File size: 19,820 Bytes
01fdb75 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 | from typing import List, Dict, Iterable, Optional
import random
import os
import glob
import torch
import webdataset as wds
import numpy
import io
from PIL import Image
import pyarrow.parquet as pq
import pyarrow.compute as pc
from torch.utils.data import IterableDataset
from torchvision.transforms import Normalize
from torchvision.transforms.functional import to_tensor
import copy
import functools
def resize(pil_image, image_size=256):
while min(*pil_image.size) >= 2 * image_size:
pil_image = pil_image.resize(
tuple(x // 2 for x in pil_image.size), resample=Image.BOX
)
scale = image_size / min(*pil_image.size)
pil_image = pil_image.resize(
tuple(round(x * scale) for x in pil_image.size), resample=Image.BICUBIC
)
return pil_image
def center_crop_fn(image, height, width):
crop_x = (image.width - width) // 2
crop_y = (image.height - height) // 2
return image.crop((crop_x, crop_y, crop_x + width, crop_y + height))
def random_crop_fn(image, height, width):
crop_x = random.randint(0, image.width - width)
crop_y = random.randint(0, image.height - height)
return image.crop((crop_x, crop_y, crop_x + width, crop_y + height))
def find_nearest_aspect_ratio_bins(aspect_ratio, aspect_ratio_bins):
min_distance = 1000000
min_index = 0
for i in range(len(aspect_ratio_bins)):
dis = abs(aspect_ratio - aspect_ratio_bins[i])
if dis < min_distance:
min_distance = dis
min_index = i
return min_index
class PackedParquetDataset(IterableDataset):
def __init__(self,
data_sources: dict,
caption_weight: dict,
resolution=256,
random_crop=False,
):
self.data_sources = data_sources
self.resolution = resolution
self.normalize = Normalize(
mean=[0.5, 0.5, 0.5],
std=[0.5, 0.5, 0.5]
)
if random_crop:
self.crop_fn = random_crop_fn
else:
self.crop_fn = center_crop_fn
self.parquet_files = []
for root, repeat in self.data_sources.items():
parquet_files = [os.path.join(root, f) for f in os.listdir(root) if f.endswith('.parquet')]
parquet_files = parquet_files * repeat
self.parquet_files.extend(parquet_files)
print("parquet files: ", len(self.parquet_files))
self.caption_weight = caption_weight
self.prefix_template = [
"A photo of ",
"A picture of ",
"A visual representation of ",
"A image of ",
"A scene of ",
"A view of ",
"A depiction of ",
]
def __iter__(self):
# when seed everything. each work, no matter local or global will have distinct seed!
worker_info = torch.utils.data.get_worker_info()
if worker_info is None:
iter_start = 0
iter_end = len(self.parquet_files)
else:
per_worker = int(len(self.parquet_files) / worker_info.num_workers)
worker_id = worker_info.id
iter_start = worker_id * per_worker
iter_end = iter_start + per_worker if worker_id < worker_info.num_workers - 1 else len(self.parquet_files)
print("iter_start: ", iter_start, "iter_end: ", iter_end, "worker_id: ", worker_info.id, "num_workers: ", worker_info.num_workers, "len: ", len(self.parquet_files))
while True:
index = random.choice(range(iter_start, iter_end))
file = self.parquet_files[index]
table = pq.read_table(file)
# random order
sampled_indices = numpy.random.choice(table.num_rows, size=table.num_rows, replace=False)
sampled_indices = sampled_indices.tolist()
for i in sampled_indices:
metadata = dict()
for cname in table.column_names:
metadata[cname] = table[cname][i].as_py()
# select a caption
caption_key = numpy.random.choice(list(self.caption_weight.keys()), p=list(self.caption_weight.values()))
if caption_key not in metadata:
continue
caption = metadata[caption_key]
# prefix template for short caption
if random.random() < 0.5 and 'long' not in caption_key:
caption = random.choice(self.prefix_template) + caption
image_bytes = metadata.pop('image')
try:
pil_image = Image.open(io.BytesIO(image_bytes))
pil_image = pil_image.convert('RGB')
height, width = pil_image.size
if min(height, width) < self.resolution:
continue
pil_image = resize(pil_image, self.resolution)
pil_image = self.crop_fn(pil_image, self.resolution, self.resolution)
raw_image = to_tensor(pil_image)
normalized_image = self.normalize(raw_image)
metadata = {
"raw_image": raw_image,
"prompt": caption,
}
data = (normalized_image, caption, metadata)
yield copy.deepcopy(data)
except:
print("not ok")
class WebDatasetPackedDataset(IterableDataset):
"""
A highly efficient WebDataset loader for large-scale pre-training.
