Sampler that restricts data loading to a subset of the dataset. 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, and load a subset of the origina
| 41 | from torch.utils.data import Sampler |
| 42 | |
| 43 | class DistributedSampler(Sampler): |
| 44 | """Sampler that restricts data loading to a subset of the dataset. |
| 45 | |
| 46 | It is especially useful in conjunction with |
| 47 | :class:`torch.nn.parallel.DistributedDataParallel`. In such case, each |
| 48 | process can pass a DistributedSampler instance as a DataLoader sampler, |
| 49 | and load a subset of the original dataset that is exclusive to it. |
| 50 | |
| 51 | .. note:: |
| 52 | Dataset is assumed to be of constant size. |
| 53 | |
| 54 | Arguments: |
| 55 | dataset: Dataset used for sampling. |
| 56 | num_replicas (optional): Number of processes participating in |
| 57 | distributed training. |
| 58 | rank (optional): Rank of the current process within num_replicas. |
| 59 | """ |
| 60 | |
| 61 | def __init__(self, dataset, pad=False, consecutive_sample=False, permutation=False, num_replicas=None, rank=None): |
| 62 | if num_replicas is None: |
| 63 | num_replicas = get_world_size() |
| 64 | if rank is None: |
| 65 | rank = get_rank() |
| 66 | self.dataset = dataset |
| 67 | self.num_replicas = num_replicas |
| 68 | self.rank = rank |
| 69 | self.epoch = 0 |
| 70 | self.consecutive_sample = consecutive_sample |
| 71 | self.permutation = permutation |
| 72 | if pad: |
| 73 | self.num_samples = int(math.ceil(len(self.dataset) * 1.0 / self.num_replicas)) |
| 74 | else: |
| 75 | self.num_samples = int(math.floor(len(self.dataset) * 1.0 / self.num_replicas)) |
| 76 | self.total_size = self.num_samples * self.num_replicas |
| 77 | |
| 78 | def __iter__(self): |
| 79 | # deterministically shuffle based on epoch |
| 80 | g = torch.Generator() |
| 81 | g.manual_seed(self.epoch) |
| 82 | |
| 83 | if self.permutation: |
| 84 | indices = list(torch.randperm(len(self.dataset), generator=g)) |
| 85 | else: |
| 86 | indices = list([x for x in range(len(self.dataset))]) |
| 87 | |
| 88 | # add extra samples to make it evenly divisible |
| 89 | if self.total_size > len(indices): |
| 90 | indices += indices[:(self.total_size - len(indices))] |
| 91 | |
| 92 | # subsample |
| 93 | if self.consecutive_sample: |
| 94 | offset = self.num_samples * self.rank |
| 95 | indices = indices[offset:offset + self.num_samples] |
| 96 | else: |
| 97 | indices = indices[self.rank:self.total_size:self.num_replicas] |
| 98 | assert len(indices) == self.num_samples |
| 99 | |
| 100 | return iter(indices) |