Trixy3/trainer/visualization.py

"""
Training visualization module for creating loss/accuracy plots and training reports.

This module provides comprehensive visualization capabilities for training metrics,
including loss curves, accuracy plots, learning rate schedules, and performance analysis.
"""

import os
import time
from pathlib import Path
from typing import Dict, List, Any, Optional, Tuple
import numpy as np

try:
    import matplotlib
    matplotlib.use('Agg')  # Use non-interactive backend
    import matplotlib.pyplot as plt
    import matplotlib.dates as mdates
    from matplotlib.gridspec import GridSpec
    MATPLOTLIB_AVAILABLE = True
except ImportError:
    MATPLOTLIB_AVAILABLE = False
    plt = None

from .base import TrainingMetrics


class TrainingVisualizer:
    """
    Creates comprehensive visualizations for training progress and results.
    
    Features:
    - Loss/accuracy curves
    - Learning rate schedules
    - Training time analysis
    - Performance metrics
    - Model comparison plots
    """
    
    def __init__(self, output_dir: str, model_name: str):
        """
        Initialize the training visualizer.
        
        Args:
            output_dir: Directory to save plots
            model_name: Name of the model for file naming
        """
        self.output_dir = Path(output_dir)
        self.model_name = model_name
        self.plots_dir = self.output_dir / "plots"
        self.plots_dir.mkdir(parents=True, exist_ok=True)
        
        if not MATPLOTLIB_AVAILABLE:
            print("Warning: matplotlib not available. Plots will not be generated.")
    
    def create_training_plots(self, metrics: TrainingMetrics, 
                            additional_metrics: Dict[str, Any] = None) -> List[str]:
        """
        Create comprehensive training visualization plots.
        
        Args:
            metrics: TrainingMetrics object with training history
            additional_metrics: Additional metrics to visualize
            
        Returns:
            List of generated plot file paths
        """
        if not MATPLOTLIB_AVAILABLE:
            return []
        
        generated_plots = []
        
        # Main training curves plot
        main_plot = self._create_main_training_plot(metrics)
        if main_plot:
            generated_plots.append(main_plot)
        
        # Learning rate schedule plot
        lr_plot = self._create_learning_rate_plot(metrics)
        if lr_plot:
            generated_plots.append(lr_plot)
        
        # Training time analysis
        time_plot = self._create_time_analysis_plot(metrics)
        if time_plot:
            generated_plots.append(time_plot)
        
        # Loss distribution plot
        loss_dist_plot = self._create_loss_distribution_plot(metrics)
        if loss_dist_plot:
            generated_plots.append(loss_dist_plot)
        
        # Additional metrics plots
        if additional_metrics:
            additional_plots = self._create_additional_metrics_plots(additional_metrics)
            generated_plots.extend(additional_plots)
        
        return generated_plots
    
    def _create_main_training_plot(self, metrics: TrainingMetrics) -> Optional[str]:
        """Create the main training curves plot (loss + accuracy)."""
        if not metrics.metrics['train_loss']:
            return None
        
        fig = plt.figure(figsize=(14, 8))
        gs = GridSpec(2, 2, hspace=0.3, wspace=0.3)
        
        epochs = list(range(len(metrics.metrics['train_loss'])))
        
        # Training and validation loss
        ax1 = fig.add_subplot(gs[0, :])
        ax1.plot(epochs, metrics.metrics['train_loss'], 'b-', label='Train Loss', linewidth=2)
        if metrics.metrics['val_loss']:
            # Filter out None values for validation loss
            val_epochs = [i for i, loss in enumerate(metrics.metrics['val_loss']) if loss is not None]
            val_losses = [loss for loss in metrics.metrics['val_loss'] if loss is not None]
            ax1.plot(val_epochs, val_losses, 'r-', label='Validation Loss', linewidth=2)
        
        ax1.set_xlabel('Epoch')
        ax1.set_ylabel('Loss')
        ax1.set_title(f'{self.model_name} - Training Progress', fontsize=14, fontweight='bold')
        ax1.legend()
        ax1.grid(True, alpha=0.3)
        ax1.set_yscale('log')  # Log scale for better loss visualization
        
