Closed namidairo777 closed 6 years ago
Hi @namidairo777 : This repo is still a work in progress, I hope to have it running soon. About the RelaxedOneHotCategorical Distribution, probs is supposed to be the prior of each class in the Gumbel-Softmax estimator.
Sorry for the late reply! Thank you so much for answering me. I got a lot of inspirations of implementing MADDPG from your repo!
Glad to hear it. I'm closing this issue for now.
out = tf.contrib.distributions.RelaxedOneHotCategorical(0.1,probs=out)