pytorch-unrolled-gans
PyTorch implementation of Unrolled Generative Adversarial Networks. The official tensorflow implementation is here.
There is an issue posted in the official repo complaining that "Couldn't reproduce mode collapse without unrolling operation".
I also found this issue (especially when using F.tanh as activation function) when implemented this repo, and found it pretty interesting.
Here are the corresponding results:
- unrolled steps = 0
- unrolled steps = 10
It's well-known that there are too many factors affect the GAN training, and I started to doubt whether the unrolled objective does help the GAN trainin or avoid mode collapse. I try to make the data distribution harder by making the weight of the mixture gaussian non-uniform, so that the generator will be harder to cover all modes.
Case 1:
- unrolled steps = 0
- unrolled steps = 10
Slight improvement can be observed. Glad to see that the unrolled objective does work at least in my experiement settings.
Case 2 (failure case):
- unrolled steps = 0
- unrolled steps = 10
Larger improvement can be oberserved in the following experiments
Other experiment settings
I found that in this eperiments, using F.relu (what the author used in the paper) results in larger gap:
uniform mixture of gaussian
- unrolled steps = 0
- unrolled steps = 10
non-uniform mixture of gaussian
- unrolled steps = 0
- unrolled steps = 10
Requirements
- PyTorch 0.4.0
- matplotlib
- seaborn
- tqdm