train.py

"""

BASED ON: "Deepnet-based surgical tools detection in laparoscopic videos"

AUTHORS: Praveen SR Konduri, G Siva Nageswara Rao

DOI: https://doi.org/10.1016/j.knosys.2025.113517



"""

import os
import torch
import torch.nn as nn
import torch.optim as optim
from torch.utils.data import DataLoader, Dataset
from torchvision import models, transforms
from PIL import Image
import pandas as pd
import numpy as np
import cv2
from sklearn.metrics import classification_report, confusion_matrix
from tqdm import tqdm
import matplotlib.pyplot as plt
import seaborn as sns


# CONFIGURATION 

BASE_PATH = r"C:\Users\anna2\ISM"  # Adjust to your path
PATH_TO_IMAGES = os.path.join(BASE_PATH, "images")
PATH_TO_TRAIN_GT = os.path.join(BASE_PATH, "Baselines", "phase_1b", "gt_for_classification_multiclass_from_filenames_0_index.csv")

MODEL_SAVE_PATH = os.path.join(BASE_PATH, "ANNA", "phase1b-6", "cvggnet_optimized_small.pth")

# Hyperparameters
VAL_FRACTION = 0.1
IMAGE_SIZE = 224  # Standard VGG input
MAX_EPOCHS = 15  # they were3 before 
BATCH_SIZE = 48 
NUM_CLASSES = 3
LEARNING_RATE = 0.0012  # Slightly reduced for stability
# da tentare dopo:  scheduler = optim.lr_scheduler.CosineAnnealingLR(
#    optimizer, T_max=MAX_EPOCHS, eta_min=1e-6)
WEIGHT_DECAY = 5e-4  # INCREASED for regularization
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"

# Features
USE_BILATERAL_FILTER = True
USE_CLASS_WEIGHTS = False
USE_EARLY_STOPPING = True
EARLY_STOP_PATIENCE = 3


#CBAM ATTENTION MODULE (section 3.3)

class ChannelAttention(nn.Module):
    """Channel Attention Module from CBAM"""
    def __init__(self, channels, reduction=16):
        super(ChannelAttention, self).__init__()
        self.avg_pool = nn.AdaptiveAvgPool2d(1)
        self.max_pool = nn.AdaptiveMaxPool2d(1)
        
        self.fc = nn.Sequential(
            nn.Conv2d(channels, channels // reduction, 1, bias=False),
            nn.ReLU(inplace=True),
            nn.Conv2d(channels // reduction, channels, 1, bias=False)
        )
        self.sigmoid = nn.Sigmoid()
    
    def forward(self, x):
        avg_out = self.fc(self.avg_pool(x))
        max_out = self.fc(self.max_pool(x))
        out = avg_out + max_out
        return self.sigmoid(out)


class SpatialAttention(nn.Module):
    """Spatial Attention Module from CBAM"""
    def __init__(self, kernel_size=7):
        super(SpatialAttention, self).__init__()
        self.conv = nn.Conv2d(2, 1, kernel_size, padding=kernel_size//2, bias=False)
        self.sigmoid = nn.Sigmoid()
    
    def forward(self, x):
        avg_out = torch.mean(x, dim=1, keepdim=True)
        max_out, _ = torch.max(x, dim=1, keepdim=True)
        x = torch.cat([avg_out, max_out], dim=1)
        x = self.conv(x)
        return self.sigmoid(x)


class CBAM(nn.Module):
    """Convolutional Block Attention Module"""
    def __init__(self, channels, reduction=16, kernel_size=7):
        super(CBAM, self).__init__()
        self.channel_attention = ChannelAttention(channels, reduction)
        self.spatial_attention = SpatialAttention(kernel_size)
    
    def forward(self, x):
        x = x * self.channel_attention(x)
        x = x * self.spatial_attention(x)
        return x


# ULTRA-OPTIMIZED CVGGNet-16 MODEL 
'''

class CVGGNet16UltraOptimized(nn.Module):

