why a customized DistributedSampler is needed?

[toyot-li]

@sshaoshuai Hello, I wonder what is the motivation for redefining a DistributedSampler here, whose __iter__ method is pretty similar to the PyTorch built-in DistributedSampler, as shown in the PyTorch source code https://pytorch.org/docs/stable/_modules/torch/utils/data/distributed.html#DistributedSampler In addition, if the tail indices are duplicated ones, would the whole evaluation process be correct? Thanks.