Pytorch Lightning implementation for "Generative Adversarial Imputation Networks (GAIN)"

An implementation of the GAIN framework for imputation using Pytorch Lightning

Table of Contents

1. About The Project
2. Built With
3. Datasets
4. Folder structure
5. How to run it
• Creation of a Docker container
• Command inputs
• Example command
6. How to replicate the results of the original paper
7. License

About The Project

Original authors: Jinsung Yoon, James Jordon, Mihaela van der Schaar

Paper: Jinsung Yoon, James Jordon, Mihaela van der Schaar, "GAIN: Missing Data Imputation using Generative Adversarial Nets," International Conference on Machine Learning (ICML), 2018.

• Original Github repository: https://github.com/jsyoon0823/GAIN
• Paper Link: http://proceedings.mlr.press/v80/yoon18a/yoon18a.pdf
• Supplementary material: http://proceedings.mlr.press/v80/yoon18a/yoon18a-supp.pdf

Built With 🔨

Datasets

This directory contains implementations of GAIN framework for imputation using the main five datasets used in the original paper:

• UCI Letter (https://archive.ics.uci.edu/ml/datasets/Letter+Recognition)
• UCI Spam (https://archive.ics.uci.edu/ml/datasets/Spambase)
• UCI Credit (https://archive.ics.uci.edu/ml/datasets/default+of+credit+card+clients)
• UCI Breast Cancer (https://archive.ics.uci.edu/ml/datasets/Breast+Cancer+Wisconsin+(Diagnostic))
• UCI Online News Popularity (https://archive.ics.uci.edu/ml/datasets/Online+News+Popularity)

Folder structure

.
├── data                        # Contains the raw data
├── docker                      # Contains the files to create a Docker container
├── src                         # Source files 
│   ├── data                    # Scripts to load and preprocess data
│   └──  models                 # Scripts that define the GAIN model, the training loop and the MLP base for GAIN
├── reports                     # Folder generated by running, contains the results of the experiments (Tensorboard logs, etc...)
├── main.py                     # Main script to run an experiment
├── replicate_table1_paper.py   # Script to replicate the results of the table 1 of the original paper, saves the results in a reports folder
├── setup.sh                    # Script that creates a Docker container
├── requirements.txt            # Requirements file
├── logo.png                    # Logo used in the README
├── LICENSE
└── README.md

How to run it

To run the pipeline for training and evaluation on GAIN framework, simply run python3 -m main.py.

Note that any model architecture can be used as the generator and discriminator model such as multi-layer perceptrons or CNNs.

Creation of a Docker container:

If you want to run the code in a Docker container, you can use the following commands:

1. Give execution permissions to the setup.sh file:

$ chmod +x setup.sh

2. Run the setup.sh file:

$ ./setup.sh

If you have exited the container, you can access it again by running the setup.sh file again.

Command inputs:

• data_name: letter, spam, credit, breast or news
• miss_rate: probability of missing components
• batch_size: batch size
• hint_rate: hint rate
• alpha: hyperparameter
• iterations: iterations

Example command

$ python3 main.py --data_name spam 
--miss_rate: 0.2 --batch_size 128 --hint_rate 0.9 --alpha 100
--iterations 10000

How to replicate the results of the original paper:

If you want to replicate the results of the table 1 of the original paper, you can use the following command:

$ python3 replicate_table1_paper.py

License

This project is licensed under the Apache License 2.0 - see the LICENSE file for details