I’m trying to evaluate which probabilistic programming package is appropriate for a commercial project, and I’m wondering if there’s an official plan for how long PyMC3 will continue to be developed for and/or supported by the dev team?
Right now I see a few things happening:
- Tensorflow & team have announced TF Probability and Edward2, which seems mostly aimed at researchers and developers who enjoy
painhaving maximum control over their tools, as opposed to ‘user focused’ packages like PyMC3.
- Uber released a probabilistic programming package Pyro sitting on torch which I don’t know a ton about
- Theano has ceased development!
It’s a bit tricky right now to figure out what to do - the TF probability ecosystem is undergoing massive change and no doubt in 6-12 months time will look pretty different to the way it does now (in terms of API, available ops/tools, bugs, etc). It seems like a bad time to adopt something like Edward1.
PyMC3 is a very mature platform with a highly active and large set of devs, but the death of Theano may have put an end-date on the tool, even if it’s not explicit. What do you guys think/feel about the current landscape? Does PyMC3 have another 12 months of energy left in it, which the rest of the tooling in the ecosystem reaches some kind of steady-state? Or is there a definite rapidly approaching ‘termination date’ for the development efforts, modulo small fixes?
Open to all opinions