search

clvrai / BicycleGAN-Tensorflow

A Tensorflow implementation of BicycleGAN.
MIT License
101 stars 31 forks source link

Keeping encoder fixed for cLR-GAN step #5

Open meowcakes opened 6 years ago

meowcakes commented 6 years ago

Hello, In the original paper the authors state they keep the encoder fixed when performing the update step for cLR-GAN. However, in your code it seems that it is being updated. Are you aware of this?

Thanks

youngwoon commented 6 years ago

Thanks for the comment.

Can you point out the description in the paper about fixing the encoder while updating cLR-GAN? I cannot find it and updating both the encoder and the generator makes more sense to me.

meowcakes commented 6 years ago

Hi, thanks for your reply. It is in section 4, under subtitle "Training details":

We only update G for the L1 loss L1latent(G, E) on the latent code (Equation 7), while keeping E fixed. We found optimizing G and E simultaneously for the loss would encourage G and E to hide the information of the latent code without learning meaningful modes.

I also just noticed another difference. In the same section they write:

For the encoder, only the predicted mean is used in cLR-GAN.

But it seems in your code you are sampling from the approximate posterior.

Thanks

youngwoon commented 6 years ago

I couldn't find it because I followed the old version of the paper.

I'll test these two changes and update the repo later.

Thanks so much.