Wasserstein Loss

[AlexanderSlivinskiy]

Discriminator and Generator seem to have similar accurracy. We want the Discriminator to outpace the Generator, and therefore implement the Wasserstein Loss, because otherwise a better discriminator would lead the Generator to benefiting.

-> Implement Wasserstein Loss in Disc
-> Make only every 10 Iterations a Gen-update

[AlexanderSlivinskiy]

Added Wasserstein loss. There's still stuff to be done

Generator:
In the generator, the adversarial loss is included only every 10 Iterations. The reconstruction losses (it seems like there are two) are included every iteration.
The 10 iterations are adjustable in the config file (gen->update_every).

Discriminator:
In the discriminator, there are already 3 other losses. It's unclear, which of them to cut out, and whether the losses have to be weighted somehow differently.

Furthermore, I'm not sure the wasserstein loss is included in the backpropagation already.

• Check whether the remaining losses in the Generator are fine
• Check the remaining Descriminator losses and which one to out.
• Check whether the Wasserstein loss is included in the optimization step

[AlexanderSlivinskiy]

Created branch for wasserstein_loss feature

• Have to check out how to replace calc_gen_loss, since it seems to give other arguments.
• Why did NaNs apprear?