CHyVAE

Code for our paper Hyperprior Induced Unsupervised Disentanglement of Latent Representations (AAAI-19). The correlated ellipses dataset used in the paper can be found here.

Requirements

• Python 3
• Tensorflow (tested on 1.10.1)
• Numpy (tested on 1.14.5)
• OpenCV (tested on 3.4.3)

Usage

Setting up the datasets

Traverse to data/ and run setup_2dshapes.sh and setup_corr-ell.sh to set up 2dshapes and correlated_ellipses datasets.

Training a model

Traverse to code/ and run

python main.py \
       --dataset [2dshapes/correlated_ellipses] \
       --z_dim [dim. of latent space] \
       --n_steps [number of training steps] \
       --nu [degrees of freedom] \
       --batch_size [batch size]

The reconstruction error and disentanglement metric will be logged at a set interval as training proceeds.

Example Run

python main.py --dataset correlated_ellipses --z_dim 10 --n_steps 150000 --nu 200 --batch_size 50

Run python main.py -h for help.

Datasets

Currently the repository includes code for experimenting on the following datasets.

• 2DShapes
• CorrelatedEllipses

Additional Results

For additonal qualitative results, please check AdditionalResults.md.

Contact

For any questions regarding the code or the paper, please email abdulfatir@u.nus.edu.

BibTeX

@inproceedings{ansari2019hyperprior,
  title={Hyperprior Induced Unsupervised Disentanglement of Latent Representations},
  author={Ansari, Abdul Fatir and Soh, Harold},
  booktitle={AAAI Conference on Artificial Intelligence},
  year={2019}
}