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.
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
This code was tested on:
- Python 3.6
- PyTorch 1.5.0
- Google Colab
Research papers included in the references
folder
- Original Paper (OpenReview)
- DCGAN (arXiv)
- WGAN with Gradient Penalty (arXiv)