app.py

import gradio as gr
import torch
import torch.nn as nn
import segmentation_models_pytorch as smp
from PIL import Image, ImageOps
import numpy as np
import albumentations as A
from albumentations.pytorch import ToTensorV2
import os

class ImprovedSkySegmentationModel(nn.Module):
    def __init__(self, encoder_name='resnet50', classes=1):
        super().__init__()
        self.model = smp.Unet(
            encoder_name=encoder_name,
            encoder_weights=None,  # Don't load pretrained weights
            classes=classes,
            activation=None,
        )

    def forward(self, x):
        output = self.model(x)
        return torch.sigmoid(output)

# Global model variable
model = None
config = None
device = None

def load_model_once():
    """Load the model once when the app starts"""
    global model, config, device
    
    if model is None:
        device = 'cuda' if torch.cuda.is_available() else 'cpu'
        
        # For Hugging Face Spaces, the model file should be in the same directory
        model_path = "sky_segmentation_model.pt"  # You'll upload this file
        
        if not os.path.exists(model_path):
            # Fallback for testing - you can remove this in production
            raise FileNotFoundError(f"Model file {model_path} not found. Please upload your trained model.")
        
        checkpoint = torch.load(model_path, map_location=device)
        
        config = checkpoint['config']
        model = ImprovedSkySegmentationModel(
            encoder_name=config['encoder_name'],
            classes=config['classes']
        )
        
        model.load_state_dict(checkpoint['model_state_dict'])
        model.eval()
        model.to(device)
        
        print(f"Model loaded successfully on {device}")

def preprocess_image(image, img_size=512):
    """Preprocess image for inference with EXIF orientation correction"""
    # Handle different input types
    if isinstance(image, str):
        # If image is a file path
        image = Image.open(image).convert('RGB')
    elif hasattr(image, 'convert'):
        # If image is already PIL Image
        image = image.convert('RGB')
    else:
        # Convert numpy array to PIL Image if necessary
        if isinstance(image, np.ndarray):
            image = Image.fromarray(image).convert('RGB')
    
    # Automatically correct orientation based on EXIF data
    image = ImageOps.exif_transpose(image)
    
    # Store original for display
    original_image = image.copy()
    
    transform = A.Compose([
        A.Resize(img_size, img_size),
        A.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
        ToTensorV2(),
    ])
    
    transformed = transform(image=np.array(image))
    return transformed['image'].unsqueeze(0), original_image

def predict_sky_mask(image_tensor):
    """Predict sky mask for an image"""
    global model, device
    
    with torch.no_grad():
        image_tensor = image_tensor.to(device)
        prediction = model(image_tensor)
        
        if prediction.dim() == 4 and prediction.size(1) == 1:
            prediction = prediction.squeeze(1)
            
        return prediction.cpu().squeeze(0).numpy()

def create_overlay(original_image, mask, alpha=0.4):
    """Create overlay of mask on original image"""
    if isinstance(original_image, Image.Image):
        original_image = np.array(original_image)
    
    # Resize mask to match original image size
    if mask.shape != original_image.shape[:2]:
        mask_resized = np.array(Image.fromarray((mask * 255).astype(np.uint8)).resize(
            (original_image.shape[1], original_image.shape[0]), Image.LANCZOS)) / 255.0
    else:
        mask_resized = mask
    
    # Create colored overlay (blue for sky areas)
    overlay = original_image.copy().astype(float)
    colored_mask = np.zeros_like(original_image, dtype=float)
    colored_mask[:, :, 2] = mask_resized * 255  # Blue channel for sky
    
    # Blend original image with colored mask
    overlay = (1 - alpha) * overlay + alpha * colored_mask
    overlay = np.clip(overlay, 0, 255).astype(np.uint8)
    
    return overlay

def segment_sky(image):
    """Main function for Gradio interface"""
    try:
        # Ensure model is loaded
        if model is None:
            load_model_once()
        
        # Preprocess image
        image_tensor, original_image = preprocess_image(image, config['img_size'])
        
        # Predict mask
        predicted_mask = predict_sky_mask(image_tensor)
        
        # Convert mask to PIL Image for display (0-255 range)
        mask_display = Image.fromarray((predicted_mask * 255).astype(np.uint8))
        
        # Create overlay
        overlay = create_overlay(original_image, predicted_mask)
        overlay_display = Image.fromarray(overlay)
        
        return original_image, mask_display, overlay_display
        
    except Exception as e:
        error_img = Image.new('RGB', (512, 512), color='red')
        return error_img, error_img, error_img

# Load model when the app starts
try:
    load_model_once()
    model_status = "✅ Model loaded successfully!"
except Exception as e:
    model_status = f"❌ Error loading model: {str(e)}"

# Create Gradio interface
with gr.Blocks(title="Sky Segmentation App", theme=gr.themes.Soft()) as demo:
    gr.Markdown("""
    # 🌤️ Sky Segmentation App
    
    Upload an image and get an AI-powered sky segmentation mask! This model identifies sky regions in your photos.
    
    **How to use:**
    1. Upload an image (JPG, PNG, etc.)
    2. The model will automatically detect sky regions
    3. View the original image, binary mask, and colored overlay
    """)
    
    # Model status
    gr.Markdown(f"**Model Status:** {model_status}")
    
    with gr.Row():
        with gr.Column(scale=1):
            # Input
            input_image = gr.Image(
                label="📁 Upload Your Image",
                type="pil",
                height=400
            )
            
            segment_btn = gr.Button("🔍 Segment Sky", variant="primary", size="lg")
        
        with gr.Column(scale=2):
            with gr.Row():
                original_output = gr.Image(label="📷 Original Image", height=300)
                mask_output = gr.Image(label="🎭 Sky Mask", height=300)
                overlay_output = gr.Image(label="🔵 Sky Overlay", height=300)
    
    # Info section
    gr.Markdown("""
    ### 📊 Understanding the Results:
    - **Original Image**: Your uploaded image
    - **Sky Mask**: Binary mask where white = sky, black = not sky
    - **Sky Overlay**: Original image with sky regions highlighted in blue
    
    ### ℹ️ About the Model:
    This model uses a U-Net architecture with ResNet50 encoder, trained specifically for sky segmentation tasks.
    The model can handle various image orientations and lighting conditions.
    
    ### 🚀 Made with:
    - PyTorch & Segmentation Models PyTorch
    - Gradio for the interface  
    - Hugging Face Spaces for hosting
    """)
    
    # Event handlers
    segment_btn.click(
        fn=segment_sky,
        inputs=[input_image],
        outputs=[original_output, mask_output, overlay_output]
    )
    
    # Also trigger on image upload
    input_image.upload(
        fn=segment_sky,
        inputs=[input_image], 
        outputs=[original_output, mask_output, overlay_output]
    )

# Launch the app
if __name__ == "__main__":
    demo.launch(
        server_name="0.0.0.0",
        server_port=7860,
        share=False
    )