File size: 17,583 Bytes
9c6594c |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 |
#
# Copyright (c) 2017-2021 NVIDIA CORPORATION. All rights reserved.
# This file is part of the WebDataset library.
# See the LICENSE file for licensing terms (BSD-style).
#
"""Train PyTorch models directly from POSIX tar archive.
Code works locally or over HTTP connections.
"""
import glob
import os
import os.path
import random
import re
import sys
import time
from dataclasses import dataclass, field
from itertools import islice
from typing import List
import braceexpand
import yaml
from . import utils
from .filters import pipelinefilter
from .pytorch import IterableDataset
from .utils import is_iterable, obsolete
def envlookup(m):
"""Look up match in the environment with prefix WDS_.
Args:
m: A match object.
Returns:
str: The value of the environment variable WDS_<m.group(1)>.
Raises:
AssertionError: If the environment variable is not found.
"""
key = m.group(1)
key = "WDS_" + key
assert key in os.environ, f"missing environment variable wds_{key}"
return os.environ[key]
def envsubst(s):
"""Substitute ${var} with the value of the environment variable WDS_var.
Args:
s (str): String to be substituted.
Returns:
str: The substituted string.
"""
return re.sub(r"\$\{(\w+)\}", envlookup, s)
def split_by_node(src, group=None):
"""Split the input sequence by PyTorch distributed rank.
Args:
src: The input sequence to be split.
group: The process group for distributed training.
Yields:
Elements from the input sequence based on the node's rank.
"""
rank, world_size, worker, num_workers = utils.pytorch_worker_info(group=group)
if world_size > 1:
yield from islice(src, rank, None, world_size)
else:
yield from src
def single_node_only(src, group=None):
"""Ensure the input sequence is not split for multi-node training.
Args:
src: The input sequence.
group: The process group for distributed training.
Yields:
Elements from the input sequence.
Raises:
ValueError: If multi-node training is detected.
"""
rank, world_size, worker, num_workers = utils.pytorch_worker_info(group=group)
if world_size > 1:
raise ValueError("you need to add an explicit nodesplitter to your input pipeline for multi-node training")
yield from src
def split_by_worker(src):
"""Split the input sequence by PyTorch DataLoader worker.
Args:
src: The input sequence to be split.
Yields:
Elements from the input sequence based on the worker's ID.
"""
rank, world_size, worker, num_workers = utils.pytorch_worker_info()
if num_workers > 1:
yield from islice(src, worker, None, num_workers)
else:
yield from src
def expand_urls(urls): # sourcery skip: for-index-underscore, last-if-guard
"""Expand the urls if they are a string.
If input is a string:
- split on '::'
- expand environment variables (using WDS_ prefix)
- expand braces
Otherwise:
- return the input as a list
Args:
urls (str or List[str]): URL list or URL string.
Returns:
List[str]: List of expanded URLs.
"""
urllist = urls.split("::")
result = []
for url in urllist:
for i in range(10):
last = url
url = envsubst(url)
if url == last:
break
result.extend(braceexpand.braceexpand(url))
return result
def expand_source(source, max_urls=int(1e9)):
"""Expand the given source into a list of URLs.
Args:
source (str or List[str] or Iterable): The source to be expanded.
max_urls (int): Maximum number of URLs to return.
Returns:
List[str]: List of expanded URLs.
Raises:
ValueError: If the source type is not supported.
"""
if isinstance(source, str):
return expand_urls(source)
elif isinstance(source, list):
return source
elif is_iterable(source):
return list(islice(source, max_urls))
else:
raise ValueError(f"cannot handle {type(source)}")
class SimpleShardList(IterableDataset):
"""An iterable dataset yielding a list of URLs."""
def __init__(self, urls, seed=None):
"""Initialize the SimpleShardList.
Args:
urls (str or List[str]): A list of URLs as a Python list or brace notation string.
seed (int or bool or None): Random seed for shuffling; if None, no shuffling is done,
if True, a random seed is generated.
"""
super().__init__()
if isinstance(urls, str):
urls = expand_urls(urls)
else:
urls = list(urls)
self.urls = urls
assert isinstance(self.urls[0], str)
if seed is True:
seed = time.time()
self.seed = seed
def __len__(self):
"""Return the number of URLs in the list.
Returns:
int: The number of URLs.
"""
return len(self.urls)
def __iter__(self):
"""Return an iterator over the shards.
Yields:
dict: A dictionary containing the URL of each shard.
"""
urls = self.urls.copy()
if self.seed is not None:
random.Random(self.seed).shuffle(urls)
for url in urls:
yield dict(url=url)
def resampled_(src, n=sys.maxsize):
"""Resample items from the source with replacement.
Args:
src: The source iterable.
n (int): The number of items to yield. Defaults to sys.maxsize.
Yields:
Randomly chosen items from the source.
