So I spent half a day seeing if I could improve performance of DataLoading by doing a batch-GPU implementation with torchvision. Once we start getting images of size 320x320 that can really help, but it turns out I can remove the CPU bottleneck by just adding num_workers=5 to the DataLoader instances. 5x speedup. Oops.
I also moved the dataset modules under classification.
So I spent half a day seeing if I could improve performance of DataLoading by doing a batch-GPU implementation with torchvision. Once we start getting images of size 320x320 that can really help, but it turns out I can remove the CPU bottleneck by just adding num_workers=5 to the DataLoader instances. 5x speedup. Oops.
I also moved the dataset modules under classification.