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.ml.linear.evaluation;
031
032import gov.sandia.cognition.math.matrix.Matrix;
033import gov.sandia.cognition.math.matrix.mtj.SparseMatrix;
034
035import java.util.List;
036
037import org.apache.log4j.Logger;
038import org.openimaj.math.matrix.CFMatrixUtils;
039import org.openimaj.ml.linear.learner.BilinearLearnerParameters;
040import org.openimaj.ml.linear.learner.BilinearSparseOnlineLearner;
041import org.openimaj.ml.linear.learner.loss.LossFunction;
042import org.openimaj.ml.linear.learner.loss.MatLossFunction;
043import org.openimaj.util.pair.Pair;
044
045public class RootMeanSumLossEvaluator extends BilinearEvaluator {
046        Logger logger = Logger.getLogger(RootMeanSumLossEvaluator.class);
047
048        @Override
049        public double evaluate(List<Pair<Matrix>> data) {
050                final Matrix u = learner.getU();
051                final Matrix w = learner.getW();
052                final Matrix bias = learner.getBias();
053                final double sumloss = sumLoss(data, u, w, bias, learner.getParams());
054                return sumloss;
055        }
056
057        public double sumLoss(List<Pair<Matrix>> pairs, Matrix u, Matrix w, Matrix bias, BilinearLearnerParameters params) {
058                LossFunction loss = params.getTyped(BilinearLearnerParameters.LOSS);
059                if(!loss.isMatrixLoss()) loss = new MatLossFunction(loss);
060                double total = 0;
061                int i = 0;
062                int ntasks = 0;
063                boolean forceSparcity = learner.getParams().getTyped(BilinearLearnerParameters.FORCE_SPARCITY);
064                if(forceSparcity){
065                        u = CFMatrixUtils.asSparseColumn(u);
066                        w = CFMatrixUtils.asSparseColumn(w);
067                        
068                }
069                for (final Pair<Matrix> pair : pairs) {
070                        final Matrix X = pair.firstObject();
071                        final Matrix Y = pair.secondObject();
072                        final SparseMatrix Yexp = BilinearSparseOnlineLearner.expandY(Y);
073                        Matrix xt = X.transpose();
074                        Matrix ut = u.transpose();
075                        final Matrix expectedAll = CFMatrixUtils.fastdot(CFMatrixUtils.fastdot(ut,xt),w);
076                        loss.setY(Yexp);
077                        loss.setX(expectedAll);
078                        if (bias != null)
079                                loss.setBias(bias);
080                        
081                        logger.debug("Testing pair: " + i);
082                        total += loss.eval(null); // Assums an identity w.
083                        i++;
084                        ntasks += Y.getNumColumns();
085                }
086                total /= ntasks;
087
088                return Math.sqrt(total);
089        }
090}