Datasets:
File size: 2,490 Bytes
9bc4c0e | 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 | """Read local or public-Hub WebDataset shards without extracting image files."""
import argparse
import io
import json
import tarfile
from pathlib import Path
from urllib.request import urlopen
from PIL import Image
def samples(fileobj):
"""Yield (PIL image, metadata, text) while retaining only one sample in RAM."""
with tarfile.open(fileobj=fileobj, mode='r|*') as archive:
pending = {}
current = None
for member in archive:
key, extension = member.name.rsplit('.', 1)
if current is not None and key != current:
if set(pending) != {'png', 'json', 'txt'}:
raise ValueError('Incomplete WebDataset sample: ' + current)
metadata = json.loads(pending['json'])
with Image.open(io.BytesIO(pending['png'])) as im:
yield im.convert('RGB'), metadata, pending['txt'].decode()
pending = {}
current = key
if extension in pending:
raise ValueError('Repeated member: ' + member.name)
pending[extension] = archive.extractfile(member).read()
if pending:
if set(pending) != {'png', 'json', 'txt'}:
raise ValueError('Incomplete WebDataset sample: ' + current)
with Image.open(io.BytesIO(pending['png'])) as im:
yield im.convert('RGB'), json.loads(pending['json']), pending['txt'].decode()
if __name__ == '__main__':
p=argparse.ArgumentParser()
p.add_argument('shard',help='Local TAR path, https URL, or shard filename in --repo')
p.add_argument('--repo',default='laion/Anim-E-Ghibli-Backgrounds-FLUX-9B')
p.add_argument('--revision',default='main')
p.add_argument('--limit',type=int,default=5)
a=p.parse_args()
if Path(a.shard).is_file():
source=open(a.shard,'rb')
else:
url=a.shard if a.shard.startswith('https://') else f'https://huggingface.co/datasets/{a.repo}/resolve/{a.revision}/data/{a.shard}'
source=urlopen(url,timeout=90)
with source:
for i,(image,metadata,text) in enumerate(samples(source)):
print(json.dumps(dict(id=metadata['id'],size=image.size,primary_class=metadata.get('primary_class'),
subcategory=metadata.get('subcategory'),personality_id=metadata.get('personality_id'),
emotion=metadata.get('emotion'),seed=metadata['seed'],prompt=text),ensure_ascii=False))
if i+1>=a.limit: break
|