j314erre
commented
2 months ago gumbel softmax((log(pi)+g)/tau), where pi are class probabilities, therefore softmax((logits + g)/tau) will suffice. https://arxiv.org/pdf/1611.01144
Open maestrojeong opened 6 years ago
j314erre
commented
2 months ago gumbel softmax((log(pi)+g)/tau), where pi are class probabilities, therefore softmax((logits + g)/tau) will suffice. https://arxiv.org/pdf/1611.01144
https://github.com/vithursant/VAE-Gumbel-Softmax/blob/master/vae_gumbel_softmax.py#L55 You use the logits, however we have to use log(logits) to perform gumbel max trick as stated in https://timvieira.github.io/blog/post/2014/07/31/gumbel-max-trick/