convert2jsonl.py

# data/Chords1217/convert2jsonl.py
"""
Use torchaudio to uniformly load audio files and extract metadata, converting
the contents of the audio/ and chordlab/ folders in the Chords1217 dataset into JSONL format.
Additionally, this version uses multiprocessing + a progress bar and
splits the final records into train/val/test (60/20/20) with seed=42.

Expected directory structure:
data/Chords1217/
    ├── audio/
    │     ├── xxx.wav
    │     ├── yyy.mp3
    │     └── ...
    └── chordlab/
          ├── xxx.lab
          ├── yyy.lab
          └── ...

One .lab file may look like this:
0.0	0.21500000000000002	N
0.21500000000000002	1.6290000000000002	F#:min7
1.629	2.966	A:maj
2.966	4.287000000000001	E:sus4
4.287	5.709	B:sus4(b7)
5.7090000000000005	7.1370000000000005	F#:min7
"""

import os
import json
import random
from functools import partial
from multiprocessing import Pool, cpu_count

import torchaudio
from tqdm import tqdm


def write_jsonl(data, output_file):
    """
    Write a list of dictionaries to a JSONL file (one JSON object per line).
    """
    with open(output_file, 'w', encoding='utf-8') as f:
        for item in data:
            json.dump(item, f, ensure_ascii=False)
            f.write('\n')


def get_audio_info(audio_path):
    """
    Use torchaudio.info to retrieve audio metadata:
      - duration: seconds (num_frames / sample_rate)
      - sample_rate: Hz
      - num_samples: total number of frames
      - bit_depth: bits per sample (may be None for some formats)
      - channels: number of channels

    torchaudio supports multiple formats (wav, mp3, flac, ogg, etc.) provided
    that the backend (sox/ffmpeg) is correctly installed.
    """
    info = torchaudio.info(audio_path)
    sample_rate = info.sample_rate
    num_samples = info.num_frames
    channels = info.num_channels

    # bits_per_sample may be None for certain file types
    bit_depth = info.bits_per_sample if hasattr(info, 'bits_per_sample') else None

    duration = num_samples / sample_rate if sample_rate else 0.0

    return duration, sample_rate, num_samples, bit_depth, channels


def parse_lab_file(lab_path):
    """
    Parse a .lab file and return a list of chord segments:
    [
        {"start_time": float, "end_time": float, "chord_str": str},
        ...
    ]
    Each line in the .lab file is expected to be:
        start_time <tab> end_time <tab> chord_str
    """
    chord_seq = []
    with open(lab_path, 'r', encoding='utf-8') as f:
        for line in f:
            line = line.strip()
            if not line:
                continue
            parts = line.split()
            if len(parts) < 3:
                # Skip lines that don't have at least 3 columns
                continue
            try:
                start_time = float(parts[0])
                end_time = float(parts[1])
                chord_str = parts[2]
            except ValueError:
                # Skip lines with invalid floats
                continue
            chord_seq.append({
                "start_time": start_time,
                "end_time": end_time,
                "chord_str": chord_str
            })
    return chord_seq


def process_single_file(fname, audio_dir, lab_dir):
    """
    Given a filename (e.g. "xxx.wav"), parse its corresponding .lab file,
    extract audio metadata, and return a record dict. If anything fails,
    return None to indicate skipping.
    """
    ext = os.path.splitext(fname)[1].lower()
    if ext not in ['.wav', '.mp3', '.flac', '.ogg', '.aac', '.m4a']:
        return None

    audio_path = os.path.join(audio_dir, fname)
    base_name = os.path.splitext(fname)[0]
    lab_path = os.path.join(lab_dir, base_name + '.lab')

    if not os.path.isfile(lab_path):
        # No corresponding .lab file → skip
        return None

    # 1. Parse the .lab file to get chord sequence
    chord_seq = parse_lab_file(lab_path)
    if not chord_seq:
        # If lab parsing yields nothing, still include an empty list
        chord_seq = []

    # 2. Get audio metadata using torchaudio
    try:
        duration, sample_rate, num_samples, bit_depth, channels = get_audio_info(audio_path)
    except Exception:
        # Failed to read audio file → skip
        return None

    # 3. Build the record dictionary
    record = {
        "audio_path": audio_path,
        "label": chord_seq,
        "duration": duration,
        "sample_rate": sample_rate,
        "num_samples": num_samples,
        "bit_depth": bit_depth,
        "channels": channels
    }
    return record


def convert_chords_to_jsonl(root_dir):
    """
    Traverse the audio/ and chordlab/ subdirectories under root_dir and build JSONL output.
    Uses multiprocessing + tqdm to speed up. After collecting all records, sorts them,
    randomly splits into train/val/test (60/20/20) with seed=42, and writes three JSONL files.
    """
    audio_dir = os.path.join(root_dir, "audio")
    lab_dir = os.path.join(root_dir, "chordlab")

    if not os.path.isdir(audio_dir):
        print(f"Error: audio directory not found: {audio_dir}")
        return
    if not os.path.isdir(lab_dir):
        print(f"Error: chordlab directory not found: {lab_dir}")
        return

    # List and sort all filenames in the audio directory
    all_files = [
        fname for fname in os.listdir(audio_dir)
        if os.path.splitext(fname)[1].lower() in ['.wav', '.mp3', '.flac', '.ogg', '.aac', '.m4a']
    ]
    all_files.sort()  # sort alphabetically to have a deterministic order before splitting

    # Prepare a partial function for multiprocessing
    worker = partial(process_single_file, audio_dir=audio_dir, lab_dir=lab_dir)

    # Use as many processes as CPU cores (minus 1 or so, but here we’ll just use all)
    num_workers = max(1, cpu_count() - 1)

    records = []
    with Pool(processes=num_workers) as pool:
        # imap_unordered to get results as they finish, and wrap with tqdm for progress bar
        for result in tqdm(pool.imap_unordered(worker, all_files), total=len(all_files),
                           desc="Processing files", unit="file"):
            if result is not None:
                records.append(result)

    if not records:
        print("No valid records were processed. Exiting.")
        return

    # Sort records by audio_path (or any other field) to make the final order deterministic
    records.sort(key=lambda x: x["audio_path"])

    # Now split into train/val/test = 60%/20%/20% with seed=42
    random.seed(42)
    random.shuffle(records)

    n = len(records)
    n_train = int(0.6 * n)
    n_val = int(0.2 * n)
    # Ensure that all records are used; leftover go to test
    n_test = n - n_train - n_val

    train_records = records[:n_train]
    val_records = records[n_train:n_train + n_val]
    test_records = records[n_train + n_val:]

    # Make sure output directory exists
    os.makedirs(root_dir, exist_ok=True)

    # Write out three separate JSONL files
    write_jsonl(train_records, os.path.join(root_dir, "Chords1217.train.jsonl"))
    write_jsonl(val_records,   os.path.join(root_dir, "Chords1217.val.jsonl"))
    write_jsonl(test_records,  os.path.join(root_dir, "Chords1217.test.jsonl"))

    print(f"Total files processed: {n}")
    print(f" → Train: {len(train_records)} records")
    print(f" → Validation: {len(val_records)} records")
    print(f" → Test: {len(test_records)} records")
    print(f"Outputs written to:\n"
          f"  {os.path.join(root_dir, 'Chords1217.train.jsonl')}\n"
          f"  {os.path.join(root_dir, 'Chords1217.val.jsonl')}\n"
          f"  {os.path.join(root_dir, 'Chords1217.test.jsonl')}")


if __name__ == "__main__":
    convert_chords_to_jsonl("data/Chords1217")