Data-efficient GANs with Adaptive Discriminator Augmentation to keras 3.0 (Tensorflow backend only)

[chunduriv]

This PR changes the Data-efficient GANs with Adaptive Discriminator Augmentation to keras 3.0 (Tensorflow backend only)

Please review the attached gist.

[fchollet]

Thanks for the PR!

I wonder whether it would be possible to make it backend agnostic by implementing def compute_loss instead of def train_step. Maybe not. I don't recall ever doing it with a GAN. It works quite well with VAEs and diffusion models though.

[chunduriv]

I think, in general it should be possible to implement a GAN using compute_loss, but in this case, we are additionally using some more sophisticated techniques. For instance, there's the line self.augmenter.update(real_logits), which we would have to put in compute_loss (because this is where we will have real_logits). This should work with tensorflow and torch backends, but not with Jax, since Jax does not allow traced code to have side effects.

[fchollet]

Ok, sounds good -- in this case, please add the generated files.