PyTorch Reimplementation - MisGAN: Learning From Incomplete Data With Generative Adversarial Networks

MSc Coursework Project in COMP6248 Deep Learning

This project reimplements MisGAN in PyTorch according to the description in the original paper. In particular, MisGAN has 2 types of architecture: convolutional (Conv-MisGAN) and fully-connected (FC-MisGAN). The reimplementation focuses on MNIST data only for a qualitative comparison between our results and original authors'. See the report for the reimplementation detail, results, and evaluation.

Running

Source code is located in the src directory. Jupyter notebooks in the test directory can also be run in isolation.

In the src directory,

Conv-MisGAN on MNIST:

python conv_misgan.py

FC-MisGAN on MNIST:

python fc_misgan.py

Requirements

This code was tested on:

• Python 3.6
• PyTorch 1.5.0
• Google Colab

References

Research papers included in the references folder

• Original Paper (OpenReview)
• DCGAN (arXiv)
• WGAN with Gradient Penalty (arXiv)