search

igul222 / improved_wgan_training

Code for reproducing experiments in "Improved Training of Wasserstein GANs"
MIT License
2.35k stars 668 forks source link

When do you start seeing positive results? #62

Closed meder411 closed 6 years ago

meder411 commented 6 years ago

I notice that on CIFAR, for example, you train for 200k generator iterations. Surely you start seeing positive results (i.e. realistic images) well before then. Around how many generator iterations into training can you tell that the network is moving in the right direction (i.e. generator outputs are looking more and more realistic)?

igul222 commented 6 years ago

Depends on your personal definition of "realistic" but I can usually tell if things are going to work out within 1000 iterations.