Closed BK1024 closed 5 years ago
xinntao
commented
5 years ago In the file https://github.com/xinntao/BasicSR/blob/master/codes/models/SRGAN_model.py (SRGAN model), the wgan-gp could work.
The interp_crit, _ = self.netD(interp) you mentioned is in the SFTGAN_ACD_model.
And SFTGAN does not support WAGN-GP.
BK1024
commented
5 years ago The following lines are the code snippet I am referring to: https://github.com/xinntao/BasicSR/blob/36984e8bb1cd4269397bb15361a4221d26d4235a/codes/models/SRGAN_model.py#L148-L157
In line 155 is the code I am referring to: interp_crit, _ = self.netD(interp). Are you saying that this part of the code will not be utilized in when training with SRGAN_model.py
xinntao
commented
5 years ago Emmm, I made a mistake....+_+
It should be interp_crit = self.netD(interp) in SRGAN model.
# maybe wrong in cls? I mean that it may be wrong in SFTGAN model.
A long time ago, I use this code to run SRGAN with WGAN-GP and it could run. But I did not observe improvements compared to vanilla GAN.
I will correct these codes. Thanks~
BK1024
commented
5 years ago Got it! Hope this helps.
Hi xinntao, in
optimize_parameters(self, step)from /BasicSR/codes/models/SRGAN_model.py, , I found python complaining about the following line from your implementation of wgan-gp:interp_crit, _ = self.netD(interp).self.netD(interp)will only return one value for the randomly interpolated image. Is it just a typo or the wgan-gp part is yet to be completed?Also, I found this line below the above line:
l_d_gp = self.l_gp_w * self.cri_gp(interp, interp_crit) # maybe wrong in cls?. This concerns me...would you mind explaining what might be wrong? Thank you!