scripts/generation/generate_hybrid_sets.py

#!/usr/bin/env python3
# pyright: reportMissingTypeStubs=false, reportUnknownMemberType=false, reportUnknownVariableType=false, reportUnusedCallResult=false, reportDeprecated=false, reportUnknownArgumentType=false
import argparse
import hashlib
import json
import logging
import random
import subprocess
import sys
import tempfile
from datetime import datetime
from pathlib import Path
from collections.abc import Mapping
from typing import cast

import librosa
import numpy as np
import soundfile as sf
from tqdm import tqdm


SEED = 42
SR = 44100
MAX_DURATION = 120
MP3_BITRATE = "192k"

BASE = Path("/ssd_data/nhn_backup/data/datasets/ai_detection_dataset")

AI_SOURCE_DIRS = [
    BASE / "fake" / "suno_v4",
    BASE / "fake" / "suno_v4_5",
    BASE / "fake" / "suno_v5",
    BASE / "fake" / "suno_v4_5_plus",
    BASE / "fake" / "suno_other",
]

HUMAN_SOURCE_DIRS = {
    "sonics": BASE / "real" / "Sonics_real",
    "mtg": BASE / "real" / "MTG",
}

OUTPUTS = {
    "B2": {"sonics": "B2_suno_sonics", "mtg": "B2_suno_mtg"},
    "B3": {"sonics": "B3_suno_sonics", "mtg": "B3_suno_mtg"},
    "B4": {"sonics": "B4_suno_sonics", "mtg": "B4_suno_mtg"},
    "B5": {"sonics": "B5_suno_sonics", "mtg": "B5_suno_mtg"},
    "B7": {"sonics": "B7_suno_sonics", "mtg": "B7_suno_mtg"},
    "B8": {"sonics": "B8_suno_sonics", "mtg": "B8_suno_mtg"},
    "C1": {"sonics": "C1_suno_sonics", "mtg": "C1_suno_mtg"},
    "C2": {"sonics": "C2_suno_sonics", "mtg": "C2_suno_mtg"},
}

B2_FILTER = "acompressor=threshold=-20dB:ratio=4:attack=5:release=50,equalizer=f=100:t=h:w=200:g=3,equalizer=f=3000:t=h:w=2000:g=2,loudnorm=I=-14:LRA=11:TP=-1"
B3_FILTER = "acompressor=threshold=-25dB:ratio=2:attack=10:release=100,equalizer=f=80:t=h:w=100:g=2,loudnorm=I=-16:LRA=9:TP=-2"
B4_FILTER = "atempo=0.95,asetrate=44100*1.02,aresample=44100,aecho=0.8:0.88:60:0.4"
B5_FILTERS = [
    "atempo=1.05,asetrate=44100*0.98,aresample=44100",
    "atempo=0.92,highpass=f=80,lowpass=f=12000",
    "chorus=0.5:0.9:50:0.4:0.25:2",
    "flanger=delay=5:depth=2:speed=0.5",
]

ALLOWED_EXTS = {".mp3", ".wav", ".flac", ".m4a", ".ogg"}


def configure_logger(category: str, source_set: str) -> logging.Logger:
    logger_name = f"hybrid_{category}_{source_set}"
    logger = logging.getLogger(logger_name)
    logger.setLevel(logging.INFO)
    logger.propagate = False

    if logger.handlers:
        return logger

    formatter = logging.Formatter("%(asctime)s | %(levelname)s | %(message)s")
    stream_handler = logging.StreamHandler(sys.stdout)
    stream_handler.setFormatter(formatter)

    log_path = BASE / f"{category}_{source_set}.log"
    file_handler = logging.FileHandler(log_path, encoding="utf-8")
    file_handler.setFormatter(formatter)

    logger.addHandler(stream_handler)
    logger.addHandler(file_handler)
    return logger


def stable_seed(category: str, source_set: str) -> int:
    digest = hashlib.md5(f"{category}:{source_set}".encode("utf-8")).hexdigest()
    return SEED + int(digest[:8], 16)


