SpatialCLAP/evaluate_similarity.py

#!/usr/bin/env python3
"""
Audio Similarity Evaluation using Spatial CLAP Model

This script evaluates the similarity between ground truth and output audio files
using the Spatial CLAP model. It processes stereo audio files, extracts embeddings,
and calculates cosine similarity scores.

Usage:
    python evaluate_similarity.py /path/to/audio/directory [options]

Expected directory structure:
    data_directory/
    ├── 000/
    │   ├── gt.wav
    │   ├── mixture.wav
    │   └── output.wav
    ├── 001/
    │   ├── gt.wav
    │   ├── mixture.wav
    │   └── output.wav
    └── ...

Requirements:
    - torch>=1.9.0
    - torchaudio>=0.9.0
    - numpy>=1.21.0
    - tqdm>=4.62.0
    - transformers>=4.20.0

Author: AI Assistant
Date: 2024
"""

import os
import sys
import torch
import torchaudio
import numpy as np
from pathlib import Path
from tqdm import tqdm
import argparse
from typing import List, Tuple, Dict
import warnings

# Suppress warnings for cleaner output
warnings.filterwarnings("ignore")

# Add the current directory to Python path to import Spatial CLAP modules
sys.path.append(os.path.dirname(os.path.abspath(__file__)))

try:
    from model import CLAPEncoder
except ImportError as e:
    print(f"Error importing Spatial CLAP model: {e}")
    print("Make sure you're running this script from the SpatialCLAP directory")
    print("and that all required dependencies are installed:")
    print("pip install torch torchaudio numpy tqdm transformers")
    sys.exit(1)


def load_and_resample_audio(file_path: str, target_sr: int = 16000) -> torch.Tensor:
    """
    Load audio file and resample to target sample rate.
    
    Args:
        file_path: Path to audio file
        target_sr: Target sample rate (default: 16000 Hz)
    
    Returns:
        Audio tensor of shape (2, time) - stereo audio
    """
    try:
        waveform, sr = torchaudio.load(file_path)
        
        # Resample if necessary
        if sr != target_sr:
            resampler = torchaudio.transforms.Resample(orig_freq=sr, new_freq=target_sr)
            waveform = resampler(waveform)
        
        # Ensure stereo format (2 channels)
        if waveform.shape[0] == 1:
            # If mono, duplicate to stereo
            waveform = waveform.repeat(2, 1)
        elif waveform.shape[0] > 2:
            # If more than 2 channels, take first 2
            waveform = waveform[:2, :]
        
        return waveform
        
    except Exception as e:
        raise RuntimeError(f"Error loading audio file {file_path}: {str(e)}")


def calculate_cosine_similarity(embedding1: torch.Tensor, embedding2: torch.Tensor) -> float:
    """
    Calculate cosine similarity between two embeddings.
    
    Args:
        embedding1: First embedding tensor
        embedding2: Second embedding tensor
    
    Returns:
        Cosine similarity score (0-1)
    """
    # Ensure embeddings are normalized
    embedding1 = torch.nn.functional.normalize(embedding1, dim=-1)
    embedding2 = torch.nn.functional.normalize(embedding2, dim=-1)
    
    # Calculate cosine similarity
    similarity = torch.cosine_similarity(embedding1, embedding2, dim=-1)
    return similarity.item()


def process_directory(data_dir: str, model: CLAPEncoder, device: str, batch_size: int = 8) -> List[Dict]:
    """
    Process all subdirectories in the data directory and calculate similarities.
    
    Args:
        data_dir: Path to directory containing subdirectories with audio files
        model: Loaded Spatial CLAP model
        device: Device to run inference on
        batch_size: Number of audio pairs to process in each batch
    
    Returns:
        List of dictionaries containing results for each subdirectory
    """
    results = []
    data_path = Path(data_dir)
    
    # Get all subdirectories
    subdirs = [d for d in data_path.iterdir() if d.is_dir()]
    subdirs.sort()  # Sort for consistent ordering
    
    print(f"Found {len(subdirs)} subdirectories to process")
    
    if len(subdirs) == 0:
        print("No subdirectories found in the specified path")
        return results
    
    # Filter subdirectories that have required files
    valid_subdirs = []
    for subdir in subdirs:
        gt_path = subdir / "gt.wav"
        output_path = subdir / "output.wav"
        if gt_path.exists() and output_path.exists():
            valid_subdirs.append(subdir)
        else:
            print(f"Warning: Missing files in {subdir.name} (skipping)")
    
    print(f"Processing {len(valid_subdirs)} valid directories in batches of {batch_size}")
    
