Is any case to prefer detach().cpu() to cpu().detach(),or vice versa?

1

Hello,they give same result,any particular case to prefer detach().cpu() to cpu().detach(),or vice versa?

2

I would prefer to use detach().cpu(), as this would detach the tensor already on the GPU and the cpu() operation won’t be tracked by Autograd. That being said, it seems to be a minor change and you shouldn’t see any differences in using any of these two approaches. :wink:

1 Like
3

Thanks,is there any case to prefer with cpu() than detach()?

4

I don’t understand the question. The cpu() operation transfers the tensor to the CPU (if not already there), while detach() cuts the computation graph. Could you explain the question a bit more?

1 Like