    """

    CVGGNet-16 with Ultra-Aggressive Optimization

    

    VGG-16 Structure (5 conv blocks):

    Block 1: conv1_1, conv1_2 (64 channels)   ← FROZEN

    Block 2: conv2_1, conv2_2 (128 channels)  ← FROZEN  

    Block 3: conv3_1, conv3_2, conv3_3 (256)  ← FROZEN

    Block 4: conv4_1, conv4_2, conv4_3 (512)  ← FROZEN (NEW)

    Block 5: conv5_1, conv5_2, conv5_3 (512)  ← TRAINABLE (only this!)

    

    Classifier: Lightweight 512→128→3 (vs original 4096→4096→3)

    

    Key Changes:

    - Freeze blocks 1-4 (only train block 5)

    - Tiny classifier (99% parameter reduction)

    - Model size: ~200MB (down from 1.6GB)

    - Trainable params: ~15% (down from 43%)

    """

    def __init__(self, num_classes=3, pretrained=True):

        super(CVGGNet16UltraOptimized, self).__init__()

        

        # Load pre-trained VGG-16

        vgg16 = models.vgg16(pretrained=pretrained)

        

        # Extract features

        self.features = vgg16.features

        

        # CBAM attention

        self.cbam = CBAM(channels=512, reduction=16)

        

        # Pooling

        self.avgpool = nn.AdaptiveAvgPool2d((7, 7))

        

        # LIGHTWEIGHT Classifier (CRITICAL FIX for model size)

        self.classifier = nn.Sequential(

            nn.Linear(512 * 7 * 7, 512),  # 25K params (vs 100M in original)

            nn.ReLU(inplace=True),

            nn.Dropout(0.6),  # INCREASED dropout for overfitting

            nn.Linear(512, 128),

            nn.ReLU(inplace=True),

            nn.Dropout(0.5),  # INCREASED dropout

            nn.Linear(128, num_classes)

        )

        

        # Apply aggressive freezing

        self._freeze_early_layers()

    

    def _freeze_early_layers(self):

        """

        ULTRA-AGGRESSIVE FREEZING: Freeze blocks 1-4, train ONLY block 5

        

        VGG-16 features structure:

        - Indices 0-4: Block 1 ← FROZEN

        - Indices 5-9: Block 2 ← FROZEN

        - Indices 10-16: Block 3 ← FROZEN

        - Indices 17-23: Block 4 ← FROZEN (NEW)

        - Indices 24-30: Block 5 ← TRAINABLE (only this!)

        """

        print("\n" + "="*70)

        print("Applying ULTRA-AGGRESSIVE Layer Freezing")

        print("="*70)

        

        # Freeze blocks 1-4 (indices 0-23)

        freeze_until_idx = 10  # Start of block 5 - MOST AGGRESSIVE

        

        for idx, layer in enumerate(self.features):

            if idx < freeze_until_idx:

                for param in layer.parameters():

                    param.requires_grad = False

        

        # Count parameters

        total_params = sum(p.numel() for p in self.parameters())

        trainable_params = sum(p.numel() for p in self.parameters() if p.requires_grad)

        frozen_params = total_params - trainable_params

        

        print(f"\nParameter Summary:")

        print(f"  Total parameters: {total_params:,}")

        print(f"  Frozen parameters: {frozen_params:,} ({100*frozen_params/total_params:.1f}%)")

        print(f"  Trainable parameters: {trainable_params:,} ({100*trainable_params/total_params:.1f}%)")

        

        print(f"\nLayer Status:")

        print(f"  ✗ FROZEN: VGG-16 Blocks 1-4 (conv1-conv4)")

        print(f"  ✓ TRAINABLE: VGG-16 Block 5 ONLY (conv5)")

        print(f"  ✓ TRAINABLE: CBAM Attention")

        print(f"  ✓ TRAINABLE: Lightweight Classifier (512→128→3)")

        

