How to understand sample from softmax?

[npuichigo]

I have no idea of this line about sampling from softmax. Can you give me a help?

sel = tf.one_hot(tf.argmax(logit_probs - tf.log(-tf.log(tf.random_uniform(logit_probs.get_shape(), minval=1e-5, maxval=1. - 1e-5, seed=seed))), 3), depth=nr_mix, dtype=tf.float32)

[powder21]

It's Gumbel-Max Trick