Add a utility to get posterior samples of discrete latent variables

[fehiepsi]

Per @martinjankowiak suggestion, after we marginalize discrete latent variables to run HMC and get posterior for continuous variables, it would be nice to have a utility to get posterior samples for discrete latent variables. That is we have a model with p(data | discrete, continuous), we marginalize and run MCMC to get p(continuous | data). This utility will get samples from p(discrete | data, continuous).

References:

• Pyro's epidemiological models
• @fritzo's answer for a question in forum

[fritzo]

This looks like another edifying application of funsors 🙂 . You may be able to simply call .sample() on the final funsor in enumeration, then unpack the resulting Delta funsors.

[fehiepsi]

This is achieved through infer_discrete.