-
Zhaoheng Ni authored
Summary: The `BucketizeBatchSampler` may return different iter_list in different node if `shuffle` is `True`, which will cause DPP training hang forever. `shuffle` in `DistributedSampler` only happens in initialization, which means it will assign the same subset to replicas in all training epochs. The PR fixes the two above issues. cc arlofaria Pull Request resolved: https://github.com/pytorch/audio/pull/3068 Reviewed By: mthrok Differential Revision: D43372110 Pulled By: nateanl fbshipit-source-id: a162728406ae995e05d2a07cfc2444fb76cf345e
2c9b3e59
