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.learner.loss; 031 032import org.apache.log4j.Logger; 033import org.openimaj.math.matrix.CFMatrixUtils; 034 035import gov.sandia.cognition.math.matrix.Matrix; 036import gov.sandia.cognition.math.matrix.Vector; 037import gov.sandia.cognition.math.matrix.mtj.SparseMatrix; 038import gov.sandia.cognition.math.matrix.mtj.SparseMatrixFactoryMTJ; 039 040public class MatSquareLossFunction extends LossFunction{ 041 Logger logger = Logger.getLogger(MatSquareLossFunction.class); 042 private SparseMatrixFactoryMTJ spf; 043 public MatSquareLossFunction() { 044 spf = SparseMatrixFactoryMTJ.INSTANCE; 045 } 046 @Override 047 public Matrix gradient(Matrix W) { 048 Matrix ret = W.clone(); 049 if(CFMatrixUtils.containsInfinity(X)){ 050 throw new RuntimeException(); 051 } 052 if(CFMatrixUtils.containsInfinity(W)){ 053 throw new RuntimeException(); 054 } 055 Matrix resid = CFMatrixUtils.fastdot(X,W); 056 if(CFMatrixUtils.containsInfinity(resid)){ 057 CFMatrixUtils.fastdot(X,W); 058 throw new RuntimeException(); 059 } 060 if(this.bias!=null) 061 { 062 resid.plusEquals(this.bias); 063 } 064 CFMatrixUtils.fastminusEquals(resid, Y); 065 if(CFMatrixUtils.containsInfinity(resid)){ 066 throw new RuntimeException(); 067 } 068 for (int t = 0; t < resid.getNumColumns(); t++) { 069 Vector xcol = this.X.getRow(t).scale(resid.getElement(t, t)).clone(); 070 CFMatrixUtils.fastsetcol(ret,t, xcol); 071 } 072 return ret; 073 } 074 @Override 075 public double eval(Matrix W) { 076 Matrix resid = null; 077 if(W == null){ 078 resid = X.clone(); 079 } else { 080 resid = CFMatrixUtils.fastdot(X,W); 081 } 082 Matrix vnobias = resid.clone(); 083 if(this.bias!=null) 084 { 085 resid.plusEquals(this.bias); 086 } 087 Matrix v = resid.clone(); 088 resid.minusEquals(Y); 089 double retval = 0; 090 091 for (int t = 0; t < resid.getNumColumns(); t++) { 092 double loss = resid.getElement(t, t); 093 retval += loss * loss; 094 logger.debug( 095 String.format( 096 "yr=%d,y=%3.2f,v=%3.2f,v(no bias)=%2.5f,error=%2.5f,serror=%2.5f", 097 t, Y.getElement(t, t), v.getElement(t, t), vnobias.getElement(t,t), loss, loss*loss 098 ) 099 ); 100 } 101 return retval; 102 } 103 104 @Override 105 public boolean test_backtrack(Matrix W, Matrix grad, Matrix prox, double eta) { 106 Matrix tmp = prox.minus(W); 107 double evalW = eval(W); 108 double evalProx = eval(prox); 109 Matrix fastdotGradTmp = CFMatrixUtils.fastdot(grad.transpose(),tmp); 110 double normGradProx = CFMatrixUtils.sum(fastdotGradTmp); 111 double normTmp = 0.5*eta*tmp.normFrobenius(); 112 return (evalProx <= evalW + normGradProx + normTmp); 113 } 114 115 @Override 116 public boolean isMatrixLoss() { 117 return true; 118 } 119}