|
14 | 14 | dataloader frequently to avoid duplicate batches. |
15 | 15 | """ |
16 | 16 |
|
17 | | -from typing import Optional |
| 17 | +import math |
| 18 | +from collections.abc import Iterator |
| 19 | +from typing import Iterable, Optional, TypeVar, Union |
18 | 20 |
|
| 21 | +import torch |
19 | 22 | import torch.distributed as dist |
20 | 23 | from torch.utils import data |
| 24 | +from torch.utils.data.dataset import Dataset |
| 25 | +from torch.utils.data.sampler import BatchSampler, Sampler |
| 26 | + |
| 27 | +_T_co = TypeVar("_T_co", covariant=True) |
| 28 | + |
| 29 | + |
| 30 | +class SkipDistributedSampler(Sampler[_T_co]): |
| 31 | + def __init__( |
| 32 | + self, |
| 33 | + dataset: Dataset, |
| 34 | + num_replicas: Optional[int] = None, |
| 35 | + rank: Optional[int] = None, |
| 36 | + shuffle: bool = True, |
| 37 | + seed: int = 0, |
| 38 | + drop_last: bool = False, |
| 39 | + skip_samples: int = 0, |
| 40 | + ) -> None: |
| 41 | + if num_replicas is None: |
| 42 | + if not dist.is_available(): |
| 43 | + raise RuntimeError("Requires distributed package to be available") |
| 44 | + num_replicas = dist.get_world_size() |
| 45 | + if rank is None: |
| 46 | + if not dist.is_available(): |
| 47 | + raise RuntimeError("Requires distributed package to be available") |
| 48 | + rank = dist.get_rank() |
| 49 | + if rank >= num_replicas or rank < 0: |
| 50 | + raise ValueError( |
| 51 | + f"Invalid rank {rank}, rank should be in the interval [0, {num_replicas - 1}]" |
| 52 | + ) |
| 53 | + self.dataset = dataset |
| 54 | + self.num_replicas = num_replicas |
| 55 | + self.rank = rank |
| 56 | + self.epoch = 0 |
| 57 | + self.drop_last = drop_last |
| 58 | + self.skip_samples = skip_samples |
| 59 | + # If the dataset length is evenly divisible by # of replicas, then there |
| 60 | + # is no need to drop any data, since the dataset will be split equally. |
| 61 | + if self.drop_last and len(self.dataset) % self.num_replicas != 0: # type: ignore[arg-type] |
| 62 | + # Split to nearest available length that is evenly divisible. |
| 63 | + # This is to ensure each rank receives the same amount of data when |
| 64 | + # using this Sampler. |
| 65 | + self.num_samples = math.ceil( |
| 66 | + (len(self.dataset) - self.skip_samples - self.num_replicas) |
| 67 | + / self.num_replicas # type: ignore[arg-type] |
| 68 | + ) |
| 69 | + else: |
| 70 | + self.num_samples = math.ceil( |
| 71 | + (len(self.dataset) - self.skip_samples) / self.num_replicas |
| 72 | + ) # type: ignore[arg-type] |
| 73 | + self.total_size = self.num_samples * self.num_replicas |
| 74 | + self.shuffle = shuffle |
| 75 | + self.seed = seed |
| 76 | + |
| 77 | + def __iter__(self) -> Iterator[_T_co]: |
| 78 | + if self.shuffle: |
| 79 | + # deterministically shuffle based on epoch and seed |
| 80 | + g = torch.Generator() |
| 81 | + g.manual_seed(self.seed + self.epoch) |
| 82 | + indices = torch.randperm(len(self.dataset), generator=g).tolist() # type: ignore[arg-type] |
| 83 | + else: |
| 84 | + indices = list(range(len(self.dataset))) # type: ignore[arg-type] |
| 85 | + |
| 86 | + if not self.drop_last: |
| 87 | + indices = indices[self.skip_samples : len(indices)] |
| 88 | + # add extra samples to make it evenly divisible |
| 89 | + padding_size = self.total_size - len(indices) |
| 90 | + if padding_size <= len(indices): |
| 91 | + indices += indices[:padding_size] |
| 92 | + else: |
| 93 | + indices += (indices * math.ceil(padding_size / len(indices)))[ |
| 94 | + :padding_size |
| 95 | + ] |
| 96 | + else: |
| 97 | + # remove tail of data to make it evenly divisible. |
| 98 | + indices = indices[self.skip_samples : self.skip_samples + self.total_size] |
| 99 | + if len(indices) != self.total_size: |
| 100 | + raise AssertionError( |
| 101 | + f"Number of indices ({len(indices)}) does not match total_size ({self.total_size})" |
| 102 | + ) |
| 103 | + |
| 104 | + # subsample |
| 105 | + indices = indices[self.rank : self.total_size : self.num_replicas] |
