1 // g2o - General Graph Optimization 2 // Copyright (C) 2011 R. Kuemmerle, G. Grisetti, W. Burgard 3 // All rights reserved. 4 // 5 // Redistribution and use in source and binary forms, with or without 6 // modification, are permitted provided that the following conditions are 7 // met: 8 // 9 // * Redistributions of source code must retain the above copyright notice, 10 // this list of conditions and the following disclaimer. 11 // * Redistributions in binary form must reproduce the above copyright 12 // notice, this list of conditions and the following disclaimer in the 13 // documentation and/or other materials provided with the distribution. 14 // 15 // THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS 16 // IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED 17 // TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A 18 // PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT 19 // HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, 20 // SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED 21 // TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR 22 // PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF 23 // LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING 24 // NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS 25 // SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 26 27 #include "optimization_algorithm_levenberg.h" 28 29 #include <iostream> 30 31 #include "g2o/stuff/timeutil.h" 32 33 #include "sparse_optimizer.h" 34 #include "solver.h" 35 #include "batch_stats.h" 36 using namespace std; 37 38 namespace g2o { 39 OptimizationAlgorithmLevenberg(std::unique_ptr<Solver> solver)40 OptimizationAlgorithmLevenberg::OptimizationAlgorithmLevenberg(std::unique_ptr<Solver> solver) 41 : OptimizationAlgorithmWithHessian(*solver.get()), 42 _currentLambda(cst(-1.)), 43 _tau(cst(1e-5)), 44 _goodStepLowerScale(cst(1. / 3.)), 45 _goodStepUpperScale(cst(2. / 3.)), 46 _ni(cst(2.)), 47 _levenbergIterations(0), 48 m_solver{std::move(solver)} { 49 _userLambdaInit = _properties.makeProperty<Property<number_t> >("initialLambda", 0.); 50 _maxTrialsAfterFailure = _properties.makeProperty<Property<int> >("maxTrialsAfterFailure", 10); 51 } 52 ~OptimizationAlgorithmLevenberg()53 OptimizationAlgorithmLevenberg::~OptimizationAlgorithmLevenberg() 54 { 55 } 56 solve(int iteration,bool online)57 OptimizationAlgorithm::SolverResult OptimizationAlgorithmLevenberg::solve(int iteration, bool online) 58 { 59 assert(_optimizer && "_optimizer not set"); 60 assert(_solver.optimizer() == _optimizer && "underlying linear solver operates on different graph"); 61 62 if (iteration == 0 && !online) { // built up the CCS structure, here due to easy time measure 63 bool ok = _solver.buildStructure(); 64 if (! ok) { 65 cerr << __PRETTY_FUNCTION__ << ": Failure while building CCS structure" << endl; 66 return OptimizationAlgorithm::Fail; 67 } 68 } 69 70 number_t t=get_monotonic_time(); 71 _optimizer->computeActiveErrors(); 72 G2OBatchStatistics* globalStats = G2OBatchStatistics::globalStats(); 73 if (globalStats) { 74 globalStats->timeResiduals = get_monotonic_time()-t; 75 t=get_monotonic_time(); 76 } 77 78 number_t currentChi = _optimizer->activeRobustChi2(); 79 80 _solver.buildSystem(); 81 if (globalStats) { 82 globalStats->timeQuadraticForm = get_monotonic_time()-t; 83 } 84 85 // core part of the Levenbarg algorithm 86 if (iteration == 0) { 87 _currentLambda = computeLambdaInit(); 88 _ni = 2; 89 } 90 91 number_t rho=0; 92 int& qmax = _levenbergIterations; 93 qmax = 0; 94 do { 95 _optimizer->push(); 96 if (globalStats) { 97 globalStats->levenbergIterations++; 98 t=get_monotonic_time(); 99 } 100 // update the diagonal of the system matrix 101 _solver.setLambda(_currentLambda, true); 102 bool ok2 = _solver.solve(); 103 if (globalStats) { 104 globalStats->timeLinearSolution+=get_monotonic_time()-t; 105 t=get_monotonic_time(); 106 } 107 _optimizer->update(_solver.x()); 108 if (globalStats) { 109 globalStats->timeUpdate = get_monotonic_time()-t; 110 } 111 112 // restore the diagonal 113 _solver.restoreDiagonal(); 114 115 _optimizer->computeActiveErrors(); 116 number_t tempChi = _optimizer->activeRobustChi2(); 117 118 if (! ok2) 119 tempChi=std::numeric_limits<number_t>::max(); 120 121 rho = (currentChi-tempChi); 122 number_t scale = computeScale(); 123 scale += cst(1e-3); // make sure it's non-zero :) 124 rho /= scale; 125 126 if (rho>0 && g2o_isfinite(tempChi)){ // last step was good 127 number_t alpha = 1.-pow((2*rho-1),3); 128 // crop lambda between minimum and maximum factors 129 alpha = (std::min)(alpha, _goodStepUpperScale); 130 number_t scaleFactor = (std::max)(_goodStepLowerScale, alpha); 131 _currentLambda *= scaleFactor; 132 _ni = 2; 133 currentChi=tempChi; 134 _optimizer->discardTop(); 135 } else { 136 _currentLambda*=_ni; 137 _ni*=2; 138 _optimizer->pop(); // restore the last state before trying to optimize 139 if (!g2o_isfinite(_currentLambda)) 140 break; 141 } 142 qmax++; 143 } while (rho<0 && qmax < _maxTrialsAfterFailure->value() && ! _optimizer->terminate()); 144 145 if (qmax == _maxTrialsAfterFailure->value() || rho==0 || !g2o_isfinite(_currentLambda)) 146 return Terminate; 147 return OK; 148 } 149 computeLambdaInit() const150 number_t OptimizationAlgorithmLevenberg::computeLambdaInit() const 151 { 152 if (_userLambdaInit->value() > 0) 153 return _userLambdaInit->value(); 154 number_t maxDiagonal=0; 155 for (size_t k = 0; k < _optimizer->indexMapping().size(); k++) { 156 OptimizableGraph::Vertex* v = _optimizer->indexMapping()[k]; 157 assert(v); 158 int dim = v->dimension(); 159 for (int j = 0; j < dim; ++j){ 160 maxDiagonal = std::max(fabs(v->hessian(j,j)),maxDiagonal); 161 } 162 } 163 return _tau*maxDiagonal; 164 } 165 computeScale() const166 number_t OptimizationAlgorithmLevenberg::computeScale() const 167 { 168 number_t scale = 0; 169 for (size_t j=0; j < _solver.vectorSize(); j++){ 170 scale += _solver.x()[j] * (_currentLambda * _solver.x()[j] + _solver.b()[j]); 171 } 172 return scale; 173 } 174 setMaxTrialsAfterFailure(int max_trials)175 void OptimizationAlgorithmLevenberg::setMaxTrialsAfterFailure(int max_trials) 176 { 177 _maxTrialsAfterFailure->setValue(max_trials); 178 } 179 setUserLambdaInit(number_t lambda)180 void OptimizationAlgorithmLevenberg::setUserLambdaInit(number_t lambda) 181 { 182 _userLambdaInit->setValue(lambda); 183 } 184 printVerbose(std::ostream & os) const185 void OptimizationAlgorithmLevenberg::printVerbose(std::ostream& os) const 186 { 187 os 188 << "\t schur= " << _solver.schur() 189 << "\t lambda= " << FIXED(_currentLambda) 190 << "\t levenbergIter= " << _levenbergIterations; 191 } 192 193 } // end namespace 194