The papar doesn't use the sigmoid function.

[dongdong092]

While in your code,i see the sigmoid function in the last layer of the discrimator. I've noticed that you may have try the model with the linear layer instead of the sigmoid function.Does it work well?

[godisboy]

Hi @dongdong092
In the original paper, the author had claimed that they used the standard objective function for GAN.

[dongdong092]

But with Cross entropy it will become JS divergence again, why not use WGAN's method to use D(X) as the loss function directly?

[zhangqianhui]

@dongdong092 It use the original loss: -log d(sigmoid(x)) = softplus(-x). You can use wgan loss. But it may not get better results.