training/main_model.py

#!/usr/bin/env python3
"""
Main file for training the CLIP model with color and hierarchy alignment.
This file centralizes all the logic for training the main model. It uses
pre-trained color and hierarchy models to guide the main model's learning
through contrastive and alignment loss functions. It handles data loading,
training with validation, and checkpoint saving.
"""

import os
# Set environment variable to disable tokenizers parallelism warnings
os.environ["TOKENIZERS_PARALLELISM"] = "false"

import pandas as pd
import numpy as np
import torch
import torch.nn.functional as F
from torch.utils.data import Dataset, DataLoader, random_split
from torchvision import transforms
from PIL import Image
import matplotlib.pyplot as plt
from transformers import CLIPProcessor, CLIPModel as CLIPModel_transformers
import warnings
from tqdm import tqdm
import config

# Suppress warnings
warnings.filterwarnings("ignore", category=FutureWarning)
warnings.filterwarnings("ignore", category=UserWarning)

# -------------------------------
# Loss Functions
# -------------------------------

def enhanced_contrastive_loss(text_features, image_features, attribute_features,
                            color_model, hierarchy_model, colors, hierarchies, temperature=0.07, alignment_weight=0.3,
                            reference_text_features=None, reference_image_features=None, reference_weight=0.1):
    """
    Enhanced contrastive loss with direct alignment between color/hierarchy models and main model.
    
    This loss combines the original triple contrastive loss with direct alignment losses
    that force the main model's color and hierarchy dimensions to align with the
    specialized color and hierarchy models.
    
    Args:
        text_features: Main model text embeddings [batch_size, embed_dim]
        image_features: Main model image embeddings [batch_size, embed_dim]
        attribute_features: Concatenated color + hierarchy features [batch_size, color_dim + hierarchy_dim]
        color_model: Pre-trained color model for extracting color embeddings
        hierarchy_model: Pre-trained hierarchy model for extracting hierarchy embeddings
        colors: List of color strings for this batch [batch_size]
        hierarchies: List of hierarchy strings for this batch [batch_size]
        temperature: Temperature scaling parameter for contrastive loss (default: 0.07)
        alignment_weight: Weight for the alignment loss component (default: 0.3)
        
    Returns:
        Tuple of (total_loss, metrics_dict) where metrics_dict contains detailed loss components
    """
    
    # Original triple contrastive loss
    text_features_norm = F.normalize(text_features, dim=-1)
    image_features_norm = F.normalize(image_features, dim=-1)
    attribute_features_norm = F.normalize(attribute_features, dim=-1)

    text_image_logits = (text_features_norm[:, config.color_emb_dim+config.hierarchy_emb_dim:] @ 
                        image_features_norm[:, config.color_emb_dim+config.hierarchy_emb_dim:].T) / temperature
    text_attr_logits = (text_features_norm[:, :config.color_emb_dim+config.hierarchy_emb_dim] @ 
                       attribute_features_norm.T) / temperature
    image_attr_logits = (attribute_features_norm @ 
                        image_features_norm[:,:config.color_emb_dim+config.hierarchy_emb_dim].T) / temperature

    # Weight distribution for original loss
    weight_text_image = 0.7
    weight_attr_based = 0.15
    
    original_logits = (weight_text_image * text_image_logits + 
                      weight_attr_based * text_attr_logits + 
                      weight_attr_based * image_attr_logits)
    
    labels = torch.arange(len(text_features)).to(text_features.device)
    original_loss = (F.cross_entropy(original_logits, labels) + 
                    F.cross_entropy(original_logits.T, labels)) / 2

    # Direct alignment loss between color model and main model first 16 dims
    with torch.no_grad():
        color_embeddings = color_model.get_text_embeddings(colors)  
        hierarchy_embeddings = hierarchy_model.get_text_embeddings(hierarchies)  
    
    # Extract color dimensions from main model embeddings
    main_color_text = text_features[:, :config.color_emb_dim]  
    main_color_image = image_features[:, :config.color_emb_dim]  
    
    # Extract hierarchy dimensions from main model embeddings
    main_hierarchy_text = text_features[:, config.color_emb_dim:config.color_emb_dim+config.hierarchy_emb_dim]  
    main_hierarchy_image = image_features[:, config.color_emb_dim:config.color_emb_dim+config.hierarchy_emb_dim]  
    
