app.py

import tensorflow as tf
import numpy as np
import cv2
import os
import pickle
from PIL import Image
import gradio as gr

# CRITICAL: Define the custom InstanceNormalization layer used in training
class InstanceNormalization(tf.keras.layers.Layer):
    def __init__(self, epsilon=1e-5, **kwargs):
        super(InstanceNormalization, self).__init__(**kwargs)
        self.epsilon = epsilon
    
    def build(self, input_shape):
        depth = input_shape[-1]
        self.scale = self.add_weight(
            shape=[depth],
            initializer=tf.random_normal_initializer(1., 0.02),
            trainable=True,
            name='scale'
        )
        self.offset = self.add_weight(
            shape=[depth],
            initializer='zeros',
            trainable=True,
            name='offset'
        )
        super().build(input_shape)
    
    def call(self, x):
        mean, variance = tf.nn.moments(x, axes=[1, 2], keepdims=True)
        inv = tf.math.rsqrt(variance + self.epsilon)
        normalized = (x - mean) * inv
        return self.scale * normalized + self.offset
    
    def get_config(self):
        config = super().get_config()
        config.update({'epsilon': self.epsilon})
        return config

# Set up TensorFlow for compatibility
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'
tf.keras.mixed_precision.set_global_policy('float32')

class MultiAttributeClassifier:
    def __init__(self):
        # Define categories for classification
        self.categories = ['content', 'style', 'time_of_day', 'weather']
        self.models = {}
        self.encoders = {}
        self.gan_models = {}
        
        # Define custom objects FIRST (before using them)
        self.custom_objects = {
            'InstanceNormalization': InstanceNormalization,
            'tf': tf
        }
        
        # Load models
        self.load_classification_models()
        self.load_gan_models()
    
    def load_classification_models(self):
        """Load all classification models and encoders"""
        print("Loading classification models...")
        print(f"📂 Looking for models in: models/classification/")
        
        # First, let's see what's actually in the classification folder
        classification_path = "models/classification"
        if os.path.exists(classification_path):
            print(f"📁 Found classification directory")
            files = os.listdir(classification_path)
            print(f"📄 Available files: {files}")
        else:
            print(f"❌ Classification directory not found: {classification_path}")
            return
        
        for category in self.categories:
            try:
                # Load model from correct path
                model_path = f"models/classification/{category}_model.h5"
                if os.path.exists(model_path):
                    print(f"🔍 Loading model: {model_path}")
                    try:
                        # Try normal loading first
                        self.models[category] = tf.keras.models.load_model(model_path)
                    except Exception as e1:
                        print(f"   ⚠️ Normal loading failed: {e1}")
                        try:
                            # Try with compile=False and custom objects
                            self.models[category] = tf.keras.models.load_model(
                                model_path, 
                                compile=False,
                                custom_objects=self.custom_objects
                            )
                            print(f"   ✅ Loaded {category} with custom_objects")
                        except Exception as e2:
                            print(f"   ❌ Failed to load {category}: {e2}")
                            continue
                    
                    print(f"✅ Loaded {category} model ({os.path.getsize(model_path)/1024/1024:.1f} MB)")
                    
                    # Load encoder
                    encoder_path = f"models/classification/{category}_encoder.pkl"
                    if os.path.exists(encoder_path):
                        with open(encoder_path, 'rb') as f:
                            encoder_data = pickle.load(f)
                        
                        # Handle different encoder formats
                        if hasattr(encoder_data, 'classes_'):
                            # Standard LabelEncoder
                            self.encoders[category] = encoder_data
                            print(f"✅ Loaded {category} encoder (LabelEncoder) - {len(encoder_data.classes_)} classes")
                        elif isinstance(encoder_data, dict):
                            # Dict format - create a wrapper
                            class EncoderWrapper:
                                def __init__(self, class_dict):
                                    if 'classes_' in class_dict:
                                        self.classes_ = class_dict['classes_']
                                    elif 'classes' in class_dict:
                                        self.classes_ = class_dict['classes']
                                    else:
                                        # Try to extract classes from dict keys/values
                                        self.classes_ = list(class_dict.keys()) if class_dict else ['unknown']
                            
