Straight Through Gumbel-Softmax Estimator implemented as per paper: Categorical Reparameterization with Gumbel-Softmax (No temperature, learning rate annealing. Hard prior used)
Code developed from:
-
Eric Jang: https://github.com/ericjang/gumbel-softmax/blob/master/gumbel_softmax_vae_v2.ipynb
-
Google Seedbank Convolutional Variational Autoencoder https://tools.google.com/seedbank/seed/5719238044024832
- Tensorflow 1.10.0
- Numpy 1.14.5
- Epochs = 10
- Temperature = 1
- Learning Rate = 0.001
- Number of categorical distributions = 30
- Anneal temperature and learning rate
- Use relaxed prior
- Increase the number of epochs
- Increase the number of categorical distributions to sample from
Example of generated MNIST images from 100 test samples