app.py

import os
import torch
import gradio as gr
import numpy as np
from PIL import Image
import torchvision.transforms.functional as TF
import torch.nn.functional as F
from collections import OrderedDict
from skimage import img_as_ubyte
import spaces

from model.CMFNet import CMFNet

# Download model weights on startup
if not os.path.exists('experiments/pretrained_models/deblur_GoPro_CMFNet.pth'):
    os.makedirs('experiments/pretrained_models', exist_ok=True)
    os.system('wget https://github.com/FanChiMao/CMFNet/releases/download/v0.0/deblur_GoPro_CMFNet.pth -P experiments/pretrained_models')

# Global model variable
model = None
device = None

def load_model():
    """Load the CMFNet model"""
    global model, device
    
    device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
    model = CMFNet()
    model = model.to(device)
    model.eval()
    
    # Load checkpoint
    weights_path = 'experiments/pretrained_models/deblur_GoPro_CMFNet.pth'
    checkpoint = torch.load(weights_path, map_location=device)
    
    try:
        model.load_state_dict(checkpoint["state_dict"])
    except:
        state_dict = checkpoint["state_dict"]
        new_state_dict = OrderedDict()
        for k, v in state_dict.items():
            name = k[7:] if k.startswith('module.') else k
            new_state_dict[name] = v
        model.load_state_dict(new_state_dict)
    
    print("Model loaded successfully!")

# Load model on startup
load_model()

@spaces.GPU
def deblur_image(image: Image.Image) -> Image.Image:
    """
    Deblur an input image using CMFNet
    
    Args:
        image: PIL Image to deblur
        
    Returns:
        PIL Image of deblurred result
    """
    if model is None:
        raise gr.Error("Model not loaded properly")
    
    try:
        # Preprocess image
        input_tensor = TF.to_tensor(image).unsqueeze(0).to(device)
        
        # Pad image to be multiple of 8
        h, w = input_tensor.shape[2], input_tensor.shape[3]
        mul = 8
        H, W = ((h + mul) // mul) * mul, ((w + mul) // mul) * mul
        padh = H - h if h % mul != 0 else 0
        padw = W - w if w % mul != 0 else 0
        input_tensor = F.pad(input_tensor, (0, padw, 0, padh), 'reflect')
        
        # Run inference
        with torch.no_grad():
            output = model(input_tensor)
            
        # Post-process
        output = torch.clamp(output, 0, 1)
        output = output[:, :, :h, :w]  # Remove padding
        output = output.squeeze(0).permute(1, 2, 0).cpu().numpy()
        output = img_as_ubyte(output)
        
        # Convert back to PIL Image
        result_image = Image.fromarray(output)
        return result_image
        
    except Exception as e:
        raise gr.Error(f"Error during inference: {str(e)}")

# Gradio interface
title = "CMFNet Image Deblurring"
description = """
# Compound Multi-branch Feature Fusion for Image Deblurring

Upload a blurry image to get a deblurred version using CMFNet. The model works best on motion blur and defocus blur.

**Note**: Images will be resized to have a maximum dimension of 512px for faster processing.
"""

article = """
<p style='text-align: center'>
<a href='https://github.com/FanChiMao/CMFNet' target='_blank'>GitHub Repository</a>
</p>
"""

# Example images
examples = [
    "images/Blur1.png",
    "images/Blur2.png", 
    "images/Blur5.png"
]

# Create Gradio interface
demo = gr.Interface(
    fn=deblur_image,
    inputs=gr.Image(type="pil", label="Upload Blurry Image"),
    outputs=gr.Image(type="pil", label="Deblurred Image"),
    title=title,
    description=description,
    article=article,
    examples=examples,
    cache_examples=True,
    theme=gr.themes.Soft(),
    allow_flagging="never"
)

if __name__ == "__main__":
    demo.launch(debug=True)