This is more of a question for the original algorithm. You might have noticed that WGAN does not converge with longer training, which is why you set the number of steps to be 1500. I changed to a larger number and then some instability happens after 1500 iterations.
Do you have any idea why this happens? WGAN paper suggests that W-distance of zero can be used as early stopping, but I think the algorithm should still converge stably afterwards.
This is more of a question for the original algorithm. You might have noticed that WGAN does not converge with longer training, which is why you set the number of steps to be 1500. I changed to a larger number and then some instability happens after 1500 iterations.
Do you have any idea why this happens? WGAN paper suggests that W-distance of zero can be used as early stopping, but I think the algorithm should still converge stably afterwards.