Non-standard likelihoods

Hi I came across pyro and numpyro and was wondering before I jump in, if they support non-standard likelihoods. For example, if I wanted to do Variational Bayes without being able to simulate from my model (as my likelihood is defined in an unusal way and I therefore can’t use pyro.sample)?

In particular, I’m interested in implementing something with likelihood tempering/whittle likelihood etc.

yes probably although you’ll need to be more specific about the exact technical requirements for your situation.

for example if you do hmc you can provide a likelihood with a trivial sampler since only the log_prob is needed for inference

I want to do this for vb and also hmc. This sounds very positive!