Spaces:
Sleeping
Sleeping
File size: 11,637 Bytes
0fd26a8 | 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 | """
WebDataset distributed utility functions, pipeline helper functions and sampler classes.
"""
from torch.utils.data import IterableDataset
import torch
import math
import random
import os
import logging
import braceexpand
def log_and_continue(exn):
"""Call in an exception handler to ignore any exception, issue a warning, and continue."""
if "No images in sample" in str(exn) or "Only one image in sample" in str(exn):
return True
if isinstance(exn, FileNotFoundError) or "FileNotFoundError" in str(type(exn)):
if os.environ.get("RANK", "0") == "0":
logging.warning(f"Handling webdataset FileNotFoundError: {exn}. Ignoring and continuing.")
return True
logging.warning(f"Handling webdataset error ({repr(exn)}). Ignoring.")
return True
# Distributed environment detection and shard allocation
def pytorch_worker_info(group=None):
"""Return node and worker info for PyTorch and some distributed environments."""
rank = 0
world_size = 1
worker = 0
num_workers = 1
if "RANK" in os.environ and "WORLD_SIZE" in os.environ:
rank = int(os.environ["RANK"])
world_size = int(os.environ["WORLD_SIZE"])
else:
try:
import torch.distributed
if torch.distributed.is_available() and torch.distributed.is_initialized():
group = group or torch.distributed.group.WORLD
rank = torch.distributed.get_rank(group=group)
world_size = torch.distributed.get_world_size(group=group)
except ModuleNotFoundError:
pass
if "WORKER" in os.environ and "NUM_WORKERS" in os.environ:
worker = int(os.environ["WORKER"])
num_workers = int(os.environ["NUM_WORKERS"])
else:
try:
import torch.utils.data
worker_info = torch.utils.data.get_worker_info()
if worker_info is not None:
worker = worker_info.id
num_workers = worker_info.num_workers
except ModuleNotFoundError:
pass
return rank, world_size, worker, num_workers
def is_multi_node_environment():
"""
check if in a multi-process (world_size > 1) environment.
"""
try:
import torch.distributed as dist
if dist.is_available() and dist.is_initialized():
if dist.get_world_size() > 1:
return True
except Exception:
pass
world_size = int(os.environ.get("WORLD_SIZE", os.environ.get("SLURM_NTASKS", "1")))
nnodes = int(os.environ.get("NNODES", os.environ.get("SLURM_NNODES", "1")))
if nnodes > 1:
return True
return world_size > 1
def split_data_by_node(urls, strategy="interleaved"):
"""split shards between nodes, even if the data is stored locally, it is recommended to use it to avoid duplicate training."""
print('*'*80)
print("split_data_by_node ing..................")
gpus_per_node = torch.cuda.device_count()
rank, world_size, worker, num_workers = pytorch_worker_info()
print("rank: {}, world_size: {}, worker: {}, num_workers: {}, gpus_per_node: {}".format(
rank, world_size, worker, num_workers, gpus_per_node))
node_rank = rank // gpus_per_node
node_world_size = world_size // gpus_per_node
if len(urls) < node_world_size:
print(f"Warning: Only {len(urls)} shards but {node_world_size} nodes. "
f"All nodes will use all shards to avoid empty assignment.")
print(f"Node {node_rank} has {len(urls)} URLs of {len(urls)} total.")
print('*'*80)
return urls
if strategy == "chunk":
urls_per_node = math.ceil(len(urls) / node_world_size)
start_idx = node_rank * urls_per_node
end_idx = min(start_idx + urls_per_node, len(urls))
node_urls = urls[start_idx:end_idx]
elif strategy == "interleaved":
node_urls = urls[node_rank::node_world_size]
elif strategy == "shuffled_chunk":
shuffled_urls = random.sample(urls, len(urls))
urls_per_node = math.ceil(len(shuffled_urls) / node_world_size)
start_idx = node_rank * urls_per_node
end_idx = min(start_idx + urls_per_node, len(urls))
node_urls = shuffled_urls[start_idx:end_idx]
else:
raise ValueError(f"Unknown strategy {strategy}")
print(f"Node {node_rank} has {len(node_urls)} URLs of {len(urls)} total.")
print('*'*80)
return node_urls
def get_dataset_size(shards, estimated_sample_per_shard=1000):
"""estimate the dataset size, based on the number of shards."""
