GPU RAM exploding on validation about srgan HOT 3 CLOSED

maybe somewhere you need detach tensor, but you haven't detach it, just copy it, then gpu will explode

from srgan.

@NicoLivesey Did you fix you problem?
I have the same problem when training with DIV2K，the code run out of memory after a few epochs. I only replaced volatile=True by torch.no_grad() and use different datasets only.

from srgan.

I've fixed it by adding with torch.no_grad(): to the entire validation for loop, probably some operations like .mean() save the entire computation graph, rather than just the values.
Reference

from srgan.