Why the gradient penalty item decreases to zero and then grows to infinity ?

igul222 / improved_wgan_training

Code for reproducing experiments in "Improved Training of Wasserstein GANs"

MIT License

2.35k stars 668 forks source link

Why the gradient penalty item decreases to zero and then grows to infinity ? #79

Open zhangshun97 opened 6 years ago

zhangshun97 commented 6 years ago

I was trying to train a DCGAN based model with WGAN and WGAN-GP improvements. However, when I print out the gradient penalty item, it seems that it will first decrease to zero and then increase to infinity. How could that happen ?