File size: 2,013 Bytes
f799bd5 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 | #!/usr/bin/env python3
"""Rebuild the per-sequence tars from the original frame tree.
The tars that get uploaded to the Hub are derived data - they are deleted locally
once uploaded. This regenerates them byte-for-byte identically.
python repack_for_upload.py --src ../../vqs_4k_kaggle --out ..
`metadata.csv` carries both ids, so the rename (legacy_seq_id -> seq_id) is applied
to the archive name AND to the paths inside it.
"""
import argparse
import csv
import os
import subprocess
from multiprocessing import Pool
from pathlib import Path
def repack(job):
src, out, kind, old, new = job
dst = Path(out) / kind / f"{new}.tar"
if dst.exists():
return None
tmp = str(dst) + ".part"
cmd = ["tar", "--sort=name", "--owner=0", "--group=0", "--mtime=UTC 2026-01-01",
f"--transform=s|^{old}|{new}|", "-cf", tmp, "-C", f"{src}/{kind}", old]
r = subprocess.run(cmd, capture_output=True)
if r.returncode != 0:
Path(tmp).exists() and os.remove(tmp)
return f"FAIL {kind}/{old}: {r.stderr.decode()[:120]}"
os.replace(tmp, dst)
return None
def main():
p = argparse.ArgumentParser()
p.add_argument("--src", required=True, help="original tree with images/<id>/ and masks/<id>/")
p.add_argument("--out", default=".", help="where to write images/*.tar and masks/*.tar")
p.add_argument("--workers", type=int, default=6)
a = p.parse_args()
meta = Path(a.out) / "metadata.csv"
rows = list(csv.DictReader(open(meta)))
for k in ("images", "masks"):
(Path(a.out) / k).mkdir(parents=True, exist_ok=True)
jobs = [(a.src, a.out, k, r["legacy_seq_id"], r["seq_id"])
for k in ("masks", "images") for r in rows]
print(f"packing {len(jobs)} archives from {a.src}")
with Pool(a.workers) as pool:
errs = [e for e in pool.imap_unordered(repack, jobs, chunksize=8) if e]
print("errors:", len(errs))
for e in errs[:5]:
print(" ", e)
if __name__ == "__main__":
main()
|