Fast.ai compatible Training Analysis Dashboard

dashboard.py

def plot_training_dashboard(learn):
    """Create a comprehensive dashboard of training metrics"""
    fig, axes = plt.subplots(2, 2, figsize=(12, 10))
    fig.suptitle('Training Analysis Dashboard', fontsize=16)
    
    # Get the values more carefully
    values = learn.recorder.values
    
    # For training loss, we want the FINAL batch loss of each epoch
    if hasattr(learn.recorder, 'losses'):
        all_train_losses = learn.recorder.losses
        n_batches = len(all_train_losses) // len(values)
        # Take the LAST training loss of each epoch (not the first)
        train_losses_per_epoch = [all_train_losses[(i+1)*n_batches-1] for i in range(len(values))]
    else:
        # Fallback: use the recorded training loss
        train_losses_per_epoch = [v[0] for v in values]
    
    valid_losses = [v[1] for v in values]
    
    # 1. Traditional loss plot
    ax1 = axes[0, 0]
    ax1.plot(train_losses_per_epoch, label='Train', color='blue', marker='o')
    ax1.plot(valid_losses, label='Valid', color='orange', marker='o')
    ax1.set_title('Training vs Validation Loss')
    ax1.set_ylabel('Loss')
    ax1.legend()
    ax1.grid(True, alpha=0.3)
    
    # 2. Loss gap (now should match the simple version)
    ax2 = axes[0, 1]
    gaps = [v - t for t, v in zip(train_losses_per_epoch, valid_losses)]
    ax2.plot(gaps, color='red', linewidth=2, marker='o')
    ax2.axhline(y=0, color='black', linestyle='--', alpha=0.3)
    ax2.set_title('Loss Gap (Valid - Train)')
    ax2.set_ylabel('Gap')
    ax2.fill_between(range(len(gaps)), 0, gaps, 
                     where=[g > 0 for g in gaps],
                     alpha=0.2, color='red')
    ax2.grid(True, alpha=0.3)
    
    # 3. Error rate or first metric
    ax3 = axes[1, 0]
    if len(learn.recorder.values[0]) > 2:  # Has metrics
        error_rates = [v[2] for v in learn.recorder.values]  # First metric after train/valid loss
        ax3.plot(error_rates, color='green', linewidth=2)
        ax3.set_title('Error Rate Evolution')
        ax3.set_ylabel('Error Rate')
    else:
        ax3.text(0.5, 0.5, 'No metrics recorded', ha='center', va='center')
        ax3.set_title('Metrics')
    ax3.set_xlabel('Epoch')
    ax3.grid(True, alpha=0.3)
    
    # 4. Loss ratio
    ax4 = axes[1, 1]
    ratios = [v/t if t > 0 else 0 for t, v in zip(train_losses_per_epoch, valid_losses)]
    ax4.plot(ratios, color='purple', linewidth=2)
    ax4.axhline(y=1, color='black', linestyle='--', alpha=0.3)
    ax4.axhline(y=2, color='orange', linestyle='--', alpha=0.3)
    ax4.set_title('Loss Ratio (Valid/Train)')
    ax4.set_ylabel('Ratio')
    ax4.set_xlabel('Epoch')
    ax4.grid(True, alpha=0.3)
    
    plt.tight_layout()
    plt.show()

# Usage
plot_training_dashboard(learn)