emiliendupont / infogan Goto Github PK

Hi Emilien,

I'm trying to understand how an infoGAN works, and I was looking at your code as an example. On line 18 in util.py, it appears that the "ent" tensor will not have any gradient with respect to the weights of the network, because it only depends on y_true, so the network should behave the same in training without it. Is this correct? Also, are you sure that on line 19 you want to subtract cross_ent, not add it? Surely if you subtract it, then as the value corresponding to the "one hot" label in y_true increases (as you want it to do) the log of it will also increase, so the crossentropy will decrease (as you want it to do). However, this will mean the loss function increases, so the opposite will tend to happen.

Cheers,
Oliver