| 106 | + if len(indices) != self.num_samples: |
| 107 | + raise AssertionError( |
| 108 | + f"Number of subsampled indices ({len(indices)}) does not match num_samples ({self.num_samples})" |
| 109 | + ) |
| 110 | + |
| 111 | + # pyrefly: ignore # bad-return |
| 112 | + return iter(indices) |
| 113 | + |
| 114 | + def __len__(self) -> int: |
| 115 | + return self.num_samples |
| 116 | + |
| 117 | + def set_epoch(self, epoch: int) -> None: |
| 118 | + r""" |
| 119 | + Set the epoch for this sampler. |
| 120 | +
|
| 121 | + When :attr:`shuffle=True`, this ensures all replicas |
| 122 | + use a different random ordering for each epoch. Otherwise, the next iteration of this |
| 123 | + sampler will yield the same ordering. |
| 124 | +
|
| 125 | + Args: |
| 126 | + epoch (int): Epoch number. |
| 127 | + """ |
| 128 | + self.epoch = epoch |
| 129 | + |
| 130 | + |
| 131 | +class DistributedBatchSampler(Sampler[list[int]]): |
| 132 | + r"""Wraps a BatchSampler to distribute batches across multiple processes in distributed training. |
| 133 | +
|
| 134 | + Each process gets a subset of batches based on its rank and the total number of replicas. |
| 135 | + This is useful for distributed training where each process should work on different batches |
| 136 | + to avoid data duplication. |
| 137 | +
|
| 138 | + Args: |
| 139 | + sampler (Sampler or Iterable): Base sampler. Can be any iterable object |
| 140 | + batch_size (int): Size of mini-batch. |
| 141 | + drop_last (bool): If ``True``, the sampler will drop the last batch if |
| 142 | + its size would be less than ``batch_size`` |
| 143 | + num_replicas (int): Number of processes participating in distributed training. |
| 144 | + rank (int): Rank of the current process within num_replicas. |
| 145 | + Should be in range [0, num_replicas - 1]. |
| 146 | + even_batches (bool): If ``True``, ensures all ranks get exactly the same number |
| 147 | + of batches by potentially dropping some batches. If ``False``, some ranks |
| 148 | + may get one extra batch. Default: ``True``. |
| 149 | +
|
| 150 | + Example: |
| 151 | + >>> # For a dataset with indices 0-20, batch_size=2, num_replicas=2 |
| 152 | + >>> # All batches would be: [[0,1], [2,3], [4,5], [6,7], [8,9], [10,11], ...] |
| 153 | + >>> |
| 154 | + >>> # With even_batches=False (original behavior): |
| 155 | + >>> # rank=0 gets batches: [[0,1], [4,5], [8,9], [12,13], [16,17], [20]] (6 batches) |
| 156 | + >>> # rank=1 gets batches: [[2,3], [6,7], [10,11], [14,15], [18,19]] (5 batches) |
| 157 | + >>> sampler_rank0 = DistributedBatchSampler( |
| 158 | + ... SequentialSampler(range(21)), batch_size=2, drop_last=False, |
| 159 | + ... num_replicas=2, rank=0, even_batches=False |
| 160 | + ... ) |
| 161 | + >>> list(sampler_rank0) |
| 162 | + [[0, 1], [4, 5], [8, 9], [12, 13], [16, 17], [20]] |
| 163 | + >>> |
| 164 | + >>> # With even_batches=True (default behavior): |
| 165 | + >>> # Both ranks get exactly 5 batches (drops the last batch [20]) |
| 166 | + >>> # rank=0 gets batches: [[0,1], [4,5], [8,9], [12,13], [16,17]] (5 batches) |
| 167 | + >>> # rank=1 gets batches: [[2,3], [6,7], [10,11], [14,15], [18,19]] (5 batches) |
| 168 | + >>> sampler_rank0_even = DistributedBatchSampler( |
| 169 | + ... SequentialSampler(range(21)), batch_size=2, drop_last=False, |
| 170 | + ... num_replicas=2, rank=0, even_batches=True |
| 171 | + ... ) |
| 172 | + >>> list(sampler_rank0_even) |
| 173 | + [[0, 1], [4, 5], [8, 9], [12, 13], [16, 17]] |
| 174 | + """ |
| 175 | + |
| 176 | + def __init__( |
| 177 | + self, |
| 178 | + sampler: Union[Sampler[int], Iterable[int]], |
| 179 | + batch_size: int, |
| 180 | + drop_last: bool, |
| 181 | + num_replicas: int = 1, |
| 182 | + rank: int = 0, |
| 183 | + even_batches: bool = True, |
| 184 | + ) -> None: |
| 185 | + # Validate batch_size |
| 186 | + if ( |
| 187 | + not isinstance(batch_size, int) |
| 188 | + or isinstance(batch_size, bool) |
| 189 | + or batch_size <= 0 |
| 190 | + ): |