                            self.encoders[category] = EncoderWrapper(encoder_data)
                            print(f"✅ Loaded {category} encoder (Dict format) - {len(self.encoders[category].classes_)} classes")
                            print(f"   Classes: {self.encoders[category].classes_}")
                        else:
                            print(f"⚠️ Unknown encoder format for {category}: {type(encoder_data)}")
                            print(f"   Content preview: {str(encoder_data)[:200]}...")
                    else:
                        print(f"⚠️ {category} encoder not found at {encoder_path}")
                else:
                    print(f"❌ {category} model not found at {model_path}")
            except Exception as e:
                print(f"❌ Failed to load {category}: {e}")
                import traceback
                traceback.print_exc()
        
        print(f"🎯 Successfully loaded {len(self.models)} classification models")
    
    def load_gan_models(self):
        """Load all GAN models for style transfer"""
        print("Loading GAN models...")
        
        # First, let's scan what's actually in the GAN folders
        gan_base_path = "models/gan"
        if os.path.exists(gan_base_path):
            print(f"📂 Found GAN models directory: {gan_base_path}")
            for folder in os.listdir(gan_base_path):
                folder_path = os.path.join(gan_base_path, folder)
                if os.path.isdir(folder_path):
                    print(f"📁 GAN folder: {folder}")
                    for file in os.listdir(folder_path):
                        print(f"   📄 {file}")
        
        # Try multiple possible file name patterns
        gan_paths = {
            # Day/Night models - try multiple naming patterns
            'day_to_night': [
                'models/gan/day_night/day_to_night_generator_final.keras',
                'models/gan/day_night/day_to_night_generator.keras',
                'models/gan/day_night/day_to_night.keras',
                'models/gan/day_night/generator_day_to_night.keras'
            ],
            'night_to_day': [
                'models/gan/day_night/night_to_day_generator_final.keras',
                'models/gan/day_night/night_to_day_generator.keras', 
                'models/gan/day_night/night_to_day.keras',
                'models/gan/day_night/generator_night_to_day.keras'
            ],
            
            # Foggy/Clear models
            'foggy_to_clear': [
                'models/gan/foggy/foggy_to_normal_generator_final.keras',
                'models/gan/foggy/foggy_to_clear_generator.keras',
                'models/gan/foggy/foggy_to_clear.keras'
            ],
            'clear_to_foggy': [
                'models/gan/foggy/normal_to_foggy_generator_final.keras',
                'models/gan/foggy/clear_to_foggy_generator.keras',
                'models/gan/foggy/clear_to_foggy.keras'
            ],
            
            # Japanese art models
            'photo_to_japanese': [
                'models/gan/japanese/photo_to_ukiyoe_generator.keras',
                'models/gan/japanese/photo_to_japanese_generator.keras',
                'models/gan/japanese/photo_to_japanese.keras'
            ],
            'japanese_to_photo': [
                'models/gan/japanese/ukiyoe_to_photo_generator.keras',
                'models/gan/japanese/japanese_to_photo_generator.keras',
                'models/gan/japanese/japanese_to_photo.keras'
            ],
            
