search

HsinYingLee / DRIT

Learning diverse image-to-image translation from unpaired data
844 stars 155 forks source link

Why use double discriminators for each domain? #32

Closed daa233 closed 5 years ago

daa233 commented 5 years ago

Thanks for your awesome work!

I see the there are two discriminators for each domain, which are disA and disA2 for domain A, disB and disB2 for domain B. I have noticed that they are updated with different inputs. However, I don't know if double discriminators are necessary here.

Have I missed something in the paper? Could you explain a little more about this?

HsinYingLee commented 5 years ago

Thank you for your interests.

It's first used in the original BiCycleGAN. One discriminator is used for the translated images, while the other is used for images generated from random noise. I think it's not theoretically necessary but it can facilitate the training process since two distributions are quite different in practice.

daa233 commented 5 years ago

I see. Thank you a lot!