Awesome

Mixtures of von Mises-Fisher Distributions (movMF)

movMF is a distribution model for D-dimensional spherical data. movMF is analogous to Gaussian distribution but in spherical space. This code is based on the article Clustering on the Unit Hypersphere using von Mises-Fisher Distributions of A. Banerjee.

Install

Get and Compile (you need boost):

git clone https://github.com/mrouvier/movMF
cd movMF
make

Generate randomly data:

python data/generate_data.py > data/vec

or

make generate_date

Program usage

To train a statistical mixture model :

./bin/movmf_train --nb_mixture 128 --nb_iteration_em 10  --train data/vec --save mixture.txt

where nb_mixture is the number of mixture, nb_iteration_em is the number of iteration of Exepctation Maximization (algorithm used to train the mixture model), train is the training file and save is a file, the program save the model to this file.

To test the mixture model :

./bin/movmf_test data/vec mixture.txt

Given a mixture model (mixture.txt) and a testing file (data/vec), the program caculate the log likelihood for each vector.