        # Training and validation accuracy
        ax2 = fig.add_subplot(gs[1, 0])
        if metrics.metrics['train_accuracy']:
            ax2.plot(epochs, metrics.metrics['train_accuracy'], 'b-', label='Train Accuracy', linewidth=2)
        if metrics.metrics['val_accuracy']:
            val_epochs = [i for i, acc in enumerate(metrics.metrics['val_accuracy']) if acc is not None]
            val_accs = [acc for acc in metrics.metrics['val_accuracy'] if acc is not None]
            ax2.plot(val_epochs, val_accs, 'r-', label='Validation Accuracy', linewidth=2)
        
        ax2.set_xlabel('Epoch')
        ax2.set_ylabel('Accuracy')
        ax2.set_title('Accuracy Progress')
        ax2.legend()
        ax2.grid(True, alpha=0.3)
        ax2.set_ylim(0, 1)
        
        # Training summary box
        ax3 = fig.add_subplot(gs[1, 1])
        ax3.axis('off')
        
        # Create summary text
        summary_text = self._create_training_summary(metrics)
        ax3.text(0.05, 0.95, summary_text, transform=ax3.transAxes, fontsize=10,
                verticalalignment='top', bbox=dict(boxstyle="round,pad=0.3", facecolor="lightblue", alpha=0.5))
        
        plt.suptitle(f'{self.model_name} Training Overview', fontsize=16, fontweight='bold')
        
        # Save plot
        plot_path = self.plots_dir / f"{self.model_name}_training_curves.png"
        plt.savefig(plot_path, dpi=300, bbox_inches='tight')
        plt.close()
        
        return str(plot_path)
    
    def _create_learning_rate_plot(self, metrics: TrainingMetrics) -> Optional[str]:
        """Create learning rate schedule visualization."""
        if not metrics.metrics['learning_rates']:
            return None
        
        fig, (ax1, ax2) = plt.subplots(2, 1, figsize=(12, 8))
        
        epochs = list(range(len(metrics.metrics['learning_rates'])))
        
        # Learning rate over time
        ax1.plot(epochs, metrics.metrics['learning_rates'], 'g-', linewidth=2)
        ax1.set_xlabel('Epoch')
        ax1.set_ylabel('Learning Rate')
        ax1.set_title(f'{self.model_name} - Learning Rate Schedule')
        ax1.grid(True, alpha=0.3)
        ax1.set_yscale('log')
        
        # Learning rate vs loss correlation
        if metrics.metrics['train_loss']:
            ax2.scatter(metrics.metrics['learning_rates'], metrics.metrics['train_loss'], 
                       alpha=0.6, c=epochs, cmap='viridis')
            ax2.set_xlabel('Learning Rate')
            ax2.set_ylabel('Training Loss')
            ax2.set_title('Learning Rate vs Training Loss')
            ax2.set_xscale('log')
            ax2.set_yscale('log')
            ax2.grid(True, alpha=0.3)
            
            # Add colorbar
            cbar = plt.colorbar(ax2.collections[0], ax=ax2)
            cbar.set_label('Epoch')
        
        plt.tight_layout()
        
        # Save plot
        plot_path = self.plots_dir / f"{self.model_name}_learning_rate.png"
        plt.savefig(plot_path, dpi=300, bbox_inches='tight')
        plt.close()
        
        return str(plot_path)
    
    def _create_time_analysis_plot(self, metrics: TrainingMetrics) -> Optional[str]:
        """Create training time analysis visualization."""
        if not metrics.metrics['epoch_times']:
            return None
        
        fig, ((ax1, ax2), (ax3, ax4)) = plt.subplots(2, 2, figsize=(14, 10))
        
        epochs = list(range(len(metrics.metrics['epoch_times'])))
        epoch_times = metrics.metrics['epoch_times']
        
        # Epoch time progression
        ax1.plot(epochs, epoch_times, 'purple', linewidth=2)
        ax1.set_xlabel('Epoch')
        ax1.set_ylabel('Time (seconds)')
        ax1.set_title('Training Time per Epoch')
        ax1.grid(True, alpha=0.3)
        
        # Cumulative time
        cumulative_time = np.cumsum(epoch_times)
        ax2.plot(epochs, cumulative_time / 3600, 'orange', linewidth=2)  # Convert to hours
        ax2.set_xlabel('Epoch')
        ax2.set_ylabel('Cumulative Time (hours)')
        ax2.set_title('Cumulative Training Time')
        ax2.grid(True, alpha=0.3)
        
