pysvihmm

Implementation of stochastic variational inference for Bayesian hidden Markov models.

Contents

HMM Classes

hmmbase.py : Abstract base class for finite variational HMMs.

hmmsvi.py : Base implementation of stochastic variational inference (SVI). Implementations that require significant changes to the logic should be based on this but broken off.

hmmbatchcd.py : Batch variational inference via coordinate ascent.

hmmbatchsgd.py : Batch VI via natural gradient.

hmmsgd_metaobs.py : SVI with batches of meta-observations. A meta-observation is a group of consecutive observations. We then form minibatches from these. The natural gradient for the global variables is computed for all observations in a meta-observation, and then those are averaged over all meta-observations in the minibatch.

hmm_fast.pyx : A fast implemenation of forward filtering backward sampling.

Test Classes

gen_synthetic.py : Functions to generate synthetic data.

test_* : Scripts to test correctness of algorithms.

Utilities

test_utitlities.py : Plotting and data generation functions used in the tests.

util.py : Miscellaneous files for HMM Classes and Test Classes.

Cython Modules

Run python setup.py build_ext --inplace to build external Cython modules.

Extensions

A C++ version can be found here

Authors

• Nick Foti
• Jason Xu
• Dillon Laird