Hi all!
I am building a WGAN-GP based CycleGAN model. On my dataset, when I use the lambda parameter for the gradient penalty as 1000, it gives a very good result (I mean it converges to zero). Yet in the original paper, the authors use lambda as 10. I am not sure whether 1000 is too much or not a a hyperparameter. But this gave me the best result so far.