"""
import random
seed = time.time()
try:
seed = open("/dev/random", "rb").read(20)
except Exception as exn:
print(repr(exn)[:50], file=sys.stderr)
rng = random.Random(seed)
print("# resampled loading", file=sys.stderr)
items = list(src)
print(f"# resampled got {len(items)} samples, yielding {n}", file=sys.stderr)
for _ in range(n):
yield rng.choice(items)
resampled = pipelinefilter(resampled_)
def non_empty(src):
"""Ensure the source yields at least one item.
Args:
src: The source iterable.
Yields:
Items from the source.
Raises:
ValueError: If the source yields no items.
"""
count = 0
for s in src:
yield s
count += 1
if count == 0:
raise ValueError("pipeline stage received no data at all and this was declared as an error")
@dataclass
class MSSource:
"""Class representing a data source."""
name: str = ""
perepoch: int = -1
resample: bool = False
urls: List[str] = field(default_factory=list)
default_rng = random.Random()
def expand(s):
"""Expand user and environment variables in a string.
Args:
s (str): The string to expand.
Returns:
str: The expanded string.
"""
return os.path.expanduser(os.path.expandvars(s))
class ResampledShards(IterableDataset):
"""An iterable dataset yielding a list of URLs with resampling."""
def __init__(
self,
urls,
nshards=sys.maxsize,
seed=0,
worker_seed=None,
deterministic=False,
max_urls=int(1e6),
empty_check=True,
):
"""Initialize the ResampledShards.
Args:
urls: A list of URLs as a Python list or brace notation string.
nshards (int): The number of shards to yield. Defaults to sys.maxsize.
seed (int): The seed for random number generation.
worker_seed (Callable or None): A function to generate worker-specific seeds.
deterministic (bool): Whether to use deterministic sampling.
max_urls (int): Maximum number of URLs to consider.
empty_check (bool): Whether to check for empty URL list.
Raises:
ValueError: If empty_check is True and no shards are found.
"""
super().__init__()
self.urls = expand_source(urls, max_urls)
if empty_check:
if len(self.urls) == 0:
raise ValueError("empty_check=True, but no shards found in ResampledShards")
assert isinstance(self.urls[0], str)
self.nshards = nshards
self.worker_seed = utils.pytorch_worker_seed if worker_seed is None else worker_seed
self.deterministic = deterministic
self.seed = seed
self.epoch = -1
def __iter__(self):
"""Return an iterator over the shards.
Yields:
dict: A dictionary containing the URL of each shard.
"""
self.epoch += 1
if self.deterministic:
seed = utils.make_seed(self.worker_seed(), self.epoch, self.seed)
else:
seed = utils.make_seed(
self.worker_seed(),
self.epoch,
self.seed,
os.getpid(),
time.time_ns(),
os.urandom(4),
)
if os.environ.get("WDS_SHOW_SEED", "0") == "1":
print(f"# ResampledShards seed {seed}")
self.rng = random.Random(seed)
for _ in range(self.nshards):
index = self.rng.randint(0, len(self.urls) - 1)
yield dict(url=self.urls[index])
ResampledShardList = ResampledShards
def check_pid_is_running(pid):
"""Check for the existence of a Unix PID.
Args:
pid (int): The process ID to check.
Returns:
bool: True if the process is running, False otherwise.
"""
try:
os.kill(pid, 0)
except OSError:
return False
else:
return True
def without_last_extension(fname):
"""Remove the last extension from a filename.
Args:
fname (str): The filename to process.
Returns:
str: The filename without the last extension.
"""
return re.sub(r"\.[^.]*$", "", fname)
def get_pid_from_filename(fname):
"""Get the PID from a filename.
Args:
fname (str): The filename to process.
Returns:
int or None: The PID if found in the filename, None otherwise.
"""
match = re.match(r"^(.*)\._(\d+)_$", fname)
if not match:
return None
return int(match.group(2))
class DirectoryShardList(IterableDataset):
"""An iterable dataset that yields shards from a directory."""
def __init__(
self,
path,
pattern="*.{tar,tgz,tar.tgz}",
poll=1,
timeout=1e12,
mode="resample",
select="random",
fate=None,
):
"""Initialize the DirectoryShardList.
Args:
path (str): The directory path to monitor for shards.
pattern (str): The glob pattern to match shard files.
poll (int): The polling interval in seconds.
timeout (float): The maximum time to wait for new shards.
mode (str): The mode for handling processed shards.
select (str): The strategy for selecting shards.
fate (Any): Currently unused parameter.
"""
assert path.endswith("/")
assert os.path.isdir(path)
self.path = path
self.poll = poll
self.pattern = pattern
self.mode = mode
self.select = select
self.fate = fate
self.timeout = timeout
def recycle(self, activename):
"""Recycle a processed shard based on the current mode.
Args:
activename (str): The name of the active shard file.