            # Season models
            'summer_to_winter': [
                'models/gan/summer_winter/summer_to_winter_generator_final.keras',
                'models/gan/summer_winter/summer_to_winter_generator.keras',
                'models/gan/summer_winter/summer_to_winter.keras'
            ],
            'winter_to_summer': [
                'models/gan/summer_winter/winter_to_summer_generator_final.keras',
                'models/gan/summer_winter/winter_to_summer_generator.keras',
                'models/gan/summer_winter/winter_to_summer.keras'
            ]
        }
        
        for model_name, possible_paths in gan_paths.items():
            model_loaded = False
            for model_path in possible_paths:
                try:
                    if os.path.exists(model_path):
                        print(f"🔍 Trying to load: {model_path}")
                        # Try loading with different compatibility options
                        try:
                            # First try: Normal loading
                            self.gan_models[model_name] = tf.keras.models.load_model(model_path)
                        except Exception as e1:
                            print(f"   ⚠️ Normal loading failed: {e1}")
                            try:
                                # Second try: Load with compile=False (ignore training config)
                                self.gan_models[model_name] = tf.keras.models.load_model(model_path, compile=False)
                                print(f"   ✅ Loaded with compile=False")
                            except Exception as e2:
                                print(f"   ⚠️ compile=False failed: {e2}")
                                try:
                                    # Third try: Load with custom objects (CRITICAL for your models)
                                    self.gan_models[model_name] = tf.keras.models.load_model(
                                        model_path, 
                                        compile=False,
                                        custom_objects=self.custom_objects
                                    )
                                    print(f"   ✅ Loaded with custom_objects (InstanceNormalization)")
                                except Exception as e3:
                                    print(f"   ❌ All loading methods failed: {e3}")
                                    # Print the actual error for debugging
                                    print(f"   🔍 Error details: {str(e3)}")
                                    raise e3
                        
                        print(f"✅ Loaded GAN: {model_name} from {model_path}")
                        model_loaded = True
                        break
                except Exception as e:
                    print(f"❌ Failed to load {model_path}: {e}")
                    continue
            
            if not model_loaded:
                print(f"⚠️ Could not load GAN model: {model_name}")
                # Let's also scan the actual directory to see what files exist
                folder_map = {
                    'day_to_night': 'day_night',
                    'night_to_day': 'day_night', 
                    'foggy_to_clear': 'foggy',
                    'clear_to_foggy': 'foggy',
                    'photo_to_japanese': 'japanese',
                    'japanese_to_photo': 'japanese',
                    'summer_to_winter': 'summer_winter',
                    'winter_to_summer': 'summer_winter'
                }
                
                if model_name in folder_map:
                    folder_path = f"models/gan/{folder_map[model_name]}"
                    if os.path.exists(folder_path):
                        print(f"   📁 Available files in {folder_path}:")
                        for file in os.listdir(folder_path):
                            if file.endswith(('.keras', '.h5')):
                                print(f"      📄 {file}")
        
        print(f"🎯 Successfully loaded {len(self.gan_models)} GAN models")
    
    def preprocess_image(self, image):
        """Preprocess image for model input"""
        if isinstance(image, str):
            image = Image.open(image)
        elif isinstance(image, np.ndarray):
            image = Image.fromarray(image)
        
        # Resize image
        image = image.resize((224, 224))
        
        # Convert to array and normalize
        img_array = np.array(image)
        if img_array.shape[-1] == 4:  # RGBA
            img_array = img_array[:, :, :3]  # Remove alpha channel
        
        # Normalize to [0, 1]
        img_array = img_array.astype(np.float32) / 255.0
        
        # Add batch dimension
        img_array = np.expand_dims(img_array, axis=0)
        
        return img_array
    
    def predict_attributes(self, image):
        """Predict multiple attributes of an image"""
        preprocessed = self.preprocess_image(image)
        predictions = {}
        
        for category in self.categories:
            if category in self.models and category in self.encoders:
                try:
                    # Get model prediction
                    pred = self.models[category].predict(preprocessed, verbose=0)
                    
                    # Get predicted class
                    predicted_class_idx = np.argmax(pred, axis=1)[0]
                    confidence = float(np.max(pred))
                    