    # Normalize for better correlation
    color_embeddings_norm = F.normalize(color_embeddings, dim=-1)
    main_color_text_norm = F.normalize(main_color_text, dim=-1)
    main_color_image_norm = F.normalize(main_color_image, dim=-1)
    
    hierarchy_embeddings_norm = F.normalize(hierarchy_embeddings, dim=-1)
    main_hierarchy_text_norm = F.normalize(main_hierarchy_text, dim=-1)
    main_hierarchy_image_norm = F.normalize(main_hierarchy_image, dim=-1)
    
    # Color alignment loss (cosine-only: more natural for normalized embeddings)
    color_text_cosine_loss = 1 - F.cosine_similarity(main_color_text_norm, color_embeddings_norm).mean()
    color_image_cosine_loss = 1 - F.cosine_similarity(main_color_image_norm, color_embeddings_norm).mean()
    color_alignment_loss = (color_text_cosine_loss + color_image_cosine_loss) / 2

    # Hierarchy alignment loss (cosine-only)
    hierarchy_text_cosine_loss = 1 - F.cosine_similarity(main_hierarchy_text_norm, hierarchy_embeddings_norm).mean()
    hierarchy_image_cosine_loss = 1 - F.cosine_similarity(main_hierarchy_image_norm, hierarchy_embeddings_norm).mean()
    hierarchy_alignment_loss = (hierarchy_text_cosine_loss + hierarchy_image_cosine_loss) / 2
    
    # Combined alignment loss
    alignment_loss = (color_alignment_loss + hierarchy_alignment_loss) / 2
    
    # Reference loss to keep embeddings close to base CLIP (preserves zero-shot capability)
    reference_loss = 0.0
    if reference_text_features is not None:
        text_ref_loss = F.mse_loss(
            F.normalize(text_features, dim=-1),
            F.normalize(reference_text_features, dim=-1)
        )
        if reference_image_features is not None:
            image_ref_loss = F.mse_loss(
                F.normalize(image_features, dim=-1),
                F.normalize(reference_image_features, dim=-1)
            )
            reference_loss = (text_ref_loss + image_ref_loss) / 2
        else:
            reference_loss = text_ref_loss
    
    # Combine losses
    total_loss = (1 - alignment_weight) * original_loss + alignment_weight * alignment_loss
    if reference_text_features is not None:
        total_loss = total_loss + reference_weight * reference_loss
    
    return total_loss, {
        'original_loss': original_loss.item(),
        'alignment_loss': alignment_loss.item(),
        'reference_loss': reference_loss if isinstance(reference_loss, float) else reference_loss.item(),
        'color_text_cosine': color_text_cosine_loss.item(),
        'color_image_cosine': color_image_cosine_loss.item(),
        'hierarchy_text_cosine': hierarchy_text_cosine_loss.item(),
        'hierarchy_image_cosine': hierarchy_image_cosine_loss.item()
    }

# -------------------------------
# Training Functions
# -------------------------------


def train_one_epoch(model, train_loader, optimizer, feature_models, color_model, hierarchy_model,
                           device, clip_processor, temperature=0.07, alignment_weight=0.3,
                           reference_model=None, reference_weight=0.1):
    """
    Enhanced training with direct color and hierarchy alignment loss.
    
    This function trains the model using the enhanced contrastive loss that includes
    direct alignment between the main model's color/hierarchy dimensions and the
    specialized color/hierarchy models.
    
    Args:
        model: Main CLIP model to train
        train_loader: DataLoader for training data
        optimizer: Optimizer instance
        feature_models: Dictionary containing color and hierarchy models
        color_model: Pre-trained color model for alignment
        hierarchy_model: Pre-trained hierarchy model for alignment
        device: Device to train on
        clip_processor: CLIP processor for text preprocessing
        temperature: Temperature scaling parameter for contrastive loss (default: 0.07)
        alignment_weight: Weight for the alignment loss component (default: 0.3)
        