"""
if self.mode == "unlink":
os.unlink(activename)
elif self.mode == "keep":
os.rename(activename, without_last_extension(activename) + "._done_")
elif self.mode == "resample":
os.rename(activename, without_last_extension(activename))
def cleanup_files_without_processes(self):
"""Clean up shard files associated with non-existent processes."""
for fname in glob.glob(os.path.join(self.path, "*._*_")):
pid = get_pid_from_filename(fname)
if pid is None:
continue
if not check_pid_is_running(pid):
self.recycle(fname)
def __iter__(self):
"""Iterate over the shards in the directory.
Yields:
dict: A dictionary containing the URL of each shard.
"""
last = time.time()
while time.time() - last < self.timeout:
candidates = sorted(glob.glob(self.path + self.pattern))
if len(candidates) == 0:
if self.poll is None:
return
time.sleep(self.poll)
continue
if self.select == "oldest":
candidate = min(candidates, key=lambda fn: os.stat(fn).st_mtime)
elif self.select == "random":
candidate = random.choice(candidates)
else:
raise ValueError(f"unknown selection strategy {self.select}")
activename = candidate + f"._{os.getpid()}_"
try:
os.rename(candidate, activename)
except FileNotFoundError:
time.sleep(self.poll)
continue
yield dict(url=activename)
self.recycle(activename)
self.cleanup_files_without_processes()
class MultiShardSample(IterableDataset):
"""An iterable dataset that samples from multiple shard sources."""
@obsolete(reason="this is going to be replaced with the WIDS JSON format")
def __init__(self, fname):
"""Initialize the MultiShardSample.
Args:
fname (str or dict): The filename of the YAML spec or a dictionary containing the spec.
"""
self.epoch = -1
self.parse_spec(fname)
def parse_spec(self, fname):
"""Parse the specification for multiple shard sources.
Args:
fname (str or dict): The filename of the YAML spec or a dictionary containing the spec.
"""
self.rng = default_rng # capture default_rng if we fork
if isinstance(fname, dict):
spec = fname
fname = "{dict}"
else:
with open(fname) as stream:
spec = yaml.safe_load(stream)
assert set(spec.keys()).issubset(set("prefix datasets buckets".split())), list(spec.keys())
prefix = expand(spec.get("prefix", ""))
self.sources = []
for ds in spec["datasets"]:
assert set(ds.keys()).issubset(set("buckets name shards resample choose".split())), list(ds.keys())
buckets = ds.get("buckets", spec.get("buckets", []))
if isinstance(buckets, str):
buckets = [buckets]
buckets = [expand(s) for s in buckets]
if buckets == []:
buckets = [""]
assert len(buckets) == 1, f"{buckets}: FIXME support for multiple buckets unimplemented"
bucket = buckets[0]
name = ds.get("name", "@" + bucket)
urls = ds["shards"]
if isinstance(urls, str):
urls = [urls]
# urls = [u for url in urls for u in braceexpand.braceexpand(url)]
urls = [prefix + os.path.join(bucket, u) for url in urls for u in braceexpand.braceexpand(expand(url))]
resample = ds.get("resample", -1)
nsample = ds.get("choose", -1)
if nsample > len(urls):
raise ValueError(f"perepoch {nsample} must be no greater than the number of shards")
if (nsample > 0) and (resample > 0):
raise ValueError("specify only one of perepoch or choose")
entry = MSSource(name=name, urls=urls, perepoch=nsample, resample=resample)
self.sources.append(entry)
print(f"# {name} {len(urls)} {nsample}", file=sys.stderr)
def set_epoch(self, seed):
"""Set the current epoch for consistent shard selection among nodes.
Args:
seed (int): The seed for the random number generator.
"""
self.rng = random.Random(seed)
def get_shards_for_epoch(self):
"""Get the list of shards for the current epoch.
Returns:
list: A list of shard URLs for the current epoch.
"""
result = []
for source in self.sources:
if source.resample > 0:
# sample with replacement
l = self.rng.choices(source.urls, k=source.resample)
elif source.perepoch > 0:
# sample without replacement
l = list(source.urls)
self.rng.shuffle(l)
l = l[: source.perepoch]
else:
l = list(source.urls)
result += l
self.rng.shuffle(result)
return result
def __iter__(self):
"""Iterate over the shards for the current epoch.
Yields:
dict: A dictionary containing the URL of each shard.
"""
shards = self.get_shards_for_epoch()
for shard in shards:
yield dict(url=shard)
def shardspec(spec):
"""Create a shard list based on the given specification.
Args:
spec (str): The specification for creating the shard list.
Returns:
IterableDataset: Either a MultiShardSample or a SimpleShardList based on the spec.
"""
if spec.endswith(".yaml"):
return MultiShardSample(spec)
else:
return SimpleShardList(spec)
|