About training code

[d12306]

@lukas-schott , thank you for your implementation, can you also show the training code in order to me to evaluate it on other datasets?

Thanks

[lukas-schott]

As our training code is intertwined with other projects, we do not plan on releasing it.

However, it should be straight forward to train one VAE for each class. All relevant hyperparameters should be in the paper. Otherwise, I am happy to answer any questions regarding the training procedure.

[d12306]

Thanks, I will try to train it and see what happens.

[JindongGu]

@lukas-schott Thanks for your awesome project and code! I am trying to train VAEs using the code. The loss function I used is ELBOs() in the abs/abs_models/loss_functions.py. However, the trained models failed to generate realistic digits. The loss function ELBOs() return elbo / (n_ch * nx * ny), which is not correct, in my opinion. There is no reason to divide elbo by the input dimensions. I appreciate your comments.

[lukas-schott]

dividing the ELBO by the input dimension in
https://github.com/bethgelab/AnalysisBySynthesis/blob/master/abs_models/loss_functions.py#L37
is equivalent to adjusting the learning rate and can be seen as a hyperparameter.