        # Time distribution histogram
        ax3.hist(epoch_times, bins=20, alpha=0.7, color='skyblue', edgecolor='black')
        ax3.set_xlabel('Epoch Time (seconds)')
        ax3.set_ylabel('Frequency')
        ax3.set_title('Epoch Time Distribution')
        ax3.grid(True, alpha=0.3)
        
        # Training efficiency (loss reduction per time)
        if metrics.metrics['train_loss'] and len(metrics.metrics['train_loss']) > 1:
            loss_improvements = []
            time_ratios = []
            
            for i in range(1, len(metrics.metrics['train_loss'])):
                if metrics.metrics['train_loss'][i-1] > 0:
                    loss_improvement = (metrics.metrics['train_loss'][i-1] - metrics.metrics['train_loss'][i]) / metrics.metrics['train_loss'][i-1]
                    time_ratio = epoch_times[i] / np.mean(epoch_times[:i+1])
                    
                    loss_improvements.append(loss_improvement)
                    time_ratios.append(time_ratio)
            
            if loss_improvements:
                ax4.scatter(time_ratios, loss_improvements, alpha=0.6, c=range(len(loss_improvements)), cmap='viridis')
                ax4.set_xlabel('Relative Epoch Time')
                ax4.set_ylabel('Loss Improvement Ratio')
                ax4.set_title('Training Efficiency')
                ax4.grid(True, alpha=0.3)
        
        plt.suptitle(f'{self.model_name} Training Time Analysis', fontsize=14, fontweight='bold')
        plt.tight_layout()
        
        # Save plot
        plot_path = self.plots_dir / f"{self.model_name}_time_analysis.png"
        plt.savefig(plot_path, dpi=300, bbox_inches='tight')
        plt.close()
        
        return str(plot_path)
    
    def _create_loss_distribution_plot(self, metrics: TrainingMetrics) -> Optional[str]:
        """Create loss distribution and convergence analysis."""
        if not metrics.metrics['train_loss']:
            return None
        
        fig, ((ax1, ax2), (ax3, ax4)) = plt.subplots(2, 2, figsize=(14, 10))
        
        train_losses = metrics.metrics['train_loss']
        epochs = list(range(len(train_losses)))
        
        # Loss convergence (moving average)
        window_size = max(5, len(train_losses) // 20)
        if len(train_losses) >= window_size:
            moving_avg = np.convolve(train_losses, np.ones(window_size)/window_size, mode='valid')
            moving_epochs = epochs[window_size-1:]
            
            ax1.plot(epochs, train_losses, alpha=0.3, color='blue', label='Raw Loss')
            ax1.plot(moving_epochs, moving_avg, color='red', linewidth=2, label=f'Moving Average ({window_size})')
            ax1.set_xlabel('Epoch')
            ax1.set_ylabel('Loss')
            ax1.set_title('Loss Convergence Analysis')
            ax1.legend()
            ax1.grid(True, alpha=0.3)
            ax1.set_yscale('log')
        
        # Loss distribution histogram
        ax2.hist(train_losses, bins=30, alpha=0.7, color='green', edgecolor='black')
        ax2.set_xlabel('Loss Value')
        ax2.set_ylabel('Frequency')
        ax2.set_title('Training Loss Distribution')
        ax2.grid(True, alpha=0.3)
        
        # Loss gradient (rate of change)
        if len(train_losses) > 1:
            loss_gradients = np.diff(train_losses)
            ax3.plot(epochs[1:], loss_gradients, color='purple', linewidth=1)
            ax3.axhline(y=0, color='red', linestyle='--', alpha=0.7)
            ax3.set_xlabel('Epoch')
            ax3.set_ylabel('Loss Change')
            ax3.set_title('Loss Gradient (Rate of Change)')
            ax3.grid(True, alpha=0.3)
        
        # Loss stability (rolling standard deviation)
        if len(train_losses) >= 10:
            window = 10
            rolling_std = []
            for i in range(window, len(train_losses)):
                std = np.std(train_losses[i-window:i])
                rolling_std.append(std)
            
            ax4.plot(epochs[window:], rolling_std, color='orange', linewidth=2)
            ax4.set_xlabel('Epoch')
            ax4.set_ylabel('Rolling Std Dev')
            ax4.set_title(f'Loss Stability (Rolling {window}-Epoch Window)')
            ax4.grid(True, alpha=0.3)
        
        plt.suptitle(f'{self.model_name} Loss Analysis', fontsize=14, fontweight='bold')
        plt.tight_layout()
        
