Add length and padding keyworks to DistributedSampler

Current implementation of `DistributedSampler` is ideal for distributed
training using map datasets, as they fit in memory and have known size.
However, it doesn't support distributed training using `IterableDataset`
datasets, as these classes do not implement `__len__`.
To fix that, a `length` keyword was added to `DistributedSampler`, which
has precedence when set.

An extra `padding=True` parameter was also added was give finer control
on whether the (returned) index list should be padded by the sampler.
This is useful for preventing duplicate reading on `IterableDataset`
datasets that do not fit in memory or which data reading or transformation
are expensive.

Finally, set_rank method was added, similarly the existing `set_epoch`,
to ease distributed training. When `DataLoader` is created with
`num_workers` > 0 and `dataset` is an instance of `ChunkDataset`,
a copy of `DistributedSampler` on each worker needs to be configured
with their new rank.

There is no back compatibility with this change.

torch/utils/data/distributed.py

@@ -9,21 +9,26 @@ class DistributedSampler(Sampler):
 
  It is especially useful in conjunction with
  :class:`torch.nn.parallel.DistributedDataParallel`. In such case, each
- process can pass a DistributedSampler instance as a DataLoader sampler,
+ process can pass a `DistributedSampler` instance as a DataLoader sampler,
  and load a subset of the original dataset that is exclusive to it.
 
  .. note::
  Dataset is assumed to be of constant size.
 
  Arguments:
- dataset: Dataset used for sampling.
- num_replicas (optional): Number of processes participating in
- distributed training.
- rank (optional): Rank of the current process within num_replicas.
- shuffle (optional): If true (default), sampler will shuffle the indices
+ dataset (Dataset): Dataset used for sampling. It can be `None` if :attr:`length` is specified
+ num_replicas (int, optional): Number of processes participating in
+ distributed training (default: `None`)
+ rank (int, optional): Rank of the current process within :attr:`num_replicas` (default: `None`)
+ shuffle (bool, optional): If `True` sampler will shuffle the indices (default: `True`)
+ length (int, optional): length of `dataset`
+ If `None`, length is calculated as `len(dataset)` (default: `None`)
+ Must be greater than 0 when :attr:`dataset` is a `IterableDataset`
+ padding (bool, optional): If `True`, the returned lists will be padded to have the same length
+ (default: `True`)
  """
 
- def __init__(self, dataset, num_replicas=None, rank=None, shuffle=True):
+ def __init__(self, dataset, num_replicas=None, rank=None, shuffle=True, length=None, padding=True):
  if num_replicas is None:
  if not dist.is_available():
  raise RuntimeError("Requires distributed package to be available")
@@ -36,23 +41,30 @@ def __init__(self, dataset, num_replicas=None, rank=None, shuffle=True):
  self.num_replicas = num_replicas
  self.rank = rank
  self.epoch = 0
- self.num_samples = int(math.ceil(len(self.dataset) * 1.0 / self.num_replicas))
+
+ self._dataset_length = length
+ if not isinstance(length, int) or length <= 0:
+ self._dataset_length = len(dataset)
+
+ self.num_samples = int(math.ceil(self._dataset_length * 1.0 / self.num_replicas))
  self.total_size = self.num_samples * self.num_replicas
  self.shuffle = shuffle
+ self.padding = padding
 
  def __iter__(self):
  # deterministically shuffle based on epoch
  g = torch.Generator()
  g.manual_seed(self.epoch)
  if self.shuffle:
- indices = torch.randperm(len(self.dataset), generator=g).tolist()
+ indices = torch.randperm(self._dataset_length, generator=g).tolist()
  else:
- indices = list(range(len(self.dataset)))
+ indices = list(range(self._dataset_length))
 
 
  # add extra samples to make it evenly divisible
- indices += indices[:(self.total_size - len(indices))]
- assert len(indices) == self.total_size
+ if self.padding:
+ indices += indices[:(self.total_size - len(indices))]
+ assert len(indices) == self.total_size
 
  # subsample
  indices = indices[self.rank:self.total_size:self.num_replicas]
@@ -63,5 +75,10 @@ def __iter__(self):
  def __len__(self):
  return self.num_samples
 
+ def set_rank(self, rank):
+ assert rank >= 0, 'rank must be >= 0'
+ assert rank < self.num_replicas, 'rank must < num_replicas'
+ self.rank = rank
+
  def set_epoch(self, epoch):
  self.epoch = epoch

torch/utils/data/distributed.pyi

@@ -3,7 +3,8 @@ from . import Sampler, Dataset
 
 T_co = TypeVar('T_co', covariant=True)
 class DistributedSampler(Sampler[T_co]):
- def __init__(self, dataset: Dataset, num_replicas: Optional[int]=..., rank: Optional[int]=...): ...
- def __iter__(self) -> Iterator[int]: ...
+ def __init__(self, dataset: Dataset, num_replicas: Optional[int]=..., rank: Optional[int]=..., shuffle: Optional[bol]=..., length: Optional[int]=..., padding: Optional[bool]=...): ...
+ def __iter__(self) -> Iterable[int]: ...
  def __len__(self) -> int: ...
+ def set_rank(self, rank: int) -> None: ...
  def set_epoch(self, epoch: int) -> None: ...