        # Calculate model size

        model_size_mb = (total_params * 4) / (1024**2)  # 4 bytes per float32

        print(f"\nEstimated Model Size:")

        print(f"  Full precision (FP32): ~{model_size_mb:.1f} MB")

        print(f"  Half precision (FP16): ~{model_size_mb/2:.1f} MB")

        print("="*70 + "\n")

    

    def forward(self, x):

        x = self.features(x)

        x = self.cbam(x)

        x = self.avgpool(x)

        x = torch.flatten(x, 1)

        x = self.classifier(x)

        return x

'''

class CVGGNetResNet50(nn.Module):
    def __init__(self, num_classes=3, pretrained=True):
        super(CVGGNetResNet50, self).__init__()
        
        # Load ResNet-50
        resnet = models.resnet50(pretrained=pretrained)
        
        # Extract feature layers
        # Index mapping:
        # 0: conv1, 1: bn1, 2: relu, 3: maxpool
        # 4: layer1, 5: layer2, 6: layer3, 7: layer4
        self.features = nn.Sequential(*list(resnet.children())[:-2])
        
        # CBAM attention on final feature maps (2048 channels)
        self.cbam = CBAM(channels=2048, reduction=16)
        
        self.avgpool = nn.AdaptiveAvgPool2d((1, 1))
        
        # Lightweight classifier
        self.classifier = nn.Sequential(
            nn.Linear(2048, 512),
            nn.ReLU(inplace=True),
            nn.Dropout(0.6),
            nn.Linear(512, 128),
            nn.ReLU(inplace=True),
            nn.Dropout(0.5),
            nn.Linear(128, num_classes)
        )
        
        # Apply freezing
        self._freeze_early_layers()
    
    def _print_freeze_summary(self):
        """Print detailed freezing summary - DEFINE THIS FIRST"""
        total_params = sum(p.numel() for p in self.parameters())
        trainable_params = sum(p.numel() for p in self.parameters() if p.requires_grad)
        frozen_params = total_params - trainable_params
        
        print(f"\nParameter Summary:")
        print(f"  Total parameters: {total_params:,}")
        print(f"  Frozen parameters: {frozen_params:,} ({100*frozen_params/total_params:.1f}%)")
        print(f"  Trainable parameters: {trainable_params:,} ({100*trainable_params/total_params:.1f}%)")
        
        print(f"\nLayer Status:")
        print(f"  ❌ FROZEN: conv1 + bn1 (initial conv)")
        print(f"  ❌ FROZEN: layer1 (3 blocks, 256 channels)")
        print(f"  ❌ FROZEN: layer2 (4 blocks, 512 channels)")
        print(f"  ✓ TRAINABLE: layer3 (6 blocks, 1024 channels)")
        print(f"  ✓ TRAINABLE: layer4 (3 blocks, 2048 channels)")
        print(f"  ✓ TRAINABLE: CBAM Attention")
        print(f"  ✓ TRAINABLE: Classifier (2048→512→128→3)")
        
        model_size_mb = (total_params * 4) / (1024**2)
        print(f"\nEstimated Model Size: ~{model_size_mb:.1f} MB")
        print("="*70 + "\n")
    
    def _freeze_early_layers(self):
        """

        RECOMMENDED: Freeze layers 1-2, train layers 3-4

        """
        print("\n" + "="*70)
        print("ResNet-50 Layer Freezing Strategy")
        print("="*70)
        
        # Freeze initial conv block
        for param in self.features[0].parameters():  # conv1
            param.requires_grad = False
        for param in self.features[1].parameters():  # bn1
            param.requires_grad = False
        
        # Freeze layer1 (early low-level features)
        for param in self.features[4].parameters():
            param.requires_grad = False
        
        # Freeze layer2 (mid-level features)
        for param in self.features[5].parameters():
            param.requires_grad = False
        
        # layer3 and layer4 remain trainable
        
        self._print_freeze_summary()
    
    def forward(self, x):
        x = self.features(x)
        x = self.cbam(x)
        x = self.avgpool(x)
        x = torch.flatten(x, 1)
        x = self.classifier(x)
        return x

