app.py

import gradio as gr
import torch
from unet import EnhancedUNet
import numpy as np
from PIL import Image
import io
import math

def initialize_model(model_path):
    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    model = EnhancedUNet(n_channels=1, n_classes=4).to(device)
    model.load_state_dict(torch.load(model_path, map_location=device))
    model.eval()
    return model, device

def extract_patches(image, patch_size=256):
    """Extract patches from the input image."""
    width, height = image.size
    patches = []
    positions = []
    
    # Calculate number of patches in each dimension
    n_cols = math.ceil(width / patch_size)
    n_rows = math.ceil(height / patch_size)
    
    # Pad image if necessary
    padded_width = n_cols * patch_size
    padded_height = n_rows * patch_size
    padded_image = Image.new('L', (padded_width, padded_height), 0)
    padded_image.paste(image, (0, 0))
    
    # Extract patches
    for i in range(n_rows):
        for j in range(n_cols):
            left = j * patch_size
            top = i * patch_size
            right = left + patch_size
            bottom = top + patch_size
            
            patch = padded_image.crop((left, top, right, bottom))
            patches.append(patch)
            positions.append((left, top, right, bottom))
    
    return patches, positions, (padded_width, padded_height), (width, height)

def stitch_patches(patches, positions, padded_size, original_size, n_classes=4):
    """Stitch patches back together into a complete mask."""
    full_mask = np.zeros((padded_size[1], padded_size[0]), dtype=np.uint8)
    
    for patch, (left, top, right, bottom) in zip(patches, positions):
        full_mask[top:bottom, left:right] = patch
        
    # Crop back to original size
    full_mask = full_mask[:original_size[1], :original_size[0]]
    return full_mask

def process_patch(patch, device):
    # Convert to grayscale if it's not already
    patch_gray = patch.convert("L")
    # Convert to numpy array and normalize
    patch_np = np.array(patch_gray, dtype=np.float32) / 255.0
    # Add batch and channel dimensions
    patch_tensor = torch.from_numpy(patch_np).float().unsqueeze(0).unsqueeze(0)
    return patch_tensor.to(device)

def create_overlay(original_image, mask, alpha=0.5):
    colors = [(0, 0, 0), (255, 0, 0), (0, 255, 0), (0, 0, 255)]  # Define colors for each class
    mask_rgb = np.zeros((*mask.shape, 3), dtype=np.uint8)
    for i, color in enumerate(colors):
        mask_rgb[mask == i] = color

    # Resize original image to match mask size
    original_image = original_image.resize((mask.shape[1], mask.shape[0]))
    original_array = np.array(original_image.convert("RGB"))

    # Create overlay
    overlay = (alpha * mask_rgb + (1 - alpha) * original_array).astype(np.uint8)
    return Image.fromarray(overlay)

def predict(input_image, model_choice):
    if input_image is None:
        return None, None
        
    model = models[model_choice]
    patch_size = 256
    
    # Extract patches
    patches, positions, padded_size, original_size = extract_patches(input_image, patch_size)
    
    # Process each patch
    predicted_patches = []
    for patch in patches:
        # Process patch
        patch_tensor = process_patch(patch, device)
        
        # Perform inference
        with torch.no_grad():
            output = model(patch_tensor)
        
        # Get prediction mask for patch
        patch_mask = torch.argmax(output, dim=1).cpu().numpy()[0]
        predicted_patches.append(patch_mask)
    
    # Stitch patches back together
    full_mask = stitch_patches(predicted_patches, positions, padded_size, original_size)
    
    # Create mask image
    mask_image = Image.fromarray((full_mask * 63).astype(np.uint8))  # Scale for better visibility
    
    # Create overlay image
    overlay_image = create_overlay(input_image, full_mask)
    
    return mask_image, overlay_image

# Initialize model (do this outside the inference function for better performance)
w_noise_model_path = "./models/best_model_w_noise.pth"
wo_noise_model_path = "./models/best_model_wo_noise.pth"
w_noise_model_v2_path = "./models/best_model_w_noise_v2.pth"

w_noise_model, device = initialize_model(w_noise_model_path)
wo_noise_model, device = initialize_model(wo_noise_model_path)
w_noise_model_v2, device = initialize_model(w_noise_model_v2_path)

models = {
    "Without Noise": wo_noise_model,
    "With Noise": w_noise_model,
    "With Noise V2": w_noise_model_v2
}

# Create Gradio interface
iface = gr.Interface(
    fn=predict,
    inputs=[
        gr.Image(type="pil"),
        gr.Dropdown(choices=["Without Noise", "With Noise", "With Noise V2"], value="With Noise V2"),
    ],
    outputs=[
        gr.Image(type="pil", label="Segmentation Mask"),
        gr.Image(type="pil", label="Overlay"),
    ],
    title="MoS2 Image Segmentation",
    description="Upload an image to get the segmentation mask and overlay visualization.",
    examples=[["./examples/image_000003.png", "With Noise"], ["./examples/image_000005.png", "Without Noise"]],
)

# Launch the interface
iface.launch(share=True)