Problem with d=3

[kylemcdonald]

I get buggy results when running with d=3. Here is some example code comparing d=2 to d=3.

%matplotlib inline
from matplotlib import pyplot as plt
import numpy as np
from tsne import bh_sne
X = np.random.random((1000,100))

X_2d = bh_sne(X, d=2)
plt.scatter(X_2d[:,0], X_2d[:,1], s=2, edgecolor='')
plt.show()

X_3d = bh_sne(X, d=3)
plt.scatter(X_3d[:,0], X_3d[:,1], s=2, edgecolor='')
plt.show()
plt.scatter(X_3d[:,0], X_3d[:,2], s=2, edgecolor='')
plt.show()

[phunterlau]

I had the same problem and I can confirm it is from the implementation bug. The scikit-learn TSNE implementation and this wrapper https://github.com/dominiek/python-bhtsne can give the correct result in 3D, and the later one runs fast (although it is gcc-6 incompatible)