snitish / libgmm

A library for training Gaussian Mixture Models written in C, with wrappers for MATLAB and Python.

libgmm

A library for training Gaussian Mixture Models written in C.

How to build

• To enable OpenMP, add the '-fopenmp' option to line 2 of Makefile

CFLAGS = -std=c99 -O3 -fopenmp

• To build the library, navigate to the libgmm directory using the terminal and type

make

• To build the MATLAB wrapper, run matlab/make.m from the MATLAB console.
• To build and the Python wrapper, navigate to the libgmm/python directory using the terminal and type

python setup.py install

Usage

• Using C API

Refer to test.c

• MATLAB wrapper

gmm = trainGMM(X, k, 'Name', 'Value', ...);

Where,
X = NxD data matrix containing N data points, each of length D
k = Number of GMM components

Optional name-value pairs:

• CovType = Covariance matrix type: "diagonal" or "spherical". (Default "diagonal")
• MaxIter = Maximum number of EM iterations. (Default 1000)
• ConvergenceTol = Convergence tolerance. (Default 1e-6)
• RegularizationValue = Regularization Value (small value added to covariance matrix to prevent it from being singular). (Default 1e-6)
• InitMethod = GMM parameter initialization method. Can be 'random' or 'kmeans'. (Default 'random')
• Python wrapper

import gmm
gmm1 = gmm.GMM(k=1, CovType='diagonal', MaxIter=1000, ConvergenceTol=1e-6, RegularizationValue=1e-6, InitMethod='random')
gmm1.fit(X)

Where,

• X = NxD numpy matrix containing N data points, each of length D
• k = Number of GMM components. (Default 1)
• CovType = Covariance matrix type: "diagonal" or "spherical". (Default "diagonal")
• MaxIter = Maximum number of EM iterations. (Default 1000)
• ConvergenceTol = Convergence tolerance. (Default 1e-6)
• RegularizationValue = Regularization Value (small value added to covariance matrix to prevent it from being singular). (Default 1e-6)
• InitMethod = GMM parameter initialization method. Can be 'random' or 'kmeans'. (Default 'random')