ha0tang / asymmetricgan Goto Github PK

Hi,

I'm trying to reproduce your results for style transfer using photo2monet, photo2vangogh, photo2ukiyoe and photo2cezanne datasets. (Collection Style)
I wanted to know if you used specific lambda values for this particular use case (as the showcased results are really good).

I have another question regarding the conditional identity preserving loss mentioned in the paper:

removing the conditional identity preserving loss, multi-scale SSIM loss and color cycle-consistency loss substantially degrades the performance, meaning that the proposed joint optimization objectives are particularly important to stabilize the training process and thus produce much better generation performance

Maybe I missed it, but I haven't found this loss anywhere in the code. Is there a practical reason for it?

Finally, I have noticed that you did not use the mentioned LSGAN loss but rather the loss behind the Wasserstein GAN GP model. What are the reasons behind it?

Thanks

I thought both of them is calculated on separate color channel, what's the difference?

This is not a implementation of the actual paper "DualGeneratorGenerativeAdversarialNetworksfor Multi-DomainImage-to-ImageTranslation" but instead "GestureGANforHandGesture-to-GestureTranslation intheWild".