inference.py

import torch
import numpy as np
from PIL import Image
from model import PretrainedUNet
from PIL import Image, ImageOps

def load_model(model_path, device):
    model = PretrainedUNet().to(device)
    checkpoint = torch.load(model_path, map_location=device)
    model.load_state_dict(checkpoint)
    model.eval()
    return model

def run_inference(model, img_a_tensor, img_b_tensor, device):
    img_a_tensor = img_a_tensor.unsqueeze(0).to(device)
    img_b_tensor = img_b_tensor.unsqueeze(0).to(device)
    
    with torch.no_grad():
        outputs = model(img_a_tensor, img_b_tensor)
        pred_mask = torch.sigmoid(outputs).cpu().squeeze().numpy()
    
    return pred_mask

def create_overlay(base_img, mask, color, threshold=0.5):
    """Create colored overlay on image"""
    base_np = np.array(base_img)
    
    if isinstance(mask, Image.Image):
        # Resize mask to match base image size
        mask = mask.resize((base_np.shape[1], base_np.shape[0]), Image.NEAREST)
        mask_np = np.array(mask.convert('L')) / 255.0
    else:
        # If numpy array, resize if needed
        if mask.shape[:2] != base_np.shape[:2]:
            mask_pil = Image.fromarray((mask * 255).astype(np.uint8))
            mask_pil = mask_pil.resize((base_np.shape[1], base_np.shape[0]), Image.NEAREST)
            mask_np = np.array(mask_pil) / 255.0
        else:
            mask_np = mask
    
    overlay = base_np.copy()
    mask_bool = mask_np > threshold
    overlay[mask_bool] = color
    
    alpha = 0.5
    result = (alpha * overlay + (1 - alpha) * base_np).astype(np.uint8)
    
    return Image.fromarray(result)