001/*
002        AUTOMATICALLY GENERATED BY jTemp FROM
003        /Users/jsh2/Work/openimaj/target/checkout/machine-learning/clustering/src/main/jtemp/org/openimaj/knn/pq/#T#ProductQuantiserUtilities.jtemp
004*/
005/**
006 * Copyright (c) 2011, The University of Southampton and the individual contributors.
007 * All rights reserved.
008 *
009 * Redistribution and use in source and binary forms, with or without modification,
010 * are permitted provided that the following conditions are met:
011 *
012 *   *  Redistributions of source code must retain the above copyright notice,
013 *      this list of conditions and the following disclaimer.
014 *
015 *   *  Redistributions in binary form must reproduce the above copyright notice,
016 *      this list of conditions and the following disclaimer in the documentation
017 *      and/or other materials provided with the distribution.
018 *
019 *   *  Neither the name of the University of Southampton nor the names of its
020 *      contributors may be used to endorse or promote products derived from this
021 *      software without specific prior written permission.
022 *
023 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
024 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
025 * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
026 * DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR
027 * ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
028 * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
029 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON
030 * ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
031 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
032 * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
033 */
034package org.openimaj.knn.pq;
035
036import org.openimaj.knn.DoubleNearestNeighboursExact;
037import org.openimaj.knn.DoubleNearestNeighboursProvider;
038import org.openimaj.ml.clustering.kmeans.DoubleKMeans;
039
040/**
041 * Utility methods for easily creating a {@link ByteProductQuantiser} using
042 * (Exact) K-Means.
043 * 
044 * @author Jonathon Hare (jsh2@ecs.soton.ac.uk)
045 * 
046 */
047public final class DoubleProductQuantiserUtilities {
048    private DoubleProductQuantiserUtilities() {
049    }
050
051        /**
052         * Learn a {@link DoubleProductQuantiser} by applying exact K-Means to
053         * sub-vectors extracted from the given data. The length of the subvectors
054         * is determined by dividing the vector length by the requested number of
055         * assigners. The number of clusters per vector subset is constant, and must
056         * be less than 256.
057         * 
058         * @param data
059         *            the data to train the {@link DoubleProductQuantiser} on.
060         * @param numAssigners
061         *            the number of sub-quantisers to learn
062         * @param K
063         *            the number of centroids per sub-quantiser
064         * @param nIter
065         *            the maximum number of iterations for each k-means clustering       
066         * 
067         * @return a trained {@link DoubleProductQuantiser}.
068         */
069        public static DoubleProductQuantiser train(double[][] data, int numAssigners, int K, int nIter) {
070                if (K > 256 || K <= 0)
071                        throw new IllegalArgumentException("0 <= K < 256");
072
073                final int subDim = data[0].length / numAssigners;
074                final double[][] tmp = new double[data.length][subDim];
075
076                final DoubleNearestNeighboursExact[] assigners = new DoubleNearestNeighboursExact[numAssigners];
077                final DoubleKMeans kmeans = DoubleKMeans.createExact(K, 100);
078
079                for (int i = 0; i < numAssigners; i++) {
080                        // copy data
081                        for (int j = 0; j < data.length; j++) {
082                                System.arraycopy(data[j], i * subDim, tmp[j], 0, subDim);
083                        }
084
085                        // kmeans
086                        final DoubleNearestNeighboursProvider centroids = (DoubleNearestNeighboursProvider) kmeans.cluster(tmp);
087
088                        assigners[i] = (DoubleNearestNeighboursExact)centroids.getNearestNeighbours();
089                }
090
091                return new DoubleProductQuantiser(assigners);
092        }
093
094        /**
095         * Learn a {@link DoubleProductQuantiser} by applying exact K-Means to
096         * sub-vectors extracted from the given data. The length of the subvectors
097         * is determined by dividing the vector length by the requested number of
098         * assigners. The number of clusters per vector subset is constant, and set
099         * at 256.
100         * 
101         * @param data
102         *            the data to train the {@link DoubleProductQuantiser} on.
103         * @param numAssigners
104         *            the number of sub-quantisers to learn
105         * @param nIter
106         *            the maximum number of iterations for each k-means clustering
107         *
108         * @return a trained {@link DoubleProductQuantiser}.
109         */
110        public static DoubleProductQuantiser train(double[][] data, int numAssigners, int nIter) {
111                return train(data, numAssigners, 256, nIter);
112        }
113}