This is a somewhat open ended question, but I was wondering what is the rationale behind supporting different backends and how to decide, as a user, which way to go.
I am starting to integrate probabilistic models in existing production systems which support the research of scientists at my institution. Although the books and tutorials available made me think that PyMC or Stan were the more mature ecosystems, I started using Pyro because we use Pytorch for other parts of the system.
Since we are just starting with PPLs, we may still have time to review our choices, but I lack advice. Has anyone on this forum some suggestions on which are the criteria to take into account to choose between Pyro and NumPyro and what will be the implications of the advancement of Funsor on these choices?
Our main constraint is that we have to make choices for our platform and we will have to maintain the code produced by PhD students and faculty several years down the road. And this is the kind of nightmare I would like to avoid.
Sorry for not being more specific and thanks for Pyro and all the work you are putting on it!