def discover_audio_files(directories: list[Path]) -> list[Path]:
    files: list[Path] = []
    for directory in directories:
        if not directory.exists():
            continue
        for path in directory.rglob("*"):
            if path.is_file() and path.suffix.lower() in ALLOWED_EXTS:
                files.append(path)
    return sorted(files)


def run_command(
    cmd: list[str], logger: logging.Logger, timeout: int = 1800
) -> tuple[bool, str]:
    try:
        completed = subprocess.run(
            cmd,
            check=False,
            capture_output=True,
            text=True,
            timeout=timeout,
        )
    except Exception as exc:  # noqa: BLE001
        return False, str(exc)

    if completed.returncode != 0:
        message = (
            completed.stderr.strip()
            or completed.stdout.strip()
            or "unknown command failure"
        )
        logger.error("Command failed: %s", " ".join(cmd))
        logger.error("Error: %s", message)
        return False, message

    return True, completed.stdout.strip()


def encode_wav_to_mp3(wav_path: Path, output_mp3: Path, logger: logging.Logger) -> bool:
    cmd = [
        "ffmpeg",
        "-y",
        "-hide_banner",
        "-loglevel",
        "error",
        "-i",
        str(wav_path),
        "-t",
        str(MAX_DURATION),
        "-ar",
        str(SR),
        "-ac",
        "2",
        "-b:a",
        MP3_BITRATE,
        str(output_mp3),
    ]
    ok, _ = run_command(cmd, logger)
    return ok and output_mp3.exists() and output_mp3.stat().st_size > 0


def ffmpeg_transform(
    input_path: Path, output_path: Path, af_filter: str, logger: logging.Logger
) -> bool:
    cmd = [
        "ffmpeg",
        "-y",
        "-hide_banner",
        "-loglevel",
        "error",
        "-i",
        str(input_path),
        "-t",
        str(MAX_DURATION),
        "-af",
        af_filter,
        "-ar",
        str(SR),
        "-ac",
        "2",
        "-b:a",
        MP3_BITRATE,
        str(output_path),
    ]
    ok, _ = run_command(cmd, logger)
    return ok and output_path.exists() and output_path.stat().st_size > 0


def trim_for_demucs(
    input_path: Path, trimmed_wav: Path, logger: logging.Logger
) -> bool:
    cmd = [
        "ffmpeg",
        "-y",
        "-hide_banner",
        "-loglevel",
        "error",
        "-i",
        str(input_path),
        "-t",
        str(MAX_DURATION),
        "-ar",
        str(SR),
        "-ac",
        "2",
        "-vn",
        "-c:a",
        "pcm_s16le",
        str(trimmed_wav),
    ]
    ok, _ = run_command(cmd, logger, timeout=300)
    return ok and trimmed_wav.exists() and trimmed_wav.stat().st_size > 0


def separate_stems_cached(
    src_audio: Path,
    cache_root: Path,
    logger: logging.Logger,
) -> tuple[Path | None, Path | None]:
    cache_key = hashlib.sha1(str(src_audio).encode("utf-8")).hexdigest()[:16]
    sample_dir = cache_root / cache_key
    stem_root = sample_dir / "htdemucs"

    vocals = None
    no_vocals = None
    if stem_root.exists():
        stem_subdirs = [p for p in stem_root.iterdir() if p.is_dir()]
        if stem_subdirs:
            candidate = stem_subdirs[0]
            vocals = candidate / "vocals.wav"
            no_vocals = candidate / "no_vocals.wav"
            if vocals.exists() and no_vocals.exists():
                return vocals, no_vocals

    sample_dir.mkdir(parents=True, exist_ok=True)
    with tempfile.TemporaryDirectory(prefix="demucs_clip_") as tmp_dir:
        clipped = Path(tmp_dir) / "clip.wav"
        if not trim_for_demucs(src_audio, clipped, logger):
            return None, None

        cmd_base = [
            sys.executable,
            "-m",
            "demucs",
            "--two-stems",
            "vocals",
            "-o",
            str(sample_dir),
            str(clipped),
        ]

