search

AAnoosheh / ComboGAN

BSD 2-Clause "Simplified" License
136 stars 25 forks source link

Explanation for forward loss #5

Closed RagMeh11 closed 5 years ago

RagMeh11 commented 5 years ago

Thanks for such a nice and clean code. Can you please explain what lambda_fwd is intended to do?

https://github.com/AAnoosheh/ComboGAN/blob/master/models/combogan_model.py#L155

Thanks.

AAnoosheh commented 5 years ago

It's not documented explicitly I think, but it's just an L1 loss on a (downsampled 2x) image pre- and post-translation.

This is not desired in most cases, but, for example, when your domains are human faces of different ages, you don't want anything but the face region changed - and even then it should be changed minimally, such that the downsampled before/after images look nearly identical.

Thanks for your interest!