It streams data from .tar files and is optimized for multi-worker data loading.
This dataset yields tuples of: (normalized_image_tensor, caption_str, metadata_dict)
Args:
urls (Iterable[str]): A list of URLs or glob patterns pointing to .tar files.
resolution (int): The target short-side resolution for image resizing.
random_crop (bool): If True, applies random cropping; otherwise, center cropping.
shuffle_buffer (int): The size of the buffer for shuffling samples. A larger buffer
provides better randomness but uses more memory. 0 to disable.
sample_shuffle (bool): If True, enables sample shuffling within the buffer.
repeat (bool): If True, the dataset will loop indefinitely. Set to True for training.
"""
def __init__(
self,
urls: Iterable[str],
resolution: int = 256,
random_crop: bool = False,
shuffle_buffer: int = 1000,
sample_shuffle: bool = True,
repeat: bool = True,
):
super().__init__()
# Ensure urls is a list, even if a single string is passed
if isinstance(urls, str):
urls = [urls]
print("INFO: Resolving dataset URLs and glob patterns...")
tar_files = []
for url in urls:
tar_files.extend(glob.glob(os.path.join(url, "**/*.tar"), recursive=True))
tar_files.extend(glob.glob(os.path.join(url, "**/*.tar.gz"), recursive=True))
num_tars = len(tar_files)
if num_tars == 0:
raise ValueError(f"No tar files found. Please check your URLs/patterns: {urls}")
print(f"INFO: Found {num_tars} tar files to stream from.")
self.urls = tar_files
self.resolution = resolution
self.normalize = Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])
self.crop_fn = random_crop_fn if random_crop else center_crop_fn
self.shuffle_buffer = shuffle_buffer
self.sample_shuffle = sample_shuffle
self.repeat = repeat
# A simple text augmentation to add variety to short captions
self.prefix_template = [
"A photo of ", "A picture of ", "A visual representation of ",
"A image of ", "A scene of ", "A view of ", "A depiction of ",
]
def _extract_image_from_sample(self, sample: Dict) -> Optional[Image.Image]:
"""Extracts the PIL image from a sample dict using the 'jpg' key."""
if 'jpg' in sample:
img_key = 'jpg'
elif 'output_image' in sample:
img_key = 'output_image'
else:
return None
image_data = sample[img_key]
if isinstance(image_data, Image.Image):
return image_data
if isinstance(image_data, (bytes, bytearray)):
try:
return Image.open(io.BytesIO(image_data))
except Exception:
return None # Return None if bytes are not a valid image
return None
def _extract_caption_from_sample(self, sample: Dict) -> str:
"""Extracts the caption from a sample dict using the 'txt' key."""
if 'txt' in sample:
text_key = 'txt'
elif "input_prompt" in sample:
text_key = 'input_prompt'
else:
return "" # Return an empty string if caption is missing
caption_data = sample[text_key]
if isinstance(caption_data, (bytes, bytearray)):
return caption_data.decode("utf-8", errors="ignore")
if isinstance(caption_data, str):
return caption_data
return str(caption_data)
def _process_pil(self, pil_image: Image.Image):
"""
Processes a PIL image: converts to RGB, resizes, crops, and normalizes.
Returns both the normalized tensor and the pre-normalization tensor.
