The tutorial on DPMMs (Dirichlet Process Mixture Models in Pyro — Pyro Tutorials 1.8.4 documentation) shows how to associate a base distribution with a cluster, in that case a multivariate Gaussian. Can I select several base distributions of different types and conduct joint inference? E.g., I need a model that, after selecting a cluster for an observation, generates a multivariate-normal value for it and one or more Bernoulli ones.
Hi @macleginn, you could try using pyro.distributions.MaskedMixture or pyro.distributions.ZeroInflatedDistribution to represent the set of base distributions, or forking and modifying their source code to fit your specific use case.
1 Like