Laurens Van Der Maaten's parametric implementation of t-SNE.
Laurens' original implementation is for Matlab, here we are running in Octave with oct2py in the notebook Parametric t-SNE (Original). This code can take hours to complete.
In the Parametric t-SNE (Keras) notebook there is an implementation of the same technique by reimplementing all functions in Python with numpy and Keras. The code runs significantly faster (on my machine, 20 minutes). There are also some work-in-progress experiments, like using the pairwise probabilty embedding to pre-train the weights of an autoencoder, which appears to converge to a lower reconstruction error than a vanilla autoencoder.
On OS X:
$ brew install octave
$ pip install -r requirements.txt
$ jupyter notebook