pclucas14/GMVAE

Implementation of Gaussian Mixture Variational Autoencoder (GMVAE) for Unsupervised Clustering

Gaussian Mixture Variational Autoencoder

Tensorflow	Pytorch

Implementation of Gaussian Mixture Variational Autoencoder (GMVAE) for Unsupervised Clustering in PyTorch and Tensorflow. The probabilistic model is based on the model proposed by Rui Shu, which is a modification of the M2 unsupervised model proposed by Kingma et al. for semi-supervised learning. Unlike other implementations that use marginalization for the categorical latent variable, we use the Gumbel-Softmax distribution, resulting in better time complexity because of the reduced number of gradient estimations.

Dependencies

1. Tensorflow. We tested our method with the 1.13.1 tensorflow version. You can Install Tensorflow by following the instructions on its website: https://www.tensorflow.org/install/pip?lang=python2.

• Caveat: Tensorflow released the 2.0 version with different changes that will not allow to execute this implementation directly. Check the migration guide for executing this implementation in the 2.0 tensorflow version.

2. PyTorch. We tested our method with the 1.3.0 pytorch version. You can Install PyTorch by following the instructions on its website: https://pytorch.org/get-started/locally/.

3. Python 3.6.8. We implemented our method with the 3.6.8 version. Additional libraries include: numpy, scipy and matplotlib.

References

• Semi-Supervised Learning with Deep Generative Models
• Categorical Reparameterization with Gumbel-Softmax
• Gaussian Mixture VAE: Lessons in Variational Inference, Generative Models, and Deep Nets