utils.py

import io
import numpy as np
from PIL import Image
from fastapi import HTTPException


LABELS = [
    "No DR",
    "Mild",
    "Moderate",
    "Severe",
    "Proliferative DR"
]


def preprocess_and_split(image_bytes: bytes):
    """
    Input: raw image bytes
    Output:
        imgs1 → for model_1
        imgs2 → for model_2
    """

    try:
        image = Image.open(io.BytesIO(image_bytes)).convert("RGB")
    except Exception:
        raise HTTPException(status_code=400, detail="Invalid image file")

    image = image.resize((224, 224))

    
    images = [
        image,  # original
        image.transpose(Image.FLIP_LEFT_RIGHT),  
    ]

    def to_array(img_list):
        return np.array([
            np.array(img, dtype=np.float32)  # matches training (0–255)
            for img in img_list
        ])

    imgs = to_array(images)

    return imgs, imgs


def predict_split(model_1, model_2, imgs1, imgs2, debug=False):
    """
    Runs inference and returns:
        label (str)
        confidence (float)
    """

    try:
        preds1 = model_1.predict(imgs1, verbose=0)
        preds2 = model_2.predict(imgs2, verbose=0)

        # Average per model
        avg1 = np.mean(preds1, axis=0)
        avg2 = np.mean(preds2, axis=0)

       
        final_pred = 0.50 * avg1 + 0.50 * avg2

        confidence = float(np.max(final_pred))
        predicted_index = int(np.argmax(final_pred))

        
        if confidence < 0.35:
            return "Uncertain", confidence

       
        if debug:
            top2 = np.argsort(final_pred)[-2:][::-1]
            print("Prediction vector:", final_pred)
            print("Top-1:", LABELS[top2[0]])
            print("Top-2:", LABELS[top2[1]])

        return LABELS[predicted_index], confidence

    except Exception as e:
        raise HTTPException(
            status_code=500,
            detail=f"Prediction failed: {str(e)}"
        )