Python TSNE implementation utilizing openmp for performance

This is based on the 10XDev/tsne fork of L.J.P. van der Maaten BH-tSNE implementation.

It has fixes to allow this to run in Python 3 and performance has been significantly increased with OpenMP parallelism. (see: tsne-perf-test)

Note: While Scikit-learn v0.17 has a tsne implementation, this implementation performs significantly faster than scikit-learn's. If you need speed, use this.

Algorithms

Barnes-Hut-SNE

A python (cython) wrapper for Barnes-Hut-SNE aka fast-tsne.

We forked 10XDev's implementation and openmp enabled the code.

Installation

This library has been added to pypi as tsne-mp

pip install tsne-mp

It requires openmp support.

• OSX - brew install libomp
• linux - 'sudo apt-get install libgomp1'
• Windows - Included with Visual Studio C++

Usage

Basic usage:

from tsne import bh_sne
X_2d = bh_sne(X)

Or, the wheels also contain an executable that can be used from the command-line as described in the original project.

Examples

• Iris
• MNIST
• word2vec on presidential speeches via @prateekpg2455

More Information

See Barnes-Hut-SNE (2013), L.J.P. van der Maaten. It is available on arxiv.