File size: 5,011 Bytes
0b32ad6 |
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 |
"""
Wrap any batch sampler for distributed training
Authors:
* Leo 2022
"""
import logging
from copy import deepcopy
from typing import Iterator, Optional, TypeVar
import torch.distributed as dist
from torch.utils.data import BatchSampler
T_co = TypeVar("T_co", covariant=True)
logger = logging.getLogger(__name__)
__all__ = [
"DistributedBatchSamplerWrapper",
]
class DistributedBatchSamplerWrapper:
def __init__(
self,
batch_sampler: BatchSampler,
num_replicas: Optional[int] = None,
rank: Optional[int] = None,
allow_duplicates: bool = False,
allow_uneven: bool = False,
) -> None:
if num_replicas is None:
if not dist.is_available():
raise RuntimeError("Requires distributed package to be available")
num_replicas = dist.get_world_size()
if rank is None:
if not dist.is_available():
raise RuntimeError("Requires distributed package to be available")
rank = dist.get_rank()
if rank >= num_replicas or rank < 0:
raise ValueError(
"Invalid rank {}, rank should be in the interval"
" [0, {}]".format(rank, num_replicas - 1)
)
self.batch_sampler = batch_sampler
self.num_replicas = num_replicas
self.rank = rank
self.allow_duplicates = allow_duplicates
self.allow_uneven = allow_uneven
def __iter__(self) -> Iterator[T_co]:
logger.info(
f"Building distributed batch sampler for rank={self.rank}, world_size={self.num_replicas}"
)
all_rank_batch_indices = list(iter(self.batch_sampler))
if len(all_rank_batch_indices) % self.num_replicas == 0:
target_batch_indices = all_rank_batch_indices
else:
num_to_halve = (
self.num_replicas - len(all_rank_batch_indices) % self.num_replicas
)
flatten_batch_indices = deepcopy(all_rank_batch_indices)
while num_to_halve > 0:
newly_flatten = []
all_cant_be_halved = True
for indices in flatten_batch_indices:
if num_to_halve > 0 and len(indices) > 1:
indices1, indices2 = (
indices[: len(indices) // 2],
indices[len(indices) // 2 :],
)
newly_flatten += [indices1, indices2]
num_to_halve -= 1
all_cant_be_halved = False
else:
newly_flatten.append(indices)
flatten_batch_indices = deepcopy(newly_flatten)
if all_cant_be_halved:
if self.allow_duplicates:
logger.warning(
"To ensure all the dataloaders in different processes get the same number "
"of batches. Some batches are duplicated. This must not happen during the "
"evaluation stage."
)
flatten_batch_indices = (
flatten_batch_indices
+ all_rank_batch_indices[:num_to_halve]
)
elif self.allow_uneven:
logger.warning(
"Total batches will not be evenly distributed across the dataloaders in "
"different processes. This must not happen during the training stage and "
"can lead to hanging, while might be okay during the evaluation stage."
)
else:
raise ValueError(
"The provided batch sampler cannot be safely wrapped for distributed training. "
"Please try increase the number of indices in each batch. Or, allowing duplicated "
"batches or uneven number of batches across dataloaders."
)
target_batch_indices = flatten_batch_indices
if not self.allow_uneven:
assert len(target_batch_indices) % self.num_replicas == 0
batch_indices = target_batch_indices[self.rank :: self.num_replicas]
return iter(batch_indices)
def __len__(self) -> int:
# Since the total number of batches dynamically depends on the current epoch,
# instead of pre-compute it which will duplicate the batch number computation logic,
# it makes no harm to simply re-compute it with __iter__ for every call, since
# __len__ is usually not frequently called and won't be the performance bottleneck
return len(list(iter(self)))
def set_epoch(self, epoch: int) -> None:
if hasattr(self.batch_sampler, "set_epoch"):
self.batch_sampler.set_epoch(epoch)
|