Join forces with PYMC? about bayespy HOT 1 OPEN

Thanks for the comment! In the long run, I definitely have a vision about a free open-source Bayesian package, which contains a large variety of different inference algorithms. PyMC implements MCMC methods and BayesPy implements VB, so by combining these two packages it would be possible to define a model once and then use VB or MCMC, or both, for the model. Sounds great, I would like to see that. When I started with BayesPy, I did consider creating a "plugin" for PyMC. However, it seemed that I really needed to create a separate package because the modifications to PyMC would have been enormous and I wanted to try to make the best choices that I can come up with, without being restricted to some framework that isn't really meant for my case. At least currently, the packages have big differences that make it quite difficult to combine them. And my feeling is that the two packages are meant for slightly different problems. But I haven't looked into this thoroughly.

So, to summarize: great idea, I support it, but doing it in practice might be very challenging.

Maybe I should some day take a closer look on what are the main obstacles in "merging" BayesPy's VB machinery into PyMC. It would be great to list those and see whether it is possible or worth it. One obstacle is that PyMC aims to be very general purpose, that is, the models and pdfs can be quite arbitrary, whereas BayesPy is currently limited to conjugate-exponential family distributions.

from bayespy.