This repository contains cleaned-up code for reproducing the quantitative experiments in Isolating Sources of Disentanglement in Variational Autoencoders [arxiv].

To train a model:

python vae_quant.py --dataset [shapes/faces] --beta 6 --tcvae

Specify --conv to use the convolutional VAE. We used a mlp for dSprites and conv for 3d faces. To see all options, use the -h flag.

The main computational difference between beta-VAE and beta-TCVAE is summarized in these lines.

To evaluate the MIG of a model:

python disentanglement_metrics.py --checkpt [checkpt]

To see all options, use the -h flag.

Download the npz file from here and place it into data/.

We cannot publicly distribute this due to the license.

Email [email protected] if you have questions about the code/data.

@inproceedings{chen2018isolating,
  title={Isolating Sources of Disentanglement in Variational Autoencoders},
  author={Chen, Ricky T. Q. and Li, Xuechen and Grosse, Roger and Duvenaud, David},
  booktitle = {Advances in Neural Information Processing Systems},
  year={2018}
}