[feat] Add performance and separation quality benchmarks

Add benchmarking for PyTorch Demucs and C++ ONNX scripts -
 - Separation quality using SI-SDR
 - Performance metrics on CPU

Signed-off-by: Anmol Mishra <anmolmishra1997@gmail.com>

Author: dhunstack

benchmark/benchmark-cpp-onnx.py

@@ -0,0 +1,65 @@
+"""
+C++ ONNX-based Demucs benchmarking script.
+
+This script uses C++ ONNX CLI tool for inference and imports common functionality
+from benchmark_common to avoid code duplication.
+"""
+
+import argparse
+import subprocess
+from pathlib import Path
+
+from benchmark_common import run_benchmark
+
+DEFAULT_CLI_PATH = '../cppscripts/build/build-cli/demucs'
+DEFAULT_MODEL_PATH = '../onnx-models/htdemucs.ort'
+
+
+def separate_cpp_onnx(mixture_path, out_dir, cli_path, model_path):
+    """Separate audio using C++ ONNX CLI tool."""
+    out_dir = Path(out_dir)
+    out_dir.mkdir(parents=True, exist_ok=True)
+    
+    cmd = [str(cli_path), str(model_path), str(mixture_path), str(out_dir)]
+    subprocess.run(cmd, check=True)
+
+
+def main():
+    parser = argparse.ArgumentParser(description='Benchmark Demucs separation using C++ ONNX CLI')
+    parser.add_argument('--musdb-root', type=Path, help='Path to MusDB root', required=True)
+    parser.add_argument('--output-root', type=Path, default=None, help='Where to store separated outputs (default: inside musdb-root)')
+    parser.add_argument('--output-dir', type=str, default='test-separated-cpp', help='Output directory name (default: test-separated-cpp)')
+    parser.add_argument('--json-out', type=str, default='benchmark_results_cpp.json', help='Output JSON file for benchmarks')
+    parser.add_argument('--cli-path', type=str, default=DEFAULT_CLI_PATH, help='Path to the C++ ONNX CLI executable')
+    parser.add_argument('--model-path', type=str, default=DEFAULT_MODEL_PATH, help='Path to the ONNX model file')
+    parser.add_argument('--force-reseparate', action='store_true', help='Force re-separation even if files already exist')
+    args = parser.parse_args()
+    
+    # Setup paths
+    musdb_root = Path(args.musdb_root)
+    output_root = Path(args.output_root) if args.output_root else musdb_root
+    out_dir = output_root / args.output_dir
+    out_dir.mkdir(parents=True, exist_ok=True)
+    
+    cli_path = Path(args.cli_path)
+    model_path = Path(args.model_path)
+    
+    print(f'Using CLI: {cli_path}')
+    print(f'Using model: {model_path}')
+
+    # Run benchmark using common flow
+    run_benchmark(
+        musdb_root=musdb_root,
+        out_dir=out_dir,
+        force_reseparate=args.force_reseparate,
+        json_out=args.json_out,
+        separate_func=separate_cpp_onnx,
+        model_identifier=str(model_path),
+        # Arguments passed to separate_cpp_onnx
+        cli_path=cli_path,
+        model_path=model_path
+    )
+
+
+if __name__ == '__main__':
+    main()

benchmark/benchmark-pytorch.py

@@ -0,0 +1,82 @@
+"""
+PyTorch-based Demucs benchmarking script.
+
+This script uses PyTorch for inference and imports common functionality
+from benchmark_common to avoid code duplication.
+"""
+
+import argparse
+from pathlib import Path
+
+import torchaudio
+from demucs.apply import apply_model
+from demucs.pretrained import get_model
+
+from benchmark_common import run_benchmark, STEM_MAP, STEM_NAMES
+
+DEFAULT_MODEL = 'htdemucs'
+
+
+def separate_pytorch(mixture_path, out_dir, model_name):
+    """Separate audio using PyTorch Demucs model."""
+    # Load model
+    model = get_model(model_name)
+    
+    # Load and preprocess audio
+    audio, rate = torchaudio.load(str(mixture_path))
+    if rate != 44100:
+        audio = torchaudio.functional.resample(audio, rate, 44100)
+    
+    # Normalize
+    ref = audio.mean(0)
+    audio = (audio - ref.mean()) / ref.std()
+    
+    # Apply model
+    sources = apply_model(model, audio[None])[0]
+    
+    # Denormalize
+    sources = sources * ref.std() + ref.mean()
+    
+    # Save stems
+    out_dir = Path(out_dir)
+    out_dir.mkdir(parents=True, exist_ok=True)
+    for target_idx in range(len(STEM_NAMES)):
+        target_name = STEM_MAP[target_idx]
+        out_audio = sources[target_idx].detach().cpu()
+        out_path = out_dir / f'target_{target_idx}_{target_name}.wav'
+        torchaudio.save(str(out_path), out_audio, sample_rate=44100)
+
+
+def main():
+    parser = argparse.ArgumentParser(description='PyTorch Demucs Benchmarking Script')
+    parser.add_argument('--musdb-root', type=Path, help='Path to MusDB root', required=True)
+    parser.add_argument('--output-root', type=Path, default=None, help='Where to store separated outputs (default: inside musdb-root)')
+    parser.add_argument('--output-dir', type=str, default='test-separated-pytorch', help='Output directory name (default: test-separated-pytorch)')
+    parser.add_argument('--json-out', type=str, default='benchmark_results_pytorch.json', help='Output JSON file for benchmarks')
+    parser.add_argument('--force-reseparate', action='store_true', help='Force re-separation even if files already exist')
+    args = parser.parse_args()
+
+    # Setup paths
+    musdb_root = Path(args.musdb_root)
+    output_root = Path(args.output_root) if args.output_root else musdb_root
+    out_dir = output_root / args.output_dir
+    out_dir.mkdir(parents=True, exist_ok=True)
+    
+    model_name = DEFAULT_MODEL
+    print(f'Using model: {model_name}')
+
+    # Run benchmark using common flow
+    run_benchmark(
+        musdb_root=musdb_root,
+        out_dir=out_dir,
+        force_reseparate=args.force_reseparate,
+        json_out=args.json_out,
+        separate_func=separate_pytorch,
+        model_identifier=model_name,
+        # Arguments passed to separate_pytorch
+        model_name=model_name
+    )
+
+
+if __name__ == '__main__':
+    main()