        ok, _ = run_command(cmd_base + ["--device", "cuda"], logger, timeout=3600)
        if not ok:
            logger.info("Demucs cuda failed; retrying on cpu for %s", src_audio.name)
            ok, _ = run_command(cmd_base + ["--device", "cpu"], logger, timeout=3600)
            if not ok:
                return None, None

    if not stem_root.exists():
        return None, None

    stem_subdirs = [p for p in stem_root.iterdir() if p.is_dir()]
    if not stem_subdirs:
        return None, None

    stem_dir = stem_subdirs[0]
    vocals = stem_dir / "vocals.wav"
    no_vocals = stem_dir / "no_vocals.wav"

    if not vocals.exists() or not no_vocals.exists():
        return None, None
    return vocals, no_vocals


def load_audio(path: Path) -> np.ndarray:
    audio, _ = librosa.load(str(path), sr=SR, mono=False, duration=MAX_DURATION)
    if audio.ndim == 1:
        audio = np.stack([audio, audio], axis=0)
    return audio


def ensure_stereo(samples: np.ndarray) -> np.ndarray:
    if samples.ndim == 1:
        return np.stack([samples, samples], axis=1)
    if samples.shape[1] == 1:
        return np.repeat(samples, 2, axis=1)
    return samples


def mix_stems(
    vocal_path: Path, instrumental_path: Path
) -> tuple[np.ndarray, int] | tuple[None, None]:
    vocal, sr_v = sf.read(str(vocal_path), always_2d=True)
    inst, sr_i = sf.read(str(instrumental_path), always_2d=True)

    if sr_v != sr_i:
        return None, None

    vocal = ensure_stereo(cast(np.ndarray, vocal))
    inst = ensure_stereo(cast(np.ndarray, inst))
    min_len = min(len(vocal), len(inst), int(SR * MAX_DURATION))
    if min_len <= 0:
        return None, None

    mixed = vocal[:min_len] * 1.0 + inst[:min_len] * 0.8
    peak = np.max(np.abs(mixed))
    if peak > 0.95:
        mixed = mixed * (0.95 / peak)

    return mixed, sr_v


def concat_mix(
    human_audio: np.ndarray, ai_audio: np.ndarray, rng: random.Random
) -> tuple[np.ndarray, dict[str, str | float]]:
    split_ratio = rng.uniform(0.3, 0.7)
    h_len = int(human_audio.shape[1] * split_ratio)
    a_len = int(ai_audio.shape[1] * (1 - split_ratio))
    ai_first = rng.random() > 0.5
    if not ai_first:
        mixed = np.concatenate([human_audio[:, :h_len], ai_audio[:, :a_len]], axis=1)
        human_start = 0.0
        human_end = h_len / SR
        ai_start = h_len / SR
        ai_end = (h_len + a_len) / SR
    else:
        mixed = np.concatenate([ai_audio[:, :a_len], human_audio[:, :h_len]], axis=1)
        ai_start = 0.0
        ai_end = a_len / SR
        human_start = a_len / SR
        human_end = (a_len + h_len) / SR

    mixed = mixed[:, : int(SR * MAX_DURATION)]
    mixed_duration = mixed.shape[1] / SR
    info: dict[str, str | float] = {
        "order": "ai_first" if ai_first else "human_first",
        "human_start_sec": round(min(human_start, mixed_duration), 3),
        "human_end_sec": round(min(human_end, mixed_duration), 3),
        "ai_start_sec": round(min(ai_start, mixed_duration), 3),
        "ai_end_sec": round(min(ai_end, mixed_duration), 3),
    }
    return mixed, info


def crossfade_mix(
    human_audio: np.ndarray, ai_audio: np.ndarray, rng: random.Random
) -> tuple[np.ndarray, dict[str, str | float]]:
    fade_samples = int(3.0 * SR)
    h_len = human_audio.shape[1]
    a_len = ai_audio.shape[1]

    if h_len < fade_samples or a_len < fade_samples:
        fade_samples = max(1, min(h_len, a_len) // 2)

    split_point = rng.randint(int(h_len * 0.3), max(int(h_len * 0.7), 1))
    split_point = min(max(split_point, fade_samples), h_len)

