Training hyperparameters

[otaviojacobi]

Hello :) thanks for this very interesting project !

I'm planning to reproduce results on One Piece dataset and then move to different datasets and do some attempts with the new technique of cycle-gan, called CUT.

Would you min sharing how did you train cycle GAN for given results ? Or the model weigths ?

Thanks and BR

[OValery16]

Hi @otaviojacobi,

I read the contrastive-unpaired-translation paper you are planning to use (by the way, thanks I didn't know it).

For the training configuration, I basically used the one described in the cycle-gan paper. I also found out that using a perceptual loss helps the GAN the converge a bit faster. As a matter of fact, this is the path also followed by the authors of CUT.

If you has some interesting results with CUT, feel free to post a link to your github here. I am also interested too. :) (I like to see interesting side projects using Deep Learning)