I am trying to make a GAN model to generate 28x28 images (just like MNIST) corresponding to each digit. The Generator takes a z vector (random noise) concatenated with the actual label vector of length 10 (one hot encoded). The output of the generator (vector of length 784) is then fed into the discriminator which gives out a length 11 vector (The first ten values correspond to the one hot encoding of the prediction and the last value tells if the input image is fake or not). MSE Loss is used for calculating the loss. The discriminator is getting trained properly but not the generator. How do I go about debugging this?
#GAN For Generating MNIST Digits
vestal talon
calm trench
@next surge Sorry for pinging, but I think you're the most experienced in this out of people I know
ruby marlin
I am trying to make a GAN model to generate 28x28 images (just like MNIST) corre...
I know you need some special setup in the training loop to optimize the generator/discriminator separately because they are adversarial. Ie if you optimize them together it’s easier for each component to break the other one than improve itself. It’s hard to guess what is wrong without more information though
vestal talon
I am using adam optim with this
