• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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: keir@google.com (Keir Mierle)
30 //         sameeragarwal@google.com (Sameer Agarwal)
31 //
32 // This tests the TrustRegionMinimizer loop using a direct Evaluator
33 // implementation, rather than having a test that goes through all the
34 // Program and Problem machinery.
35 
36 #include <cmath>
37 #include "ceres/cost_function.h"
38 #include "ceres/dense_qr_solver.h"
39 #include "ceres/dense_sparse_matrix.h"
40 #include "ceres/evaluator.h"
41 #include "ceres/internal/port.h"
42 #include "ceres/linear_solver.h"
43 #include "ceres/minimizer.h"
44 #include "ceres/problem.h"
45 #include "ceres/trust_region_minimizer.h"
46 #include "ceres/trust_region_strategy.h"
47 #include "gtest/gtest.h"
48 
49 namespace ceres {
50 namespace internal {
51 
52 // Templated Evaluator for Powell's function. The template parameters
53 // indicate which of the four variables/columns of the jacobian are
54 // active. This is equivalent to constructing a problem and using the
55 // SubsetLocalParameterization. This allows us to test the support for
56 // the Evaluator::Plus operation besides checking for the basic
57 // performance of the trust region algorithm.
58 template <bool col1, bool col2, bool col3, bool col4>
59 class PowellEvaluator2 : public Evaluator {
60  public:
PowellEvaluator2()61   PowellEvaluator2()
62       : num_active_cols_(
63           (col1 ? 1 : 0) +
64           (col2 ? 1 : 0) +
65           (col3 ? 1 : 0) +
66           (col4 ? 1 : 0)) {
67     VLOG(1) << "Columns: "
68             << col1 << " "
69             << col2 << " "
70             << col3 << " "
71             << col4;
72   }
73 
~PowellEvaluator2()74   virtual ~PowellEvaluator2() {}
75 
76   // Implementation of Evaluator interface.
CreateJacobian() const77   virtual SparseMatrix* CreateJacobian() const {
78     CHECK(col1 || col2 || col3 || col4);
79     DenseSparseMatrix* dense_jacobian =
80         new DenseSparseMatrix(NumResiduals(), NumEffectiveParameters());
81     dense_jacobian->SetZero();
82     return dense_jacobian;
83   }
84 
Evaluate(const Evaluator::EvaluateOptions & evaluate_options,const double * state,double * cost,double * residuals,double * gradient,SparseMatrix * jacobian)85   virtual bool Evaluate(const Evaluator::EvaluateOptions& evaluate_options,
86                         const double* state,
87                         double* cost,
88                         double* residuals,
89                         double* gradient,
90                         SparseMatrix* jacobian) {
91     const double x1 = state[0];
92     const double x2 = state[1];
93     const double x3 = state[2];
94     const double x4 = state[3];
95 
96     VLOG(1) << "State: "
97             << "x1=" << x1 << ", "
98             << "x2=" << x2 << ", "
99             << "x3=" << x3 << ", "
100             << "x4=" << x4 << ".";
101 
102     const double f1 = x1 + 10.0 * x2;
103     const double f2 = sqrt(5.0) * (x3 - x4);
104     const double f3 = pow(x2 - 2.0 * x3, 2.0);
105     const double f4 = sqrt(10.0) * pow(x1 - x4, 2.0);
106 
107     VLOG(1) << "Function: "
108             << "f1=" << f1 << ", "
109             << "f2=" << f2 << ", "
110             << "f3=" << f3 << ", "
111             << "f4=" << f4 << ".";
112 
113     *cost = (f1*f1 + f2*f2 + f3*f3 + f4*f4) / 2.0;
114 
115     VLOG(1) << "Cost: " << *cost;
116 
117     if (residuals != NULL) {
118       residuals[0] = f1;
119       residuals[1] = f2;
120       residuals[2] = f3;
121       residuals[3] = f4;
122     }
123 
124     if (jacobian != NULL) {
125       DenseSparseMatrix* dense_jacobian;
126       dense_jacobian = down_cast<DenseSparseMatrix*>(jacobian);
127       dense_jacobian->SetZero();
128 
129       ColMajorMatrixRef jacobian_matrix = dense_jacobian->mutable_matrix();
130       CHECK_EQ(jacobian_matrix.cols(), num_active_cols_);
131 
132       int column_index = 0;
133       if (col1) {
134         jacobian_matrix.col(column_index++) <<
135             1.0,
136             0.0,
137             0.0,
138             sqrt(10.0) * 2.0 * (x1 - x4) * (1.0 - x4);
139       }
140       if (col2) {
