Only import torch.distributed when needed (#882)

* Fix an import error for `get_world_size` and `get_rank`

* Only import torch.distributed when needed

torch.distributed is only used in DistributedGroupSampler

* use `get_dist_info` to obtain world size and rank

`get_dist_info` from `mmcv.runner.utils` handles the problem of `distributed_c10d` doesn't exist.

mmdet/datasets/loader/sampler.py

@@ -4,7 +4,7 @@ import math
 import torch
 import numpy as np
 
-from torch.distributed import get_world_size, get_rank
+from mmcv.runner.utils import get_dist_info
 from torch.utils.data import Sampler
 from torch.utils.data import DistributedSampler as _DistributedSampler
 
@@ -95,10 +95,11 @@ class DistributedGroupSampler(Sampler):
                  samples_per_gpu=1,
                  num_replicas=None,
                  rank=None):
+        _rank, _num_replicas = get_dist_info()
         if num_replicas is None:
-            num_replicas = get_world_size()
+            num_replicas = _num_replicas
         if rank is None:
-            rank = get_rank()
+            rank = _rank
         self.dataset = dataset
         self.samples_per_gpu = samples_per_gpu
         self.num_replicas = num_replicas