From eb4ac2d1b42388da3a1f849bfdff7eddc79841bc Mon Sep 17 00:00:00 2001 From: Kyle McDonald Date: Tue, 30 Aug 2016 16:04:13 -0400 Subject: [PATCH] Update readme.md --- readme.md | 8 +++++--- 1 file changed, 5 insertions(+), 3 deletions(-) diff --git a/readme.md b/readme.md index 276afb6..f142d5c 100644 --- a/readme.md +++ b/readme.md @@ -1,8 +1,10 @@ #Parametric t-SNE -Laurens Van Der Maaten's [parametric implementation](https://lvdmaaten.github.io/publications/papers/AISTATS_2009.pdf) of t-SNE. The original implementation is for Matlab, here we are running in Octave with oct2py. +Laurens Van Der Maaten's [parametric implementation](https://lvdmaaten.github.io/publications/papers/AISTATS_2009.pdf) of t-SNE. -I'm currently working on porting the technique to a more recent toolkit. +Laurens' original implementation is for Matlab, here we are running in Octave with oct2py in the notebook `Parametric t-SNE (Original)`. This code can take hours to complete. + +In the `Parametric t-SNE (Keras)` notebook there is an implementation of the same technique by reimplementing all functions in Python with numpy and [Keras](https://keras.io/). The code runs significantly faster (on my machine, 20 minutes). There are also some work-in-progress experiments, like using the pairwise probabilty embedding to pre-train the weights of an autoencoder, which appears to converge to a lower reconstruction error than a vanilla autoencoder. ## Setup @@ -12,4 +14,4 @@ On OS X: $ brew install octave $ pip install -r requirements.txt $ jupyter notebook -``` \ No newline at end of file +```