[MEANS, COVARIANCES, PRIORS] = VL_GMM(X, NUMCLUSTERS) fits a GMM with NUMCLUSTERS components to the data X. Each column of X represent a sample point. X may be either SINGLE or DOUBLE. MEANS, COVARIANCES, and PRIORS are respectively the means, the diagonal covariances, and the prior probabilities of the Guassian modes. MEANS and COVARIANCES have the same number of rows as X and NUMCLUSTERS columns with one column per mode. PRIORS is a row vector with NUMCLUSTER entries summing to one.
[MEANS, COVARIANCES, PRIORS, LL] = VL_GMM(...) returns the loglikelihood (LL) of the model as well.
[MEANS, COVARIANCES, PRIORS, LL, POSTERIORS] = VL_GMM(...) returns the posterior probabilities POSTERIORS of the Gaussian modes given each data point. The POSTERIORS matrix has NUMCLUSTERS rows and NUMDATA columns.
VL_GMM() supports different initialization and optimization methods. Specifically, the following options are supported:
- Verbose
Increase the verbosity level (may be specified multiple times).
- Initialization RAND
RAND initializes the means as random data poitns and the covaraince matrices as the covariance of X. CUSTOM allow specifying the initial means, covariances, and prior probabilities.
- InitMeans none
Specify the initial means (size(X,1)-by-NUMCLUSTERS matrix).
- InitPriors none
Specify the initial weights (a vector of dimension NUMCLUSTER).
- InitCovariances none
Specify the initial diagonal covariance matrices
- NumRepetitions 1
Number of times to restart EM. The solution with maximum loglikelihood is returned.
- CovarianceBound 10e-6
Set the lower bound on the diagonal covariance values. The bound can be either a scalar or a vector with one entry per dimension. Using null bounds is possible, but may yield degenerate solutions, including NaNs.
- Example
VL_GMM(X, 10, 'verbose', 'MaxNumIterations', 20) learns a mixture of ten Gaussians using at most twenty iterations of the algorithm.