This is an improved implementation of the paper Stochastic Gradient VB and the Variational Auto-Encoder by D. Kingma and Prof. Dr. M. Welling. This code uses ReLUs and the adam optimizer, instead of sigmoids and adagrad. These changes make the network converge much faster.
To run the MNIST experiment:
###NB: This code is not as nicely polished as the Torch7 and Theano version. It is mainly for playing around with TensorFlow, which is why I tried to add as many of its bells and whistles as possible. PRs to make it more "TensorFlowy" are welcomed! Specifically if I made a mistake that causes a slow down.
There is no continuous version for now, but there will probably be one in the near future.
The code is MIT licensed.