- Code for IO-HMMs is built up on SSM (https://github.com/lindermanlab/ssm) toolbox; we added functions on top of the available repo. Installation of this package requires running the following on terminal:
pip install -e .
-
Packages required to run scripts in this repo are available in al.yaml.
-
Note: For Gibbs sampling using polya-gamma augmentation, we used code from https://github.com/slinderman/pypolyagamma. It is part of al.yaml but can also be installed using:
pip install pypolyagamma
- run_iohmm.py trains IO-HMMs using different input selection schemes (infomax_gibbs/ infomax_VI/ random) and stores all values (error/entropy/selected inputs) to plot in Results_IOHMM. This also provides different variants of Gibbs sampling for fitting the model: our laplace-based Gibbs (called 'gibbs'), Gibbs sampling with Polya-Gamma augmentation (gibbs_PG), and laplace-based Gibbs with parallel chains (called 'gibbs_parallel), which we compare in the supplement. Example:
python run_iohmms.py --seed 1 --input_selection infomax_gibbs --fitting_method gibbs
- run_modelmismatch.py trains IO-HMMs using the inputs selected by a single GLM (the model-mismatch analysis) and stores all values to plot in Results_modelmismatch. It takes a value for random seed. Example:
python run_modelmismatch.py --seed 1
-
state_predictions.py performs the state inference analysis.
-
plotting_nbs/plotIOHMM.ipynb plots results stored after running run_iohmm.py, and replicates Fig 5 from the paper. plotting_nbs/Suppfigs_IOHMM.ipynb replicates supplemental IO-HMM figures.
-
run_mglm.py trains MGLMs using active learning as well as random sampling and stores all values (error/entropy) to plot in Results_MGLM. It takes a value for random seed. Example:
python run_mglm.py --seed 1 -
plotting_nbs/plotMGLM.ipynb plots results stored after running run_mglm.py
(Note that code for MGLM is independent of the SSM toolbox and only requries numpy/scipy/autograd)