    # Process in batches
    for batch_start in tqdm(range(0, len(valid_subdirs), batch_size), desc="Processing batches"):
        batch_end = min(batch_start + batch_size, len(valid_subdirs))
        batch_subdirs = valid_subdirs[batch_start:batch_end]
        
        try:
            # Load all audio files in the batch
            gt_audios = []
            output_audios = []
            batch_info = []
            
            for subdir in batch_subdirs:
                gt_path = subdir / "gt.wav"
                output_path = subdir / "output.wav"
                
                # Load and resample audio files
                gt_audio = load_and_resample_audio(str(gt_path))
                output_audio = load_and_resample_audio(str(output_path))
                
                # Ensure both audios have the same length (pad with zeros if necessary)
                max_len = max(gt_audio.shape[1], output_audio.shape[1])
                
                if gt_audio.shape[1] < max_len:
                    gt_audio = torch.nn.functional.pad(gt_audio, (0, max_len - gt_audio.shape[1]))
                if output_audio.shape[1] < max_len:
                    output_audio = torch.nn.functional.pad(output_audio, (0, max_len - output_audio.shape[1]))
                
                gt_audios.append(gt_audio)
                output_audios.append(output_audio)
                batch_info.append({
                    'directory': subdir.name,
                    'gt_path': str(gt_path),
                    'output_path': str(output_path)
                })
            
            # Stack into batch tensors
            gt_batch = torch.stack(gt_audios).to(device)  # (batch_size, 2, time)
            output_batch = torch.stack(output_audios).to(device)  # (batch_size, 2, time)
            
            # Get embeddings for the entire batch
            with torch.no_grad():
                gt_embeddings = model.embed_audio(gt_batch)
                output_embeddings = model.embed_audio(output_batch)
            
            # Calculate similarities for the batch
            for i, info in enumerate(batch_info):
                similarity = calculate_cosine_similarity(gt_embeddings[i], output_embeddings[i])
                
                result = {
                    'directory': info['directory'],
                    'similarity': similarity,
                    'gt_path': info['gt_path'],
                    'output_path': info['output_path']
                }
                results.append(result)
                
        except Exception as e:
            print(f"Error processing batch {batch_start}-{batch_end}: {str(e)}")
            # Process individual files in the batch if batch processing fails
            for subdir in batch_subdirs:
                try:
                    gt_path = subdir / "gt.wav"
                    output_path = subdir / "output.wav"
                    
                    gt_audio = load_and_resample_audio(str(gt_path))
                    output_audio = load_and_resample_audio(str(output_path))
                    
                    # Ensure both audios have the same length
                    max_len = max(gt_audio.shape[1], output_audio.shape[1])
                    
                    if gt_audio.shape[1] < max_len:
                        gt_audio = torch.nn.functional.pad(gt_audio, (0, max_len - gt_audio.shape[1]))
                    if output_audio.shape[1] < max_len:
                        output_audio = torch.nn.functional.pad(output_audio, (0, max_len - output_audio.shape[1]))
                    
                    # Add batch dimension and move to device
                    gt_audio = gt_audio.unsqueeze(0).to(device)
                    output_audio = output_audio.unsqueeze(0).to(device)
                    
                    # Get embeddings
                    with torch.no_grad():
                        gt_embedding = model.embed_audio(gt_audio)
                        output_embedding = model.embed_audio(output_audio)
                    
                    # Calculate cosine similarity
                    similarity = calculate_cosine_similarity(gt_embedding, output_embedding)
                    
                    # Store results
                    result = {
                        'directory': subdir.name,
                        'similarity': similarity,
                        'gt_path': str(gt_path),
                        'output_path': str(output_path)
                    }
                    results.append(result)
                    
                except Exception as individual_error:
                    print(f"Error processing {subdir.name}: {str(individual_error)}")
                    continue
    
    return results


def save_results(results: List[Dict], output_file: str):
    """
    Save results to a text file.
    
    Args:
        results: List of result dictionaries
        output_file: Path to output file
    """
    try:
        with open(output_file, 'w') as f:
            f.write("Directory\tSimilarity\tGT_Path\tOutput_Path\n")
            for result in results:
                f.write(f"{result['directory']}\t{result['similarity']:.6f}\t{result['gt_path']}\t{result['output_path']}\n")
        print(f"Results saved to: {output_file}")
    except Exception as e:
        print(f"Error saving results: {str(e)}")


def print_summary(results: List[Dict]):
    """
    Print summary statistics.
    