                    # Get class name from encoder - handle different formats
                    try:
                        if hasattr(self.encoders[category], 'classes_'):
                            classes = self.encoders[category].classes_
                            if predicted_class_idx < len(classes):
                                class_name = classes[predicted_class_idx]
                            else:
                                class_name = f"class_{predicted_class_idx}"
                        else:
                            class_name = f"class_{predicted_class_idx}"
                    except Exception as e:
                        print(f"Error getting class name for {category}: {e}")
                        class_name = f"class_{predicted_class_idx}"
                    
                    predictions[category] = {
                        'class': class_name,
                        'confidence': confidence
                    }
                except Exception as e:
                    print(f"Error predicting {category}: {e}")
                    predictions[category] = {
                        'class': 'unknown',
                        'confidence': 0.0
                    }
            else:
                # Fallback predictions if models not loaded
                fallback_predictions = {
                    'content': {'class': 'outdoor', 'confidence': 0.6},
                    'style': {'class': 'realistic', 'confidence': 0.7},
                    'time_of_day': {'class': 'day', 'confidence': 0.8},
                    'weather': {'class': 'clear', 'confidence': 0.8}
                }
                predictions[category] = fallback_predictions.get(category, {'class': 'unknown', 'confidence': 0.0})
        
        return predictions
    
    def apply_style_transfer(self, image, transformation):
        """Apply style transfer using trained GAN models"""
        if transformation not in self.gan_models:
            return None, f"GAN model '{transformation}' not available"
        
        try:
            # Preprocess image for GAN (256x256, normalized to [-1, 1])
            if isinstance(image, str):
                img = Image.open(image)
            elif isinstance(image, np.ndarray):
                img = Image.fromarray(image)
            else:
                img = image
            
            # Resize to 256x256 for GAN
            img = img.resize((256, 256))
            img_array = np.array(img)
            
            if img_array.shape[-1] == 4:  # RGBA
                img_array = img_array[:, :, :3]  # Remove alpha channel
            
            # Normalize to [-1, 1] for GAN
            img_array = (img_array.astype(np.float32) / 127.5) - 1.0
            img_array = np.expand_dims(img_array, axis=0)
            
            # Apply transformation
            model = self.gan_models[transformation]
            generated = model.predict(img_array, verbose=0)
            
            # Denormalize and convert back to image
            generated = (generated[0] + 1.0) * 127.5
            generated = np.clip(generated, 0, 255).astype(np.uint8)
            
            return generated, "Transformation completed!"
            
        except Exception as e:
            print(f"Error in style transfer: {e}")
            return None, f"Error: {str(e)}"
    
    def get_style_recommendations(self, predictions):
        """Get style transfer recommendations based on predictions"""
        recommendations = []
        
        # Time-based recommendations
        if 'time_of_day' in predictions:
            time_pred = predictions['time_of_day']
            if time_pred['class'] == 'day' and time_pred['confidence'] > 0.7:
                recommendations.append({
                    'transformation': 'day_to_night',
                    'confidence': time_pred['confidence'],
                    'description': f"Transform scene to night with {time_pred['confidence']*100:.0f}% confidence"
                })
            elif time_pred['class'] == 'night' and time_pred['confidence'] > 0.7:
                recommendations.append({
                    'transformation': 'night_to_day',
                    'confidence': time_pred['confidence'],
                    'description': f"Transform scene to day with {time_pred['confidence']*100:.0f}% confidence"
                })
        
        # Weather-based recommendations
        if 'weather' in predictions:
            weather_pred = predictions['weather']
            if weather_pred['class'] == 'clear' and weather_pred['confidence'] > 0.6:
                recommendations.append({
                    'transformation': 'clear_to_foggy',
                    'confidence': weather_pred['confidence'],
                    'description': f"Add fog atmosphere with {weather_pred['confidence']*100:.0f}% confidence"
                })
            elif weather_pred['class'] == 'foggy' and weather_pred['confidence'] > 0.6:
                recommendations.append({
                    'transformation': 'foggy_to_clear',
                    'confidence': weather_pred['confidence'],
                    'description': f"Clear fog from scene with {weather_pred['confidence']*100:.0f}% confidence"
                })
        
