Conditional

[Shahryar-Noei]

Great repo!
Any hint on how to extend this to conditional?

[explainingai-code]

Hello @Shahryar-Noei ,
One thing you can try with is concatenating label information along channel dimension in both generator and discriminator.
Specifically for generator, create representation from noise using a convtranspose layer and also a representation from label (using conv transpose) and then concat both of these together. The second generator layer operators on this concatenated output, and every other layer of the generator network remains the same
Similarly for the discriminator, have a num_classxHxW label input(class label channel being all 1 and rest all 0) , and instead of the first discriminator layer, process input image, and label input via two different conv2d layers , and again like generator concat both along channel dimension and for all other discriminator layers, nothing changes.