convert2jsonl.py

#!/usr/bin/env python3
import os
import json
import argparse

import numpy as np
import pandas as pd
import soundfile as sf
from tqdm import tqdm

def generate_split_jsonl(data_path: str, split_name: str, tags: np.ndarray, binary: np.ndarray):
    """
    Reads {split_name}.tsv, processes all .mp3 files, and generates MTT.{split_name}.jsonl.
    Includes error handling, a progress bar, and bitrate calculation for MP3s.
    """
    mt_dir = data_path
    tsv_path = os.path.join(mt_dir, f'{split_name}.tsv')
    if split_name == "valid":
        out_split_name = "val"
    else:
        out_split_name = split_name
    out_path = os.path.join(mt_dir, f'MTT.{out_split_name}.jsonl')
    fail_log_path = os.path.join(mt_dir, f'fail.{out_split_name}.txt')

    # Read index and filenames
    df = pd.read_csv(tsv_path, sep='\t', header=None, names=['idx', 'title'])

    failed_count = 0
    failed_records = []

    print(f"Processing split: {split_name}")
    with open(out_path, 'w', encoding='utf-8') as fw:
        # Use tqdm for a progress bar
        for _, row in tqdm(df.iterrows(), total=df.shape[0], desc=f"-> Generating {split_name}.jsonl"):
            try:
                i = int(row['idx'])
                title = row['title']  # e.g., "48/948.low.mp3"
                audio_path = os.path.join(mt_dir, 'mp3', title)

                # Read audio metadata (supports .mp3)
                info = sf.info(audio_path)
                duration = info.frames / info.samplerate
                num_samples = info.frames
                sample_rate = info.samplerate
                channels = info.channels

                # --- MODIFICATION START ---
                # Calculate bitrate, which is more meaningful for compressed formats like MP3
                bitrate = None
                # Check if duration is valid to avoid division by zero
                if duration > 0:
                    try:
                        file_size_bytes = os.path.getsize(audio_path)
                        # Bitrate in bits per second (bps)
                        bitrate = int((file_size_bytes * 8) / duration)
                    except OSError:
                        # File might not exist or other OS-level error
                        pass
                
                # Infer bit depth from subtype, will be None for mp3
                bit_depth = None
                if hasattr(info, 'subtype') and info.subtype and info.subtype.startswith('PCM_'):
                    try:
                        bit_depth = int(info.subtype.split('_', 1)[1])
                    except (ValueError, IndexError):
                        pass # Could not parse bit depth from subtype
                # --- MODIFICATION END ---

                # Get the list of labels for this sample
                labels = tags[binary[i].astype(bool)].tolist()

                # Assemble the JSON object and write to file
                record = {
                    "audio_path":  audio_path,
                    "label":       labels,
                    "duration":    duration,
                    "sample_rate": sample_rate,
                    "num_samples": num_samples,
                    "bit_depth":   bit_depth,   # This will be null for MP3 files
                    "bitrate":     bitrate,     # This is the newly added field
                    "channels":    channels
                }
                fw.write(json.dumps(record, ensure_ascii=False) + "\n")

            except Exception as e:
                # If any error occurs, log it and skip the file
                failed_count += 1
                failed_records.append(f"File: {title}, Error: {str(e)}")
                continue

    print(f"Successfully generated {out_path}")

    # After the loop, report and log any failures
    if failed_count > 0:
        print(f"Skipped {failed_count} corrupted or problematic files for split '{split_name}'.")
        # Append failures to fail.txt
        with open(fail_log_path, 'a', encoding='utf-8') as f_fail:
            f_fail.write(f"--- Failures for split: {split_name} ({failed_count} files) ---\n")
            for record in failed_records:
                f_fail.write(record + "\n")
            f_fail.write("\n")

def main():
    parser = argparse.ArgumentParser(
        description="Generate JSONL files for MTT dataset splits (train/valid/test) for .mp3 files.")
    parser.add_argument(
        "data_path",
        help="Root directory of the MTT dataset, containing annotations, tags, labels, and tsv splits.")
    args = parser.parse_args()

    mt_dir = args.data_path

    # Use a generic failure log name that doesn't need to be cleaned for each split
    fail_log_path = os.path.join(mt_dir, 'processing_failures.log')
    if os.path.exists(fail_log_path):
        os.remove(fail_log_path)
        print(f"Removed old log file: {fail_log_path}")

    try:
        # Load tags and binary label matrix
        tags = np.load(os.path.join(mt_dir, 'tags.npy'))
        binary = np.load(os.path.join(mt_dir, 'binary_label.npy'))
    except FileNotFoundError as e:
        print(f"Error: Could not find required .npy file. {e}")
        return

    # Generate JSONL for each split
    for split in ['train', 'valid', 'test']:
        generate_split_jsonl(args.data_path, split, tags, binary)

    print("\nProcessing complete.")
    if os.path.exists(fail_log_path):
        print(f"A log of all failed files has been saved to: {fail_log_path}")

if __name__ == "__main__":
    main()