Variational Autoencoder with Implicit Optimal Priors

This is a pytorch implementation of the following paper [AAAI] [arXiv]:

@inproceedings{takahashi2019variational,
  title={Variational Autoencoder with Implicit Optimal Priors},
  author={Takahashi, Hiroshi and Iwata, Tomoharu and Yamanaka, Yuki and Yamada, Masanori and Yagi, Satoshi},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  volume={33},
  pages={5066--5073},
  year={2019}
}

Please read license.txt before reading or using the files.

Prerequisites

Please install python>=3.6, torch, torchvision, numpy and scipy.

Usage

usage: main.py [-h] [--dataset DATASET] [--prior PRIOR]
               [--learning_rate LEARNING_RATE] [--seed SEED]

You can choose the dataset from following four image datasets: MNIST, OMNIGLOT, Histopathology and FreyFaces.
You can choose the prior of the VAE from normal (standard Gaussian prior) or iop (implicit optimal prior).
You can also change the random seed of the training and learning_rate of the optimizer (Adam).

Example

MNIST with standard Gaussian prior:

python main.py --dataset MNIST --prior normal

MNIST with implicit optimal prior:

python main.py --dataset MNIST --prior iop

Output

After the training, the mean of log-likelihood for test dataset will be displayed.
The detailed information of the training and test will be saved in npy directory.