I have multiple discrete random variables. For example I have one Bernoulli and one Categorical random variable. Should I use 2 different data dependent baselines, one for each? Or should I create a single data dependent baseline and specify it in both `pyro.sample`

statements?

probably the former; but without any knowledge of the dependency structure of your model and guide it’s impossible to say

1 Like

My model is very similar to the AIR example, just imagine that you also predict a class for each digit (adding a categorical latent variable) the categorical latent variable will then just be used as an additional input to the z_what encoder and decoder. Kind of turning the VAE part of AIR into a Conditional VAE.

I also observe the sequence of digit classes.

in general the baseline will work well if it can approximately model the downstream cost at a sample site. so you should condition on any upstream information that you suspect would be helpful in doing so.

1 Like