Discriminant loss does not decrease

[WelkinYang]

Thanks for your excellent contribution, I am trying to apply this work to other domains, but unfortunately I found that the loss function of the discriminator could not be reduced in the experiments and the Bce loss was always higher than 0.6. I used VPSDE and generated 10 steps results with a maximum likelihood based SDE solver to serve as fake data for the discriminator. I used the same Unet as in the repository. When i change the Unet encoder to the whole Unet, the loss become more low than the loss of Unet encoder. Looking forward to get some advice. Thank you.