Barnes-Hut t-SNE for Python

This is a fork of a fork of a fork. I have not looked at most of the code in this repository, I just use it.

This project, also known as "Python-TSNE", may be the only working Python 3 implementation of Barnes-Hut t-SNE. It's a sadly unloved project, perhaps because everyone expects it to be superseded by scikit-learn any day now.

But, as of May 2017, scikit-learn's t-SNE has a bug that causes it to get no benefit from Barnes-Hut; it uses as much space as plain t-SNE, and runs out of memory on moderately large data. So, given that:

• osdf made a Cython wrapper of the original C++ code by L. J. P. van der Maaten.
• danielfrg made it into a Python package that could be installed with pip, using a GitHub URL.
• alexisbcook fixed namespace problems that were causing danielfrg's package to fail to import in some situations.
• I made alexisbcook's version into a package on PyPI with a coherent version number, so that other code can depend on it.

I have no intention of continuing to maintain this project. If something about it doesn't quite work for you, I suggest you continue the tradition and make your own fork.

Requirements

• numpy > =1.7.1
• scipy >= 0.12.0
• cython >= 0.19.1
• cblas or openblas.

Anaconda is one way to get these requirements. (I don't use it myself.)

Installation

The point of this particular fork is to be installable through PyPI:

pip install barnes-hut-tsne

Usage

Basic usage:

from tsne import bh_sne
X_2d = bh_sne(X)

I found this docstring:

bh_sne(data)
Run Barnes-Hut T-SNE on _data_.

@param data         The data.

@param pca_d        The dimensionality of data is reduced via PCA
                    to this dimensionality.

@param d            The embedding dimensionality. Must be fixed to
                    2.

@param perplexity   The perplexity controls the effective number of
                    neighbors.

@param theta        If set to 0, exact t-SNE is run, which takes
                    very long for dataset > 5000 samples.

@param random_state A numpy RandomState object; if None, use
                    the numpy.random singleton. Init the RandomState
                    with a fixed seed to obtain consistent results
                    from run to run.

@param copy_data    Copy the data to prevent it from being modified
                    by the C code

Examples

• Iris
• MNIST

More Information

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