dev4488 / VAE_gumble_softmax

pytorch implementation of VAE-Gumble-Softmax

VAE with Gumbel-Softmax in Pytorch

Pytorch implementation of a Variational Autoencoder with Gumbel-Softmax Distribution. Refer to the following paper:

• Categorical Reparametrization with Gumbel-Softmax by Jang, Gu and Poole

Table of Contents

• Installation
• Ananconda
• Results

Installation

The program requires the following dependencies (easy to install using pip or Ananconda):

• python 2.7/3.5
• pytorch (version 0.3.1)
• numpy

Anaconda: Train

Train VAE-Gumbel-Softmax model on the local machine using MNIST dataset:

python vae_gumbel_softmax.py

Results

Hyperparameters

Batch Size:                         128
Learning Rate:                      0.0001
Initial Temperature:                1.0
Minimum Temperature:                0.5
Anneal Rate:                        0.00003
Learnable Temperature:              False

MNIST

Ground Truth/Reconstructions	Generated Samples