Code for our paper -- Hyperprior Induced Unsupervised Disentanglement of Latent Representations (AAAI 2019)
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.
Traverse to data/
and run setup_2dshapes.sh
and setup_corr-ell.sh
to set up 2dshapes
and correlated_ellipses
datasets.
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.
Currently the repository includes code for experimenting on the following datasets.
For additonal qualitative results, please check AdditionalResults.md.
For any questions regarding the code or the paper, please email abdulfatir@u.nus.edu.
@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}
}