Datasets:
Download scripts/parquet_to_tree.py from PrentisAI/ScreenRef-Annotations: direct link, hf CLI and curl.
- Browser
- Download file 3.45 kB
-
https://huggingface.co/datasets/PrentisAI/ScreenRef-Annotations/resolve/main/scripts/parquet_to_tree.py
- Command line
-
hf download hf://datasets/PrentisAI/ScreenRef-Annotations/scripts/parquet_to_tree.py
-
curl -L -o parquet_to_tree.py https://huggingface.co/datasets/PrentisAI/ScreenRef-Annotations/resolve/main/scripts/parquet_to_tree.py
3.45 kB
| #!/usr/bin/env python3 | |
| """Unpack the ScreenRef Parquet shards back into the images/ tree used by meta/. | |
| Each Parquet row stores the screenshot bytes plus its path relative to the task's | |
| image root; the root itself comes from meta/meta_full.json. The file lands at | |
| <out>/<meta root>/<path>, which is exactly the path recorded in | |
| stats/image_manifest.tsv.gz. | |
| python scripts/parquet_to_tree.py --shards /path/to/ScreenRef/data \ | |
| --meta meta/meta_full.json --out . # writes ./images/... | |
| python scripts/parquet_to_tree.py ... --verify stats/image_manifest.tsv.gz | |
| A screenshot referenced by several rows is written once. Existing files are skipped, | |
| so an interrupted run can simply be restarted. | |
| """ | |
| import argparse, glob, gzip, hashlib, json, os, sys | |
| import pyarrow.parquet as pq | |
| def main(): | |
| ap = argparse.ArgumentParser() | |
| ap.add_argument('--shards', required=True, help='directory holding train-*.parquet') | |
| ap.add_argument('--meta', default='meta/meta_full.json') | |
| ap.add_argument('--out', default='.', help='images/ is created under this directory') | |
| ap.add_argument('--verify', metavar='MANIFEST', | |
| help='check size and sha256 of every written file against the manifest') | |
| args = ap.parse_args() | |
| roots = {task: ent['root'] for task, ent in json.load(open(args.meta)).items()} | |
| manifest = None | |
| if args.verify: | |
| with gzip.open(args.verify, 'rt') as fh: | |
| next(fh) | |
| manifest = {} | |
| for line in fh: | |
| p, s, h = line.rstrip('\n').split('\t') | |
| manifest[p] = (int(s), h) | |
| shards = sorted(glob.glob(os.path.join(args.shards, 'train-*.parquet'))) | |
| if not shards: | |
| sys.exit(f'no train-*.parquet under {args.shards}') | |
| written = skipped = rows = bad = 0 | |
| for i, shard in enumerate(shards, 1): | |
| pf = pq.ParquetFile(shard) | |
| for rg in range(pf.num_row_groups): | |
| for r in pf.read_row_group(rg, columns=['image', 'task']).to_pylist(): | |
| rows += 1 | |
| root = roots.get(r['task']) | |
| if root is None: | |
| sys.exit(f'task {r["task"]!r} not in {args.meta}') | |
| rel = os.path.join(root, r['image']['path']) | |
| dst = os.path.join(args.out, rel) | |
| if os.path.exists(dst): | |
| skipped += 1 | |
| continue | |
| raw = r['image']['bytes'] | |
| if manifest is not None: | |
| key = os.path.relpath(rel, 'images') | |
| want = manifest.get(key) | |
| if want is None or want != (len(raw), hashlib.sha256(raw).hexdigest()): | |
| print(f'MISMATCH {key}', file=sys.stderr) | |
| bad += 1 | |
| continue | |
| os.makedirs(os.path.dirname(dst), exist_ok=True) | |
| tmp = dst + '.part' | |
| with open(tmp, 'wb') as fo: | |
| fo.write(raw) | |
| os.replace(tmp, dst) | |
| written += 1 | |
| print(f'[{i}/{len(shards)}] {os.path.basename(shard)} ' | |
| f'rows {rows:,} written {written:,} already present {skipped:,}', flush=True) | |
| print(f'done: {rows:,} rows, {written:,} files written, {skipped:,} skipped' | |
| + (f', {bad:,} mismatches' if manifest is not None else '')) | |
| return 1 if bad else 0 | |
| if __name__ == '__main__': | |
| sys.exit(main()) | |