handler.py

import torch
from PIL import Image
import io
import base64
import torchvision.transforms as T

from model import MedSAM2Model

class EndpointHandler:
    def __init__(self, path=""):
        self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
        self.model = MedSAM2Model().to(self.device)
        self.model.eval()


    def preprocess(self, inputs):
        # Unwrap if "inputs" key exists
        if "inputs" in inputs:
            inputs = inputs["inputs"]
        
        image_b64 = inputs.get("image")
        if not image_b64:
            raise ValueError("Missing 'image' field in input.")

        image_bytes = base64.b64decode(image_b64)
        image = Image.open(io.BytesIO(image_bytes)).convert("RGB")

        # Transform PIL image to tensor and normalize (example)
        transform = T.Compose([
            T.ToTensor(),  # Converts to tensor and scales pixels to [0,1]
            # Add normalization if your model requires it, e.g.:
            # T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
        ])

        image_tensor = transform(image).unsqueeze(0)  # Add batch dim: [1, 3, H, W]

        return image_tensor.to(self.device)


    def postprocess(self, output):
        return {"output": output.cpu().tolist()}

    def __call__(self, inputs):
        x = self.preprocess(inputs)
        with torch.no_grad():
            output = self.model(x)
        return self.postprocess(output)