| |
| """Build a ready-to-train text view from a local references-only release.""" |
|
|
| import argparse |
| import itertools |
| import os |
| from pathlib import Path |
| import tempfile |
|
|
| import pyarrow as pa |
|
|
| from finepdf_release import FinePDFResolver, iter_release_rows, iter_training_examples |
| from release_data import ShardedWriter, write_json |
|
|
|
|
| def main(argv=None): |
| p = argparse.ArgumentParser(description=__doc__) |
| p.add_argument("--release-dir", type=Path, required=True) |
| p.add_argument("--config", help="Language config, e.g. eng_Latn or fra_Latn; required at the combined root") |
| p.add_argument("--output-dir", type=Path, required=True) |
| p.add_argument("--local-source-root", type=Path) |
| p.add_argument("--cache-dir") |
| p.add_argument("--num-negatives", type=int, default=8) |
| p.add_argument("--all-queries", action="store_true") |
| p.add_argument("--limit", type=int, help="Maximum emitted examples, including additional queries") |
| p.add_argument("--rows-per-shard", type=int, default=100_000) |
| args = p.parse_args(argv) |
| if args.num_negatives < 0 or args.rows_per_shard <= 0 or (args.limit is not None and args.limit <= 0): |
| p.error("Invalid negative count, shard size, or limit") |
| if args.output_dir.exists(): |
| p.error("Output already exists; choose a new directory") |
| args.output_dir.parent.mkdir(parents=True, exist_ok=True) |
| staging = Path(tempfile.mkdtemp(prefix=args.output_dir.name + ".partial-", dir=args.output_dir.parent)) |
| schema = pa.schema([(name, pa.string()) for name in |
| ["query", "positive"] + ["negative_{}".format(i) for i in range(args.num_negatives)]]) |
| writer = ShardedWriter(staging, "train", schema, args.rows_per_shard) |
| with FinePDFResolver(args.release_dir, args.local_source_root, args.cache_dir, config=args.config) as resolver: |
| examples = iter_training_examples(iter_release_rows(args.release_dir, config=args.config), resolver, |
| args.num_negatives, args.all_queries) |
| if args.limit is not None: |
| examples = itertools.islice(examples, args.limit) |
| while True: |
| batch = list(itertools.islice(examples, 256)) |
| if not batch: |
| break |
| writer.write(pa.Table.from_pylist(batch, schema=schema)) |
| manifest = writer.close() |
| if manifest["num_rows"] == 0: |
| raise ValueError("No examples to materialize") |
| write_json(staging / "manifest.json", {"train": manifest, "num_negatives": args.num_negatives, |
| "all_queries": args.all_queries}) |
| write_json(staging / "_SUCCESS", {"num_rows": manifest["num_rows"]}) |
| os.rename(staging, args.output_dir) |
| print("Wrote {} examples to {}".format(manifest["num_rows"], args.output_dir)) |
|
|
|
|
| if __name__ == "__main__": |
| main() |
|
|
|
|