        # Save plot
        plot_path = self.plots_dir / f"{self.model_name}_loss_analysis.png"
        plt.savefig(plot_path, dpi=300, bbox_inches='tight')
        plt.close()
        
        return str(plot_path)
    
    def _create_additional_metrics_plots(self, additional_metrics: Dict[str, Any]) -> List[str]:
        """Create plots for additional metrics like speaker verification, etc."""
        generated_plots = []
        
        # Speaker verification metrics (if available)
        if 'speaker_verification' in additional_metrics:
            plot_path = self._create_speaker_verification_plot(additional_metrics['speaker_verification'])
            if plot_path:
                generated_plots.append(plot_path)
        
        # Embedding analysis (if available)
        if 'embedding_analysis' in additional_metrics:
            plot_path = self._create_embedding_analysis_plot(additional_metrics['embedding_analysis'])
            if plot_path:
                generated_plots.append(plot_path)
        
        return generated_plots
    
    def _create_speaker_verification_plot(self, verification_metrics: Dict[str, Any]) -> Optional[str]:
        """Create speaker verification performance visualization."""
        fig, ((ax1, ax2), (ax3, ax4)) = plt.subplots(2, 2, figsize=(14, 10))
        
        # EER visualization
        eer = verification_metrics.get('equal_error_rate', 0)
        eer_threshold = verification_metrics.get('eer_threshold', 0)
        
        ax1.bar(['EER'], [eer], color='red', alpha=0.7)
        ax1.set_ylabel('Equal Error Rate')
        ax1.set_title(f'Speaker Verification EER: {eer:.4f}')
        ax1.grid(True, alpha=0.3)
        
        # Similarity distributions
        pos_sim = verification_metrics.get('mean_positive_similarity', 0)
        neg_sim = verification_metrics.get('mean_negative_similarity', 0)
        
        ax2.bar(['Positive Pairs', 'Negative Pairs'], [pos_sim, neg_sim], 
               color=['green', 'red'], alpha=0.7)
        ax2.set_ylabel('Mean Similarity')
        ax2.set_title('Speaker Similarity Distribution')
        ax2.grid(True, alpha=0.3)
        
        # Threshold visualization
        thresholds = np.linspace(neg_sim - 0.2, pos_sim + 0.2, 100)
        ax3.axvline(x=eer_threshold, color='red', linestyle='--', linewidth=2, label=f'EER Threshold: {eer_threshold:.3f}')
        ax3.axvline(x=eer_threshold + 0.1, color='orange', linestyle='--', alpha=0.7, label='Conservative: +0.1')
        ax3.axvline(x=eer_threshold - 0.1, color='blue', linestyle='--', alpha=0.7, label='Liberal: -0.1')
        ax3.set_xlabel('Similarity Threshold')
        ax3.set_title('Recommended Thresholds')
        ax3.legend()
        ax3.grid(True, alpha=0.3)
        
        # Performance summary
        ax4.axis('off')
        summary_text = f"""Speaker Verification Summary:
        
EER: {eer:.4f}
EER Threshold: {eer_threshold:.4f}
Positive Pairs: {verification_metrics.get('num_positive_pairs', 0):,}
Negative Pairs: {verification_metrics.get('num_negative_pairs', 0):,}
Mean Positive Similarity: {pos_sim:.4f}
Mean Negative Similarity: {neg_sim:.4f}
Separability: {pos_sim - neg_sim:.4f}
        """
        
        ax4.text(0.05, 0.95, summary_text, transform=ax4.transAxes, fontsize=12,
                verticalalignment='top', bbox=dict(boxstyle="round,pad=0.3", facecolor="lightgreen", alpha=0.5))
        
        plt.suptitle(f'{self.model_name} Speaker Verification Analysis', fontsize=14, fontweight='bold')
        plt.tight_layout()
        
        # Save plot
        plot_path = self.plots_dir / f"{self.model_name}_speaker_verification.png"
        plt.savefig(plot_path, dpi=300, bbox_inches='tight')
        plt.close()
        
        return str(plot_path)
    
    def _create_embedding_analysis_plot(self, embedding_metrics: Dict[str, Any]) -> Optional[str]:
        """Create embedding quality analysis visualization."""
        fig, ((ax1, ax2), (ax3, ax4)) = plt.subplots(2, 2, figsize=(14, 10))
        
