| |
| """ |
| Verify the published HuggingFace dataset against the original archive. |
| |
| Downloads a random sample of parquet shards from the Hub, expands them the same |
| way `restore_datasets.py` does, and compares every row against the corresponding |
| entry read straight out of `VLadapter.zip`. Feature arrays must match bit-for-bit |
| and annotation files byte-for-byte. |
| |
| python verify_published.py --zip VLadapter.zip --shards 6 --annotations 5 |
| |
| Each shard is deleted after checking, so disk use stays around one shard. |
| """ |
| import argparse |
| import io |
| import random |
| import sys |
| import zipfile |
| from pathlib import Path |
|
|
| import h5py |
| import numpy as np |
| import pyarrow.parquet as pq |
|
|
| ROOT = "vlt5_dataset/" |
| FEAT_SHAPE = (49, 2048) |
|
|
|
|
| def drop(path): |
| try: |
| real = Path(path).resolve() |
| Path(path).unlink(missing_ok=True) |
| real.unlink(missing_ok=True) |
| except OSError: |
| pass |
|
|
|
|
| def main(): |
| p = argparse.ArgumentParser(description=__doc__, |
| formatter_class=argparse.RawDescriptionHelpFormatter) |
| p.add_argument("--zip", required=True, help="the original VLadapter.zip") |
| p.add_argument("--repo-id", default="ylsung/VL-Adapter-datasets") |
| p.add_argument("--shards", type=int, default=6, help="random feature shards to check") |
| p.add_argument("--annotations", type=int, default=5, help="random annotation files to check") |
| p.add_argument("--seed", type=int, default=None) |
| args = p.parse_args() |
|
|
| from huggingface_hub import HfApi, hf_hub_download |
|
|
| rng = random.Random(args.seed) |
| api = HfApi() |
| files = api.list_repo_files(args.repo_id, repo_type="dataset") |
| shards = sorted(f for f in files if f.startswith("features/") and f.endswith(".parquet")) |
| anns = sorted(f for f in files if f.startswith("annotations/")) |
| print(f"[hub] {len(shards)} feature shards, {len(anns)} annotation files") |
|
|
| zf = zipfile.ZipFile(args.zip) |
| by_name = {i.filename: i for i in zf.infolist()} |
|
|
| rows = mismatches = missing = 0 |
|
|
| for shard in rng.sample(shards, min(args.shards, len(shards))): |
| _, dataset, variant, _ = shard.split("/") |
| local = hf_hub_download(args.repo_id, shard, repo_type="dataset") |
| table = pq.read_table(local) |
| ids = table.column("id").to_pylist() |
| feats = table.column("features") |
| bad = 0 |
| for i, img_id in enumerate(ids): |
| orig_name = f"{ROOT}{dataset}/clip_features/{variant}/{img_id}.h5" |
| info = by_name.get(orig_name) |
| if info is None: |
| missing += 1 |
| continue |
| with h5py.File(io.BytesIO(zf.read(info)), "r") as fh: |
| ref = fh[f"{img_id}/features"][...] |
| got = np.asarray(feats[i].as_py(), dtype=np.float16).reshape(FEAT_SHAPE) |
| if not (got.dtype == ref.dtype and np.array_equal(got, ref)): |
| bad += 1 |
| rows += 1 |
| mismatches += bad |
| print(f" {shard}: {len(ids)} rows, {bad} mismatched") |
| drop(local) |
|
|
| ann_bad = 0 |
| for ann in rng.sample(anns, min(args.annotations, len(anns))): |
| orig_name = ROOT + ann[len("annotations/"):] |
| info = by_name.get(orig_name) |
| if info is None: |
| print(f" {ann}: NOT IN ARCHIVE") |
| missing += 1 |
| continue |
| local = hf_hub_download(args.repo_id, ann, repo_type="dataset") |
| same = Path(local).read_bytes() == zf.read(info) |
| ann_bad += (not same) |
| print(f" {ann}: {info.file_size/1e6:.1f} MB, {'identical' if same else 'DIFFERS'}") |
| drop(local) |
|
|
| print() |
| print(f"feature rows compared : {rows}") |
| print(f" bit-exact : {rows - mismatches}") |
| print(f" mismatched : {mismatches}") |
| print(f"annotations compared : {min(args.annotations, len(anns))}, {ann_bad} differing") |
| print(f"missing from archive : {missing}") |
| ok = not (mismatches or ann_bad or missing) |
| print("RESULT:", "PASS" if ok else "FAIL") |
| return 0 if ok else 1 |
|
|
|
|
| if __name__ == "__main__": |
| sys.exit(main()) |
|
|
