train_class_gmmT_train_class_gmmtrain_class_gmmTrainClassGmmTrainClassGmm (Operator)

Name

train_class_gmmT_train_class_gmmtrain_class_gmmTrainClassGmmTrainClassGmm — Train a Gaussian Mixture Model.

Signature

Description

train_class_gmmtrain_class_gmmtrain_class_gmmTrainClassGmmTrainClassGmm trains the Gaussian Mixture Model (GMM) referenced by GMMHandleGMMHandleGMMHandleGMMHandleGMMHandle. Before the GMM can be trained, all training samples to be used for the training must be stored in the GMM using add_sample_class_gmmadd_sample_class_gmmadd_sample_class_gmmAddSampleClassGmmAddSampleClassGmm, add_samples_image_class_gmmadd_samples_image_class_gmmadd_samples_image_class_gmmAddSamplesImageClassGmmAddSamplesImageClassGmm, or read_samples_class_gmmread_samples_class_gmmread_samples_class_gmmReadSamplesClassGmmReadSamplesClassGmm. After the training, new training samples can be added to the GMM and the GMM can be trained again.

During the training, the error that results from the GMM applied to the training vectors will be minimized with the expectation maximization (EM) algorithm.

MaxIterMaxIterMaxIterMaxItermaxIter specifies the maximum number of iterations per class for the EM algorithm. In practice, values between 20 and 200 should be sufficient for most problems. ThresholdThresholdThresholdThresholdthreshold specifies a threshold for the relative changes of the error. If the relative change in error exceeds the threshold after MaxIterMaxIterMaxIterMaxItermaxIter iterations, the algorithm will be canceled for this class. Because the algorithm starts with the maximum specified number of centers (parameter NumCentersNumCentersNumCentersNumCentersnumCenters in create_class_gmmcreate_class_gmmcreate_class_gmmCreateClassGmmCreateClassGmm), in case of a premature termination the number of centers and the error for this class will not be optimal. In this case, a new training with different parameters (e.g. another value for RandSeedRandSeedRandSeedRandSeedrandSeed in create_class_gmmcreate_class_gmmcreate_class_gmmCreateClassGmmCreateClassGmm) can be tried.

ClassPriorsClassPriorsClassPriorsClassPriorsclassPriors specifies the method of calculation of the class priors in GMM. If 'training'"training""training""training""training" is specified, the priors of the classes are taken from the proportion of the corresponding sample data during training. If 'uniform'"uniform""uniform""uniform""uniform" is specified, the priors are set equal to 1/NumClasses for all classes.

RegularizeRegularizeRegularizeRegularizeregularize is used to regularize (nearly) singular covariance matrices during the training. A covariance matrix might collapse to singularity if it is trained with linearly dependent data. To avoid this, a small value specified by RegularizeRegularizeRegularizeRegularizeregularize is added to each main diagonal element of the covariance matrix, which prevents this element from becoming smaller than RegularizeRegularizeRegularizeRegularizeregularize. A recommended value for RegularizeRegularizeRegularizeRegularizeregularize is 0.0001. If RegularizeRegularizeRegularizeRegularizeregularize is set to 0.0, no regularization is performed.

The centers are initially randomly distributed. In individual cases, relatively high errors will result from the algorithm because the initial random values determined by RandSeedRandSeedRandSeedRandSeedrandSeed in create_class_gmmcreate_class_gmmcreate_class_gmmCreateClassGmmCreateClassGmm lead to local minima. In this case, a new GMM with a different value for RandSeedRandSeedRandSeedRandSeedrandSeed should be generated to test whether a significantly smaller error can be obtained.

It should be noted that, depending on the number of centers, the type of covariance matrix, and the number of training samples, the training can take from a few seconds to several hours.

On output, train_class_gmmtrain_class_gmmtrain_class_gmmTrainClassGmmTrainClassGmm returns in CentersCentersCentersCenterscenters the number of centers per class that have been found to be optimal by the EM algorithm. This values can be used as a reference in NumCentersNumCentersNumCentersNumCentersnumCenters (in create_class_gmmcreate_class_gmmcreate_class_gmmCreateClassGmmCreateClassGmm) for future GMMs. If the number of centers found by training a new GMM on integer training data is unexpectedly high, this might be corrected by adding a RandomizeRandomizeRandomizeRandomizerandomize noise to the training data in add_sample_class_gmmadd_sample_class_gmmadd_sample_class_gmmAddSampleClassGmmAddSampleClassGmm. IterIterIterIteriter contains the number of performed iterations per class. If a value in IterIterIterIteriter equals MaxIterMaxIterMaxIterMaxItermaxIter, the training algorithm has been terminated prematurely (see above).

Parallelization

• Multithreading type: exclusive (runs in parallel only with independent operators).
• Multithreading scope: global (may be called from any thread).
• Automatically parallelized on internal data level.

Parameters

Example (HDevelop)

create_class_gmm (NumDim, NumClasses, [1,5], 'full', 'none', 0, 42,\
                  GMMHandle)
* Add the training data
read_samples_class_gmm (GMMHandle, 'samples.gsf')
* Train the GMM
train_class_gmm (GMMHandle, 100, 1e-4, 'training', 1e-4, Centers, Iter)
* Write the Gaussian Mixture Model to file
write_class_gmm (GMMHandle, 'gmmclassifier.gmm')
clear_class_gmm (GMMHandle)

Result

If the parameters are valid, the operator train_class_gmmtrain_class_gmmtrain_class_gmmTrainClassGmmTrainClassGmm returns the value 2 (H_MSG_TRUE). If necessary an exception is raised.

Possible Predecessors

add_sample_class_gmmadd_sample_class_gmmadd_sample_class_gmmAddSampleClassGmmAddSampleClassGmm, read_samples_class_gmmread_samples_class_gmmread_samples_class_gmmReadSamplesClassGmmReadSamplesClassGmm

Possible Successors

evaluate_class_gmmevaluate_class_gmmevaluate_class_gmmEvaluateClassGmmEvaluateClassGmm, classify_class_gmmclassify_class_gmmclassify_class_gmmClassifyClassGmmClassifyClassGmm, write_class_gmmwrite_class_gmmwrite_class_gmmWriteClassGmmWriteClassGmm, create_class_lut_gmmcreate_class_lut_gmmcreate_class_lut_gmmCreateClassLutGmmCreateClassLutGmm

Alternatives

read_class_gmmread_class_gmmread_class_gmmReadClassGmmReadClassGmm

See also

create_class_gmmcreate_class_gmmcreate_class_gmmCreateClassGmmCreateClassGmm

References

Christopher M. Bishop: “Neural Networks for Pattern Recognition”; Oxford University Press, Oxford; 1995.
Mario A.T. Figueiredo: “Unsupervised Learning of Finite Mixture Models”; IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol. 24, No. 3; March 2002.

Module

Foundation