    Returns:
        Tuple of (average_loss, metrics_dict) where metrics_dict contains detailed loss components
    """
    model.train()
    total_loss = 0.0
    total_metrics = {
        'original_loss': 0.0,
        'alignment_loss': 0.0,
        'reference_loss': 0.0,
        'color_text_cosine': 0.0,
        'color_image_cosine': 0.0,
        'hierarchy_text_cosine': 0.0,
        'hierarchy_image_cosine': 0.0
    }
    num_batches = 0
    
    pbar = tqdm(train_loader, desc="Training Enhanced", leave=False)
    
    for batch_idx, (images, texts, colors, hierarchy) in enumerate(pbar):
        # Move data to device
        images = images.to(device)
        images = images.expand(-1, 3, -1, -1)
        
        # Process text inputs
        text_inputs = clip_processor(text=texts, padding=True, return_tensors="pt")
        text_inputs = {k: v.to(device) for k, v in text_inputs.items()}
        
        # Reference features to keep embeddings close to base CLIP
        reference_text_features = None
        reference_image_features = None
        if reference_model is not None:
            with torch.no_grad():
                reference_text_features = reference_model.get_text_features(**text_inputs)
                reference_image_features = reference_model.get_image_features(pixel_values=images)

        # Forward pass
        optimizer.zero_grad()
        outputs = model(**text_inputs, pixel_values=images)

        text_features = outputs.text_embeds
        image_features = outputs.image_embeds

        # Get feature embeddings
        if hasattr(feature_models[config.color_column], 'get_color_name_embeddings'):
            color_features = feature_models[config.color_column].get_color_name_embeddings(colors)
        else:
            color_features = feature_models[config.color_column].get_text_embeddings(colors)
        hierarchy_features = feature_models[config.hierarchy_column].get_text_embeddings(hierarchy)
        concat_features = torch.cat((color_features, hierarchy_features), dim=1)

        # Calculate enhanced loss with hierarchy alignment
        loss, metrics = enhanced_contrastive_loss(
            text_features, image_features, concat_features,
            color_model, hierarchy_model, colors, hierarchy, temperature, alignment_weight,
            reference_text_features=reference_text_features,
            reference_image_features=reference_image_features,
            reference_weight=reference_weight
        )
        
        # Backward pass
        loss.backward()
        
        # Gradient clipping to prevent exploding gradients
        torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm=1.0)
        
        optimizer.step()
        
        total_loss += loss.item()
        for key, value in metrics.items():
            total_metrics[key] += value
        num_batches += 1
        
        # Update progress bar
        pbar.set_postfix({
            'Loss': f'{loss.item():.4f}',
            'Align': f'{metrics["alignment_loss"]:.4f}',
            'ColCos': f'{metrics["color_text_cosine"]:.3f}',
            'HierCos': f'{metrics["hierarchy_text_cosine"]:.3f}'
        })
    
    avg_metrics = {key: value / num_batches for key, value in total_metrics.items()}
    return total_loss / num_batches, avg_metrics

def valid_one_epoch(model, val_loader, feature_models, device, clip_processor, temperature=0.07, alignment_weight=0.3,
                    reference_model=None, reference_weight=0.1):
    """
    Validate the model for one epoch using enhanced contrastive loss.
    
    Args:
        model: Main CLIP model to validate
        val_loader: DataLoader for validation data
        feature_models: Dictionary containing color and hierarchy models
        device: Device to validate on
        clip_processor: CLIP processor for text preprocessing
        temperature: Temperature scaling parameter for contrastive loss (default: 0.07)
        alignment_weight: Weight for the alignment loss component (default: 0.3)
        
    Returns:
        Average validation loss for the epoch
    """
    model.eval()
    total_loss = 0.0
    num_batches = 0
    
    # Extract color and hierarchy models
    color_model = feature_models[config.color_column]
    hierarchy_model = feature_models[config.hierarchy_column]
    
    # Create progress bar for validation
    pbar = tqdm(val_loader, desc="Validation", leave=False)
    
    with torch.no_grad():
        for batch_idx, (images, texts, colors, hierarchy) in enumerate(pbar):
            # Move data to device
            images = images.to(device)
            images = images.expand(-1, 3, -1, -1)  # Ensure 3 channels
            
            # Process text inputs
            text_inputs = clip_processor(text=texts, padding=True, return_tensors="pt")
            text_inputs = {k: v.to(device) for k, v in text_inputs.items()}
            
            # Reference features to keep embeddings close to base CLIP
            reference_text_features = None
            reference_image_features = None
            if reference_model is not None:
                reference_text_features = reference_model.get_text_features(**text_inputs)
                reference_image_features = reference_model.get_image_features(pixel_values=images)

