MINIMALIST

Written by Giulio Isacchini, MPIDS Göttingen - ENS Paris and Natanael Spisak, ENS Paris

The code is written in Python3. Last updated on 24-05-2021

Reference: MINIMALIST: Mutual INformatIon Maximization for Amortized Likelihood Inference from Sampled Trajectories, Giulio Isacchini, Natanael Spisak, Armita Nourmohammad, Thierry Mora and Aleksandra M. Walczak

To reproduce the figures

In order to reproduce the plots you need to run the following commands.

1. Install the mimsbi package

Enter the mimsbi folder and run the command python setup.py install

2. Run analysis

To run the analysis for the the 4 task run the script run_analysis.py with the options: ou,bd,sir and/or lorenz.

3. Plot the results.

For Figure 2, run fig2.py

For Figure 3 run fig3.py

For Figure 4, run fig4.py

Extended usage

This directory includes a stable version of the mimsbi package.

The package allows to infer the likelihood-to-evidence ratio model using one of three objective functions: MINE, FDIV or BCE. The package has implemented simulators for the processes studied in the MINIMALIST paper: Ornstein-Uhlenbeck, birth-death, SIR and Lorenz processes. To add another functionality one needs to add a new Simulator class to mimsbi/models. Then, inference can be performed using the DensityRatioEstimator class. For example of usage go to the scripts directory where separate files can be used to

1. simulate the data scripts/simulate.py
2. tune network hyperparameters scripts/infer_hyperpars.py
3. likelihood-to-evidence ratio inference scripts/infer_estimators.py
4. posterior evaluation scripts/compare_estimators.py

To use the above scripts with a new model, its specifications need to be added in scripts/utils.py return_pars function. A simple data generation to posterior evaluation protocol is also available in the mimsbi/tutorial.ipynb Jupyter notebook.

Requisites

• tensorflow>2.1
• numpy
• pandas
• scipy
• matplotlib
• tqdm