Dear @martinarjovsky,
I am currently working on a project with MRI data.
I was using WGAN -GP loss on 2D implementation, with hyperparameters proposed in WGAN-GP paper - everything worked smoothly.
Now I switched to 3D implementation and started facing issues.
The G loss explodes to extremely high values(10^7), while D loss goes really low(-10^6).
I understand that for WGAN to work the critic needs to be near optima. However if done so, the Critic keeps producing high output for fake images which makes G loss skyrocket. My patch size is (176,144,16), in 2d it was (176,144).
1)I tried adding layer normalization to Critic, even though the loss values do not explode, the GAN fails to converge.
2) I tried tinkering the learning rate.
2.1.) High learning obviously make it even worse
2.2.) With low learning rates this explosion still happens but later in training.
3) I tried changing number of C iterations
3.1.) The more of Critic iteration I do - the faster it skyrockets.
3.2.) If i do same number of Critic/Generator iterations(1:1) the loss stays in normal margins, but the net does not converge to anything reasonable.
Any idea what could be the cause?
Thank you!
Dear @martinarjovsky, I am currently working on a project with MRI data. I was using WGAN -GP loss on 2D implementation, with hyperparameters proposed in WGAN-GP paper - everything worked smoothly. Now I switched to 3D implementation and started facing issues. The G loss explodes to extremely high values(10^7), while D loss goes really low(-10^6). I understand that for WGAN to work the critic needs to be near optima. However if done so, the Critic keeps producing high output for fake images which makes G loss skyrocket. My patch size is (176,144,16), in 2d it was (176,144). 1)I tried adding layer normalization to Critic, even though the loss values do not explode, the GAN fails to converge. 2) I tried tinkering the learning rate. 2.1.) High learning obviously make it even worse 2.2.) With low learning rates this explosion still happens but later in training. 3) I tried changing number of C iterations 3.1.) The more of Critic iteration I do - the faster it skyrockets. 3.2.) If i do same number of Critic/Generator iterations(1:1) the loss stays in normal margins, but the net does not converge to anything reasonable. Any idea what could be the cause? Thank you!