            # Forward pass
            outputs = model(**text_inputs, pixel_values=images)

            text_features = outputs.text_embeds
            image_features = outputs.image_embeds

            # Get feature embeddings
            if hasattr(feature_models[config.color_column], 'get_color_name_embeddings'):
                color_features = feature_models[config.color_column].get_color_name_embeddings(colors)
            else:
                color_features = feature_models[config.color_column].get_text_embeddings(colors)
            hierarchy_features = feature_models[config.hierarchy_column].get_text_embeddings(hierarchy)
            concat_features = torch.cat((color_features, hierarchy_features), dim=1)

            # Calculate loss with all required arguments
            loss, metrics = enhanced_contrastive_loss(
                text_features, image_features, concat_features,
                color_model, hierarchy_model, colors, hierarchy,
                temperature, alignment_weight,
                reference_text_features=reference_text_features,
                reference_image_features=reference_image_features,
                reference_weight=reference_weight
            )
            
            total_loss += loss.item()
            num_batches += 1
            
            # Update progress bar
            pbar.set_postfix({
                'Loss': f'{loss.item():.4f}',
                'Avg Loss': f'{total_loss/num_batches:.4f}'
            })
    
    return total_loss / num_batches

# -------------------------------
# Dataset
# -------------------------------

class CustomDataset(Dataset):
    """
    Custom dataset for main model training.
    
    Handles loading images from local paths, extracting text descriptions,
    and applying appropriate transformations for training and validation.
    """
    
    def __init__(self, dataframe, use_local_images=True, image_size=224):
        """
        Initialize the custom dataset.
        
        Args:
            dataframe: DataFrame with columns for image paths, text descriptions, colors, and hierarchy labels
            use_local_images: Whether to use local images (default: True)
            image_size: Size of images after resizing (default: 224)
        """
        self.dataframe = dataframe
        self.use_local_images = use_local_images
        self.image_size = image_size
        
        # Transforms with augmentation for training (increased augmentation to reduce overfitting)
        self.transform = transforms.Compose([
            transforms.Resize((image_size, image_size)),
            transforms.RandomHorizontalFlip(p=0.5),
            transforms.RandomRotation(15),  # Increased for more variation
            transforms.ColorJitter(brightness=0.3, contrast=0.3, saturation=0.3, hue=0.15),  # Increased intensity
            transforms.RandomAffine(degrees=0, translate=(0.1, 0.1), scale=(0.9, 1.1)),  # Increased transform range
            transforms.RandomApply([transforms.GaussianBlur(kernel_size=3, sigma=(0.1, 2.0))], p=0.2),  # Add blur
            transforms.ToTensor(),
            transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
        ])
        
        # Transforms for validation (no augmentation)
        self.val_transform = transforms.Compose([
            transforms.Resize((image_size, image_size)),
            transforms.ToTensor(),
            transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
        ])
        
        self.training_mode = True
        
    def set_training_mode(self, training=True):
        """
        Switch between training and validation transforms.
        
        Args:
            training: If True, use training transforms with augmentation; if False, use validation transforms
        """
        self.training_mode = training

    def __len__(self):
        """Return the number of samples in the dataset."""
        return len(self.dataframe)

    def __getitem__(self, idx):
        """
        Get a sample from the dataset.
        
        Args:
            idx: Index of the sample
            
        Returns:
            Tuple of (image_tensor, description_text, color_label, hierarchy_label)
        """
        row = self.dataframe.iloc[idx]
        
        image_data = row[config.column_local_image_path]
        image = Image.open(image_data).convert("RGB")
        
        # Apply appropriate transform
        if self.training_mode:
            image = self.transform(image)
        else:
            image = self.val_transform(image)

        # Get text and labels
        description = row[config.text_column]
        color = row[config.color_column]
        hierarchy = row[config.hierarchy_column]

        return image, description, color, hierarchy

# -------------------------------
# Model Loading
# -------------------------------

def load_models():
    """
    Load color and hierarchy models from checkpoints.

