Code for Robust-KSD Test

This repository contains the code for reproducing the experiments in the paper

• Liu, X. and Briol, F.X., 2024. On the Robustness of Kernel Goodness-of-Fit Tests. arXiv preprint arXiv:2408.05854.

How to install?

Install as a package

Before running any scripts, run the following to install the current package and the dependencies.

pip install git+https://github.com/XingLLiu/robust-kernel-test

Install dependencies only

Alternatively, to install only the dependencies but not as a package, run

pip install -r requirements.txt

Example

After installing as a package, it can be loaded as a Python module using

import rksd

See notebooks/example.ipynb for an example of how to use this module to perform the robust-KSD test.

Reproducibility

To reproduce all figures in the paper,

1. Run sh sh_scripts/all.sh to generate all results. This can take ~7 hours.
2. Run the corresponding notebooks in notebooks/ to generate the plots.

You can also reproduce an individual experiment by running its corresponding shell script. E.g., to reproduce the Gaussian-Bernoulli Restricted Boltzmann Machine experiment, run sh sh_scripts/rbm.sh.

Folder structure

.
├── rksd                          # Source files for robust-KSD test and benchmarks
├── sh_scripts                    # Shell scripts to run experiments
├── data                          # Folder to store Galaxy data and experimental results
├── figs                          # Folder to store figures
├── experiments                   # Scripts for experiments
├── notebooks                     # Jupyter notebooks for tutorial and generating plots
├── setup.py                      # Setup file for easy-install of rksd
├── requirements.txt              # Package dependencies 
└── README.md