Possible error in code about pytorch-gan HOT 3 CLOSED

Just to clarify, Mirtha's question is about the implementation of infoGAN.

Also I noticed that at no point the (infoGAN) network is given information that allows it to associate labels to digits, it looks like "labels" is being treated as an additional code variable (but categorical rather than numeric). So the network manages to cluster the digits into 10 classes, but with no control of what digit goes to which class, they end up in a random order - and what is worse, the same class may contain hybrid versions of digits that look similar such as 4 and 9. This can be fixed with small changes in the loss functions for generator and discriminator - if I have a chance I may post my suggested changes at a later time.

from pytorch-gan.

@mlucas491 Thanks, that should definitely be np.random.uniform.

@mlerma54 Your observation is correct, and that is also one of the interesting properties with InfoGAN. The network finds it's own clustering based on the properties of the data, in a totally unsupervised manner.

from pytorch-gan.

See fix.

from pytorch-gan.