File size: 12,607 Bytes
5f5554c | 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 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 | #!/usr/bin/env python3
"""Pack GLAMI product images into Hub-ready Parquet shards.
Reads images straight out of glami_images.tar.gz in a single streaming pass --
the 1.3M files are never unpacked onto disk -- and writes
images/shard-00000.parquet, images/shard-00001.parquet, ...
with the `datasets` Image() feature, so the Hub's Dataset Viewer renders them
and `load_dataset(..., streaming=True)` works without a loading script.
Images are keyed by itemId only. They are deliberately NOT split into
train/test: the splits live in items_train.csv / items_test.csv / *_split.csv,
which reference itemId, so a single image table stays joinable by all of them
and never stores the same picture twice.
Usage:
python scripts/build_image_shards.py --inspect # peek inside the tarball first
python scripts/build_image_shards.py --limit 5000 # small trial run
python scripts/build_image_shards.py # build locally (resumable)
python scripts/build_image_shards.py \
--upload-repo zidcenek/GLAMIDuplicationDetection # build AND upload in one pass
With --upload-repo each batch of shards is pushed and then deleted locally, so
packing 1.3M images needs a couple of GB of free disk rather than ~16GB.
Requires: pip install datasets Pillow (Pillow is needed to encode the Image feature)
"""
from __future__ import annotations
import argparse
import json
import os
import re
import sys
import tarfile
import time
from array import array
from pathlib import Path
REPO_ROOT = Path(__file__).resolve().parent.parent
IMAGE_SUFFIXES = {".jpg", ".jpeg", ".png", ".webp", ".gif", ".bmp"}
# Magic bytes, checked on every image. Truncated downloads and HTML error
# pages that got saved with a .jpg name are common in a 1.3M-file crawl and
# would otherwise poison the shard at load time.
MAGIC = (
(b"\xff\xd8\xff", "jpeg"),
(b"\x89PNG\r\n\x1a\n", "png"),
(b"GIF87a", "gif"),
(b"GIF89a", "gif"),
(b"BM", "bmp"),
)
def sniff(data: bytes) -> str | None:
for prefix, kind in MAGIC:
if data.startswith(prefix):
return kind
if data[:4] == b"RIFF" and data[8:12] == b"WEBP":
return "webp"
return None
def human(n: float) -> str:
for unit in ("B", "KB", "MB", "GB", "TB"):
if n < 1024:
return f"{n:.1f}{unit}"
n /= 1024
return f"{n:.1f}PB"
def iter_members(source: Path):
"""Yield (name, bytes) for every regular file in a tarball or directory."""
if source.is_dir():
for path in sorted(source.rglob("*")):
if path.is_file():
yield str(path.relative_to(source)), path.read_bytes()
return
# "r|gz" is the streaming mode: sequential, no seeking, constant memory.
with tarfile.open(source, "r|gz") as tar:
for member in tar:
if not member.isfile():
continue
handle = tar.extractfile(member)
if handle is None:
continue
yield member.name, handle.read()
def make_id_extractor(pattern: str | None):
"""itemId from a filename. Default: the whole stem if numeric, else its last digit run."""
if pattern:
rx = re.compile(pattern)
def extract(stem: str):
m = rx.search(stem)
return int(m.group(1)) if m else None
return extract
trailing = re.compile(r"(\d+)(?!.*\d)")
def extract(stem: str):
if stem.isdigit():
return int(stem)
m = trailing.search(stem)
return int(m.group(1)) if m else None
return extract
def inspect(source: Path, extract, count: int) -> None:
print(f"First {count} entries in {source.name}:\n")
seen = 0
for name, data in iter_members(source):
if Path(name).name.startswith("._") or name.startswith("__MACOSX/"):
continue
stem = Path(name).stem
print(
f" {name}\n"
f" itemId={extract(stem)} format={sniff(data)} size={human(len(data))}"
)
seen += 1
if seen >= count:
break
print(
"\nCheck that itemId matches the itemId column in items_train.csv."
"\nIf it does not, pass --id-regex with a capture group, e.g."
'\n --id-regex "item_(\\d+)_thumb"'
)
def main() -> int:
p = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
p.add_argument("--source", type=Path, default=REPO_ROOT / "glami_images.tar.gz",
help="tarball or already-extracted directory of images")
p.add_argument("--out", type=Path, default=REPO_ROOT / "images",
help="output directory for the parquet shards")
p.add_argument("--shard-mb", type=int, default=400,
help="target uncompressed image bytes per shard (default: 400)")
p.add_argument("--id-regex", default=None,
help="regex with one capture group extracting itemId from the filename stem")
p.add_argument("--limit", type=int, default=None, help="stop after N images (trial run)")
p.add_argument("--inspect", nargs="?", type=int, const=15, default=None,
help="list the first N tar entries and exit, without writing anything")
p.add_argument("--allow-duplicates", action="store_true",
help="keep repeated itemIds instead of skipping them")
p.add_argument("--upload-repo", default=None, metavar="REPO_ID",
help="upload shards to this Hub dataset as they are built, then delete them "
"locally (e.g. zidcenek/GLAMIDuplicationDetection). Keeps peak disk at "
"--upload-every-mb instead of the full dataset size.")
p.add_argument("--upload-every-mb", type=int, default=2000,
help="upload once this many MB of shards have accumulated (default: 2000)")
args = p.parse_args()
if not args.source.exists():
sys.exit(f"source not found: {args.source}")
if args.source.is_file() and args.source.stat().st_size == 0:
sys.exit(f"{args.source} is empty -- is the download still running?")
