DataLoader Usage

[ericspod]

Adding persistent_workers=True to the arguments for any DataLoader object will speed the training process since it won't have to recreate processes at each epoch. This helped a lot in Windows and may be less helpful elsewhere, and it should be tested with notebooks and scripts. ThreadDataLoader may also provide some additional improvement.

[danieltudosiu]

I agree with persistent_workers and even pin_memory but my experience with ThreadDataLoader is patchy at best I would stir away from recommending it but I would be fine just mentioning it as an alternative.

[danieltudosiu]

Also, should we use 'non_blocking=True' as well?

[Warvito]

Adding persistent_workers=True to the arguments for any DataLoader object will speed the training process since it won't have to recreate processes at each epoch. This helped a lot in Windows and may be less helpful elsewhere, and it should be tested with notebooks and scripts. ThreadDataLoader may also provide some additional improvement.

Thanks for the suggestion, I am adopting persistent_workers=True in the new tutorials. I am not including pin_memory, but it is often used in monai tutorials (https://github.com/Project-MONAI/tutorials/search?p=1&q=pin_memory), so feel free to use it in the tutorials that you create.

[ericspod]

We've discussed pinned memory and non-blocking before, they do help at least sometimes but I feel we need somewhere to explain when and why a bit better. Try with them and see how they do in the tutorials.