Replicating PyMC3 mixture model of different distributions

[akotlar]

Hi, thanks for the great work on the package. Trying to learn how to use Pyro in the context of mixtures of (potentially) different distributions. I found an example from a PyMC3 tutorial, and was curious about the recommended approach to replicating it.

In PyMC3 one can provide a list of distributions to a “mixture” distribution (see screenshot, which comes from https://psyarxiv.com/aes5f/download)

What is the canonical way of doing this in Pyro today? Is it this: Arbitrary mixture models and discrete latent variable enumeration (there are noted performance disadvantages there)

[fritzo]

Yes, your linked solution is still canonical.

Note there is a more recent MixtureSameFamily distribution that will be available in the next PyTorch release, however as the name indicates it restricts to mixtures within a single family (e.g. Normal distributions of different parameters).