"""
pil_image = pil_image.convert('RGB')
# Skip images that are smaller than the target resolution
if min(*pil_image.size) < self.resolution:
return None
# Resize, crop, and convert to tensor
pil_image = resize(pil_image, self.resolution)
pil_image = self.crop_fn(pil_image, self.resolution, self.resolution)
raw_image_tensor = to_tensor(pil_image)
normalized_image_tensor = self.normalize(raw_image_tensor)
return normalized_image_tensor, raw_image_tensor
def _make_pipeline(self, worker_id: int, num_workers: int):
"""
Creates the data loading pipeline using webdataset.
This pipeline is optimized for performance in a multi-worker setup.
"""
# `resampled=self.repeat` is a robust way to handle repeating datasets
# and shuffling the order of shards for each epoch.
handler = wds.warn_and_continue
dataset = wds.DataPipeline(
wds.SimpleShardList(self.urls),
# at this point we have an iterator over all the shards
# this shuffles the shards
wds.shuffle(100),
# add wds.split_by_node here if you are using multiple nodes
wds.split_by_worker,
# at this point, we have an iterator over the shards assigned to each worker
wds.tarfile_to_samples(),
# this shuffles the samples in memory
wds.shuffle(self.shuffle_buffer),
# this decodes the images and json
wds.decode("pil", handler=handler),
)
return dataset
def __iter__(self):
"""The iterator method that yields data samples."""
worker_info = torch.utils.data.get_worker_info()
if worker_info is None:
# Single-process data loading
worker_id, num_workers = 0, 1
else:
# Multi-process data loading
worker_id = worker_info.id
num_workers = worker_info.num_workers
# Create a unique pipeline for each worker
pipeline = self._make_pipeline(worker_id, num_workers)
for sample in pipeline:
try:
pil_image = self._extract_image_from_sample(sample)
if pil_image is None:
print("skip image")
continue # Skip if sample has no valid image
processed_data = self._process_pil(pil_image)
if processed_data is None:
continue # Skip if image processing fails (e.g., too small)
img_tensor, raw_image = processed_data
caption = self._extract_caption_from_sample(sample)
# Optional: Add a random prefix to short captions for augmentation
if random.random() < 0.5 and len(caption.split()) < 30:
caption = random.choice(self.prefix_template) + caption
metadata = {
"raw_image": raw_image, # Tensor before normalization
"prompt": caption,
}
yield (img_tensor, caption, metadata)
except GeneratorExit:
# This exception is raised when the consumer of the iterator stops.
raise
except Exception:
# Catch-all for any other errors in a sample to make the stream robust.
# In a real-world scenario, you might want to log this error.
# e.g., print(f"Warning: Skipping a bad sample due to {e}")
print("fail")
continue
class WebDatasetPackedDataset_gpt(IterableDataset):
"""
Stream webdataset (.tar/.gz/...) files using webdataset library.
Yields tuples: (normalized_image_tensor, caption_str, metadata_dict)
Arguments:
- urls: list[str] or str (glob, tar path, or list)
- caption_weight: dict mapping caption-field-name -> probability. If provided, tries to select field accordingly.
- resolution: int target short-side before crop (same semantics as original)
- random_crop: bool
- shuffle_buffer: int for wds.shuffle (0 means no shuffle)
- sample_shuffle: bool when True will call .shuffle() in pipeline
"""
def __init__(
self,
urls: Iterable[str],
caption_weight: Optional[Dict[str, float]] = None,
resolution: int = 256,
random_crop: bool = False,
shuffle_buffer: int = 1000,
sample_shuffle: bool = True,
repeat: bool = True,
):
super().__init__()
if isinstance(urls, str):
urls = [urls]
self.urls = list(urls)
if len(self.urls) == 0:
raise ValueError("No webdataset urls provided.")
