3 ms·
No, GANs train exactly one discriminator, jointly with the generator. There's no guarantee that you can't train another good discriminator out of band. Further
by sdenton4 7y ago
No, GANs train exactly one discriminator, jointly with the generator. There's no guarantee that you can't train another good discriminator out of band.
Furthermore, GAN discriminators are (as I understand it) often hobbled a bit to ensure that the generator can make progress on the loss function. An always-correct D doesn't provide a useful gradient.