# RAPID BILATERAL FILTER (section 3.2 of paper)
# ref: "Bilateral Filtering: Theory and Applications"
# By Sylvain Paris, Pierre Kornprobst, Jack Tumblin and Frédo Durand
# DOI: 10.1561/0600000020 

def rapid_bilateral_filter(image, radius=5, sigma_color=150, sigma_space=8):
    """Rapid Bilateral Filter for image's contrast

      enhancement. Returns smoothened images where 

      important image features are enhanced and non 

      relevant features are eliminated"""
    if isinstance(image, Image.Image):
        image = np.array(image)
    
    filtered = cv2.bilateralFilter(image, radius, sigma_color, sigma_space)
    return filtered


# DATASET 

class SurgicalToolDataset(Dataset):
    """Dataset with optional Rapid Bilateral Filter preprocessing"""
    
    def __init__(self, img_dir, annotation_file, transform=None, 

                 validation_set=False, use_bilateral_filter=True):
        gt = pd.read_csv(annotation_file)
        
        if validation_set:
            self.img_labels = gt[gt["validation_set"] == 1]
        else:
            self.img_labels = gt[gt["validation_set"] == 0]
        
        self.img_dir = img_dir
        self.transform = transform
        self.use_bilateral_filter = use_bilateral_filter
        
        self.images = self.img_labels["file_name"].values
        self.labels = self.img_labels["category_id"].values
    
    def __len__(self):
        return len(self.img_labels)
    
    def __getitem__(self, idx):
        img_path = os.path.join(self.img_dir, self.images[idx])
        image = Image.open(img_path).convert('RGB')
        
        if self.use_bilateral_filter:
            image = rapid_bilateral_filter(image)
            image = Image.fromarray(image)
        
        label = self.labels[idx]
        
        if self.transform:
            image = self.transform(image)
        
        return image, label


# EARLY STOPPING

class EarlyStopping:
    """Early stopping to prevent overfitting"""
    def __init__(self, patience=3, min_delta=0.001):
        self.patience = patience
        self.min_delta = min_delta
        self.counter = 0
        self.best_loss = None
        
    def __call__(self, val_loss):
        if self.best_loss is None:
            self.best_loss = val_loss
        elif val_loss > self.best_loss - self.min_delta:
            self.counter += 1
            if self.counter >= self.patience:
                return True
        else:
            self.best_loss = val_loss
            self.counter = 0
        return False


#TRAINING FUNCTIONS

def compute_class_weights(labels, num_classes):
    """Compute class weights for imbalanced datasets"""
    class_counts = np.bincount(labels, minlength=num_classes)
    total_samples = len(labels)
    weights = total_samples / (num_classes * class_counts)
    weights = torch.FloatTensor(weights)
    print(f"\nClass weights computed: {weights.numpy()}")
    return weights


def train_epoch(model, train_loader, criterion, optimizer, device, class_weights=None):
    """Train for one epoch"""
    model.train()
    running_loss = 0.0
    correct = 0
    total = 0
    
    if class_weights is not None:
        criterion = nn.CrossEntropyLoss(weight=class_weights.to(device))
    
    pbar = tqdm(train_loader, desc="Training", leave=False)
    for images, labels in pbar:
        images, labels = images.to(device), labels.to(device)
        
        optimizer.zero_grad()
        outputs = model(images)
        loss = criterion(outputs, labels)
        loss.backward()
        torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm=1.0)
        optimizer.step()
        
        running_loss += loss.item()
        _, predicted = torch.max(outputs.data, 1)
        total += labels.size(0)
        correct += (predicted == labels).sum().item()
        
        pbar.set_postfix({'loss': f'{loss.item():.4f}', 
                         'acc': f'{100.*correct/total:.2f}%'})
    
    epoch_loss = running_loss / len(train_loader)
    epoch_acc = 100. * correct / total
    
    return epoch_loss, epoch_acc


def validate(model, val_loader, criterion, device):
    """Validate the model"""
    model.eval()
    running_loss = 0.0
    all_predictions = []
    all_labels = []
    
    with torch.no_grad():
        for images, labels in tqdm(val_loader, desc="Validating", leave=False):
            images, labels = images.to(device), labels.to(device)
            
            outputs = model(images)
            loss = criterion(outputs, labels)
            
            running_loss += loss.item()
            
