Multiprocessing slows dataloaders at the each beginning epoch

4

update pytorch to 1.7, then the arg persistent_workers will be helped. See this link What are the (dis) advantages of persistent_workers

1 Like