1 // Ceres Solver - A fast non-linear least squares minimizer
2 // Copyright 2012 Google Inc. All rights reserved.
3 // http://code.google.com/p/ceres-solver/
4 //
5 // Redistribution and use in source and binary forms, with or without
6 // modification, are permitted provided that the following conditions are met:
7 //
8 // * Redistributions of source code must retain the above copyright notice,
9 // this list of conditions and the following disclaimer.
10 // * Redistributions in binary form must reproduce the above copyright notice,
11 // this list of conditions and the following disclaimer in the documentation
12 // and/or other materials provided with the distribution.
13 // * Neither the name of Google Inc. nor the names of its contributors may be
14 // used to endorse or promote products derived from this software without
15 // specific prior written permission.
16 //
17 // THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
18 // AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19 // IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20 // ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
21 // LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
22 // CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
23 // SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
24 // INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
25 // CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
26 // ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
27 // POSSIBILITY OF SUCH DAMAGE.
28 //
29 // Author: sameeragarwal@google.com (Sameer Agarwal)
30
31 #include "ceres/internal/eigen.h"
32 #include "ceres/internal/scoped_ptr.h"
33 #include "ceres/levenberg_marquardt_strategy.h"
34 #include "ceres/linear_solver.h"
35 #include "ceres/trust_region_strategy.h"
36 #include "glog/logging.h"
37 #include "gmock/gmock.h"
38 #include "gmock/mock-log.h"
39 #include "gtest/gtest.h"
40
41 using testing::AllOf;
42 using testing::AnyNumber;
43 using testing::HasSubstr;
44 using testing::ScopedMockLog;
45 using testing::_;
46
47 namespace ceres {
48 namespace internal {
49
50 const double kTolerance = 1e-16;
51
52 // Linear solver that takes as input a vector and checks that the
53 // caller passes the same vector as LinearSolver::PerSolveOptions.D.
54 class RegularizationCheckingLinearSolver : public DenseSparseMatrixSolver {
55 public:
RegularizationCheckingLinearSolver(const int num_cols,const double * diagonal)56 RegularizationCheckingLinearSolver(const int num_cols, const double* diagonal)
57 : num_cols_(num_cols),
58 diagonal_(diagonal) {
59 }
60
~RegularizationCheckingLinearSolver()61 virtual ~RegularizationCheckingLinearSolver() {}
62
63 private:
SolveImpl(DenseSparseMatrix * A,const double * b,const LinearSolver::PerSolveOptions & per_solve_options,double * x)64 virtual LinearSolver::Summary SolveImpl(
65 DenseSparseMatrix* A,
66 const double* b,
67 const LinearSolver::PerSolveOptions& per_solve_options,
68 double* x) {
69 CHECK_NOTNULL(per_solve_options.D);
70 for (int i = 0; i < num_cols_; ++i) {
71 EXPECT_NEAR(per_solve_options.D[i], diagonal_[i], kTolerance)
72 << i << " " << per_solve_options.D[i] << " " << diagonal_[i];
73 }
74 return LinearSolver::Summary();
75 }
76
77 const int num_cols_;
78 const double* diagonal_;
79 };
80
TEST(LevenbergMarquardtStrategy,AcceptRejectStepRadiusScaling)81 TEST(LevenbergMarquardtStrategy, AcceptRejectStepRadiusScaling) {
82 TrustRegionStrategy::Options options;
83 options.initial_radius = 2.0;
84 options.max_radius = 20.0;
85 options.min_lm_diagonal = 1e-8;
86 options.max_lm_diagonal = 1e8;
87
88 // We need a non-null pointer here, so anything should do.
89 scoped_ptr<LinearSolver> linear_solver(
90 new RegularizationCheckingLinearSolver(0, NULL));
91 options.linear_solver = linear_solver.get();
92
93 LevenbergMarquardtStrategy lms(options);
94 EXPECT_EQ(lms.Radius(), options.initial_radius);
95 lms.StepRejected(0.0);
96 EXPECT_EQ(lms.Radius(), 1.0);
97 lms.StepRejected(-1.0);
98 EXPECT_EQ(lms.Radius(), 0.25);
99 lms.StepAccepted(1.0);
100 EXPECT_EQ(lms.Radius(), 0.25 * 3.0);
101 lms.StepAccepted(1.0);
102 EXPECT_EQ(lms.Radius(), 0.25 * 3.0 * 3.0);
103 lms.StepAccepted(0.25);
104 EXPECT_EQ(lms.Radius(), 0.25 * 3.0 * 3.0 / 1.125);
105 lms.StepAccepted(1.0);
106 EXPECT_EQ(lms.Radius(), 0.25 * 3.0 * 3.0 / 1.125 * 3.0);
107 lms.StepAccepted(1.0);
108 EXPECT_EQ(lms.Radius(), 0.25 * 3.0 * 3.0 / 1.125 * 3.0 * 3.0);
109 lms.StepAccepted(1.0);
110 EXPECT_EQ(lms.Radius(), options.max_radius);
111 }
112
TEST(LevenbergMarquardtStrategy,CorrectDiagonalToLinearSolver)113 TEST(LevenbergMarquardtStrategy, CorrectDiagonalToLinearSolver) {
114 Matrix jacobian(2, 3);
115 jacobian.setZero();
116 jacobian(0, 0) = 0.0;
117 jacobian(0, 1) = 1.0;
118 jacobian(1, 1) = 1.0;
119 jacobian(0, 2) = 100.0;
120
121 double residual = 1.0;
122 double x[3];
123 DenseSparseMatrix dsm(jacobian);
124
125 TrustRegionStrategy::Options options;
126 options.initial_radius = 2.0;
127 options.max_radius = 20.0;
128 options.min_lm_diagonal = 1e-2;
129 options.max_lm_diagonal = 1e2;
130
131 double diagonal[3];
132 diagonal[0] = options.min_lm_diagonal;
133 diagonal[1] = 2.0;
134 diagonal[2] = options.max_lm_diagonal;
135 for (int i = 0; i < 3; ++i) {
136 diagonal[i] = sqrt(diagonal[i] / options.initial_radius);
137 }
138
139 RegularizationCheckingLinearSolver linear_solver(3, diagonal);
140 options.linear_solver = &linear_solver;
141
142 LevenbergMarquardtStrategy lms(options);
143 TrustRegionStrategy::PerSolveOptions pso;
144
145 {
146 ScopedMockLog log;
147 EXPECT_CALL(log, Log(_, _, _)).Times(AnyNumber());
148 EXPECT_CALL(log, Log(WARNING, _,
149 HasSubstr("Failed to compute a finite step.")));
150
151 TrustRegionStrategy::Summary summary =
152 lms.ComputeStep(pso, &dsm, &residual, x);
153 EXPECT_EQ(summary.termination_type, LINEAR_SOLVER_FAILURE);
154 }
155 }
156
157 } // namespace internal
158 } // namespace ceres
159