Code for "Distilling importance sampling" paper

This code is to reproduce the experiments in the paper Distilling importance sampling for likelihood-free inference.

Version history

• v1.0 is Tensorflow code used in versions 1-3 of the arxiv paper
• v1.1 adds some bug fixes
• v2.0 is PyTorch code for version 4 of the arxiv paper
• V2.1 has some minor modifications to plotting code
• V2.2 is PyTorch code for version 5 of the arxiv paper
• V2.3 is PyTorch code for version 6 of the arxiv paper

Python package versions

For the paper the code was run using python 3.8.10 and the following package versions:

• matplotlib 3.5.1
• networkx 2.7.1
• nflows 0.14
• numpy 1.22.1
• pandas 1.4.0
• torch 1.10.2
• UMNN 1.68

Contents

General DIS files

• DIS.py contains the inference algorithms. The train method performs the Distilling Importance Sampling algorithm.
• The models directory contains code for various example models. These encapsulate prior + model + tempering scheme. There is also some general code for storing and manipulating weighted samples.
• utils.py has some utility functions used in various places in the code.

DIS experiment scripts

The distilled importance sampling experiments in the paper can be reproduced using the scripts

• MG1_comparison.py
• sin_example.py
• SInetwork_example1.py (this also runs a likelihood-based analysis)
• SInetwork_example2.py

Note .out files contains terminal output from running the corresponding script. This often includes timing information.

Other experiments

The abc folder contains python code for the ABC-PMC analyses. The experiments can be run using the script MG1_ABC.py.

Plots

These scripts create plots used in the paper. Some other plots are created in within the experiment scripts.

• MG1_post_plots.py - plots for the MG1 example results
• abc/MG1_ABC_plots.py - plots for ABC baseline on MG1 example

The scripts above require various results files, which are the ad hoc collection of .npy, .pkl, .csv and .mat files in the repository.

Miscellaneous

• LikelihoodBased.py and SInetworkLikelihood.py help perform likelihood based calculations for the SI network example.