search

lindermanlab / ssm

Bayesian learning and inference for state space models
MIT License
559 stars 197 forks source link

Exact entropy and Exact M-Step #96

Closed bantin closed 4 years ago

bantin commented 4 years ago

This PR has the following main changes: 1) Update AR observations to use an exact M-Step when lags = 1 2) Make entropy calculation for SLDSTridiagVariationalPosterior exact, instead of approximated via sampling. 3) Refactor variational params of SLDSTridiagVariationalPosterior to consistently use all parameters.

We also add some tests for the entropy calculation and the analytical hessian calculation for linear-gaussian dynamics.

@slinderman There are a whole lot of commits here -- I think many were duplicated when I rebased off of master. If needed I could try to squash some of them together before merging.

bantin commented 4 years ago

@slinderman I think I've addressed most of the comments in the last couple commits. Let me know if you see anything else that needs attention!