        # Content-based recommendations
        if 'content' in predictions:
            content_pred = predictions['content']
            if content_pred['class'] in ['outdoor', 'landscape'] and content_pred['confidence'] > 0.6:
                recommendations.extend([
                    {
                        'transformation': 'summer_to_winter',
                        'confidence': 0.8,
                        'description': f"Transform scene to winter with snow and cold atmosphere"
                    },
                    {
                        'transformation': 'winter_to_summer',
                        'confidence': 0.8,
                        'description': f"Transform scene to summer with warm, lush atmosphere"
                    }
                ])
        
        # Style-based recommendations
        if 'style' in predictions:
            style_pred = predictions['style']
            if style_pred['class'] == 'realistic' and style_pred['confidence'] > 0.6:
                recommendations.append({
                    'transformation': 'photo_to_japanese',
                    'confidence': style_pred['confidence'],
                    'description': f"Transform to Japanese ukiyo-e art style"
                })
        
        return recommendations

# Initialize classifier globally
print("🚀 Starting StyleTransfer App...")
classifier = MultiAttributeClassifier()
print(f"🎯 Initialization complete!")
print(f"   📊 Classification models loaded: {len(classifier.models)}")
print(f"   🎨 GAN models loaded: {len(classifier.gan_models)}")
if len(classifier.models) > 0:
    print(f"   ✅ Available categories: {list(classifier.models.keys())}")
if len(classifier.gan_models) > 0:
    print(f"   ✅ Available transformations: {list(classifier.gan_models.keys())}")
print("="*50)

def analyze_image(image):
    """Analyze uploaded image and provide style recommendations"""
    if image is None:
        choices_with_labels = [(transformation_labels[t], t) for t in available_transformations]
        return "Please upload an image first.", gr.update(choices=choices_with_labels, value=None, visible=True), []
    
    try:
        # Get predictions for all attributes
        predictions = classifier.predict_attributes(image)
        
        # Format analysis results
        analysis_text = "## 🔍 Image Analysis Results\n\n"
        for category, pred in predictions.items():
            confidence_pct = pred['confidence'] * 100
            analysis_text += f"**{category.replace('_', ' ').title()}:** {pred['class']} ({confidence_pct:.1f}% confidence)\n\n"
        
        # Get style recommendations
        recommendations = classifier.get_style_recommendations(predictions)
        
        # Format recommendations for display  
        if recommendations:
            analysis_text += "## 🎨 AI Suggestions\n\n"
            for rec in recommendations:
                analysis_text += f"**{rec['transformation'].replace('_', ' → ').title()}** ({rec['confidence']*100:.0f}%) {rec['description']}\n\n"
        else:
            analysis_text += "## 🎨 AI Suggestions\n\nNo specific recommendations - but feel free to try any transformation!\n\n"
        
        analysis_text += "---\n**Choose any transformation(s) below - you're not limited to the suggestions!**"
        
        # Always return ALL available transformations, regardless of analysis
        choices_with_labels = [(transformation_labels[t], t) for t in available_transformations]
        return analysis_text, gr.update(choices=choices_with_labels, value=None, visible=True), []
        
    except Exception as e:
        print(f"Error in analysis: {e}")
        import traceback
        traceback.print_exc()
        # Even if analysis fails, still show all transformations
        choices_with_labels = [(transformation_labels[t], t) for t in available_transformations]
        return f"Error analyzing image: {str(e)}\n\n**All transformations still available below:**", gr.update(choices=choices_with_labels, value=None, visible=True), []

def apply_transformations(image, selected_transformations):
    """Apply selected style transformations"""
    if image is None:
        return "Please upload an image first.", []
    
    if not selected_transformations:
        return "Please select at least one transformation.", []
    
    results = []
    status_messages = []
    
