Code for Multivariate soft rank via entropic optimal transport: sample efficiency and generative modeling

This repository provides two applications of novel multivariate soft rank energy (sRE) and soft rank mmd (sRMMD). (a) Developing a generative model using sRE and sRMMD as the loss functions to produce MNIST-digits, (b) utilizing sRMMD as the loss in a deep generative model to produce valid knockoffs in order to select statistically significant features.

Package Dependencies to sRMMMD-based knockoff filter

• 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

How to run the code

1. To reproduce the MNIST results from the paper:
   

   • Figure 1(b)- run 'mnist_figures_geneartion.py'
     
   • Figure 1(a)- use lossType = 'mmd' and run 'mnist_figures_geneartion.py'
     
   • Figure 1(c)- use lossType = 'sRMMD' and run 'mnist_figures_geneartion.py'
     

2. To reproduce knockoff figures from the paper
   

   • Extra package dependencies for other benchmarks
     
     • DDLK : install the package from https://github.com/rajesh-lab/ddlk
     • KnockoffGAN : install Tensorflow v2 and use code from https://bitbucket.org/mvdschaar/mlforhealthlabpub/src/master/
   • Reproducing Figure 2(c)- run 'knockoff_figures_geneartion.py
   • Figure 2(a)- use distType = 'GaussianAR1' and run 'knockoff_figures_geneartion.py
   • Figure 2(b)- use distType = 'GaussianMixtureAR1' and run 'knockoff_figures_geneartion.py
   • Figure 2(d)- use distType = 'SparseGaussian' and run 'knockoff_figures_geneartion.py
     

3. To reproduce Table 1 from the paper
   

   • run real_dataset.py

   N:B: In case of any error regarding package dependices while running 'mnist_figures_geneartion.py' and 'real_dataset.py', run each method separately.

Demo notebooks

These notebooks provide an overall view how sRMMD-knockoff filter works on synthetic and real data

• Examples/knockoff_synthetic_settings.ipynb code to generate valid knockoffs using sRMMD.
• Examples/knockoff_real_data.ipynb metabolites selection using sRMMD knockoffs on the real data set available in dataset/Real dataset