HexaGAN: Generative Adversarial Nets for Real World Classification

A Tensorflow Implementation of HexaGAN (Pytorch version will be uploaded)

When dealing with the real world data, we encounter three problems such as 1) missing data, 2) class imbalance, and 3) missing label problems. In this paper, we propose HexaGAN, a generative adversarial network framework that shows promising classification performance for all three problems.

• Reference: Uiwon Hwang, Dahuin Jung, Sungroh Yoon. “HexaGAN: Generative Adversarial Nets for Real World Classification.” International Conference on Machine Learning (ICML), 2019

• Paper URL: http://proceedings.mlr.press/v97/hwang19a/hwang19a.pdf

• Appendix URL: http://proceedings.mlr.press/v97/hwang19a/hwang19a-supp.pdf

• Presentation PPT: https://icml.cc/media/Slides/icml/2019/halla(11-14-00)-11-15-15-4629-hexagan_genera.pdf

Files

• ops.py: various operations for building neural networks and data loading

• ops_cnn.py: various operations for convolutional neural networks (for the MNIST dataset)

• model.py: HexaGAN model (for the breast dataset)

• train_breast.py: classification on the breast dataset with 20% missingness

• train_mnist.py: missing data imputation on the MNIST dataset with 50% missingness (including HexaGAN model)