search

SKTBrain / DiscoGAN

Official implementation of "Learning to Discover Cross-Domain Relations with Generative Adversarial Networks"
773 stars 170 forks source link

gen loss formula #1

Closed edgarriba closed 7 years ago

edgarriba commented 7 years ago

in lines 251-252 you apply curriculum learning to compute the total gan loss. What's the motivation for that? Cannot find it in the paper.

Besides, is there any authorship relation between this repo and https://github.com/carpedm20/DiscoGAN-pytorch ?

jazzsaxmafia commented 7 years ago

We found out that in some cases assigning the same ratio to GAN loss and reconstruction loss fails to generate decent results. We guessed this occurred since training with GAN loss is much harder than training with reconstruction loss, especially in the early phase, and this curriculum strategy worked for those cases.

Carpedm20's code was independently implemented before we made this repo public, and his and ours have no authorship relation.

edgarriba commented 7 years ago

so then I guess that the formulas from the paper need a slight modification. Besides, starting_rate and default_rate have both the same value. Is it correct?

hshyunsookim commented 7 years ago

Thanks for pointing it out. We will consider making that change in our v2 paper.

starting_rate and default_rate are equal only in angle_pairing.py.

edgarriba commented 7 years ago

great!