            _, predicted = torch.max(outputs.data, 1)
            all_predictions.extend(predicted.cpu().numpy())
            all_labels.extend(labels.cpu().numpy())
    
    val_loss = running_loss / len(val_loader)
    
    return val_loss, all_predictions, all_labels


def plot_confusion_matrix(labels, predictions, save_path):
    """Plot confusion matrix"""
    cm = confusion_matrix(labels, predictions)
    
    plt.figure(figsize=(8, 6))
    sns.heatmap(cm, annot=True, fmt='d', cmap='Blues', 
                xticklabels=[f'Class {i}' for i in range(len(cm))],
                yticklabels=[f'Class {i}' for i in range(len(cm))])
    plt.title('Confusion Matrix')
    plt.ylabel('True Label')
    plt.xlabel('Predicted Label')
    plt.tight_layout()
    plt.savefig(save_path, dpi=300, bbox_inches='tight')
    plt.close()
    print(f"✓ Confusion matrix saved to {save_path}")


def plot_training_history(train_losses, val_losses, train_accs, val_accs, save_path):
    """Plot training history"""
    epochs = range(1, len(train_losses) + 1)
    
    fig, (ax1, ax2) = plt.subplots(1, 2, figsize=(14, 5))
    
    # Loss plot
    ax1.plot(epochs, train_losses, 'b-o', label='Train Loss', linewidth=2)
    ax1.plot(epochs, val_losses, 'r-s', label='Val Loss', linewidth=2)
    ax1.set_xlabel('Epoch', fontsize=12)
    ax1.set_ylabel('Loss', fontsize=12)
    ax1.set_title('Training and Validation Loss', fontsize=14, fontweight='bold')
    ax1.legend(fontsize=11)
    ax1.grid(True, alpha=0.3)
    
    # Accuracy plot
    ax2.plot(epochs, train_accs, 'b-o', label='Train Acc', linewidth=2)
    ax2.plot(epochs, val_accs, 'r-s', label='Val Acc', linewidth=2)
    ax2.set_xlabel('Epoch', fontsize=12)
    ax2.set_ylabel('Accuracy (%)', fontsize=12)
    ax2.set_title('Training and Validation Accuracy', fontsize=14, fontweight='bold')
    ax2.legend(fontsize=11)
    ax2.grid(True, alpha=0.3)
    
    plt.tight_layout()
    plt.savefig(save_path, dpi=300, bbox_inches='tight')
    plt.close()
    print(f"✓ Training history saved to {save_path}")


# MAIN TRAINING FUNCTION 

def main():
    """Main training pipeline"""
    
    # Set seeds for reproducibility
    torch.manual_seed(543)
    np.random.seed(543)
    
    print("="*70)
    print("CVGGNet-16 ULTRA-OPTIMIZED Training")
    print("Strategy: Ultra-Aggressive Freezing + Tiny Classifier")
    print("="*70)
    print(f"Device: {DEVICE}")
    print(f"Batch size: {BATCH_SIZE}")
    print(f"Max epochs: {MAX_EPOCHS} (REDUCED to prevent overfitting)")
    print(f"Learning rate: {LEARNING_RATE}")
    print(f"Weight decay: {WEIGHT_DECAY} (INCREASED for regularization)")
    print(f"Bilateral filter: {USE_BILATERAL_FILTER}")
    print(f"Early stopping: {USE_EARLY_STOPPING} (patience={EARLY_STOP_PATIENCE})")
    print("="*70 + "\n")
    
