inference.py

import base64
import io

import numpy as np
from PIL import Image
import torch
import torchxrayvision as xrv

def init():
    """
    Called once at container startup.
    Loads the DenseNet model from torchxrayvision (using HF Hub weights)
    and sets up the crop transform.
    """
    global model, transform
    model_name = "densenet121-res224-chex"
    model = xrv.models.get_model(model_name, from_hf_hub=True)
    model.eval()
    # Center‐crop to a square patch around the lung
    transform = xrv.datasets.XRayCenterCrop(pad=32)

def predict(request):
    """
    Called on each inference request.
    Expects a JSON payload like {"image": "data:image/jpeg;base64,/9j/4AAQ..."}.
    Returns a dict with scores and labels.
    """
    # 1) Decode base64 Data URI
    data_uri = request.json.get("image", "")
    b64 = data_uri.split(",")[-1]
    img = Image.open(io.BytesIO(base64.b64decode(b64))).convert("RGB")

    # 2) To numpy array & normalize
    arr = np.array(img)
    arr = xrv.datasets.normalize(arr, 255)  # scale pixel values

    # 3) Center crop & to tensor
    arr = transform(arr)  # H×W → cropped H×W
    tensor = torch.tensor(arr).permute(2, 0, 1).float().unsqueeze(0)

    # 4) Inference
    with torch.no_grad():
        scores = model(tensor).tolist()

    # 5) Return scores + pathologies
    return {"scores": scores, "labels": model.pathologies}