Ask for help about GAN use conv

[Liujingxiu23]

Does anyone use gan with conv?

In the paper "Generative adversarial network-based postfilter for statistical parametric speech synthesis", they use the following net structure:

I add this stucture in gantts, however, the training of gan_d_warmup always fails. the loss of trainning is reduced but the loss of testting shaked， the gan model tends to judge every sample (no matter real or fake) just as real.
My code reference https://github.com/bajibabu/postfilt_gan/blob/master/models.py .
I use crop=58, and split the feat frames to [0, 58] [ 58, 2*58] ....[(n-1)58, n58]... and do discriminator to every fragment.

Doest any one tried conv gan, how is the performance?