    Returns:
        Dictionary mapping model names to model instances:
        - 'color': ColorCLIP model instance
        - 'hierarchy': HierarchyModel instance
    """
    from training.color_model import ColorCLIP
    from training.hierarchy_model import HierarchyModel

    # --- Color model ---
    print("Loading ColorCLIP (CLIP-backbone) ...")
    color_model = ColorCLIP.from_checkpoint(config.color_model_path, device=config.device)
    color_model.eval()
    color_model.name = config.color_column

    # --- Hierarchy model ---
    print("Loading HierarchyModel (CLIP-backbone) ...")
    hierarchy_model = HierarchyModel.from_checkpoint(config.hierarchy_model_path, device=config.device)
    hierarchy_model.eval()
    hierarchy_model.name = config.hierarchy_column

    feature_models = {model.name: model for model in [color_model, hierarchy_model]}
    return feature_models

# -------------------------------
# Main Training Function
# -------------------------------

def train_model(model, train_loader, val_loader, feature_models, device, 
                      num_epochs=20, learning_rate=1e-5, temperature=0.07, 
                      save_path=config.main_model_path, alignment_weight=0.3, 
                      color_alignment_model=None, weight_decay=3e-4,
                      reference_model=None, reference_weight=0.1):
    """
    Custom training loop using train_one_epoch and valid_one_epoch functions.
    
    This function handles the complete training process including:
    - Training and validation loops
    - Learning rate scheduling
    - Early stopping
    - Model checkpointing
    - Training curve visualization
    
    Args:
        model: Main CLIP model to train
        train_loader: DataLoader for training data
        val_loader: DataLoader for validation data
        feature_models: Dictionary containing color and hierarchy models
        device: Device to train on
        num_epochs: Number of training epochs (default: 20)
        learning_rate: Learning rate for optimizer (default: 1e-5)
        temperature: Temperature scaling parameter for contrastive loss (default: 0.07)
        save_path: Path to save model checkpoints (default: main_model_path)
        alignment_weight: Weight for alignment loss component if using enhanced loss (default: 0.3)
        color_alignment_model: Optional color model for alignment (default: None, uses feature_models)
        weight_decay: L2 regularization weight (default: 3e-4, increased to reduce overfitting)
        
    Returns:
        Tuple of (training_losses, validation_losses) lists
    """
    model = model.to(device)
    # Use AdamW with weight decay for better regularization (reduces overfitting)
    optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate, weight_decay=weight_decay)
    scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer, mode='min', patience=3, factor=0.5)
    
    train_losses = []
    val_losses = []
    best_val_loss = float('inf')
    patience_counter = 0
    patience = 7  # Increased from 5 to 7 for better convergence
    
    print(f"Starting training for {num_epochs} epochs...")
    print(f"Learning rate: {learning_rate}")
    print(f"Temperature: {temperature}")
    print(f"Weight decay: {weight_decay}")
    print(f"Alignment weight: {alignment_weight}")
    print(f"Device: {device}")
    print(f"Training samples: {len(train_loader.dataset)}")
    print(f"Validation samples: {len(val_loader.dataset)}")
    print(f"Batch size: {train_loader.batch_size}")
    print(f"Estimated time per epoch: ~{len(train_loader) * 2 / 60:.1f} minutes")
    
    # Create processor once for efficiency
    processor = CLIPProcessor.from_pretrained('laion/CLIP-ViT-B-32-laion2B-s34B-b79K')
    
    # Freeze and move reference model (used for text-space regularization)
    if reference_model is not None:
        reference_model = reference_model.to(device)
        reference_model.eval()
        for param in reference_model.parameters():
            param.requires_grad = False
    
    # Create progress bar for epochs
    epoch_pbar = tqdm(range(num_epochs), desc="Training Progress", position=0)
    
    for epoch in epoch_pbar:
        # Update epoch progress bar
        epoch_pbar.set_description(f"Epoch {epoch+1}/{num_epochs}")
        
        # Training
        if color_alignment_model is None:
            color_alignment_model = feature_models[config.color_column]
        hierarchy_model = feature_models[config.hierarchy_column]
        train_loss, align_metrics = train_one_epoch(
            model, train_loader, optimizer, feature_models, color_alignment_model, hierarchy_model,
            device, processor, temperature, alignment_weight,
            reference_model=reference_model, reference_weight=reference_weight
        )
        train_losses.append(train_loss)
        
        # Validation
        val_loss = valid_one_epoch(
            model, val_loader, feature_models, device, processor, 
            temperature=temperature, alignment_weight=alignment_weight,
            reference_model=reference_model, reference_weight=reference_weight
        )
        val_losses.append(val_loss)
        