if ',' in shards:
shards_list = []
for pattern in shards.split(','):
pattern = pattern.strip()
if not pattern:
continue
shards_list.extend(list(braceexpand.braceexpand(pattern)))
else:
shards_list = list(braceexpand.braceexpand(shards))
num_shards = len(shards_list)
total_size = num_shards * estimated_sample_per_shard
print(f"Estimating dataset size: {total_size} samples ({num_shards} shards * {estimated_sample_per_shard} samples/shard)")
return total_size, num_shards
# Pipeline helper functions (module level, supports pickle/spawn)
def nodesplitter_identity(urls):
return urls
def handle_reconstruction_task(sample, handler=log_and_continue):
in_key = None
if "in.png" in sample:
in_key = "in.png"
elif "in.jpg" in sample:
in_key = "in.jpg"
out_key = None
if "out.png" in sample:
out_key = "out.png"
elif "out.jpg" in sample:
out_key = "out.jpg"
if in_key and not out_key:
if in_key == "in.png":
sample["out.png"] = sample["in.png"]
else:
sample["out.jpg"] = sample["in.jpg"]
return sample
def extract_fields_to_tuple(sample, handler=log_and_continue):
in_img = sample.get("in.png") or sample.get("in.jpg")
out_img = sample.get("out.png") or sample.get("out.jpg")
if out_img is None and in_img is not None:
out_img = in_img
sample_type = sample.get("type", None)
return (in_img, out_img, sample_type)
def identity_function(x, handler=log_and_continue):
return x
def has_input_image(sample):
return "in.png" in sample or "in.jpg" in sample
class WeightedRoundRobinSampler(IterableDataset):
def __init__(self, pipelines, weights):
super().__init__()
if len(weights) != len(pipelines):
raise ValueError(f"number of weights ({len(weights)}) must be equal to the number of pipelines ({len(pipelines)})")
self.pipelines = pipelines
self.weights = weights
total_weight = sum(weights)
normalized_weights = [w / total_weight for w in weights]
max_decimal_places = max(len(str(w).split('.')[-1]) if '.' in str(w) else 0 for w in normalized_weights)
scale_factor = 10 ** max_decimal_places
int_weights = [int(w * scale_factor) for w in normalized_weights]
def gcd(a, b):
while b:
a, b = b, a % b
return a
def gcd_list(nums):
result = nums[0]
for num in nums[1:]:
result = gcd(result, num)
return result
common_divisor = gcd_list(int_weights)
int_weights = [w // common_divisor for w in int_weights]
self.sampling_sequence = []
for i, weight in enumerate(int_weights):
self.sampling_sequence.extend([i] * weight)
def __iter__(self):
import itertools
iterators = [iter(p) for p in self.pipelines]
sequence_iter = itertools.cycle(self.sampling_sequence)
active = [True] * len(iterators)
while True:
if not any(active):
break
idx = next(sequence_iter)
if active[idx]:
try:
yield next(iterators[idx])
except StopIteration:
active[idx] = False
if not any(active):
break
continue
class StrictProportionalBatchSampler(IterableDataset):
"""
a strictly proportional batch sampler (适用于 resampled=True)
ensure that the samples in each batch are strictly allocated according to the weight ratio
"""
def __init__(self, pipelines, weights, batch_size):
super().__init__()
if len(weights) != len(pipelines):
raise ValueError(f"number of weights ({len(weights)}) must be equal to the number of pipelines ({len(pipelines)})")
self.pipelines = pipelines
self.weights = weights
self.batch_size = batch_size
total_weight = sum(weights)
normalized_weights = [w / total_weight for w in weights]
self.samples_per_pipeline = []
float_counts = [batch_size * w for w in normalized_weights]
int_counts = [round(c) for c in float_counts]
current_sum = sum(int_counts)
diff = batch_size - current_sum
if diff != 0:
errors = [(float_counts[i] - int_counts[i], i) for i in range(len(int_counts))]
errors.sort(reverse=(diff > 0))
for _ in range(abs(diff)):
_, idx = errors.pop(0)
int_counts[idx] += 1 if diff > 0 else -1
self.samples_per_pipeline = int_counts
weight_strs = [f"{w*100:.1f}%" for w in normalized_weights]
sample_strs = [f"{count}" for count in self.samples_per_pipeline]
actual_ratios = [f"{count/batch_size*100:.1f}%" for count in self.samples_per_pipeline]
print(f"Strict proportional batch sampling enabled:")
print(f" Target weights: {' : '.join(weight_strs)}")
print(f" Actual samples per batch: {' : '.join(sample_strs)} (total={batch_size})")
print(f" Actual ratios: {' : '.join(actual_ratios)}")
def __iter__(self):
import random as _random
iterators = [iter(p) for p in self.pipelines]
while True:
batch_samples = []
for idx, count in enumerate(self.samples_per_pipeline):
for _ in range(count):
sample = next(iterators[idx])
batch_samples.append(sample)
_random.shuffle(batch_samples)
normalized_samples = []
for sample in batch_samples:
if len(sample) == 3:
normalized_samples.append((sample[0], sample[1], sample[2], None))
elif len(sample) == 4:
normalized_samples.append(sample)
else:
raise ValueError(f"Unexpected sample length: {len(sample)}")
batch_transposed = list(zip(*normalized_samples))
batch_results = []
for idx, items in enumerate(batch_transposed):
if idx < 3:
filtered_items = [item for item in items if item is not None]
if len(filtered_items) != len(items):
raise ValueError(f"Found None in tensor items at index {idx}")
batch_results.append(torch.stack(list(filtered_items)))
else:
type_list = list(items)
batch_results.append(type_list)
yield tuple(batch_results)
|