Discriminator given the labels?

Is the Discriminator given the target labels? How can it tell whether the image is real or fake in the first instance?

In general there are two phases in GAN training:

  • 1 Train discriminator: Here real and fake data is used. So yes, Labels (whether real or fake) are used. Generators weights are frozen in this step.

  • 2 Train generator: here, discriminators weights are frozen in this step.

These two steps are repeated afterwards.

The generator’s goal is to trick the discriminator which has to assess whether it’s real or fake.

So if you have a super powerful generator which makes almost „close to perfect“ fakes, basically the discriminators assessment will be random resp. around 50% to detect the fake.

Here also some useful information:

Best regards


Got it!. Thanks @Christian_Simonis