        # Learning rate scheduling
        scheduler.step(val_loss)
        
        # Calculate overfitting gap
        overfitting_gap = val_loss - train_loss
        
        # Update epoch progress bar with metrics
        postfix = {
            'Train Loss': f'{train_loss:.4f}',
            'Val Loss': f'{val_loss:.4f}',
            'Gap': f'{overfitting_gap:.4f}',
            'LR': f'{optimizer.param_groups[0]["lr"]:.2e}',
            'Best Val': f'{best_val_loss:.4f}'
        }
        if align_metrics is not None:
            postfix.update({
                'Align': f"{align_metrics['alignment_loss']:.3f}", 
                'ColCos': f"{align_metrics['color_text_cosine']:.3f}",
                'HierCos': f"{align_metrics['hierarchy_text_cosine']:.3f}"
            })
        epoch_pbar.set_postfix(postfix)
        
        # Warning if overfitting is detected
        if overfitting_gap > 0.15 and epoch > 3:
            print(f"\n⚠️  Warning: Significant overfitting detected at epoch {epoch+1} (gap={overfitting_gap:.4f})")
        
        # Save best model
        if val_loss < best_val_loss:
            best_val_loss = val_loss
            patience_counter = 0
            
            # Save checkpoint
            torch.save({
                'epoch': epoch,
                'model_state_dict': model.state_dict(),
                'optimizer_state_dict': optimizer.state_dict(),
                'train_loss': train_loss,
                'val_loss': val_loss,
                'best_val_loss': best_val_loss,
            }, save_path)
        else:
            patience_counter += 1
        
        # Early stopping
        if patience_counter >= patience:
            print(f"\n🛑 Early stopping triggered after {patience_counter} epochs without improvement")
            break
    
    # Plot training curves with overfitting analysis
    plt.figure(figsize=(15, 5))
    
    # Plot 1: Training and Validation Loss
    plt.subplot(1, 3, 1)
    plt.plot(train_losses, label='Train Loss', color='blue', linewidth=2)
    plt.plot(val_losses, label='Val Loss', color='red', linewidth=2)
    plt.title('Training and Validation Loss', fontsize=12, fontweight='bold')
    plt.xlabel('Epoch')
    plt.ylabel('Loss')
    plt.legend()
    plt.grid(True, alpha=0.3)
    
    # Plot 2: Overfitting Gap (Val Loss - Train Loss)
    plt.subplot(1, 3, 2)
    gap = [val_losses[i] - train_losses[i] for i in range(len(train_losses))]
    plt.plot(gap, label='Overfitting Gap', color='purple', linewidth=2)
    plt.axhline(y=0, color='black', linestyle='--', alpha=0.3)
    plt.axhline(y=0.1, color='red', linestyle='--', alpha=0.3, label='Warning threshold')
    plt.title('Overfitting Gap (Val - Train)', fontsize=12, fontweight='bold')
    plt.xlabel('Epoch')
    plt.ylabel('Gap')
    plt.legend()
    plt.grid(True, alpha=0.3)
    
    # Plot 3: Loss comparison
    plt.subplot(1, 3, 3)
    epochs = list(range(len(train_losses)))
    plt.plot(epochs, train_losses, 'o-', label='Train Loss', color='blue', linewidth=2)
    plt.plot(epochs, val_losses, 's-', label='Val Loss', color='red', linewidth=2)
    plt.fill_between(epochs, train_losses, val_losses, alpha=0.2, color='red')
    plt.title('Loss Comparison', fontsize=12, fontweight='bold')
    plt.xlabel('Epoch')
    plt.ylabel('Loss')
    plt.legend()
    plt.grid(True, alpha=0.3)
    
    plt.tight_layout()
    curves_path = str(config.ROOT_DIR / "figures" / "training_curves.png")
    plt.savefig(curves_path, dpi=300, bbox_inches='tight')
    plt.close()

    print(f"\nTraining completed!")
    print(f"Best validation loss: {best_val_loss:.4f}")
    print(f"Final model saved to: {save_path}")
    print(f"Training curves saved to: {curves_path}")
    
    return train_losses, val_losses

# -------------------------------
# Main Function
# -------------------------------

def main():
    print("="*80)
    print("🚀 Training of the model with alignment color and hierarchy")
    print("="*80)
    
