functional as F
+from sam3.perflib.masks_ops import mask_iou
+from scipy.optimize import linear_sum_assignment
+
+
+def associate_det_trk(
+    det_masks,
+    track_masks,
+    iou_threshold=0.5,
+    iou_threshold_trk=0.5,
+    det_scores=None,
+    new_det_thresh=0.0,
+):
+    """
+    Optimized implementation of detection <-> track association that minimizes DtoH syncs.
+
+    Args:
+        det_masks: (N, H, W) tensor of predicted masks
+        track_masks: (M, H, W) tensor of track masks
+
+    Returns:
+        new_det_indices: list of indices in det_masks considered 'new'
+        unmatched_trk_indices: list of indices in track_masks considered 'unmatched'
+    """
+    with torch.autograd.profiler.record_function("perflib: associate_det_trk"):
+        assert isinstance(det_masks, torch.Tensor), "det_masks should be a tensor"
+        assert isinstance(track_masks, torch.Tensor), "track_masks should be a tensor"
+        if det_masks.size(0) == 0 or track_masks.size(0) == 0:
+            return list(range(det_masks.size(0))), [], {}, {}  # all detections are new
+
+        if list(det_masks.shape[-2:]) != list(track_masks.shape[-2:]):
+            # resize to the smaller size to save GPU memory
+            if torch.numel(det_masks[-2:]) < torch.numel(track_masks[-2:]):
+                track_masks = (
+                    F.interpolate(
+                        track_masks.unsqueeze(1).float(),
+                        size=det_masks.shape[-2:],
+                        mode="bilinear",
+                        align_corners=False,
+                    ).squeeze(1)
+                    > 0
+                )
+            else:
+                # resize detections to track size
+                det_masks = (
+                    F.interpolate(
+                        det_masks.unsqueeze(1).float(),
+                        size=track_masks.shape[-2:],
+                        mode="bilinear",
+                        align_corners=False,
+                    ).squeeze(1)
+                    > 0
+                )
+
+        det_masks = det_masks > 0
+        track_masks = track_masks > 0
+
+        iou = mask_iou(det_masks, track_masks)  # (N, M)
+        igeit = iou >= iou_threshold
+        igeit_any_dim_1 = igeit.any(dim=1)
+        igeit_trk = iou >= iou_threshold_trk
+
+        iou_list = iou.cpu().numpy().tolist()
+        igeit_list = igeit.cpu().numpy().tolist()
+        igeit_any_dim_1_list = igeit_any_dim_1.cpu().numpy().tolist()
+        igeit_trk_list = igeit_trk.cpu().numpy().tolist()
+
+        det_scores_list = (
+            det_scores
+            if det_scores is None
+            else det_scores.cpu().float().numpy().tolist()
+        )
+
+        # Hungarian matching for tracks (one-to-one: each track matches at most one detection)
+        # For detections: allow many tracks to match to the same detection (many-to-one)
+
+        # If either is empty, return all detections as new
+        if det_masks.size(0) == 0 or track_masks.size(0) == 0:
+            return list(range(det_masks.size(0))), [], {}
+
+        # Hungarian matching: maximize IoU for tracks
+        cost_matrix = 1 - iou.cpu().numpy()  # Hungarian solves for minimum cost
+        row_ind, col_ind = linear_sum_assignment(cost_matrix)
+
+        def branchy_hungarian_better_uses_the_cpu(
+            cost_matrix, row_ind, col_ind, iou_list, det_masks, track_masks
+        ):
+            matched_trk = set()
+            matched_det = set()
+            matched_det_scores = {}  # track index -> [det_score, det_score * iou] det score of matched detection mask
+            for d, t in zip(row_ind, col_ind):
+                matched_det_scores[t] = [
+                    det_scores_list[d],
+                    det_scores_list[d] * iou_list[d][t],
+                ]
+                if igeit_trk_list[d][t]:
+                    matched_trk.add(t)
+                    matched_det.add(d)
+
+            # Tracks not matched by Hungarian assignment above threshold are unmatched
+            unmatched_trk_indices = [
+                t for t in range(track_masks.size(0)) if t not in matched_trk
+            ]
+
+            # For detections: allow many tracks to match to the same detection (many-to-one)
+            # So, a detection is 'new' if it does not match any track above threshold
+            assert track_masks.size(0) == igeit.size(
+                1
+            )  # Needed for loop optimizaiton below
+            new_det_indices = []
+            for d in range(det_masks.size(0)):
+                if not igeit_any_dim_1_list[d]:
+                    if det_scores is not None and det_scores[d] >= new_det_thresh:
+                        new_det_indices.append(d)
+
+            # for each detection, which tracks it matched to (above threshold)
+            det_to_matched_trk = defaultdict(list)
+            for d in range(det_masks.size(0)):
+                for t in range(track_masks.size(0)):
+                    if igeit_list[d][t]:
+                        det_to_matched_trk[d].append(t)
+
+            return (
+                new_det_indices,
+                unmatched_trk_indices,
+                det_to_matched_trk,
+                matched_det_scores,
+            )
+
+        return (branchy_hungarian_better_uses_the_cpu)(
+            cost_matrix, row_ind, col_ind, iou_list, det_masks, track_masks
+        )