ad_gmix_trainscript: Interface to Baggenstoss'es EM-Algorithm to calculate...

In Mthrun/AdaptGauss2D: Gaussian Mixture Modeling in 2D

Interface to Baggenstoss'es EM-Algorithm to calculate Gaussian Mixture Model

Description

Interface to Baggenstoss'es EM-Algorithm to calculate Gaussian Mixture Model for given data.

Usage

ad_gmix_trainscript(Parm, Data, Nit, SamplesPerMode=NULL, Bias=0, Maxclose=NULL,
Addmodes=1, KurtosisThreshold=1.0, Verbose=0)

Arguments

Parm	Nested list with parameters for GMM. Features carrying permanent values. Features$name [1:d] String vector with feature names. Features$min_std [1:NMODE] Vector of covariance constraints. Modes carrying modifyable values. Modes$cholesky_covar [d*NMODE, d] Numerical matrix with NMODE many square matrices stacked vertically with the covariance matrix. Modes$mean [1:NMODE, d] Numerical matrix with nmode different means and d feature dimensions. Modes$weight [1, 1:NMODE] Numerical matrix with weights for each mean.
Data	[1:n,1:d] Numerical matrix with normalized data. N samples with DIM feature dimensions.
Nit	Numerical value: max number of iterations.
SamplesPerMode	Optional: Numerical value: Samples-per-mode (minimum for pruning). Default: 4*d.
Bias	Optional: Binary value: Covariance constraint method. Choose: 1=BIAS, 0=CONSTRAINT. Default=0.
Maxclose	Optional: Numerical value: Maximum mode closeness. Use more negative values to promote mode consolidation. Use higher values for larger dimension (Default: Suggestion: -2 * d).
Addmodes	Optional: 0 or 1 value: If set to 1, will use kurt.m to split modes. Default = 1.
KurtosisThreshold	Optional: Default=1.0. Numerical value: Kurtosis and Skew threshold for mode splitting. Should be about 1.0. Higher values (i.e. 1.2) will make mode splitting less likely.
Verbose	Optional: Optional: 0, 1 or 2. 0 = no output, 1 = messages, 2 = messages + plot. Default ==0. Print some outputs.

Value

List with multiple elements of different data structures:

EMmean	[1:l,1:d] Numerical matrix carrying L means row wise.
EMCov	[1:(l*d),1:d] Numerical matrix with all covariance matrices appended beneath each other.
EMalpha	[1:l] Vector with wights for all L modes.
EMParams	Nested list with parameters for GMM. EMParams is the adapted or trained input variable Parm, see input argument Parm.

Author(s)

Quirin Stier

References

Baggenstoss, Paul M., and T. E. Luginbuhl.: An EM algorithm for joint model estimation. IEEE International Conference on Acoustics, Speech, and Signal Processing. Proceedings. ICASSP99 (Cat. No.99CH36258), Phoenix, AZ, USA, 1999, pp. 1825-1828 vol.4, IEEE, doi:10.1109/ICASSP.1999.758276, 1999.

Mthrun/AdaptGauss2D documentation built on July 19, 2022, 3:11 a.m.