gan training

[gongshichina]

https://github.com/hbutsuak95/monolayout/blob/5339d5f7e8f7fbc8272bd96abba16d6128b42098/train.py#L252

In the classical GAN training, we avoid letting the loss_D to influence the weight updating of G networks(because loss_D is the reverse training objective of G). However,
here the loss_D will accumulate gradients of G network, and update the weights of G networks by the following self.model_optimizer.step().
Could you please comment on it? Thanks!

[manila95]

Here we are using the discriminator as a regularization technique, unlike GAN where you have a generator and a discriminator, the task of the discriminator there is to just distinguish b/w real and fake images and the generator constantly updates itself indirectly to fool the discriminator. Here we have an AutoEncoder which outputs a layout and the discriminator makes sure that this layout looks similar to a real layout of an autonomous driving scene.