    # DATA PREPARATION 
    
    # Create validation split
    df = pd.read_csv(PATH_TO_TRAIN_GT)
    if "validation_set" not in df.columns:
        df["validation_set"] = 0
        val_indices = df.sample(frac=VAL_FRACTION, random_state=42).index
        df.loc[val_indices, "validation_set"] = 1
        df.to_csv(PATH_TO_TRAIN_GT, index=False)
        print(f"✓ Created validation split ({VAL_FRACTION*100:.0f}%)\n")
    
    # REDUCED Data Augmentation (was too aggressive)
    train_transform = transforms.Compose([
        transforms.Resize((IMAGE_SIZE, IMAGE_SIZE)),
        transforms.RandomHorizontalFlip(p=0.5),  # REDUCED from 0.5
        transforms.RandomRotation(degrees=15),
        #transforms.AugMix(severity=2),    # REDUCED from 15
        # REMOVED ColorJitter - too aggressive for surgical images
        transforms.ToTensor(),
        transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
    ])
    
    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])
    ])
    
    # Create datasets
    train_dataset = SurgicalToolDataset(
        img_dir=PATH_TO_IMAGES,
        annotation_file=PATH_TO_TRAIN_GT,
        transform=train_transform,
        validation_set=False,
        use_bilateral_filter=USE_BILATERAL_FILTER
    )
    
    val_dataset = SurgicalToolDataset(
        img_dir=PATH_TO_IMAGES,
        annotation_file=PATH_TO_TRAIN_GT,
        transform=val_transform,
        validation_set=True,
        use_bilateral_filter=USE_BILATERAL_FILTER
    )
    
    # Create dataloaders
    train_loader = DataLoader(train_dataset, batch_size=BATCH_SIZE, 
                             shuffle=True, num_workers=6, pin_memory=True)
    val_loader = DataLoader(val_dataset, batch_size=BATCH_SIZE, 
                           shuffle=False, num_workers=6, pin_memory=True)
    
    print(f"Dataset sizes:")
    print(f"  Training: {len(train_dataset)} images")
    print(f"  Validation: {len(val_dataset)} images")
    print(f"  Batches per epoch: {len(train_loader)} (train), {len(val_loader)} (val)")
    
    # Compute class weights
    class_weights = None
    if USE_CLASS_WEIGHTS:
        class_weights = compute_class_weights(train_dataset.labels, NUM_CLASSES)
    
    # MODEL SETUP 
    
    print(f"\nCreating CVGGNet-Resnet Ultra-Optimized model...")
    model = CVGGNetResNet50(num_classes=NUM_CLASSES, pretrained=True).to(DEVICE)
    
    # Loss and optimizer
    criterion = nn.CrossEntropyLoss()
    
    # Optimizer - only for trainable parameters
    optimizer = optim.AdamW(
        filter(lambda p: p.requires_grad, model.parameters()),
        lr=LEARNING_RATE,
        weight_decay=WEIGHT_DECAY
    )
    
    # Learning rate scheduler
    scheduler = optim.lr_scheduler.ReduceLROnPlateau(
        optimizer, mode='min', factor=0.5, patience=2, verbose=True
    )  #DA CAPIRE 
    
    # Early stopping
    early_stopping = None
    if USE_EARLY_STOPPING:
        early_stopping = EarlyStopping(patience=EARLY_STOP_PATIENCE, min_delta=0.001)
    
    # TRAINING LOOP
    
    best_val_loss = float('inf')
    best_val_acc = 0.0
    train_losses, val_losses = [], []
    train_accs, val_accs = [], []
    
    print("\n" + "="*70)
    print("Starting Training")
    print("="*70 + "\n")
    
    import time
    training_start_time = time.time()
    
    for epoch in range(MAX_EPOCHS):
        epoch_start_time = time.time()
        
        print(f"\nEpoch [{epoch+1}/{MAX_EPOCHS}]")
        print("-" * 70)
        
