001/**
002 * Copyright (c) 2011, The University of Southampton and the individual contributors.
003 * All rights reserved.
004 *
005 * Redistribution and use in source and binary forms, with or without modification,
006 * are permitted provided that the following conditions are met:
007 *
008 *   *  Redistributions of source code must retain the above copyright notice,
009 *      this list of conditions and the following disclaimer.
010 *
011 *   *  Redistributions in binary form must reproduce the above copyright notice,
012 *      this list of conditions and the following disclaimer in the documentation
013 *      and/or other materials provided with the distribution.
014 *
015 *   *  Neither the name of the University of Southampton nor the names of its
016 *      contributors may be used to endorse or promote products derived from this
017 *      software without specific prior written permission.
018 *
019 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
020 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
021 * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
022 * DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR
023 * ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
024 * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
025 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON
026 * ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
027 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
028 * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
029 */
030package org.openimaj.workinprogress.optimisation;
031
032import java.util.Random;
033
034import org.openimaj.data.DataSource;
035import org.openimaj.data.DoubleArrayBackedDataSource;
036import org.openimaj.workinprogress.optimisation.params.Parameters;
037import org.openimaj.workinprogress.optimisation.params.VectorParameters;
038
039import scala.actors.threadpool.Arrays;
040
041public class SGD<MODEL, DATATYPE, PTYPE extends Parameters<PTYPE>> {
042        public int maxEpochs = 100;
043        public int batchSize = 1;
044        public LearningRate<PTYPE> learningRate;
045        public MODEL model;
046        public DifferentiableObjectiveFunction<MODEL, DATATYPE, PTYPE> fcn;
047
048        public void train(DataSource<DATATYPE> data) {
049                final DATATYPE[] batch = data.createTemporaryArray(batchSize);
050
051                for (int e = 0; e < maxEpochs; e++) {
052                        for (int i = 0; i < data.size(); i += batchSize) {
053                                final int currentBatchStop = Math.min(data.size(), i + batchSize);
054                                final int currentBatchSize = currentBatchStop - i;
055                                data.getData(i, currentBatchStop, batch);
056
057                                final PTYPE grads = fcn.derivative(model, batch[0]);
058                                for (int j = 1; j < currentBatchSize; j++) {
059                                        grads.addInplace(fcn.derivative(model, batch[j]));
060                                }
061                                grads.multiplyInplace(learningRate.getRate(e, i, grads));
062                                fcn.updateModel(model, grads);
063                        }
064                }
065        }
066
067        public double value(MODEL model, DATATYPE data) {
068                return 0;
069        }
070
071        public static void main(String[] args) {
072                final double[][] data = new double[1000][2];
073                final Random rng = new Random();
074                for (int i = 0; i < data.length; i++) {
075                        final double x = rng.nextDouble();
076                        data[i][0] = x;
077                        data[i][1] = 0.3 * x + 20 + (rng.nextGaussian() * 0.01);
078                }
079                final DoubleArrayBackedDataSource ds = new DoubleArrayBackedDataSource(data);
080
081                final double[] model = { 0, 0 };
082
083                final DifferentiableObjectiveFunction<double[], double[], VectorParameters> fcn = new DifferentiableObjectiveFunction<double[], double[], VectorParameters>()
084                {
085                        @Override
086                        public double value(double[] model, double[] data) {
087                                final double diff = data[1] - (model[0] * data[0] + model[1]);
088                                return diff * diff;
089                        }
090
091                        @Override
092                        public VectorParameters derivative(double[] model, double[] data) {
093                                final double[] der = {
094                                                2 * data[0] * (-data[1] + model[0] * data[0] + model[1]),
095                                                2 * (-data[1] + model[0] * data[0] + model[1])
096                                };
097
098                                return new VectorParameters(der);
099                        }
100
101                        @Override
102                        public void updateModel(double[] model, VectorParameters weights) {
103                                model[0] -= weights.vector[0];
104                                model[1] -= weights.vector[1];
105                        }
106                };
107
108                final SGD<double[], double[], VectorParameters> sgd = new SGD<double[], double[], VectorParameters>();
109                sgd.model = model;
110                sgd.fcn = fcn;
111                sgd.learningRate = new StaticLearningRate<VectorParameters>(0.01);
112                sgd.batchSize = 1;
113                sgd.maxEpochs = 10;
114
115                sgd.train(ds);
116
117                System.out.println(Arrays.toString(model));
118        }
119}