This is a list of important changes to PyMC 3 relative to the 2.x releases.
- Added gradient-based MCMC samplers: Hamiltonian MC (
HMC
) and No-U-Turn Sampler (NUTS
) - Added slice sampler
- Automatic gradient calculations using Theano
- Convenient generalized linear model specification using Patsy formulae
- Parallel sampling via
multiprocessing
(IPython parallel support planned) - New model specification using context managers
- New Automatic Differentiation Variational Inference[AVDI] (http://arxiv.org/abs/1506.03431) (
ADVI
) allowing faster sampling thanHMC
for some problems. - Model evaluation like Deviance Information Critertion
DIC
andWAIC
- Numerous docs, examples and API documentation highlighting the power of this model specification API.