search

YongfeiYan / Gumbel_Softmax_VAE

PyTorch implementation of a Variational Autoencoder with Gumbel-Softmax Distribution
203 stars 37 forks source link

hard Gumbel softmax = ST Gumbel softmax? #5

Closed CXX1113 closed 5 years ago

YongfeiYan commented 5 years ago

Yeah, if hard =True, the samples will be categorical, and the gradients are copied from y_hard to y(straight through estimator).