| """Verify a downloaded copy of this dataset is complete and internally consistent. |
| |
| uv run python scripts/verify_download.py [DATASET_ROOT] |
| |
| Written for the case where `huggingface-cli download` died partway -- e.g. on an HTTP 429 |
| rate limit. A partial download leaves a tree that looks plausible: lerobot will happily open |
| it and train on whatever episodes happen to be present. This checks counts and cross-references |
| before you spend GPU hours. |
| |
| Needs only stdlib + pandas/pyarrow (pandas only for the --deep row-count pass). |
| """ |
| import argparse |
| import glob |
| import json |
| import os |
| import pathlib |
| import sys |
|
|
|
|
| def main() -> int: |
| ap = argparse.ArgumentParser() |
| ap.add_argument("root", nargs="?", default=str(pathlib.Path(__file__).resolve().parent.parent)) |
| ap.add_argument("--deep", action="store_true", |
| help="also open every parquet and check row counts against episodes.jsonl") |
| a = ap.parse_args() |
| root = pathlib.Path(a.root) |
| fail = [] |
|
|
| for rel in ("meta/info.json", "meta/episodes.jsonl", "meta/episodes_stats.jsonl", "meta/tasks.jsonl"): |
| if not (root / rel).exists(): |
| fail.append(f"missing {rel}") |
| if fail: |
| for f in fail: |
| print("FAIL", f) |
| return 1 |
|
|
| info = json.load(open(root / "meta/info.json")) |
| eps = {json.loads(l)["episode_index"]: json.loads(l) for l in open(root / "meta/episodes.jsonl")} |
| stats = [json.loads(l)["episode_index"] for l in open(root / "meta/episodes_stats.jsonl")] |
| cams = sorted(k.split("observation.images.")[1] for k in info["features"] if info["features"][k]["dtype"] == "video") |
| parquets = sorted(glob.glob(str(root / "data/chunk-000/episode_*.parquet"))) |
| videos = glob.glob(str(root / "videos/chunk-000/*/*.mp4")) |
| incomplete = glob.glob(str(root / ".cache/huggingface/download/**/*.incomplete"), recursive=True) |
|
|
| exp_e, exp_f, exp_v = info["total_episodes"], info["total_frames"], info["total_videos"] |
| print(f"root {root}") |
| print(f"cameras {cams}") |
| print(f"episodes info={exp_e} parquet={len(parquets)} episodes.jsonl={len(eps)} stats={len(stats)}") |
| print(f"videos info={exp_v} on disk={len(videos)}") |
| print(f"frames info={exp_f} episodes.jsonl sum={sum(e['length'] for e in eps.values())}") |
| print(f"in-flight {len(incomplete)} .incomplete file(s)") |
|
|
| if len(parquets) != exp_e: |
| fail.append(f"{len(parquets)} parquet files but info.total_episodes={exp_e}") |
| if len(eps) != exp_e: |
| fail.append(f"episodes.jsonl has {len(eps)} rows but info.total_episodes={exp_e}") |
| if sorted(eps) != list(range(exp_e)): |
| fail.append("episodes.jsonl indices are not contiguous 0..N-1") |
| if stats != list(range(exp_e)): |
| fail.append("episodes_stats.jsonl is not aligned 0..N-1") |
| if len(videos) != exp_v: |
| fail.append(f"{len(videos)} mp4 files but info.total_videos={exp_v}") |
| if sum(e["length"] for e in eps.values()) != exp_f: |
| fail.append("episodes.jsonl lengths do not sum to info.total_frames") |
| if incomplete: |
| fail.append(f"{len(incomplete)} .incomplete file(s) -- the download did not finish") |
|
|
| |
| missing, tiny = [], [] |
| for ei in range(exp_e): |
| if not (root / f"data/chunk-000/episode_{ei:06d}.parquet").exists(): |
| missing.append(f"parquet ep{ei}") |
| for c in cams: |
| p = root / f"videos/chunk-000/observation.images.{c}/episode_{ei:06d}.mp4" |
| if not p.exists(): |
| missing.append(f"{c} ep{ei}") |
| elif p.stat().st_size < 4096: |
| tiny.append(f"{c} ep{ei} ({p.stat().st_size} B)") |
| if missing: |
| fail.append(f"{len(missing)} missing file(s), first 10: {missing[:10]}") |
| if tiny: |
| fail.append(f"{len(tiny)} suspiciously small video(s): {tiny[:10]}") |
|
|
| if a.deep: |
| import pandas as pd |
| bad = [] |
| for ei in range(exp_e): |
| n = len(pd.read_parquet(root / f"data/chunk-000/episode_{ei:06d}.parquet", columns=["frame_index"])) |
| if n != eps[ei]["length"]: |
| bad.append((ei, eps[ei]["length"], n)) |
| print(f"deep {exp_e} parquets opened, {len(bad)} row-count mismatch(es)") |
| if bad: |
| fail.append(f"row-count mismatches, first 10: {bad[:10]}") |
|
|
| print() |
| if fail: |
| for f in fail: |
| print("FAIL", f) |
| print("\nIncomplete or inconsistent. Re-run the download -- it resumes:") |
| print(" export HF_HUB_DISABLE_XET=1") |
| print(" uv run huggingface-cli download $REPO_ID --repo-type dataset \\") |
| print(" --local-dir $HF_LEROBOT_HOME/$REPO_ID --max-workers 4") |
| return 1 |
| print("PASS -- download is complete and self-consistent.") |
| return 0 |
|
|
|
|
| if __name__ == "__main__": |
| sys.exit(main()) |
|
|
