GMM SVI with relaxed categorical as latent variables

[marting]

Hi @sopchi,

I ran into similar issues and found this thread. The code appears to work somewhat for two clusters with the relaxed Bernoulli.

Hope that helps.