self.resolution = resolution
self.normalize = Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])
self.crop_fn = random_crop_fn if random_crop else center_crop_fn
self.caption_weight = caption_weight or {}
self.prefix_template = [
"A photo of ",
"A picture of ",
"A visual representation of ",
"A image of ",
"A scene of ",
"A view of ",
"A depiction of ",
]
self.shuffle_buffer = shuffle_buffer
self.sample_shuffle = sample_shuffle
self.repeat = repeat
def _extract_image_from_sample(self, sample: Dict):
# webdataset with .decode("pil") will decode images into PIL objects for image keys.
# We'll prefer already-decoded images; otherwise try bytes->PIL.
# Look for common keys:
possible_image_keys = ['jpg', 'png', 'jpeg', 'image', 'img', 'data']
for k in possible_image_keys:
if k in sample:
v = sample[k]
if isinstance(v, Image.Image):
return v
if isinstance(v, (bytes, bytearray)):
try:
return Image.open(io.BytesIO(v))
except Exception:
pass
# fallback: find first PIL or bytes among values
for v in sample.values():
if isinstance(v, Image.Image):
return v
if isinstance(v, (bytes, bytearray)):
try:
return Image.open(io.BytesIO(v))
except Exception:
continue
# nothing found
return None
def _extract_caption_from_sample(self, sample: Dict):
# If caption_weight provided, try to choose a key according to probabilities.
if self.caption_weight:
keys = list(self.caption_weight.keys())
probs = list(self.caption_weight.values())
# choose key (might not exist in sample)
chosen_key = random.choices(keys, weights=probs, k=1)[0]
if chosen_key in sample:
val = sample[chosen_key]
if isinstance(val, (bytes, bytearray)):
return val.decode("utf-8", errors="ignore")
return str(val)
# if chosen not present, fallthrough to generic search
# generic search for text-like fields
text_keys = ['txt', 'caption', 'text', 'json', 'meta', 'label']
for k in text_keys:
if k in sample:
v = sample[k]
if isinstance(v, (bytes, bytearray)):
return v.decode("utf-8", errors="ignore")
return str(v)
# fallback: any string/bytes value
for v in sample.values():
if isinstance(v, (bytes, bytearray)):
return v.decode("utf-8", errors="ignore")
if isinstance(v, str):
return v
return "" # empty caption if none found
def _process_pil(self, pil_image: Image.Image):
pil_image = pil_image.convert('RGB')
if min(*pil_image.size) < self.resolution:
return None # skip
pil_image = resize(pil_image, self.resolution)
pil_image = self.crop_fn(pil_image, self.resolution, self.resolution)
raw_image = to_tensor(pil_image)
normalized_image = self.normalize(raw_image)
return normalized_image, raw_image
def _make_pipeline(self, worker_id: int, num_workers: int):
# Build webdataset pipeline; shard by worker_id/num_workers; optional shuffle; decode images to PIL
urls = self.urls
ds = wds.WebDataset(urls).decode("pil")
# shard for multi-worker
if num_workers > 1:
ds = ds.shard(worker_id, num_workers)
if self.sample_shuffle and self.shuffle_buffer > 0:
ds = ds.shuffle(self.shuffle_buffer)
if self.repeat:
ds = ds.repeat() # infinite stream
return ds
def __iter__(self):
worker_info = torch.utils.data.get_worker_info()
if worker_info is None:
worker_id = 0
num_workers = 1
else:
worker_id = worker_info.id
num_workers = worker_info.num_workers
ds = self._make_pipeline(worker_id, num_workers)
# iterate forever (or until process killed)
for sample in ds:
# sample is a dict-like mapping keys->values
try:
pil_image = self._extract_image_from_sample(sample)
if pil_image is None:
continue
caption = self._extract_caption_from_sample(sample)
# optionally add prefix for short captions (replicates original behavior)
if random.random() < 0.5 and len(caption.split()) < 30:
caption = random.choice(self.prefix_template) + caption
img_tensor, raw_image = self._process_pil(pil_image)
if img_tensor is None:
continue
metadata = {"raw_image": raw_image, "prompt": caption,}
yield (img_tensor, caption, metadata)
except GeneratorExit:
raise
except Exception:
# ignore single-bad-sample errors to keep stream robust
continue
|