gradient penalty of sigmoid instead of logits.

[d-michele]

WGAN-GP-tensorflow/src/wgan_gp.py

Line 94 in 9f27dd9

gradients = tf.gradients(self.discriminator(interpolates, is_reuse=True), [interpolates])[0]

Hi, I would ask you an explanation of why did you calculate the gradient on the sigmoid instead of logits. Thanks!

[ChengBinJin]

@d-michele
According to the theory of WGAN and WGAN-GP, the Wasserstein distance is calculated using the output of the discriminator. And, a sigmoid function is not utilized at the end of the discriminator. Therefore, when calculating the gradient penalty from the discriminator, there is also no sigmoid function.

[opetliak]

@ChengBinJin Hi, why do you return two outputs (sigmoid and linear) in basicDiscriminator?

[ChengBinJin]

@lenpetlyak The linear value is used by calculating the Wasserstein distance, and the sigmoid value is just for debugging.