Closed vzinche closed 5 years ago
there to many changes unrelated to backprop every n steps to merge this
@vzinche In general, I think it's a good idea. Could you reopen the PR with only the changes required for backprop every n steps (i.e. without the transforms, etc)?
added an option to choose the frequency of backprop (if one can't afford to train in batches due to memory limitations, but doesn't want to backprop every sample)