This repository provides a Python package for sampling approximate model-X knockoffs using deep generative models.
Accompanying paper: https://arxiv.org/abs/1811.06687. Published in the Journal of the American Statistical Association (https://doi.org/10.1080/01621459.2019.1660174)
To learn more about the algorithm implemented in this package, visit https://web.stanford.edu/group/candes/deep-knockoffs/ and read the accompanying paper.
To learn more about the broader framework of knockoffs, visit https://web.stanford.edu/group/candes/knockoffs/.
The code contained in this repository was tested on the following configuration of Python:
- python=3.6.5
- numpy=1.14.0
- scipy=1.0.0
- pytorch=0.4.1
- cvxpy=1.0.10
- cvxopt=1.2.0
- pandas=0.23.4
cd DeepKnockoffs
python setup.py install --user- examples/toy-example.ipynb A usage example on a toy problem with multivariate Gaussian variables is available in the form of a Jupyter Notebook.
- examples/experiments-1.ipynb Code to train the machine used in the paper.
- examples/experiments-2.ipynb Code to compute the goodness-of-fit diagnostics for the machine used in the paper.
- examples/experiments-3.ipynb Code to perform the controlled variable selection experiments in the paper.
- examples/data-preprocessing.ipynb Example of how to pre-process data containing extremely correlated variables.
This software is distributed under the GPLv3 license and it comes with ABSOLUTELY NO WARRANTY.