Losses become nan after some number of epochs about sdgym

[Baukebrenninkmeijer]

Successfully reproduced on the adult dataset with your specified preprocessing. As verification, the changes in the last few commits in project/scaffolding have only been syntactic, right?

[Baukebrenninkmeijer]

Tried it on both the Census as well as the Adult dataset. Also tried using both the GMMTransformer and the BGMTransformer. Neither change seems to make a difference.

[Baukebrenninkmeijer]

Narrowed the problem down to something specifically with TGAN. Using the latest project_scaffolding the problem persists with both the census and adult dataset. TableGAN does not seem to have the problem. MedGAN also doesn't seem to suffer from this problem.

[ManuelAlvarezC]

Hi @Baukebrenninkmeijer,

Thanks for reporting this, we will have a look and see how can it be fixed.

[Baukebrenninkmeijer]

Good to hear. I have the old TGAN working so I don't necessarily need this TGAN. Thus, I likely won't investigate any further into this issue either.

[kveerama]

@Baukebrenninkmeijer we just released the paper about SDGym here. This explains the latest TGAN model which is in this current SDGym repo (the older version was in this repo) This paper creates a new much approach (we should really call it CTGAN to differentiate it from TGAN- I can see how that can be confusing). Also, the paper explains the reasons behind SDGym and why we created it.

[Baukebrenninkmeijer]

Haha you guys beat me to it! I was hoping your paper would be finished after my thesis.

Could you maybe elaborate on the usage of the Gumbel softmax? I've been confused by when to and when not to use adapted softmax functions to overcome the differentiation problem. In the old TGAN paper you stated that it was not necessary to use the Gumbel softmax because the number of categories is fairly small, so the probability distribution can be generated directly with softmax. However, in the new TGAN paper you decide to use it to overcome the same stated problem.

[Baukebrenninkmeijer]

@kveerama