I'm currently trying to implement the Pix2Pix algorithm which is a GAN Structure, but i have some issues with the convergence of the discriminator and the output pictures of the generator...
1) Convergence Problem :
It seems that the discriminator doesn't converge at all. When i print the loss of the generator, it seems to work very well :
But when i print the loss of the discriminator, i have the following plot :
Or more precisely :
Do you know what are the possible reasons of such a behavior ? How can i stabilize the learning of the discriminator ?
2) Chromatic aberrations
I have also some problems with the generated pictures. Indeed i have often a total saturation of the colors, the printed objects have only one color like that :
The solution seems to train the discriminator every 200 steps for example, in this case i obtain something like that :
But it is not satisfying at all...
(I precise that the first column is the input of the generator, the second one is the output of the generator and the third one is the target picture, for the moment i'm only trying my network to reproduce the same picture... it should be easy...)
NB : the initialization seems also to play a really important role on the colors, indeed, with the exact same parameters, i obtained after thousands of steps really different results.
Has someone an idea to explain these phenomena ?
Thanks a lot for your reading and your potential help !
