Blame for scripts/extract_features.py

# /// script
# requires-python = ">=3.10"
# dependencies = [
#   "numpy>=1.26",
#   "pandas>=2",
#   "librosa>=0.10",
#   "torch>=2.4",
#   "transformers>=4.45,<5",
#   "nnAudio>=0.3.3",
#   "laion-clap==1.1.5",
#   "torchvision", "torchaudio",  # used by laion_clap but not declared
#   "huggingface_hub",
# ]
# ///
"""Extract one feature vector per FreeTZAN clip, for scripts/run_baselines.py.

Representations:
  mfcc  20 MFCCs and their deltas (librosa defaults at 22.05 kHz), mean and std over time: 80 dims
  fma   FMA's precomputed librosa feature set (fma_metadata/features.csv; chroma, tonnetz, MFCC,
        spectral and zero-crossing statistics), computed by FMA on the same 30 s clips: 518 dims
  clap  LAION-CLAP audio embedding (laion_clap package, HTSAT-base music checkpoint
        music_audioset_epoch_15_esc_90.14.pt), L2-normalized and averaged over three 10 s windows: 512 dims.
        (The Hugging Face transformers port, laion/larger_clap_music, gives embeddings unrelated to this
        checkpoint's and chance-level zero-shot accuracy on FreeTZAN, so the reference code is used.)
  mert  MERT-v1-95M hidden states, averaged over time, three 10 s windows and all 13 layers: 768 dims

Output: <out-dir>/<name>.npz with arrays track_id (n,) and X (n, d), rows in metadata.csv order.
Existing files are skipped unless --overwrite is given.

Usage: uv run scripts/extract_features.py --fma-dir /path/to/fma [--features mfcc fma clap mert] [--out-dir ...]
"""
import argparse
import subprocess
from concurrent.futures import ProcessPoolExecutor, ThreadPoolExecutor
from pathlib import Path

import numpy as np
import pandas as pd

REPO = Path(__file__).resolve().parent.parent
N_WINDOWS, WINDOW_S = 3, 10
# Pinned model revisions (MERT runs code from its model repository).
CLAP = ('lukewys/laion_clap', 'b3708341862f581175dba5c356a4ebf74a9b6651', 'music_audioset_epoch_15_esc_90.14.pt')
MERT = ('m-a-p/MERT-v1-95M', '12af15fef9d0ac838c3f475bfbbf26d2060dd4f5')


def load_audio(path, sr):
    p = subprocess.run(['ffmpeg', '-nostdin', '-v', 'error', '-i', str(path), '-ac', '1', '-ar', str(sr),
                        '-f', 'f32le', '-'], capture_output=True, check=True)
    return np.frombuffer(p.stdout, np.float32)


def windows(x, sr):
    """N_WINDOWS evenly spaced WINDOW_S windows spanning the clip (clips are 29-30 s long)."""
    w = WINDOW_S * sr
    starts = np.linspace(0, max(len(x) - w, 0), N_WINDOWS).astype(int)
    return [x[s:s + w] for s in starts]


def mfcc_one(path):
    import librosa
    y = load_audio(path, 22050)
    m = librosa.feature.mfcc(y=y, sr=22050, n_mfcc=20)
    f = np.concatenate([m, librosa.feature.delta(m)])
    return np.concatenate([f.mean(1), f.std(1)])


def extract_mfcc(paths, args):
    with ProcessPoolExecutor(args.workers) as ex:
        return np.stack(list(ex.map(mfcc_one, paths, chunksize=8)))


def extract_fma(md, args):
    f = pd.read_csv(Path(args.fma_dir) / 'fma_metadata' / 'features.csv', index_col=0, header=[0, 1, 2])
    missing = set(md.track_id) - set(f.index)
    assert not missing, f'{len(missing)} tracks missing from features.csv'
    return f.loc[md.track_id].to_numpy(np.float32)


def embed_all(paths, sr, embed, args):
    """Decode in background threads while embedding on the GPU."""
    out = []
    with ThreadPoolExecutor(args.workers) as ex:
        for i, x in enumerate(ex.map(lambda p: load_audio(p, sr), paths)):
            out.append(embed(windows(x, sr)))
            if (i + 1) % 200 == 0:
                print(f'  {i + 1}/{len(paths)}', flush=True)
    return np.stack(out)


