modal_inference.py

import modal
import os

app = modal.App("surgisight")

image = (
    modal.Image.debian_slim(python_version="3.11")
    .apt_install("libgl1", "libglib2.0-0")
    .pip_install(
        "ultralytics",
        "pillow",
        "numpy",
        "opencv-python-headless",
        "huggingface_hub",
    )
)

# Cache the model weights inside the Modal image so it doesn't re-download every call
with image.imports():
    from ultralytics import YOLO
    from PIL import Image as PILImage
    import numpy as np
    import cv2
    import io


@app.cls(gpu="T4", image=image, secrets=[modal.Secret.from_name("hf-secret")])
class SurgiSightDetector:

    @modal.enter()
    def load_model(self):
        from huggingface_hub import hf_hub_download
        model_path = hf_hub_download(
            repo_id="sugan04/cholec-yolo26n-seg",
            filename="best.pt",
            token=os.environ.get("HF_TOKEN")
        )
        self.model = YOLO(model_path)

    @modal.method()
    def run(self, image_bytes: bytes, conf_threshold: float = 0.25):
        nparr = np.frombuffer(image_bytes, np.uint8)
        frame = cv2.imdecode(nparr, cv2.IMREAD_COLOR)

        results = self.model(frame, task="segment", conf=conf_threshold)
        annotated = results[0].plot()

        # Encode annotated image back to bytes
        _, buffer = cv2.imencode(".png", annotated)
        annotated_bytes = buffer.tobytes()

        # Extract detections
        boxes = results[0].boxes
        detections = []
        if boxes is not None and len(boxes) > 0:
            for box in boxes:
                detections.append({
                    "cls_id": int(box.cls[0]),
                    "conf": float(box.conf[0])
                })

        return {"annotated_bytes": annotated_bytes, "detections": detections}


# For local testing
@app.local_entrypoint()
def main():
    from PIL import Image as PILImage
    import io
    detector = SurgiSightDetector()
    print("Modal SurgiSight detector ready.")