bdecost / ntsne

A simple numpy wrapper for the reference implementation of Barnes-Hut t-SNE

ntsne.py

a python3 numpy wrapper for bh_tsne (https://github.com/lvdmaaten/bhtsne)

Usage

ntsne uses bh_tsne via python's subprocess library. If bh_tsne is not on the system path or in the current working directory, ntsne will attempt to clone and compile the bhtsne binary under ~/.ntsne.

Just pass a numpy array containing your high-dimensional data matrix to ntsne.tsne, which returns a numpy array containing the t-SNE map coordinates:

import ntsne
import matplotlib.pyplot as plt
from sklearn.datasets import load_digits

digits = load_digits()
X_tsne = ntsne.tsne(digits.data, perplexity=30, theta=0.5)

plt.scatter(X_tsne[:,0], X_tsne[:,1], c=digits.target)
plt.show()

ntsne.best_tsne will run bh_tsne multiple times with the same parameters, returning the t-SNE results with the lowest KL-divergence.

X_tsne = ntsne.best_tsne(digits.data, perplexity=30, theta=0.5, n_repeats=3)

plt.scatter(X_tsne[:,0], X_tsne[:,1], c=digits.target)
plt.show()