    part1 = human_audio[:, :split_point]
    part2 = ai_audio

    fade_out = np.linspace(1.0, 0.0, fade_samples)
    fade_in = np.linspace(0.0, 1.0, fade_samples)

    overlap = part1[:, -fade_samples:] * fade_out + part2[:, :fade_samples] * fade_in
    mixed = np.concatenate(
        [
            part1[:, :-fade_samples],
            overlap,
            part2[:, fade_samples:],
        ],
        axis=1,
    )
    mixed = mixed[:, : int(SR * MAX_DURATION)]
    peak = np.max(np.abs(mixed)) if mixed.size else 0.0
    if peak > 0.95:
        mixed = mixed * (0.95 / peak)
    mixed_duration = mixed.shape[1] / SR
    crossfade_start = (split_point - fade_samples) / SR
    crossfade_end = split_point / SR
    info: dict[str, str | float] = {
        "order": "human_first",
        "human_only_start_sec": 0.0,
        "human_only_end_sec": round(min(crossfade_start, mixed_duration), 3),
        "crossfade_start_sec": round(min(crossfade_start, mixed_duration), 3),
        "crossfade_end_sec": round(min(crossfade_end, mixed_duration), 3),
        "crossfade_duration_sec": round(fade_samples / SR, 3),
        "ai_only_start_sec": round(min(crossfade_end, mixed_duration), 3),
        "ai_only_end_sec": round(mixed_duration, 3),
    }
    return mixed, info


def write_metadata(
    output_mp3: Path, record: Mapping[str, str | float], metadata_jsonl: Path
) -> None:
    sidecar = output_mp3.with_suffix(".json")
    sidecar.write_text(
        json.dumps(record, ensure_ascii=False, indent=2), encoding="utf-8"
    )
    with metadata_jsonl.open("a", encoding="utf-8") as handle:
        handle.write(json.dumps(record, ensure_ascii=False) + "\n")


def generate_track(
    category: str,
    ai_file: Path,
    human_file: Path,
    output_mp3: Path,
    rng: random.Random,
    demucs_cache: Path,
    logger: logging.Logger,
) -> tuple[bool, dict[str, str | float] | None]:
    try:
        if category == "B2":
            return ffmpeg_transform(human_file, output_mp3, B2_FILTER, logger), None

        if category == "B3":
            return ffmpeg_transform(ai_file, output_mp3, B3_FILTER, logger), None

        if category == "B4":
            return ffmpeg_transform(ai_file, output_mp3, B4_FILTER, logger), None

        if category == "B5":
            af_filter = rng.choice(B5_FILTERS)
            return ffmpeg_transform(ai_file, output_mp3, af_filter, logger), None

        if category in {"B7", "B8"}:
            logger.warning("B7/B8 use YouTube crawling, not generate_track()")
            return False, None

        if category in {"C1", "C2"}:
            human_audio = load_audio(human_file)
            ai_audio = load_audio(ai_file)
            if category == "C1":
                mixed, mix_info = concat_mix(human_audio, ai_audio, rng)
            else:
                mixed, mix_info = crossfade_mix(human_audio, ai_audio, rng)

            if mixed.size == 0:
                return False, None

            with tempfile.TemporaryDirectory(prefix="mix_c1c2_") as tmp_dir:
                temp_wav = Path(tmp_dir) / "mixed.wav"
                sf.write(str(temp_wav), mixed.T, SR)
                ok = encode_wav_to_mp3(temp_wav, output_mp3, logger)
                return ok, mix_info if ok else None

        raise ValueError(f"Unsupported category: {category}")

    except Exception as exc:  # noqa: BLE001
        logger.exception(
            "Track generation failed for %s (%s): %s", output_mp3.name, category, exc
        )
        return False, None


def process_category_source(category: str, source_set: str, target: int) -> None:
    logger = configure_logger(category, source_set)
    rng = random.Random(stable_seed(category, source_set))

