Disentangling the latent space of Variational Autoencoder

Disentangling the latent space learned by a VAE has been subject to a lot of research in recent years [1,2]. This repository provides a simple tensorflow implementation of a Variational Autoencoder and the objective function proposed by [1]:

Experiments showed that this modification to the objective function improves the tradeoff between reconstruction quality and disentanglement.

Usage

1. Run pip install -r requirements.txt
2. Either run python visualize.py to visualize an existing model or python train.py to train a new model.
3. The Dsprites dataset used in the repository can be found here:

Demo

Trained this model for 1200 epochs.

References

[1] HIGGINS, I., MATTHEY, L., PAL, A., BURGESS, C., GLOROT, X., BOTVINICK, M., MOHAMED, S., AND LERCHNER, A. beta-vae: Learning basic visual concepts with a constrained variational framework. In ICLR (2017).
[2] BURGESS, C. P., HIGGINS, I., PAL, A., MATTHEY, L., WATTERS, N., DESJARDINS, G., AND LERCHNER, A. Understanding disentangling in β-VAE. ArXiv abs/1804.03599 (2018).