| |
| """Rename <task>-NNNN.parquet shards to the Hub convention data/train-XXXXX-of-YYYYY.parquet. |
| |
| The Hub infers the split from this naming; without it load_dataset will not find |
| `train`. Renaming is an in-place os.rename, no data is copied. |
| |
| Before renaming, every task in the meta file must have at least one shard. A |
| missing task means an export job failed silently, which is very hard to notice |
| after upload (the dataset looks normal, it just lacks a whole subset). |
| """ |
| import argparse, glob, json, os, re |
|
|
| ap = argparse.ArgumentParser() |
| ap.add_argument("--dir", default="parquet_out/data") |
| ap.add_argument("--meta", default="meta/meta_full.json") |
| ap.add_argument("--skip-completeness", action="store_true", |
| help="skip the every-task-has-a-shard check (only when exporting a subset on purpose)") |
| ap.add_argument("--apply", action="store_true", help="without this flag, only print the plan") |
| a = ap.parse_args() |
|
|
| files = sorted(f for f in glob.glob(os.path.join(a.dir, "*.parquet")) |
| if not re.match(r"train-\d{5}-of-\d{5}\.parquet$", os.path.basename(f))) |
| n = len(files) |
| if n == 0: |
| raise SystemExit("no shards to rename") |
|
|
| if not a.skip_completeness: |
| meta = json.load(open(a.meta)) |
| have = {os.path.basename(f).rsplit("-", 1)[0] for f in files} |
| missing = sorted(set(meta) - have) |
| if missing: |
| raise SystemExit( |
| f"{len(missing)} task(s) have no shard; rerun them before finalizing:\n " |
| + "\n ".join(missing)) |
|
|
| total_b = sum(os.path.getsize(f) for f in files) |
| print(f"{n} shards, {total_b / 2**30:.1f} GiB in total") |
| for i, src in enumerate(files): |
| dst = os.path.join(a.dir, f"train-{i:05d}-of-{n:05d}.parquet") |
| if a.apply: |
| os.rename(src, dst) |
| elif i < 3 or i == n - 1: |
| print(f" {os.path.basename(src)} -> {os.path.basename(dst)}") |
| print("renamed" if a.apply else "(dry run; add --apply to rename)") |
|
|