    # Configuration (tuned for zero-shot + separation balance)
    num_epochs = 10
    learning_rate = 1.5e-5
    temperature = 0.09
    alignment_weight = 0.10    # reduced from 0.2: softer alignment preserves CLIP zero-shot
    reference_weight = 0.25    # increased from 0.1: stronger regularization toward base CLIP
    weight_decay = 1e-3        # increased from 5e-4: better generalization
    batch_size = 128
    subset_size = 100000

    # Load the data
    print(f"\n📂 Loading the data...")
    df = pd.read_csv(config.local_dataset_path)
    print(f"  Data downloaded: {len(df)} samples")

    # filter the rows with NaN values
    df_clean = df.dropna(subset=[config.column_local_image_path])
    df_clean = df_clean[df_clean[config.column_local_image_path].astype(str).str.len() > 0]
    print(f"  After filtering NaN: {len(df_clean)} samples")

    # Creation of datasets
    dataset = CustomDataset(df_clean)

    # Sample 100k for training
    subset_size = min(subset_size, len(dataset))
    train_size = int(0.8 * subset_size)
    val_size = subset_size - train_size

    np.random.seed(42)
    subset_indices = np.random.choice(len(dataset), subset_size, replace=False)
    subset_dataset = torch.utils.data.Subset(dataset, subset_indices)

    train_dataset, val_dataset = random_split(
        subset_dataset,
        [train_size, val_size],
        generator=torch.Generator().manual_seed(42)
    )
    
    # Creation of dataloaders
    train_loader = DataLoader(
        train_dataset, 
        batch_size=batch_size, 
        shuffle=True, 
        num_workers=2, 
        pin_memory=True if torch.cuda.is_available() else False
    )
    val_loader = DataLoader(
        val_dataset, 
        batch_size=batch_size, 
        shuffle=False, 
        num_workers=2, 
        pin_memory=True if torch.cuda.is_available() else False
    )
    
    print(f"  Train: {len(train_dataset)} samples")
    print(f"  Validation: {len(val_dataset)} samples")
    
    # Loading models
    print(f"\n🔧 Loading models...")
    feature_models = load_models()
    
    # Load or create the main model
    print(f"\n📦 Loading main model...")
    clip_model = CLIPModel_transformers.from_pretrained(
        'laion/CLIP-ViT-B-32-laion2B-s34B-b79K'
    )
    # Frozen reference CLIP to regularize text space (improves cross-domain generalization)
    reference_clip = CLIPModel_transformers.from_pretrained(
        'laion/CLIP-ViT-B-32-laion2B-s34B-b79K'
    )
    
    # Move the model on the device
    clip_model = clip_model.to(config.device)
    reference_clip = reference_clip.to(config.device)
    reference_clip.eval()
    for param in reference_clip.parameters():
        param.requires_grad = False
    
    # Training with enhanced loss
    print(f"\n🎯 Beginning training...")
    print(f"\n" + "="*80)
    
    train_losses, val_losses = train_model(
        model=clip_model,
        train_loader=train_loader,
        val_loader=val_loader,
        feature_models=feature_models,
        device=config.device,
        num_epochs=num_epochs,
        learning_rate=learning_rate,
        temperature=temperature,
        save_path=config.main_model_path,
        alignment_weight=alignment_weight,
        color_alignment_model=feature_models[config.color_column],
        weight_decay=weight_decay,
        reference_model=reference_clip,
        reference_weight=reference_weight
    )
    
    print("\n" + "="*80)
    print("✅ Training finished!")
    print(f"  Model saved: {config.main_model_path}")
    print(f"  Training curves: figures/training_curves.png")
    print("\n📊 Final results:")
    print(f"  Last train loss: {train_losses[-1]:.4f}")
    print(f"  Last validation loss: {val_losses[-1]:.4f}")
    print(f"  Best validation loss: {min(val_losses):.4f}")
    print(f"  Overfitting gap (val-train): {val_losses[-1] - train_losses[-1]:.4f}")
    if val_losses[-1] - train_losses[-1] > 0.1:
        print("  ⚠️  Warning: Significant overfitting detected!")
    elif val_losses[-1] - train_losses[-1] < 0.05:
        print("  ✅ Good generalization!")
    print("="*80)

if __name__ == "__main__":
    main()