|
|
import json |
|
|
from collections import defaultdict |
|
|
from pathlib import Path |
|
|
|
|
|
from datasets import ( |
|
|
DatasetInfo, |
|
|
DownloadManager, |
|
|
Features, |
|
|
GeneratorBasedBuilder, |
|
|
Image, |
|
|
Split, |
|
|
SplitGenerator, |
|
|
Value, |
|
|
) |
|
|
|
|
|
_URLS = [f"data/{i:0>5}.tar" for i in range(2123)] |
|
|
|
|
|
|
|
|
class Laion2bMultiKoreanSubsetWithImage(GeneratorBasedBuilder): |
|
|
|
|
|
VERSION = "0.1.3" |
|
|
|
|
|
def _info(self): |
|
|
return DatasetInfo( |
|
|
description="laion2b_multi_korean_subset data with images", |
|
|
features=Features( |
|
|
{ |
|
|
"image": Image(), |
|
|
"text": Value("string"), |
|
|
"width": Value("int32"), |
|
|
"height": Value("int32"), |
|
|
} |
|
|
), |
|
|
supervised_keys=None, |
|
|
license="cc-by-4.0", |
|
|
version=self.VERSION, |
|
|
) |
|
|
|
|
|
def _split_generators(self, dl_manager: DownloadManager): |
|
|
downloaded = dl_manager.download(_URLS) |
|
|
return [ |
|
|
SplitGenerator( |
|
|
name=Split.TRAIN, |
|
|
gen_kwargs={ |
|
|
"archives": [ |
|
|
dl_manager.iter_archive(archive) for archive in downloaded |
|
|
] |
|
|
}, |
|
|
) |
|
|
] |
|
|
|
|
|
def _generate_examples(self, archives): |
|
|
idx = 0 |
|
|
temp = defaultdict(dict) |
|
|
|
|
|
for archive in archives: |
|
|
iter_archive = iter(archive) |
|
|
while True: |
|
|
|
|
|
try: |
|
|
path, file = next(iter_archive) |
|
|
except Exception: |
|
|
break |
|
|
|
|
|
if path.endswith(".txt"): |
|
|
continue |
|
|
|
|
|
file_id = Path(path).stem |
|
|
|
|
|
if path.endswith(".json"): |
|
|
try: |
|
|
meta = json.loads(file.read()) |
|
|
except Exception: |
|
|
continue |
|
|
|
|
|
temp[file_id]["text"] = meta["caption"] |
|
|
temp[file_id]["width"] = meta["width"] |
|
|
temp[file_id]["height"] = meta["height"] |
|
|
|
|
|
if "image" in temp[file_id]: |
|
|
yield idx, temp.pop(file_id) |
|
|
idx += 1 |
|
|
|
|
|
elif path.endswith(".webp"): |
|
|
try: |
|
|
temp[file_id]["image"] = {"path": path, "bytes": file.read()} |
|
|
except Exception: |
|
|
continue |
|
|
|
|
|
if "text" in temp[file_id]: |
|
|
yield idx, temp.pop(file_id) |
|
|
idx += 1 |
|
|
|
|
|
temp.clear() |
|
|
|
