I've tried to combine improved WGAN with pix2pix, but I got a very unstable training even got exploded ( NaN). I have also implemented original WGAN with pix2pix, and it works well. It seems like the clipping way does make sure the weight not exploded, but the gradient penalty doesn't work well.
Does anyone have any idea or experience to combine these two?
I've tried to combine improved WGAN with pix2pix, but I got a very unstable training even got exploded ( NaN). I have also implemented original WGAN with pix2pix, and it works well. It seems like the clipping way does make sure the weight not exploded, but the gradient penalty doesn't work well. Does anyone have any idea or experience to combine these two?