glmhmm

This package provides flexible and easy-to-use code for fitting Generalized Linear Models (GLMs), Hidden Markov Models (HMMs), and combination GLM-HMMs (also known as Input-Output HMMs or IO-HMMs). The GLM code include options for fitting observational data of either Bernoulli or Multinomial distributions. HMM inference is done using the Expectation Maximization (EM) algorithm, with an optional mode for performing deterministic-annealing EM (DAEM), which is useful for hard to fit datasets.

Package Contents

glmhmm

• glm.py: GLM class fitting code
• hmm.py: HMM class fitting code
• glm-hmm.py: GLM-HMM class fitting code
• init_params.py: a script for defining initialization options for different model parameters
• observations.py: a script for defining distribution options for GLM observations
• utils.py: a script containing miscellaneous helper functions
• analysis.py: a script containing post-fitting analysis functions used in Bolkan, Stone et al 2021
• visualize.py: a script containing functions for plotting the figures seen in Bolkan, Stone et al 2021

examples

• fit-glm.ipynb: a simple example of fitting a GLM to simulated data
• fit-hmm.ipynb: a simple example of fitting an HMM to simulated data
• fit-hmm-DAEM.ipynb: an example illustrating the benefits of deterministic annealing EM (DAEM)
• fit-glm-hmm.ipynb: an example of fitting GLM-HMMs to simulated data

figures

Each jupyter notebook in the figures folder recreates the plots from a specified figure in Bolkan, Stone et al 2021

• data: a folder includeing pre-formatted design matrices and choice behavior for fitting the three mouse data sets described in Bolkan, Stone et al 2021
• matlab: a folder containing MATLAB scripts used for fitting the psychometric curve plots shown in Bolkan, Stone et al 2021
• fit models: a folder containing the model objects from several models fit during the figure generation process.
• fig4.ipynb: fits a GLM to real data and interprets results
• fig5.ipynb: compares model performance between a standard Bernoulli GLM and a 3-state GLM-HMM
• fig6.ipynb: fits a 3-state GLM-HMM to real data and interpret results
• fig7.ipynb: analyzes how the three states identified by the GLM-HMM manifest in the data
• extdatafig7: describes model selection and control analyses
• extdatafig9: shows how model simulations recapitulate characteristics of the real data
• suppfig4: shows how individual mice occupy different states for each session of the task

Fork installation

Create their environment, then install the requirements, but it includes GoalSelection, which is the package I'm writing for the setup data (maybe in the future flexivexi, when we merge?)

You will need access to GoalSelection (or its successor), which is private by now.

conda env create -n glmhmm --file env.yml 
pip install jaxlib==0.1.71 -f https://storage.googleapis.com/jax-releases/jax_releases.html 
pip install -e . 

Installation

For easy installation, download the code using the link above or type the following into a terminal window:

git clone https://github.com/irisstone/glmhmm.git

For convenience, we recommend setting up a virtual environment before running the code, to avoid any unpleasant version control issues or interactions with other projects you're working on. See the env.yml file for configuration details. Note the package requires python 3.7 to run.