features/extractor.py

"""Feature extraction into a compact, streamable binary store.

Layout (written under ``features/``):

* ``train.bin`` — a flat little-endian float32 blob. Audio samples and the
  flattened mel spectrogram for every clip are appended back to back.
* ``train.dat`` — a JSONL index/manifest. One record per clip giving byte
  offsets and shapes into ``train.bin`` plus ``duration`` and ``sample_rate``.

The writer streams one clip at a time, so memory usage is independent of the
dataset size (works for datasets far larger than RAM).
"""

from __future__ import annotations

import json
from dataclasses import dataclass
from pathlib import Path
from typing import Iterable, Optional

import numpy as np

from video_vec2wav2_tokenizer.audio.extractor import load_wav
from video_vec2wav2_tokenizer.utils.io import ensure_dir
from video_vec2wav2_tokenizer.utils.logging import get_logger

logger = get_logger(__name__)

_FLOAT_BYTES = 4  # float32


def compute_mel_spectrogram(
    samples: np.ndarray,
    sample_rate: int,
    *,
    n_mels: int = 80,
    n_fft: int = 400,
    hop_length: int = 160,
) -> np.ndarray:
    """Return a log-mel spectrogram of shape ``(n_mels, n_frames)`` (float32).

    Uses librosa when available; otherwise falls back to a NumPy STFT + a
    triangular mel filterbank so the feature path has no hard dependency.
    """
    try:
        import librosa  # type: ignore

        mel = librosa.feature.melspectrogram(
            y=samples, sr=sample_rate, n_mels=n_mels, n_fft=n_fft, hop_length=hop_length
        )
        log_mel = librosa.power_to_db(mel, ref=np.max)
        return log_mel.astype(np.float32)
    except ImportError:
        return _numpy_log_mel(
            samples, sample_rate, n_mels=n_mels, n_fft=n_fft, hop_length=hop_length
        )


def _hz_to_mel(hz: np.ndarray) -> np.ndarray:
    return 2595.0 * np.log10(1.0 + hz / 700.0)


def _mel_to_hz(mel: np.ndarray) -> np.ndarray:
    return 700.0 * (10.0 ** (mel / 2595.0) - 1.0)


def _mel_filterbank(sr: int, n_fft: int, n_mels: int) -> np.ndarray:
    n_freqs = n_fft // 2 + 1
    fft_freqs = np.linspace(0, sr / 2, n_freqs)
    mel_min, mel_max = _hz_to_mel(np.array([0.0])), _hz_to_mel(np.array([sr / 2]))
    mel_points = np.linspace(mel_min[0], mel_max[0], n_mels + 2)
    hz_points = _mel_to_hz(mel_points)

    fb = np.zeros((n_mels, n_freqs), dtype=np.float32)
    for m in range(1, n_mels + 1):
        f_left, f_center, f_right = hz_points[m - 1], hz_points[m], hz_points[m + 1]
        left = (fft_freqs - f_left) / max(f_center - f_left, 1e-9)
        right = (f_right - fft_freqs) / max(f_right - f_center, 1e-9)
        fb[m - 1] = np.clip(np.minimum(left, right), 0.0, None)
    return fb


def _numpy_log_mel(
    samples: np.ndarray, sr: int, *, n_mels: int, n_fft: int, hop_length: int
) -> np.ndarray:
    if samples.size < n_fft:
        samples = np.pad(samples, (0, n_fft - samples.size))
    window = np.hanning(n_fft).astype(np.float32)
    n_frames = 1 + (len(samples) - n_fft) // hop_length
    stft = np.empty((n_fft // 2 + 1, n_frames), dtype=np.complex64)
    for i in range(n_frames):
        frame = samples[i * hop_length : i * hop_length + n_fft] * window
        stft[:, i] = np.fft.rfft(frame)
    power = (np.abs(stft) ** 2).astype(np.float32)
    fb = _mel_filterbank(sr, n_fft, n_mels)
    mel = fb @ power
    log_mel = 10.0 * np.log10(np.maximum(mel, 1e-10))
    return log_mel.astype(np.float32)


@dataclass(slots=True)
class FeatureExtractor:
    """Stream clips into ``train.bin`` / ``train.dat``."""

    sample_rate: int = 16000
    n_mels: int = 80
    n_fft: int = 400
    hop_length: int = 160
    split: str = "train"

    def extract_dataset(
        self,
        clip_paths: Iterable[str | Path],
        features_dir: str | Path,
    ) -> Path:
        """Write features for all ``clip_paths``. Returns the manifest path."""
        features_dir = ensure_dir(features_dir)
        bin_path = features_dir / f"{self.split}.bin"
        dat_path = features_dir / f"{self.split}.dat"

        offset = 0
        n = 0
        with bin_path.open("wb") as bin_fh, dat_path.open("w", encoding="utf-8") as dat_fh:
            for clip_path in clip_paths:
                clip_path = Path(clip_path)
                samples, sr = load_wav(clip_path, sample_rate=self.sample_rate)
                mel = compute_mel_spectrogram(
                    samples, sr, n_mels=self.n_mels, n_fft=self.n_fft, hop_length=self.hop_length
                )

                audio_f32 = samples.astype("<f4", copy=False)
                mel_f32 = mel.astype("<f4", copy=False)

                audio_bytes = audio_f32.tobytes()
                mel_bytes = mel_f32.tobytes()
                bin_fh.write(audio_bytes)
                bin_fh.write(mel_bytes)

                audio_offset = offset
                mel_offset = offset + len(audio_bytes)
                record = {
                    "id": clip_path.stem,
                    "audio": clip_path.name,
                    "sample_rate": sr,
                    "duration": round(samples.shape[0] / sr, 4),
                    "audio_offset": audio_offset,
                    "audio_length": int(audio_f32.size),
                    "mel_offset": mel_offset,
                    "mel_shape": list(mel_f32.shape),
                    "dtype": "float32",
                }
                dat_fh.write(json.dumps(record))
                dat_fh.write("\n")

                offset = mel_offset + len(mel_bytes)
                n += 1

        logger.info("Extracted features for %d clips -> %s", n, bin_path)
        return dat_path


def read_feature_record(
    bin_path: str | Path, record: dict
) -> tuple[np.ndarray, np.ndarray]:
    """Read one clip's ``(audio, mel)`` arrays back from the binary store."""
    with Path(bin_path).open("rb") as fh:
        fh.seek(record["audio_offset"])
        audio = np.frombuffer(
            fh.read(record["audio_length"] * _FLOAT_BYTES), dtype="<f4"
        ).copy()
        fh.seek(record["mel_offset"])
        mel_count = int(np.prod(record["mel_shape"]))
        mel = np.frombuffer(
            fh.read(mel_count * _FLOAT_BYTES), dtype="<f4"
        ).reshape(record["mel_shape"]).copy()
    return audio, mel