        # Train
        train_loss, train_acc = train_epoch(
            model, train_loader, criterion, optimizer, DEVICE, class_weights
        )
        
        print(f"Train Loss: {train_loss:.4f}, Train Accuracy: {train_acc:.2f}%")
        
        # Validate
        val_loss, val_predictions, val_labels = validate(
            model, val_loader, criterion, DEVICE
        )
        
        val_acc = 100. * np.sum(np.array(val_predictions) == np.array(val_labels)) / len(val_labels)
        
        print(f"Val Loss: {val_loss:.4f}, Val Accuracy: {val_acc:.2f}%")
        
        # Classification report
        print("\nValidation Metrics:")
        report = classification_report(val_labels, val_predictions, 
                                      target_names=[f'Class {i}' for i in range(NUM_CLASSES)],
                                      digits=4)
        print(report)
        
        # Save history
        train_losses.append(train_loss)
        val_losses.append(val_loss)
        train_accs.append(train_acc)
        val_accs.append(val_acc)
        
        # Learning rate scheduling
        scheduler.step(val_loss)
        
        # Save best model
        if val_acc > best_val_acc:
            best_val_acc = val_acc
            best_val_loss = val_loss
            torch.save({
                'epoch': epoch,
                'model_state_dict': model.state_dict(),
                'optimizer_state_dict': optimizer.state_dict(),
                'val_acc': val_acc,
                'val_loss': val_loss,
                'train_acc': train_acc,
                'train_loss': train_loss,
            }, MODEL_SAVE_PATH)
            print(f"\n✓ Best model saved! (Val Acc: {val_acc:.2f}%)")
        
        # Early stopping check
        if early_stopping is not None:
            if early_stopping(val_loss):
                print(f"\n⚠️ Early stopping at epoch {epoch+1}")
                break
        
        epoch_time = time.time() - epoch_start_time
        print(f"\nEpoch time: {epoch_time/60:.2f} minutes")
        print(f"Current LR: {optimizer.param_groups[0]['lr']:.6f}")
    
    training_time = time.time() - training_start_time
    
    # FINAL EVALUATION 
    
    print("\n" + "="*70)
    print("Training Complete!")
    print("="*70)
    print(f"Total training time: {training_time/60:.2f} minutes")
    print(f"Best Validation Accuracy: {best_val_acc:.2f}%")
    print(f"Best Validation Loss: {best_val_loss:.4f}")
    print(f"Model saved to: {MODEL_SAVE_PATH}")
    
    # Check model size
    model_size_bytes = os.path.getsize(MODEL_SAVE_PATH)
    model_size_mb = model_size_bytes / (1024**2)
    print(f"Model file size: {model_size_mb:.1f} MB")
    
    if model_size_mb > 500:
        print("⚠️ WARNING: Model still large (>500MB). Check classifier architecture.")
    else:
        print("✓ Model size is good for HuggingFace upload!")
    
    # Load best model for final evaluation
    checkpoint = torch.load(MODEL_SAVE_PATH)
    model.load_state_dict(checkpoint['model_state_dict'])
    
    # Final validation
    _, final_predictions, final_labels = validate(model, val_loader, criterion, DEVICE)
    
    # Plot confusion matrix
    cm_path = os.path.join(BASE_PATH, 'confusion_matrix_ultra_optimized.png')
    plot_confusion_matrix(final_labels, final_predictions, cm_path)
    
    # Plot training history
    history_path = os.path.join(BASE_PATH, 'training_history_ultra_optimized.png')
    plot_training_history(train_losses, val_losses, train_accs, val_accs, history_path)
    
    # Final metrics
    print("\n" + "="*70)
    print("Final Validation Metrics:")
    print("="*70)
    final_report = classification_report(final_labels, final_predictions,
                                         target_names=[f'Class {i}' for i in range(NUM_CLASSES)],
                                         digits=4)
    print(final_report)
    
    print(f"\n✓ All done! Results saved in {BASE_PATH}")
    print("="*70)
    
    return model


if __name__ == "__main__":
    model = main()