Found a Bug with going_modular.data_setup.create_dataloaders!

[luxexistentia]

my env: Local Window PC with Ryzen5600x & rtx3070 & pytorch2.4.0

I was using the going_modular files and I didn't know why but it took too long to train/test models and gpu usage was almost 0% while cpu is almost 100% constantly, even tho device was set to "cuda".

But fortunately I found out what the problem was. It was problem with dataloader. I tried different number of num_workers manually and If it was set to 0, It seems like automatically pytorch optimize multi threading.

So In my test, By changing num_workers=0, speed to load a batch(32 imgs) went from 8.26 s ± 253 ms per loop -> 111 ms ± 7.13 ms per loop Around 74.5X Faster

And while Cpu reaching under 30%. And finally my training time got so much faster and almost fully use my graphic card!(almost reach to 100% during training)

Conclusion: Changing dataloader setting can boost up around 75 times. So I hope you to change the file:)

Ps. Thanks for ur Lecture! I finished it and it helped me a lot!

[mrdbourke]

Hi @luxexistentia ,

Great find!

Thank you for sharing.

This is interesting as I've found on different hardware tweaking the num_workers parameter can significantly improve loading times (e.g. when num_workers=0, data loading is slow but when num_workers > 0, data loading improves up to a threshold).

Good work troubleshooting this though.

Perhaps your CPU is best when using num_workers=0 rather than increasing it.