Which implementation of t-sne?

[vu-minh]

Implementation of t-sne

• scikit-learn
• Van de matten (python): https://lvdmaaten.github.io/tsne/code/tsne_python.zip
• Van de matten Barnes-Hut t-SNE (C++) : https://github.com/lvdmaaten/bhtsne/
• Multicore-TSNE (only for bookmark): https://github.com/DmitryUlyanov/Multicore-TSNE

params of t-sne we should choose?

• fix learning_rate = 100.0
• perplexities: which values should we use (#5 )?

[vu-minh]

MNIST, 2500 examples:

1. Scikit-leanr t-SNE: code runs in a Single core

   real 0m47,447s user 0m45,440s sys 0m2,004s (In fact with the default param of sklearn TSNE package, it uses barnes-hut implementation with Cython)

2. Van de matten implementation: The code can run with Multi cores (numpy mechanism ?) but very slow:

   real 8m8,653s user 15m36,796s sys 1m51,731s

3. C++ Barnes-Hut with python wrapper runs in Single core:

   real 0m17,285s user 0m17,859s sys 0m0,113s