    Args:
        results: List of result dictionaries
    """
    if not results:
        print("No results to summarize")
        return
    
    similarities = [r['similarity'] for r in results]
    
    print(f"\n" + "="*60)
    print(f"SUMMARY STATISTICS")
    print(f"="*60)
    print(f"Total processed: {len(results)}")
    print(f"Mean similarity: {np.mean(similarities):.6f}")
    print(f"Std similarity:  {np.std(similarities):.6f}")
    print(f"Min similarity:  {np.min(similarities):.6f}")
    print(f"Max similarity:  {np.max(similarities):.6f}")
    print(f"Median similarity: {np.median(similarities):.6f}")
    print(f"="*60)
    
    # Print some example results
    print(f"\nFirst 10 results:")
    print(f"{'#':<3} {'Directory':<12} {'Similarity':<12}")
    print(f"{'-'*3} {'-'*12} {'-'*12}")
    for i, result in enumerate(results[:10]):
        print(f"{i+1:<3} {result['directory']:<12} {result['similarity']:.6f}")
    
    if len(results) > 10:
        print(f"... and {len(results) - 10} more results")


def check_dependencies():
    """Check if all required dependencies are available."""
    required_modules = ['torch', 'torchaudio', 'numpy', 'tqdm', 'transformers']
    missing_modules = []
    
    for module in required_modules:
        try:
            __import__(module)
        except ImportError:
            missing_modules.append(module)
    
    if missing_modules:
        print("❌ Missing required dependencies:")
        for module in missing_modules:
            print(f"   - {module}")
        print("\nPlease install missing dependencies:")
        print("pip install torch torchaudio numpy tqdm transformers")
        return False
    
    return True


def main():
    """Main function to run the audio similarity evaluation."""
    
    parser = argparse.ArgumentParser(
        description="Evaluate audio similarity using Spatial CLAP model",
        formatter_class=argparse.RawDescriptionHelpFormatter,
        epilog="""
Examples:
  python evaluate_similarity.py /path/to/audio/directory
  python evaluate_similarity.py /path/to/audio/directory --output results.txt
  python evaluate_similarity.py /path/to/audio/directory --device cuda --batch_size 16
  python evaluate_similarity.py /path/to/audio/directory --batch_size 4 --device cpu
        """
    )
    
    parser.add_argument(
        "data_dir", 
        type=str, 
        help="Path to directory containing subdirectories with gt.wav, mixture.wav, output.wav files"
    )
    parser.add_argument(
        "--output", 
        type=str, 
        default=None,
        help="Output file for results (default: similarity_results.txt in the data directory)"
    )
    parser.add_argument(
        "--device", 
        type=str, 
        default="auto",
        choices=["auto", "cpu", "cuda"],
        help="Device to use (default: auto)"
    )
    parser.add_argument(
        "--batch_size", 
        type=int, 
        default=8,
        help="Batch size for processing (default: 8)"
    )
    parser.add_argument(
        "--verbose", 
        action="store_true",
        help="Enable verbose output"
    )
    
    args = parser.parse_args()
    
    # Print header
    print("🎵 Spatial CLAP Audio Similarity Evaluation")
    print("=" * 50)
    
    # Check dependencies
    if not check_dependencies():
        sys.exit(1)
    
    # Set device
    if args.device == "auto":
        device = "cuda" if torch.cuda.is_available() else "cpu"
    else:
        device = args.device
    
    print(f"Using device: {device}")
    
    # Check if data directory exists
    if not os.path.exists(args.data_dir):
        print(f"❌ Error: Data directory {args.data_dir} does not exist")
        sys.exit(1)
    
    # Set default output path to be in the data directory
    if args.output is None:
        args.output = os.path.join(args.data_dir, "similarity_results.txt")
    
    # Load Spatial CLAP model
    print("🔄 Loading Spatial CLAP model...")
    try:
        model = CLAPEncoder()
        model.load_pretrained()
        model.to(device)
        model.eval()
        print("✅ Model loaded successfully")
    except Exception as e:
        print(f"❌ Error loading model: {str(e)}")
        print("Make sure you're running from the SpatialCLAP directory")
        sys.exit(1)
    
    # Process directories
    print(f"🔄 Processing audio files in: {args.data_dir}")
    print(f"Using batch size: {args.batch_size}")
    results = process_directory(args.data_dir, model, device, args.batch_size)
    
    if not results:
        print("❌ No results generated")
        print("Check that your directory structure contains subdirectories with gt.wav and output.wav files")
        sys.exit(1)
    
    # Save results
    save_results(results, args.output)
    
    # Print summary
    print_summary(results)
    
    print(f"\n✅ Evaluation completed successfully!")
    print(f"📊 Results saved to: {args.output}")


if __name__ == "__main__":
    main()