Handles stochastic gates during training and deterministic during validation/testing?
#1
by bryanjmo - opened
I am missing how the functions are currently able to handle the gates being stochastic during training and deterministic during testing. In the original repo this is handled during forward() with z = self.sample_z(input.size(0), sample=self.training). But in this L0_Regularizer version it seems that sample is always set to True. Am I missing something?