        # Distance comparison
        intra_dist = embedding_metrics.get('mean_intra_distance', 0)
        inter_dist = embedding_metrics.get('mean_inter_distance', 0)
        separability = embedding_metrics.get('separability_ratio', 0)
        
        distances = ['Intra-Speaker', 'Inter-Speaker']
        values = [intra_dist, inter_dist]
        colors = ['blue', 'red']
        
        bars = ax1.bar(distances, values, color=colors, alpha=0.7)
        ax1.set_ylabel('Distance')
        ax1.set_title('Speaker Embedding Distances')
        ax1.grid(True, alpha=0.3)
        
        # Add value labels on bars
        for bar, value in zip(bars, values):
            height = bar.get_height()
            ax1.text(bar.get_x() + bar.get_width()/2., height + height*0.01,
                    f'{value:.4f}', ha='center', va='bottom')
        
        # Separability ratio
        ax2.bar(['Separability Ratio'], [separability], color='green', alpha=0.7)
        ax2.set_ylabel('Ratio (Inter/Intra)')
        ax2.set_title(f'Embedding Separability: {separability:.2f}')
        ax2.grid(True, alpha=0.3)
        
        # Add separability quality indicator
        if separability > 2.0:
            quality = "Excellent"
            color = 'green'
        elif separability > 1.5:
            quality = "Good"
            color = 'orange'
        else:
            quality = "Poor"
            color = 'red'
        
        ax2.text(0, separability + separability*0.05, quality, ha='center', 
                fontweight='bold', color=color, fontsize=14)
        
        # Speaker count information
        num_speakers = embedding_metrics.get('num_speakers', 0)
        ax3.pie([num_speakers, max(1, 10 - num_speakers)], labels=['Trained Speakers', 'Potential'], 
               autopct='%1.0f', startangle=90, colors=['lightblue', 'lightgray'])
        ax3.set_title(f'Speaker Coverage ({num_speakers} speakers)')
        
        # Quality assessment
        ax4.axis('off')
        
        # Determine overall quality
        if separability > 2.0 and intra_dist < inter_dist:
            overall_quality = "EXCELLENT"
            quality_color = "lightgreen"
        elif separability > 1.5:
            overall_quality = "GOOD"
            quality_color = "lightyellow"
        else:
            overall_quality = "NEEDS IMPROVEMENT"
            quality_color = "lightcoral"
        
        summary_text = f"""Embedding Quality Assessment:
        
Overall Quality: {overall_quality}

Metrics:
• Number of Speakers: {num_speakers}
• Intra-Speaker Distance: {intra_dist:.4f}
• Inter-Speaker Distance: {inter_dist:.4f}
• Separability Ratio: {separability:.2f}

Quality Indicators:
• Distance Separation: {'✓' if inter_dist > intra_dist else '✗'}
• Good Separability: {'✓' if separability > 1.5 else '✗'}
• Excellent Separability: {'✓' if separability > 2.0 else '✗'}
        """
        
        ax4.text(0.05, 0.95, summary_text, transform=ax4.transAxes, fontsize=11,
                verticalalignment='top', bbox=dict(boxstyle="round,pad=0.3", facecolor=quality_color, alpha=0.7))
        
        plt.suptitle(f'{self.model_name} Embedding Quality Analysis', fontsize=14, fontweight='bold')
        plt.tight_layout()
        
        # Save plot
        plot_path = self.plots_dir / f"{self.model_name}_embedding_analysis.png"
        plt.savefig(plot_path, dpi=300, bbox_inches='tight')
        plt.close()
        
        return str(plot_path)
    
    def _create_training_summary(self, metrics: TrainingMetrics) -> str:
        """Create a text summary of training results."""
        total_epochs = len(metrics.metrics['train_loss'])
        final_train_loss = metrics.metrics['train_loss'][-1] if metrics.metrics['train_loss'] else 0
        final_train_acc = metrics.metrics['train_accuracy'][-1] if metrics.metrics['train_accuracy'] else 0
        
        val_losses = [loss for loss in metrics.metrics['val_loss'] if loss is not None]
        final_val_loss = val_losses[-1] if val_losses else 0
        
        val_accs = [acc for acc in metrics.metrics['val_accuracy'] if acc is not None]
        final_val_acc = val_accs[-1] if val_accs else 0
        
        total_time = sum(metrics.metrics['epoch_times']) if metrics.metrics['epoch_times'] else 0
        avg_epoch_time = total_time / max(1, total_epochs)
        
        summary = f"""Training Summary:
        
