Timeline

PyMC Timeline

This is extremely preliminary and very subject to change. The goal here, as with PyMC, is to reason about uncertainty, and help users decide on a wise path forward. In particular, the point release estimates are bound to be wildly inaccurate, and are specifically not goals or promises, but guesses.

Currently (May 2018), we have no problem enthusiastically recommending PyMC3 for new projects, and plan to continue adding features.

1. April 2018 - PyMC3 3.4
2. June 2018 - Confirm new target backend for PyMC4
3. November 2018 - PyMC3 3.5.1 - Deprecate plotting, stats and diagnostics modules, switching to arviz
4. November 2018 - Theano support ends
5. PyMC4 alpha
   • Python 3 only
   • Tensorflow backend
   • "alpha" means:
     • There is an API that allows models to be specified,
     • NUTS usually works
     • Inference plays well with diagnostics (arviz) This probably means dict of numpy arrays or pandas dataframe.
6. PyMC3 3.6
   • Remove plotting, stats and diagnostics modules in favor of arviz
7. PyMC4 beta
   • "beta" means:
     • API is pretty settled
     • VI usually works
     • feature parity with PyMC3 (of those features that are used)
8. PyMC3 3.7
   • Should cut a last release, and promise stability for at least two more years
9. Active development on PyMC3 stops:
   • PRs still reviewed and accepted, but development encouraged to PyMC4
   • This means next GSoC (2019) is on PyMC4
10. PyMC4 4.0
    • Means everything is pretty stable?