org.openimaj.ml.kernel

Class HomogeneousKernelMap

java.lang.Object

org.openimaj.ml.kernel.HomogeneousKernelMap

@References(references={@Reference(type=Article,author={"Vedaldi, A.","Zisserman, A."},title="Efficient Additive Kernels via Explicit Feature Maps",year="2012",journal="Pattern Analysis and Machine Intelligence, IEEE Transactions on",pages={"480","492"},number="3",volume="34",customData={"keywords","approximation theory;computer vision;data handling;feature extraction;learning (artificial intelligence);spectral analysis;support vector machines;Nystrom approximation;additive homogeneous kernels;approximate finite-dimensional feature maps;approximation error;computer vision;data dependency;explicit feature maps;exponential decay;large scale nonlinear support vector machines;linear SVM;spectral analysis;Additives;Approximation methods;Histograms;Kernel;Measurement;Support vector machines;Training;Kernel methods;feature map;large scale learning;object detection.;object recognition","doi","10.1109/TPAMI.2011.153","ISSN","0162-8828"}),@Reference(type=Inproceedings,author={"A. Vedaldi","A. Zisserman"},title="Efficient Additive Kernels via Explicit Feature Maps",year="2010",booktitle="Proceedings of the IEEE Conf. on Computer Vision and Pattern Recognition (CVPR)")})
public class HomogeneousKernelMap
extends Object

Implementation of the Homogeneous Kernel Map. The Homogeneous Kernel Map transforms data into a compact linear representation such that applying a linear SVM can approximate to a high degree of accuracy the application of a non-linear SVM to the original data. Additive kernels including Chi2, intersection, and Jensen-Shannon are supported.

This implementation is based directly on the VLFeat implementation written by Andrea Verdaldi, although it has been refactored to better fit with Java conventions.

Author:

Jonathon Hare (jsh2@ecs.soton.ac.uk), Based on code originally written by Andrea Verdaldi

See Also:

"http://www.vlfeat.org/api/homkermap.html", "http://www.robots.ox.ac.uk/~vgg/software/homkermap/"

Nested Class Summary

Nested Classes

Modifier and Type	Class and Description
static class	HomogeneousKernelMap.ExtractorWrapper<T> Helper implementation of a FeatureExtractor that wraps another FeatureExtractor and then applies the HomogeneousKernelMap to the output before returning the vector.
static class	HomogeneousKernelMap.KernelType Types of supported kernel for the HomogeneousKernelMap
static class	HomogeneousKernelMap.WindowType Types of window supported by the HomogeneousKernelMap.

Constructor Summary

Constructors

Constructor and Description
HomogeneousKernelMap(HomogeneousKernelMap.KernelType kernelType, double gamma, HomogeneousKernelMap.WindowType windowType) Construct with the given kernel, gamma and window.
HomogeneousKernelMap(HomogeneousKernelMap.KernelType kernelType, double gamma, int order, double period, HomogeneousKernelMap.WindowType windowType) Construct with the given kernel, gamma, order, period and window.
HomogeneousKernelMap(HomogeneousKernelMap.KernelType kernelType, double gamma, int order, HomogeneousKernelMap.WindowType windowType) Construct with the given kernel, gamma, order and window.
HomogeneousKernelMap(HomogeneousKernelMap.KernelType kernelType, HomogeneousKernelMap.WindowType windowType) Construct with the given kernel and window.

Method Summary

Methods

Modifier and Type	Method and Description
<T> FeatureExtractor<DoubleFV,T>	createWrappedExtractor(FeatureExtractor<? extends FeatureVector,T> inner) Construct a new HomogeneousKernelMap.ExtractorWrapper that applies the map to features extracted by an internal extractor.
void	evaluate(double[] destination, int stride, int offset, double x) Evaluate the kernel for the given x value.
DoubleFV	evaluate(DoubleFV in) Compute the Homogeneous Kernel Map approximation of the given feature vector

Methods inherited from class java.lang.Object

clone, equals, finalize, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait

Constructor Detail

HomogeneousKernelMap

public HomogeneousKernelMap(HomogeneousKernelMap.KernelType kernelType,
                    HomogeneousKernelMap.WindowType windowType)

Construct with the given kernel and window. The Gamma and order values are set at their defaults of 1. The period is computed automatically.

Parameters:

kernelType - the type of kernel

windowType - the type of window (use HomogeneousKernelMap.WindowType.Rectangular if unsure)

HomogeneousKernelMap

public HomogeneousKernelMap(HomogeneousKernelMap.KernelType kernelType,
                    double gamma,
                    HomogeneousKernelMap.WindowType windowType)

Construct with the given kernel, gamma and window. The period is computed automatically and the approximation order is set to 1.

Parameters:

kernelType - the type of kernel

gamma - the gamma value. the standard kernels are 1-homogeneous, but smaller values can work better in practice.

windowType - the type of window (use HomogeneousKernelMap.WindowType.Rectangular if unsure)

HomogeneousKernelMap

public HomogeneousKernelMap(HomogeneousKernelMap.KernelType kernelType,
                    double gamma,
                    int order,
                    HomogeneousKernelMap.WindowType windowType)

Construct with the given kernel, gamma, order and window. The period is computed automatically.

Parameters:

kernelType - the type of kernel

gamma - the gamma value. the standard kernels are 1-homogeneous, but smaller values can work better in practice.

order - the approximation order (usually 1 is enough)

windowType - the type of window (use HomogeneousKernelMap.WindowType.Rectangular if unsure)

HomogeneousKernelMap

public HomogeneousKernelMap(HomogeneousKernelMap.KernelType kernelType,
                    double gamma,
                    int order,
                    double period,
                    HomogeneousKernelMap.WindowType windowType)

Construct with the given kernel, gamma, order, period and window. If the period is negative, it will be replaced by the default.

Parameters:

kernelType - the type of kernel

gamma - the gamma value. the standard kernels are 1-homogeneous, but smaller values can work better in practice.

order - the approximation order (usually 1 is enough)

period - the periodicity of the kernel spectrum

windowType - the type of window (use HomogeneousKernelMap.WindowType.Rectangular if unsure)

Method Detail

evaluate

public void evaluate(double[] destination,
            int stride,
            int offset,
            double x)

Evaluate the kernel for the given x value. The output values will be written into the destination array at offset + j*stride intervals where j is between 0 and 2 * order + 1.

Parameters:

destination - the destination array

stride - the stride

offset - the offset

x - the value to compute the kernel approximation for

evaluate

public DoubleFV evaluate(DoubleFV in)

Compute the Homogeneous Kernel Map approximation of the given feature vector

Parameters:

in - the feature vector

Returns:

the expanded feature vector

createWrappedExtractor

public <T> FeatureExtractor<DoubleFV,T> createWrappedExtractor(FeatureExtractor<? extends FeatureVector,T> inner)

Construct a new HomogeneousKernelMap.ExtractorWrapper that applies the map to features extracted by an internal extractor.

Type Parameters:

T - Type of object that features can be extracted from

Parameters:

inner - the internal extractor

Returns:

the wrapped FeatureExtractor