Hi, we recently submitted an article with this new inference algorithm called the Marginal Unbiased Score Expansion (MUSE). In short, its a VI competitor, and could kind of be considered a type of SBI. The initial code is in Julia and plugs into Turing.jl, one of Julia’s PPLs (example of how it looks and my post on the Julia forums).
I’d like to write a PyMC interface as well to reach a broader audience. I’ve never used PyMC before but have been familiarizing myself with it the last few days. I was wondering if someone could offer advice on two things:
As I’ve learned, there’s just about to be an under-the-hood transition between PyMC3 and PyMC4 (not yet released). Any advice on which I should target? Perhaps relevant is that MUSE really shines on high dimensional problems, so perhaps the more performant PyMC4 would make sense? If the answer is that I may as well go for PyMC4, should I just check out the master branch of pymc and that’s it? Is there any other dependent packages I need?
Any advice or examples I can follow on writing a custom “sampler”. MUSE ultimately just needs forward simulations and gradients of the joint likelihood, and is otherwise quite simple.
Thanks for any help.