Total Epochs: {total_epochs}
Total Time: {total_time/3600:.2f} hours
Avg Time/Epoch: {avg_epoch_time:.2f} sec

Final Metrics:
Train Loss: {final_train_loss:.6f}
Train Accuracy: {final_train_acc:.4f}
Val Loss: {final_val_loss:.6f}
Val Accuracy: {final_val_acc:.4f}

Best Validation:
Loss: {metrics.best_val_loss:.6f}
Epoch: {metrics.best_epoch}
        """
        
        return summary
    
    def create_comparison_plot(self, multiple_metrics: Dict[str, TrainingMetrics], 
                             title: str = "Model Comparison") -> Optional[str]:
        """
        Create comparison plots for multiple training runs.
        
        Args:
            multiple_metrics: Dictionary mapping model names to their metrics
            title: Title for the comparison plot
            
        Returns:
            Path to generated comparison plot
        """
        if not MATPLOTLIB_AVAILABLE or not multiple_metrics:
            return None
        
        fig, ((ax1, ax2), (ax3, ax4)) = plt.subplots(2, 2, figsize=(16, 12))
        
        colors = plt.cm.Set1(np.linspace(0, 1, len(multiple_metrics)))
        
        for (model_name, metrics), color in zip(multiple_metrics.items(), colors):
            epochs = list(range(len(metrics.metrics['train_loss'])))
            
            # Training loss comparison
            ax1.plot(epochs, metrics.metrics['train_loss'], 
                    label=model_name, color=color, linewidth=2)
            
            # Validation loss comparison
            if metrics.metrics['val_loss']:
                val_epochs = [i for i, loss in enumerate(metrics.metrics['val_loss']) if loss is not None]
                val_losses = [loss for loss in metrics.metrics['val_loss'] if loss is not None]
                ax2.plot(val_epochs, val_losses, 
                        label=model_name, color=color, linewidth=2, linestyle='--')
            
            # Training accuracy comparison
            if metrics.metrics['train_accuracy']:
                ax3.plot(epochs, metrics.metrics['train_accuracy'], 
                        label=model_name, color=color, linewidth=2)
            
            # Training time comparison
            if metrics.metrics['epoch_times']:
                ax4.plot(epochs, np.cumsum(metrics.metrics['epoch_times']) / 3600, 
                        label=model_name, color=color, linewidth=2)
        
        # Configure subplots
        ax1.set_xlabel('Epoch')
        ax1.set_ylabel('Training Loss')
        ax1.set_title('Training Loss Comparison')
        ax1.legend()
        ax1.grid(True, alpha=0.3)
        ax1.set_yscale('log')
        
        ax2.set_xlabel('Epoch')
        ax2.set_ylabel('Validation Loss')
        ax2.set_title('Validation Loss Comparison')
        ax2.legend()
        ax2.grid(True, alpha=0.3)
        ax2.set_yscale('log')
        
        ax3.set_xlabel('Epoch')
        ax3.set_ylabel('Accuracy')
        ax3.set_title('Training Accuracy Comparison')
        ax3.legend()
        ax3.grid(True, alpha=0.3)
        
        ax4.set_xlabel('Epoch')
        ax4.set_ylabel('Cumulative Time (hours)')
        ax4.set_title('Training Time Comparison')
        ax4.legend()
        ax4.grid(True, alpha=0.3)
        
        plt.suptitle(title, fontsize=16, fontweight='bold')
        plt.tight_layout()
        
        # Save plot
        plot_path = self.plots_dir / f"model_comparison_{int(time.time())}.png"
        plt.savefig(plot_path, dpi=300, bbox_inches='tight')
        plt.close()
        
        return str(plot_path)


def create_training_visualizer(output_dir: str, model_name: str) -> TrainingVisualizer:
    """
    Factory function to create a training visualizer.
    
    Args:
        output_dir: Directory to save plots
        model_name: Name of the model
        
    Returns:
        TrainingVisualizer instance
    """
    return TrainingVisualizer(output_dir=output_dir, model_name=model_name)