Usage of KL divergence about tgan

[leix28]

Hi,

(1) There's no specific reason in choosing KL divergence against other loss functions. You can definitely try other distance metrics or feed extra features to the discriminator.
For example, MedGAN feeds the mean and standard deviation of a batch as features to the discriminator. TableGAN explicitly optimizes the each column's mean and standard deviation for generator.

(2) Not really. Empirically, we found that simply optimize GAN loss will leads to mode collapse.

(3) Please check our new project, SDGym. Our new TGAN model uses WGAN-GP and PacGAN. We apply similar preprocessing method. A paper about SDGym will be available soon.

[Baukebrenninkmeijer]

Ah, this is exactly what I've been experimenting with. Using the WGAN-GP architecture. However, in my testing this hasn't improved results, giving worse covariance matrices and worse per column distributions compared to the real dataset. Have you gotten improved results with this?

The SDGym is exactly something I was thinking of and will be working on for my Thesis, which I'm doing now. I'm very interested in the results of SDGym.