Double tf.log()'s when calculating CategoricalPd.sample() ?

[keven425]

https://github.com/openai/baselines/blob/4993286230ac92ead39a66005b7042b56b8598b0/baselines/common/distributions.py#L155

On this line above, it takes log twice on u. Why is that the case? I thought it should only take log once, like this:

return tf.argmax(self.logits -tf.log(u), axis=-1)

@joschu

[unixpickle]

-log(-log(uniform)) is the Gumbel distribution. This use a Gumbel-Softmax to sample from a uniform distribution.