prep/build.py

import argparse
import gc
import json
import os
import shutil
from datetime import datetime
from pathlib import Path
from typing import Optional

from datasets import DatasetDict, load_dataset

from .lang_def import WIKIPEDIA_LANGUAGES

_MAX_SHARD_SIZE = 256 * 1024 * 1024


def main(
    languages: list[str],
    date: str,
    dest_dirpath: Path,
    cache_dirpath: Optional[Path],
    mirror_url: str,
    override: bool = False,
):
    for language in languages:
        print("=====================")
        print(f"start processing `{language}`")
        start_time = datetime.now()

        lang_dest_dirpath = dest_dirpath / language
        if lang_dest_dirpath.exists():
            if override:
                print(f"overriding `{language}` at {lang_dest_dirpath}")
                shutil.rmtree(lang_dest_dirpath)
            else:
                print(
                    f"skipping `{language}` because {lang_dest_dirpath} already exists"
                )
                print(f"done  processing `{language}` ✅")
                print("=====================\n\n")
                continue

        build_kwargs = {
            "language": language.replace("-", "_"),
            "date": date,
            "mirror_url": mirror_url,
            "beam_runner": "DirectRunner",
            # "beam_runner": "apache_beam.runners.dask.dask_runner.DaskRunner",
        }
        if cache_dirpath:
            build_kwargs["cache_dir"] = str(cache_dirpath)

        try:
            dsd = load_dataset(
                "./prep/ds_script.py",
                **build_kwargs,
            )
            elapsed = datetime.now() - start_time
            assert isinstance(dsd, DatasetDict)
            print(f"done  processing `{language}` (elapsed: {elapsed})")

            for split_name, split in dsd.items():
                # With a beam runner, the dataset_size property is not actually
                # the number of bytes.

                dataset_size = split.download_size
                assert isinstance(dataset_size, int)
                num_shards = 1 + dataset_size // _MAX_SHARD_SIZE
                print(f"start splitting `{language}` into {num_shards} shards")

                shard_filenames = []
                for shard_index in range(num_shards):
                    ds_shard = split.shard(num_shards, shard_index)

                    shard_suffix = f"{shard_index+1:04d}-of-{num_shards:04d}"
                    dest_filename = f"{split_name}-{shard_suffix}.parquet"
                    shard_filenames.append(dest_filename)

                    os.makedirs(lang_dest_dirpath, exist_ok=True)
                    ds_shard.to_parquet(lang_dest_dirpath / dest_filename)

                # Create the info.json file
                info = {"shards": shard_filenames}
                with open(lang_dest_dirpath / "info.json", "w") as f:
                    json.dump(info, f)

            del dsd

        except Exception as e:
            print("❌❌❌❌")
            print(f"error with {language}: {e}")
            print("❌❌❌❌")

        gc.collect()
        print(f"done  processing `{language}` ✅")
        print("=====================\n\n")


if __name__ == "__main__":

    def formatter(prog):
        return argparse.HelpFormatter(prog, max_help_position=52)

    parser = argparse.ArgumentParser(
        prog="Wikipedia Builder",
        description="Prepares the Wikipedia dataset for each language",
        formatter_class=formatter,
    )
    parser.add_argument(
        "--date",
        help="Wikipedia dump date (e.g. 20230601)",
        default="20230601",
        type=str,
    )
    parser.add_argument(
        "--language",
        help="Language code (e.g. en). Default to processing all languages",
        type=str,
        nargs="*",
        metavar="LANG",
    )
    parser.add_argument(
        "--cache-dir",
        help="Cache directory for 🤗 Datasets",
        default=None,
        metavar="DIR",
        type=str,
    )
    parser.add_argument(
        "--mirror-url",
        help="Mirror URL",
        default="https://dumps.wikimedia.org",
        type=str,
    )
    parser.add_argument(
        "--override", help="Override existing files", action="store_true"
    )
    args = parser.parse_args()

    # Prepare the arguments
    if not args.language:
        languages = WIKIPEDIA_LANGUAGES[args.date]
    else:
        languages = args.language
    print("languages to process:", languages)
    dest_dirpath = Path("./data") / args.date

    main(
        languages,
        args.date,
        dest_dirpath,
        cache_dirpath=args.cache_dir,
        mirror_url=args.mirror_url,
        override=args.override,
    )