    ai_files = discover_audio_files(AI_SOURCE_DIRS)
    human_files = discover_audio_files([HUMAN_SOURCE_DIRS[source_set]])
    rng.shuffle(ai_files)
    rng.shuffle(human_files)

    if not ai_files or not human_files:
        logger.error(
            "No source files found for category=%s source_set=%s", category, source_set
        )
        return

    output_dir = BASE / "fake" / OUTPUTS[category][source_set]
    output_dir.mkdir(parents=True, exist_ok=True)
    metadata_jsonl = output_dir / "metadata.jsonl"
    demucs_cache = BASE / "demucs_separated_hybrid" / source_set
    demucs_cache.mkdir(parents=True, exist_ok=True)

    existing = [p for p in output_dir.glob("*.mp3") if p.stat().st_size > 0]
    if len(existing) >= target:
        logger.info(
            "%s/%s already has %d files (target=%d), skipping",
            category,
            source_set,
            len(existing),
            target,
        )
        return

    logger.info(
        "Start %s/%s | ai=%d human=%d existing=%d target=%d",
        category,
        source_set,
        len(ai_files),
        len(human_files),
        len(existing),
        target,
    )

    required = target - len(existing)
    created = 0
    attempts = 0
    file_index = 0
    max_attempts = max(required * 20, required + 100)

    progress = tqdm(total=required, desc=f"{category}_{source_set}", unit="track")

    while created < required and attempts < max_attempts:
        output_name = f"{category.lower()}_{source_set}_{file_index:05d}.mp3"
        output_mp3 = output_dir / output_name
        file_index += 1

        if output_mp3.exists() and output_mp3.stat().st_size > 0:
            continue

        idx = attempts % min(len(ai_files), len(human_files))
        ai_file = ai_files[idx % len(ai_files)]
        human_file = human_files[idx % len(human_files)]
        attempts += 1

        ok, mix_info = generate_track(
            category=category,
            ai_file=ai_file,
            human_file=human_file,
            output_mp3=output_mp3,
            rng=rng,
            demucs_cache=demucs_cache,
            logger=logger,
        )
        if not ok:
            if output_mp3.exists() and output_mp3.stat().st_size == 0:
                output_mp3.unlink(missing_ok=True)
            continue

        if not output_mp3.exists() or output_mp3.stat().st_size == 0:
            logger.error("Generated output is missing or empty: %s", output_mp3)
            if output_mp3.exists():
                output_mp3.unlink(missing_ok=True)
            continue

        record: dict[str, str | float] = {
            "source_ai": ai_file.name,
            "source_human": human_file.name,
            "processing_method": category,
            "source_set": source_set,
            "timestamp": datetime.now().isoformat(),
            "output_filename": output_mp3.name,
        }
        if mix_info:
            record.update(mix_info)
        write_metadata(output_mp3, record, metadata_jsonl)

        created += 1
        progress.update(1)

    progress.close()
    logger.info(
        "Finished %s/%s | created=%d required=%d attempts=%d",
        category,
        source_set,
        created,
        required,
        attempts,
    )


def parse_args() -> argparse.Namespace:
    parser = argparse.ArgumentParser(
        description="Generate B-hybrid and C-mixing datasets."
    )
    parser.add_argument(
        "--category",
        type=str,
        default="all",
        choices=["B2", "B3", "B4", "B5", "B7", "B8", "C1", "C2", "all"],
        help="Category to generate",
    )
    parser.add_argument(
        "--source-set",
        type=str,
        default="all",
        choices=["sonics", "mtg", "all"],
        help="Human source set",
    )
    parser.add_argument(
        "--target",
        type=int,
        default=2000,
        help="Target number of tracks per output folder",
    )
    return parser.parse_args()


def main() -> None:
    random.seed(SEED)
    args = parse_args()

    categories = (
        [args.category]
        if args.category != "all"
        else ["B2", "B3", "B4", "B5", "B7", "B8", "C1", "C2"]
    )
    source_sets = [args.source_set] if args.source_set != "all" else ["sonics", "mtg"]

    for category in categories:
        for source_set in source_sets:
            process_category_source(category, source_set, args.target)


if __name__ == "__main__":
    main()