data_generator.py

import argparse
from datasets import load_dataset
import os

def save_dataset_to_parquet(dataset, output_path, dataset_name):
    os.makedirs(output_path, exist_ok=True)
    for split in dataset:
        split_dataset = dataset[split]
        
        # Save the entire split as a single Parquet file
        split_path = os.path.join(output_path, dataset_name, f"{split}.parquet")
        os.makedirs(os.path.dirname(split_path), exist_ok=True)  # Create directories if needed
        split_dataset.to_parquet(split_path)

def main(dataset_name):
    num_proc = os.cpu_count()
    dataset_paths = {
        "speech_vqav2": "./hf_generator_vqav2.py",
        "speech_vg": "./hf_generator_vg.py",
        "speech_asr": "./hf_generator_asr.py",
        "speech_tts": "./hf_generator_tts.py",
        "laion": "./hf_generator_laion.py",
    }

    if dataset_name not in dataset_paths:
        raise ValueError(f"Dataset name '{dataset_name}' not recognized. Available options: {list(dataset_paths.keys())}")

    ds = load_dataset(
        dataset_paths[dataset_name],
        trust_remote_code=True,
        cache_dir=f"~/.cache/{dataset_name}",
        num_proc=num_proc
    )
    save_dataset_to_parquet(ds, "./hf_datasets/", dataset_name)

if __name__ == "__main__":
    parser = argparse.ArgumentParser(description="Save specified Hugging Face dataset to Parquet format.")
    parser.add_argument("dataset_name", type=str, help="Name of the dataset to load and save to Parquet. Options: speech_vqav2, speech_vg, speech_asr, speech_tts, laion")
    args = parser.parse_args()
    
    main(args.dataset_name)


# from datasets import load_dataset
# import os

# def save_dataset_to_parquet(dataset, output_path, dataset_name):
#     os.makedirs(output_path, exist_ok=True)
#     for split in dataset:
#         split_dataset = dataset[split]
        
#         # Save the entire split as a single Parquet file
#         split_path = os.path.join(output_path, dataset_name, f"{split}.parquet")
#         os.makedirs(os.path.dirname(split_path), exist_ok=True)  # Create directories if needed
#         split_dataset.to_parquet(split_path)

# num_proc = os.cpu_count()
# ds = load_dataset("./hf_generator_vqav2.py", trust_remote_code=True, cache_dir="~/.cache/speech_vqav2", num_proc=num_proc)
# save_dataset_to_parquet(ds, "./hf_datasets/", "speech_vqav2")

# ds = load_dataset("./hf_generator_vg.py", trust_remote_code=True, cache_dir="~/.cache/speech_vg", num_proc=num_proc)
# save_dataset_to_parquet(ds, "./hf_datasets/", "speech_vg")

# ds = load_dataset("./hf_generator_asr.py", trust_remote_code=True, cache_dir="~/.cache/speech_asr", num_proc=64)
# save_dataset_to_parquet(ds, "./hf_datasets/", "speech_asr")

# ds = load_dataset("./hf_generator_tts.py", trust_remote_code=True, cache_dir="~/.cache/speech_tts", num_proc=64)
# save_dataset_to_parquet(ds, "./hf_datasets/", "speech_tts")

# ds = load_dataset("./hf_generator_laion.py", trust_remote_code=True, cache_dir="~/.cache/speech_laion", num_proc=64)
# save_dataset_to_parquet(ds, "./hf_datasets/", "laion")