HNPE: Hierarchical Neural Posterior Estimation

What you will find here

This repository contains code related to the method we have proposed in our NeurIPS 2021 paper

P Rodrigues, T Moreau, G Louppe, A Gramfort "HNPE: Leveraging Global Parameters for Neural Posterior Estimation". Proc. Advances in Neural Information Processing Systems (NeurIPS) 34, 2021

We have included basic code for reproducing the two numerical illustrations presented in the paper:

• A toy model for which we can derive all analytic properties of the posterior distribution
• The Jansen-Rit neural mass model that we have used to relate physiological parameters to real EEG signals

Most of our code is based on the wonderful sbi package.

Setup

Clone the GitHub repository on your local computer, and install the package by running:

pip install -e .

All dependencies with pinned versions are listed in requirements.txt if you want to reproduce the exact environment we used to run the original computations.

To ensure that your code finds the right scripts, run:

python -c 'import hnpe'

Note that you might want to create a conda environment before doing all these installations, e.g.:

conda create -n hnpe_env python=3.7

Usage

Example 1 - Toy Model

To obtain an approximation of the posterior distribution of our toy model when only one observation is available and no noise is added to it, you should go to Ex1-ToyModel and enter in your terminal

python inference.py

This creates an approximation of the posterior distribution and stores its parameters in /results

To check the results, you can simply enter

python inference.py --viz

Now if you would like to see what happens to the posterior distribution when N = 10 extra observations are available, you should enter

python inference.py --nextra 10

To include noise in the observations, you should run, for instance,

python inference.py --nextra 10 --noise 0.05

Example 2 - Jansen & Rit Neural Mass Model

The way of doing things for Ex2-JRNMM is exactly the same, except for a few choices of input parameters.

Please do

python inference.py --help

for a list of all options available for each example.

Important: To run the examples in Ex2-JRNMM you will need to make sure that the R code in https://github.com/massimilianotamborrino/sdbmpABC runs on your computer! This is a C++ implementation of the Jansen-Rit neural mass model compiled for R and which we bind to python.