continuous latent code: more dimensions or more code?

[jiadingfang]

Hello,

Thanks for the amazing implementation. I have a question about how to set up the continuous latent code. In the mnist experiment, you use two 1-dimensional uniform distributed code as your continuous latent code, but can I simply use one 2-dimensional uniform distributed latent code instead? I think it doesn't make much difference since each log likelihood for continuous latent code is computed in the same way and they sum up to the cross entropy term with the same weight.

I would like to hear your voice about why you set up your continuous latent code this way and what is your reason to design like this.

Big thanks in advance!