| 191 | + raise ValueError( |
| 192 | + f"batch_size should be a positive integer value, but got batch_size={batch_size}" |
| 193 | + ) |
| 194 | + |
| 195 | + # Validate drop_last |
| 196 | + if not isinstance(drop_last, bool): |
| 197 | + raise ValueError( |
| 198 | + f"drop_last should be a boolean value, but got drop_last={drop_last}" |
| 199 | + ) |
| 200 | + |
| 201 | + # Validate num_replicas |
| 202 | + if not isinstance(num_replicas, int) or num_replicas <= 0: |
| 203 | + raise ValueError( |
| 204 | + f"num_replicas should be a positive integer value, but got num_replicas={num_replicas}" |
| 205 | + ) |
| 206 | + |
| 207 | + # Validate rank |
| 208 | + if not isinstance(rank, int) or rank < 0 or rank >= num_replicas: |
| 209 | + raise ValueError( |
| 210 | + f"rank should be an integer in range [0, {num_replicas - 1}], but got rank={rank}" |
| 211 | + ) |
| 212 | + |
| 213 | + # Validate even_batches |
| 214 | + if not isinstance(even_batches, bool): |
| 215 | + raise ValueError( |
| 216 | + f"even_batches should be a boolean value, but got even_batches={even_batches}" |
| 217 | + ) |
| 218 | + |
| 219 | + self.sampler = sampler |
| 220 | + self.batch_size = batch_size |
| 221 | + self.drop_last = drop_last |
| 222 | + self.num_replicas = num_replicas |
| 223 | + self.rank = rank |
| 224 | + self.even_batches = even_batches |
| 225 | + |
| 226 | + # Create a BatchSampler to generate all batches |
| 227 | + self.batch_sampler = BatchSampler(sampler, batch_size, drop_last) |
| 228 | + |
| 229 | + def __iter__(self) -> Iterator[list[int]]: |
| 230 | + if self.even_batches: |
| 231 | + # When even_batches=True, ensure all ranks get the same number of batches |
| 232 | + # by potentially dropping some batches |
| 233 | + all_batches = list(self.batch_sampler) |
| 234 | + total_batches = len(all_batches) |
| 235 | + |
| 236 | + # Calculate how many batches each rank should get to make them even |
| 237 | + batches_per_rank = total_batches // self.num_replicas |
| 238 | + |
| 239 | + # Only consider the first batches_per_rank * num_replicas batches |
| 240 | + # This ensures even distribution |
| 241 | + total_even_batches = batches_per_rank * self.num_replicas |
| 242 | + |
| 243 | + batch_idx = 0 |
| 244 | + for batch in all_batches: |
| 245 | + if batch_idx >= total_even_batches: |
| 246 | + # Stop yielding once we've exhausted the even batches |
| 247 | + break |
| 248 | + # Only yield batches that belong to current rank |
| 249 | + if batch_idx % self.num_replicas == self.rank: |
| 250 | + yield batch |
| 251 | + batch_idx += 1 |
| 252 | + else: |
| 253 | + # Original behavior when even_batches=False |
| 254 | + batch_idx = 0 |
| 255 | + for batch in self.batch_sampler: |
| 256 | + # Only yield batches that belong to current rank |
| 257 | + if batch_idx % self.num_replicas == self.rank: |
| 258 | + yield batch |
| 259 | + batch_idx += 1 |
| 260 | + |
| 261 | + def __len__(self) -> int: |
| 262 | + # Calculate total number of batches from BatchSampler |
| 263 | + total_batches = len(self.batch_sampler) # type: ignore[arg-type] |
| 264 | + |
| 265 | + if self.even_batches: |
| 266 | + # When even_batches=True, all ranks get exactly the same number of batches |
| 267 | + return total_batches // self.num_replicas |
| 268 | + else: |
| 269 | + # Original behavior when even_batches=False |
| 270 | + # Each rank gets approximately total_batches // num_replicas batches |
| 271 | + # The remaining batches are distributed among the first few ranks |
| 272 | + batches_per_rank = total_batches // self.num_replicas |
| 273 | + remaining_batches = total_batches % self.num_replicas |
| 274 | + |
| 275 | + # Current rank gets one extra batch if it's among the first 'remaining_batches' ranks |
| 276 | + if self.rank < remaining_batches: |
| 277 | + return batches_per_rank + 1 |
| 278 | + else: |
| 279 | + return batches_per_rank |
21 | 280 |
|
22 | 281 |
|
23 | 282 | # pyre-fixme[24]: expected generic parameter |
|
0 commit comments