Lipschitz standardization

This repository contains an implementation of the paper Lipschitz standardization for multivariate learning. The implementation is mostly based on Pytorch and Ignite and the repository is self-contained, meaning that it includes the list of dependencies, code, and datasets needed to run the experiments of the paper.

How to run the code

Everything needed to run the experiments shown in the paper is contained in the script main.py.

A Conda environment.yml file is provided with all the dependencies needed to run the code. In order to install all the dependencies (assuming that Conda is already installed), just run conda env create -f environment.yml. A new environment called lip-std will appeared, to activate it execute conda activate lip-std.

Running the algorithm is quite straightforward, a command line interface is provided and help is provided through

python main.py --help

As an example, we can run on the Breast dataset using the second mask that has 10% of missing data (located in dataset/Breast) by calling

python main.py -seed=7 -model=vae -dataset=datasets/Breast/ -miss-perc=10 -miss-suffix=2 -trick=gamma 

The previous command uses the Gamma trick (ours-gamma). Other options are -trick=bern for the Bernoulli trick (ours-bern) and -trick=none to only treat continuous variables (ours-none). Besides, standardization and normalization are supported via the -std-none and -max options. For example:

python main.py -seed=7 -model=vae -dataset=datasets/Breast/ -miss-perc=10 -miss-suffix=2 -std-none

For the sake of completeness, here is the output of the argument --help:

usage:  [-h] [-seed SEED] [-root ROOT] [-to-file] [-batch-size BATCH_SIZE]
        [-learning-rate LEARNING_RATE] [-max-epochs MAX_EPOCHS]
        [-print-every PRINT_EVERY] -model {mm,vae,mf}
        [-latent-size LATENT_SIZE] [-num-clusters NUM_CLUSTERS]
        [-hidden-size HIDDEN_SIZE] -dataset DATASET -miss-perc MISS_PERC
        -miss-suffix MISS_SUFFIX [-trick {gamma,bern,none}] [-max] [-std-none]

optional arguments:
  -h, --help            show this help message and exit
  -seed SEED
  -root ROOT            Output folder (default: results)
  -to-file              Redirect output to 'stdout.txt'
  -batch-size BATCH_SIZE
                        Batch size (default: 1024)
  -learning-rate LEARNING_RATE
                        Learning rate (default: 1e-2 if MF, 1e-3 otherwise)
  -max-epochs MAX_EPOCHS
                        Max epochs (default: as described in the appendix)
  -print-every PRINT_EVERY
                        Interval to print (default: 25)
  -model {mm,vae,mf}    Model to use: Mixture Model (mm), Matrix Factorization
                        (mf), or VAE (vae)
  -latent-size LATENT_SIZE
  -num-clusters NUM_CLUSTERS
  -hidden-size HIDDEN_SIZE
                        Size of the hidden layers (VAE)
  -dataset DATASET      Dataset to use (path to folder)
  -miss-perc MISS_PERC  Missing percentage
  -miss-suffix MISS_SUFFIX
                        Suffix of the missing percentage file
  -trick {gamma,bern,none}
                        Trick to use (if any)
  -max                  Normalize data
  -std-none             Standardize data