def extract_clap(paths, args):
    import sys
    import torch
    from huggingface_hub import hf_hub_download
    argv, sys.argv = sys.argv, sys.argv[:1]  # laion_clap parses sys.argv on import
    import laion_clap
    sys.argv = argv
    repo, rev, fname = CLAP
    scalar = (np._core if hasattr(np, '_core') else np.core).multiarray.scalar
    # weights_only load; the checkpoint's optimizer state also contains numpy float64 scalars
    with torch.serialization.safe_globals([(scalar, 'numpy.core.multiarray.scalar'), np.dtype,
                                           type(np.dtype(np.float64))]):
        ck = torch.load(hf_hub_download(repo, fname, revision=rev), map_location='cpu', weights_only=True)
    sd = ck['state_dict']
    sd = {k.removeprefix('module.'): v for k, v in sd.items()}
    sd.pop('text_branch.embeddings.position_ids', None)  # buffer that newer transformers no longer stores
    model = laion_clap.CLAP_Module(enable_fusion=False, amodel='HTSAT-base', device=args.device)
    model.model.load_state_dict(sd, strict=True)
    model.eval()

    @torch.no_grad()
    def embed(wins):
        x = np.stack(wins).clip(-1, 1)
        x = (x * 32767).astype(np.int16).astype(np.float32) / 32767  # 16-bit quantization, as in laion_clap's usage
        return model.get_audio_embedding_from_data(x=x).mean(0)  # windows are L2-normalized by laion_clap

    return embed_all(paths, 48000, embed, args)


def extract_mert(paths, args):
    import torch
    from transformers import AutoModel, Wav2Vec2FeatureExtractor
    name, rev = MERT
    fe = Wav2Vec2FeatureExtractor.from_pretrained(name, revision=rev, trust_remote_code=True)
    model = AutoModel.from_pretrained(name, revision=rev, trust_remote_code=True).to(args.device).eval()

    @torch.no_grad()
    def embed(wins):
        inp = fe(wins, sampling_rate=fe.sampling_rate, return_tensors='pt')
        out = model(**{k: v.to(args.device) for k, v in inp.items()}, output_hidden_states=True)
        return torch.stack(out.hidden_states).mean(dim=(0, 1, 2)).cpu().numpy()  # layers x windows x time

    return embed_all(paths, fe.sampling_rate, embed, args)


EXTRACTORS = {'mfcc': extract_mfcc, 'fma': extract_fma, 'clap': extract_clap, 'mert': extract_mert}


def main():
    ap = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
    ap.add_argument('--features', nargs='+', default=list(EXTRACTORS), choices=list(EXTRACTORS))
    ap.add_argument('--fma-dir', help='FMA root with fma_metadata/; needed for "fma"')
    ap.add_argument('--out-dir', default=str(Path.home() / '.cache' / 'freetzan' / 'features'))
    ap.add_argument('--device', default='cuda')
    ap.add_argument('--workers', type=int, default=16)
    ap.add_argument('--overwrite', action='store_true')
    args = ap.parse_args()

    if 'fma' in args.features and not args.fma_dir:
        ap.error('--fma-dir is required for the "fma" features')
    md = pd.read_csv(REPO / 'metadata.csv')
    paths = [REPO / p for p in md.path]
    out_dir = Path(args.out_dir)
    out_dir.mkdir(parents=True, exist_ok=True)
    for name in args.features:
        dst = out_dir / f'{name}.npz'
        if dst.exists() and not args.overwrite:
            print(f'{name}: exists, skipping ({dst})')
            continue
        print(f'{name}: extracting', flush=True)
        X = EXTRACTORS[name](md if name == 'fma' else paths, args)
        assert X.shape[0] == len(md) and np.isfinite(X).all(), name
        np.savez(dst, track_id=md.track_id.to_numpy(), X=X.astype(np.float32))
        print(f'{name}: {X.shape} -> {dst}', flush=True)


if __name__ == '__main__':
    main()