ericjang / gumbel-softmax

categorical variational autoencoder using the Gumbel-Softmax estimator
MIT License
425 stars 101 forks source link

logits_py incorrect in gumbel_softmax_vae_v2 #9

Open backpropper opened 6 years ago

backpropper commented 6 years ago

logits_py should be the log of the current logits_py or we can just give it as probs and not logits to the corresponding distribution.

BobXWu commented 5 years ago

Hi, do you know the meaning of the APIs log_prob and prob in tf.distributions or tf.contrib.distributions? They are quite confusing.