extract = make_id_extractor(args.id_regex)
if args.inspect is not None:
inspect(args.source, extract, args.inspect)
return 0
from datasets import Dataset, Features, Image, Value # imported late so --inspect needs no deps
from datasets.utils.logging import disable_progress_bar
api = None
if args.upload_repo:
from huggingface_hub import HfApi
from huggingface_hub.utils import HfHubHTTPError
api = HfApi()
try:
print(f"Uploading to {args.upload_repo} as {api.whoami()['name']}.")
except Exception:
sys.exit("not authenticated -- run 'hf auth login' (or export HF_TOKEN) first")
disable_progress_bar() # one bar per shard is noise across thousands of shards
features = Features({"itemId": Value("int64"), "image": Image()})
args.out.mkdir(parents=True, exist_ok=True)
state_path = args.out / ".build_state.json"
# Resume support. The gzip stream cannot be seeked, so resuming still has to
# decompress its way back to where it stopped -- but it does not re-encode
# or re-upload any shard that already exists.
state = json.loads(state_path.read_text()) if state_path.exists() else {}
shard_no = state.get("next_shard", 0)
resume_from = state.get("images_done", 0)
skip = resume_from
# Every packed itemId is appended here as a flat int64 array. Duplicate
# detection therefore survives both a resume and the shards being deleted
# after they are uploaded, which is how build_and_upload.py keeps peak disk
# down to one wave.
seen_path = args.out / ".seen_ids.bin"
seen_ids: set[int] = set()
if resume_from:
print(f"Resuming: {resume_from:,} images already packed into {shard_no} shard(s).")
if not args.allow_duplicates and seen_path.exists():
packed = array("q")
packed.frombytes(seen_path.read_bytes())
seen_ids.update(packed)
print(f" {len(seen_ids):,} itemIds loaded back for duplicate checking.")
pending: list[Path] = []
ids: list[int] = []
blobs: list[dict] = []
pending_bytes = 0
target = args.shard_mb * 1024 * 1024
done = skipped_nonimage = skipped_noid = skipped_corrupt = skipped_dupe = 0
started = time.time()
def flush() -> None:
nonlocal shard_no, ids, blobs, pending_bytes
if not ids:
return
out = args.out / f"shard-{shard_no:05d}.parquet"
Dataset.from_dict({"itemId": ids, "image": blobs}, features=features).to_parquet(out)
if not args.allow_duplicates:
with open(seen_path, "ab") as fh:
array("q", ids).tofile(fh)
shard_no += 1
rate = done / max(time.time() - started, 1e-9)
print(f" wrote {out.name} {len(ids):,} images {human(pending_bytes)} "
f"[{done:,} packed, {rate:.0f} img/s]", flush=True)
state_path.write_text(json.dumps({"next_shard": shard_no, "images_done": resume_from + done}))
pending.append(out)
ids, blobs, pending_bytes = [], [], 0
def push() -> None:
"""Upload the shards built since the last push, then delete them locally."""
nonlocal pending
if not api or not pending:
return
size = sum(f.stat().st_size for f in pending)
print(f" uploading {len(pending)} shard(s), {human(size)} ...", flush=True)
try:
api.upload_folder(
repo_id=args.upload_repo,
repo_type="dataset",
folder_path=str(args.out),
path_in_repo="images",
allow_patterns=["shard-*.parquet"],
commit_message=f"images: shards through {resume_from + done:,} images",
)
except HfHubHTTPError as exc:
sys.exit(f"upload failed: {exc}\n"
"Rerun the same command -- packed shards are on disk and the build resumes.")
# Only reached on a successful commit. The itemIds survive in
# .seen_ids.bin, so deleting the shards does not weaken dedup.
for f in pending:
f.unlink()
print(f" uploaded, freed {human(size)}", flush=True)
pending = []
for name, data in iter_members(args.source):
base = Path(name).name
# macOS `tar` writes an AppleDouble sidecar next to every file; it shares
# the real name, so it would otherwise claim the itemId first.
if base.startswith("._") or name.startswith("__MACOSX/"):
skipped_nonimage += 1
continue
suffix = Path(name).suffix.lower()
if suffix not in IMAGE_SUFFIXES:
skipped_nonimage += 1
continue
# Everything past here counts as an image, so the resume counter stays
# aligned with the tar order regardless of what we decide to drop.
if skip:
skip -= 1
continue
item_id = extract(Path(name).stem)
if item_id is None:
skipped_noid += 1
continue
if sniff(data) is None:
skipped_corrupt += 1
continue
if not args.allow_duplicates:
if item_id in seen_ids:
skipped_dupe += 1
continue
seen_ids.add(item_id)
ids.append(item_id)
blobs.append({"bytes": data, "path": f"{item_id}{suffix}"})
pending_bytes += len(data)
done += 1
if pending_bytes >= target:
flush()
if sum(f.stat().st_size for f in pending) >= args.upload_every_mb * 1024 * 1024:
push()
if args.limit and done >= args.limit:
break
flush()
push()
where = args.upload_repo or f"{args.out}/"
print(f"\nDone: {done:,} images this run, {shard_no} shard(s) total -> {where}")
for label, n in (("non-image entries", skipped_nonimage), ("no itemId in filename", skipped_noid),
("corrupt / not an image", skipped_corrupt), ("duplicate itemId", skipped_dupe)):
if n:
print(f" skipped {n:,} -- {label}")
if skipped_noid or skipped_corrupt:
print("\nA large skip count usually means --id-regex is wrong or the crawl "
"saved error pages. Re-check with --inspect before uploading.")
return 0
if __name__ == "__main__":
raise SystemExit(main())
|