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.
In my other repository the implementation is in Torch7 (lua), this version is based on Theano (Python). To run the MNIST experiment:
Setting the continuous boolean to true will make the script run the freyfaces experiment. It is necessary to tweak the batch_size and learning rate parameter for this to run smoothly.
There used to be a scikit-learn implementation too, but it was very slow and outdated. You can still find it by looking at the code at this commit
The code is MIT licensed.