About num_workers > 1

[PSacfc]

Thanks for your excellent work.
I got a question about why performance will decrease when num_workers > 1. What if only one single dataset is used

[tung-nd]

We implemented a highly customized IterableDataset for dealing with the large dataset size, and num_workers works differently compared to a standard pytorch dataset. We did not find data loading to be the bottleneck so we always set num_workers=1.