loss-landscape-anim icon indicating copy to clipboard operation
loss-landscape-anim copied to clipboard

Published 20 hours ago • verified publisher icon logancyang

[CUDA out of memory] ResNet18 & Cifar100

Open lhchau opened this issue 6 months ago • 2 comments

Hi logancyang, I have successfully run your code using ResNet18 on Cifar100 with bs=16, epochs=200. However, during the 4th epoch, the GPU VRAM overflowed. Upon reviewing the code, I found that this line is causing the GPU RAM to accumulate at each step. Do you have a solution for this problem?

https://github.com/logancyang/loss-landscape-anim/blob/a8d5610ff433b94aca68af42952f0bad8e0e34b3/loss_landscape_anim/model.py#L180

lhchau avatar Feb 16 '24 09:02 lhchau

hi @lhchau, thanks for checking this out. This project is not actively maintained at the moment, but feel free to make your own fork if you have time to implement a solution.

logancyang avatar Feb 16 '24 18:02 logancyang

I also had the same issue. any solution @lhchau

alqurri77 avatar Mar 07 '24 04:03 alqurri77