    for transformation in selected_transformations:
        try:
            transformed_img, message = classifier.apply_style_transfer(image, transformation)
            if transformed_img is not None:
                results.append(transformed_img)
                status_messages.append(f"✅ {transformation.replace('_', ' → ').title()}: {message}")
            else:
                status_messages.append(f"❌ {transformation.replace('_', ' → ').title()}: {message}")
        except Exception as e:
            status_messages.append(f"❌ {transformation}: Error - {str(e)}")
    
    status_text = "\n".join(status_messages)
    return status_text, results

# Available transformations for manual selection - show user-friendly names
available_transformations = [
    "day_to_night", "night_to_day", 
    "clear_to_foggy", "foggy_to_clear",
    "photo_to_japanese", "japanese_to_photo",
    "summer_to_winter", "winter_to_summer"
]

# User-friendly transformation names
transformation_labels = {
    "day_to_night": "🌅→🌙 Day to Night",
    "night_to_day": "🌙→🌅 Night to Day", 
    "clear_to_foggy": "☀️→🌫️ Clear to Foggy",
    "foggy_to_clear": "🌫️→☀️ Foggy to Clear",
    "photo_to_japanese": "📷→🎨 Photo to Japanese Art",
    "japanese_to_photo": "🎨→📷 Japanese Art to Photo",
    "summer_to_winter": "🌿→❄️ Summer to Winter",
    "winter_to_summer": "❄️→🌿 Winter to Summer"
}

# Create Gradio interface
with gr.Blocks(title="Intelligent Multi-Attribute Style Transfer", theme=gr.themes.Soft()) as demo:
    gr.Markdown("# 🎨 Intelligent Multi-Attribute Style Transfer")
    gr.Markdown("Upload an image and our AI will analyze it to provide smart suggestions - **but you can choose ANY transformation you want!**")
    gr.Markdown("💡 **Tip:** You can skip analysis and apply transformations directly!")
    
    # Show available transformations
    gr.Markdown("## Available Transformations:")
    gr.Markdown("• 🌅 Day ↔ Night conversion (CycleGAN)")
    gr.Markdown("• 🎨 Photo ↔ Japanese ukiyo-e art style (CycleGAN)")
    gr.Markdown("• 🌫️ Foggy ↔ Clear weather transformation (CycleGAN)")
    gr.Markdown("• 🌿 Summer ↔ Winter seasonal atmosphere (CycleGAN)")
    gr.Markdown("---")
    
    with gr.Row():
        with gr.Column(scale=1):
            image_input = gr.Image(label="📤 Upload Your Image", type="pil")
            analyze_btn = gr.Button("🔍 Analyze Image (Optional)", variant="primary")
            
        with gr.Column(scale=1):
            analysis_output = gr.Markdown("## 📊 Image Analysis Results", label="Analysis Results")
            recommendations = gr.CheckboxGroup(
                choices=[(transformation_labels[t], t) for t in available_transformations], 
                label="🎨 Choose Transformations (All Available)",
                visible=True,
                value=None
            )
    
    with gr.Row():
        with gr.Column():
            apply_btn = gr.Button("🎯 Apply Selected Transfers", variant="secondary")
            
    with gr.Row():
        status_output = gr.Textbox(label="📋 Applied Transformations", interactive=False)
        
    with gr.Row():
        results_gallery = gr.Gallery(
            label="🖼️ Transformed Images", 
            show_label=True, 
            elem_id="gallery",
            columns=2,
            rows=2,
            height="auto"
        )
    
    # Event handlers
    analyze_btn.click(
        fn=analyze_image,
        inputs=[image_input],
        outputs=[analysis_output, recommendations, results_gallery]
    )
    
    apply_btn.click(
        fn=apply_transformations,
        inputs=[image_input, recommendations],
        outputs=[status_output, results_gallery]
    )

if __name__ == "__main__":
    demo.launch(server_name="0.0.0.0", server_port=7860)