File size: 6,823 Bytes
158b61b |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 |
/*
Moses - statistical machine translation system
Copyright (C) 2005-2015 University of Edinburgh
This library is free software; you can redistribute it and/or
modify it under the terms of the GNU Lesser General Public
License as published by the Free Software Foundation; either
version 2.1 of the License, or (at your option) any later version.
This library is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
Lesser General Public License for more details.
You should have received a copy of the GNU Lesser General Public
License along with this library; if not, write to the Free Software
Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
*/
#include "ExpectedBleuOptimizer.h"
namespace ExpectedBleuTraining
{
void ExpectedBleuOptimizer::AddTrainingInstance(const size_t nBestSizeCount,
const std::vector<float>& sBleu,
const std::vector<double>& overallScoreUntransformed,
const std::vector< boost::unordered_map<size_t, float> > &sparseScore,
bool maintainUpdateSet)
{
// compute xBLEU
double sumUntransformedScores = 0.0;
for (std::vector<double>::const_iterator overallScoreUntransformedIt=overallScoreUntransformed.begin();
overallScoreUntransformedIt!=overallScoreUntransformed.end(); ++overallScoreUntransformedIt)
{
sumUntransformedScores += *overallScoreUntransformedIt;
}
double xBleu = 0.0;
assert(nBestSizeCount == overallScoreUntransformed.size());
std::vector<double> p;
for (size_t i=0; i<nBestSizeCount; ++i)
{
if (sumUntransformedScores != 0) {
p.push_back( overallScoreUntransformed[i] / sumUntransformedScores );
} else {
p.push_back( 0 );
}
xBleu += p.back() * sBleu[ i ];
}
for (size_t i=0; i<nBestSizeCount; ++i)
{
double D = sBleu[ i ] - xBleu;
for (boost::unordered_map<size_t, float>::const_iterator sparseScoreIt=sparseScore[i].begin();
sparseScoreIt!=sparseScore[i].end(); ++sparseScoreIt)
{
const size_t name = sparseScoreIt->first;
float N = sparseScoreIt->second;
if ( std::fpclassify( p[i] * N * D ) == FP_SUBNORMAL )
{
m_err << "Error: encountered subnormal value: p[i] * N * D= " << p[i] * N * D
<< " with p[i]= " << p[i] << " N= " << N << " D= " << D << '\n';
m_err.flush();
exit(1);
} else {
m_gradient[name] += p[i] * N * D;
if ( maintainUpdateSet )
{
m_updateSet.insert(name);
}
}
}
}
m_xBleu += xBleu;
}
void ExpectedBleuOptimizer::InitSGD(const std::vector<float>& sparseScalingFactor)
{
const size_t nFeatures = sparseScalingFactor.size();
memcpy(&m_previousSparseScalingFactor.at(0), &sparseScalingFactor.at(0), nFeatures);
m_gradient.resize(nFeatures);
}
float ExpectedBleuOptimizer::UpdateSGD(std::vector<float>& sparseScalingFactor,
size_t batchSize,
bool useUpdateSet)
{
float xBleu = m_xBleu / batchSize;
// update sparse scaling factors
if (useUpdateSet) {
for (std::set<size_t>::const_iterator it = m_updateSet.begin(); it != m_updateSet.end(); ++it)
{
size_t name = *it;
UpdateSingleScalingFactorSGD(name, sparseScalingFactor, batchSize);
}
m_updateSet.clear();
} else {
for (size_t name=0; name<sparseScalingFactor.size(); ++name)
{
UpdateSingleScalingFactorSGD(name, sparseScalingFactor, batchSize);
}
}
m_xBleu = 0;
m_gradient.clear();
return xBleu;
}
void ExpectedBleuOptimizer::UpdateSingleScalingFactorSGD(size_t name,
std::vector<float>& sparseScalingFactor,
size_t batchSize)
{
// regularization
if ( m_regularizationParameter != 0 )
{
m_gradient[name] = m_gradient[name] / m_xBleu - m_regularizationParameter * 2 * sparseScalingFactor[name];
} else {
// need to normalize by dividing by batchSize
m_gradient[name] /= batchSize;
}
// the actual update
sparseScalingFactor[name] += m_learningRate * m_gradient[name];
// discard scaling factors below a threshold
if ( fabs(sparseScalingFactor[name]) < m_floorAbsScalingFactor )
{
sparseScalingFactor[name] = 0;
}
}
void ExpectedBleuOptimizer::InitRPROP(const std::vector<float>& sparseScalingFactor)
{
const size_t nFeatures = sparseScalingFactor.size();
m_previousSparseScalingFactor.resize(nFeatures);
memcpy(&m_previousSparseScalingFactor.at(0), &sparseScalingFactor.at(0), nFeatures);
m_previousGradient.resize(nFeatures);
m_gradient.resize(nFeatures);
m_stepSize.resize(nFeatures, m_initialStepSize);
}
float ExpectedBleuOptimizer::UpdateRPROP(std::vector<float>& sparseScalingFactor,
const size_t batchSize)
{
float xBleu = m_xBleu / batchSize;
// update sparse scaling factors
for (size_t name=0; name<sparseScalingFactor.size(); ++name)
{
// Sum of gradients. All we need is the sign. Don't need to normalize by dividing by batchSize.
// regularization
if ( m_regularizationParameter != 0 )
{
m_gradient[name] = m_gradient[name] / m_xBleu - m_regularizationParameter * 2 * sparseScalingFactor[name];
}
// step size
int sign = Sign(m_gradient[name]) * Sign(m_previousGradient[name]);
if (sign > 0) {
m_stepSize[name] *= m_increaseRate;
} else if (sign < 0) {
m_stepSize[name] *= m_decreaseRate;
}
if (m_stepSize[name] < m_minStepSize) {
m_stepSize[name] = m_minStepSize;
}
if (m_stepSize[name] > m_maxStepSize) {
m_stepSize[name] = m_maxStepSize;
}
// the actual update
m_previousGradient[name] = m_gradient[name];
if (sign >= 0) {
if (m_gradient[name] > 0) {
m_previousSparseScalingFactor[name] = sparseScalingFactor[name];
sparseScalingFactor[name] += m_stepSize[name];
} else if (m_gradient[name] < 0) {
m_previousSparseScalingFactor[name] = sparseScalingFactor[name];
sparseScalingFactor[name] -= m_stepSize[name];
}
} else {
sparseScalingFactor[name] = m_previousSparseScalingFactor[name];
// m_previousGradient[name] = 0;
}
// discard scaling factors below a threshold
if ( fabs(sparseScalingFactor[name]) < m_floorAbsScalingFactor )
{
sparseScalingFactor[name] = 0;
}
}
m_xBleu = 0;
m_gradient.clear();
return xBleu;
}
}
|