What if the output layer of the generator is activated by "tanh" ?

zsyzzsoft / co-mod-gan

[ICLR 2021, Spotlight] Large Scale Image Completion via Co-Modulated Generative Adversarial Networks

Other

445 stars 67 forks source link

What if the output layer of the generator is activated by "tanh" ? #28

Open MenSanYan opened 3 years ago

MenSanYan commented 3 years ago

Hi, I notice that the last layer of the generator is activated by "linear", etc., the range of pixel values might stretch across (-infinite, +infinite) interval. Is there any advice or reasons for using "tanh" instead of "linear", so that pixel values is limited to (-1, +1) interval.

zsyzzsoft commented 3 years ago

I remember that StyleGAN's authors have tried tanh but they did not find improvements.

zengxianyu commented 3 years ago

I saw an explanation there looks reasonable https://github.com/taki0112/StyleGAN-Tensorflow/issues/5