141         jacobian_matrix.col(column_index++) <<
142             10.0,
143             0.0,
144             2.0*(x2 - 2.0*x3)*(1.0 - 2.0*x3),
145             0.0;
146       }
147 
148       if (col3) {
149         jacobian_matrix.col(column_index++) <<
150             0.0,
151             sqrt(5.0),
152             2.0*(x2 - 2.0*x3)*(x2 - 2.0),
153             0.0;
154       }
155 
156       if (col4) {
157         jacobian_matrix.col(column_index++) <<
158             0.0,
159             -sqrt(5.0),
160             0.0,
161             sqrt(10.0) * 2.0 * (x1 - x4) * (x1 - 1.0);
162       }
163       VLOG(1) << "\n" << jacobian_matrix;
164     }
165 
166     if (gradient != NULL) {
167       int column_index = 0;
168       if (col1) {
169         gradient[column_index++] = f1  + f4 * sqrt(10.0) * 2.0 * (x1 - x4);
170       }
171 
172       if (col2) {
173         gradient[column_index++] = f1 * 10.0 + f3 * 2.0 * (x2 - 2.0 * x3);
174       }
175 
176       if (col3) {
177         gradient[column_index++] =
178             f2 * sqrt(5.0) + f3 * (2.0 * 2.0 * (2.0 * x3 - x2));
179       }
180 
181       if (col4) {
182         gradient[column_index++] =
183             -f2 * sqrt(5.0) + f4 * sqrt(10.0) * 2.0 * (x4 - x1);
184       }
185     }
186 
187     return true;
188   }
189 
Plus(const double * state,const double * delta,double * state_plus_delta) const190   virtual bool Plus(const double* state,
191                     const double* delta,
192                     double* state_plus_delta) const {
193     int delta_index = 0;
194     state_plus_delta[0] = (col1  ? state[0] + delta[delta_index++] : state[0]);
195     state_plus_delta[1] = (col2  ? state[1] + delta[delta_index++] : state[1]);
196     state_plus_delta[2] = (col3  ? state[2] + delta[delta_index++] : state[2]);
197     state_plus_delta[3] = (col4  ? state[3] + delta[delta_index++] : state[3]);
198     return true;
199   }
200 
NumEffectiveParameters() const201   virtual int NumEffectiveParameters() const { return num_active_cols_; }
NumParameters() const202   virtual int NumParameters()          const { return 4; }
NumResiduals() const203   virtual int NumResiduals()           const { return 4; }
204 
205  private:
206   const int num_active_cols_;
207 };
208 
209 // Templated function to hold a subset of the columns fixed and check
210 // if the solver converges to the optimal values or not.
211 template<bool col1, bool col2, bool col3, bool col4>
IsTrustRegionSolveSuccessful(TrustRegionStrategyType strategy_type)212 void IsTrustRegionSolveSuccessful(TrustRegionStrategyType strategy_type) {
213   Solver::Options solver_options;
214   LinearSolver::Options linear_solver_options;
215   DenseQRSolver linear_solver(linear_solver_options);
216 
217   double parameters[4] = { 3, -1, 0, 1.0 };
218 
219   // If the column is inactive, then set its value to the optimal
220   // value.
221   parameters[0] = (col1 ? parameters[0] : 0.0);
222   parameters[1] = (col2 ? parameters[1] : 0.0);
223   parameters[2] = (col3 ? parameters[2] : 0.0);
224   parameters[3] = (col4 ? parameters[3] : 0.0);
225 
226   PowellEvaluator2<col1, col2, col3, col4> powell_evaluator;
227   scoped_ptr<SparseMatrix> jacobian(powell_evaluator.CreateJacobian());
228 
229   Minimizer::Options minimizer_options(solver_options);
230   minimizer_options.gradient_tolerance = 1e-26;
231   minimizer_options.function_tolerance = 1e-26;
232   minimizer_options.parameter_tolerance = 1e-26;
233   minimizer_options.evaluator = &powell_evaluator;
234   minimizer_options.jacobian = jacobian.get();
235 
236   TrustRegionStrategy::Options trust_region_strategy_options;
237   trust_region_strategy_options.trust_region_strategy_type = strategy_type;
238   trust_region_strategy_options.linear_solver = &linear_solver;
239   trust_region_strategy_options.initial_radius = 1e4;
240   trust_region_strategy_options.max_radius = 1e20;
241   trust_region_strategy_options.min_lm_diagonal = 1e-6;
242   trust_region_strategy_options.max_lm_diagonal = 1e32;
243   scoped_ptr<TrustRegionStrategy> strategy(
244       TrustRegionStrategy::Create(trust_region_strategy_options));
245   minimizer_options.trust_region_strategy = strategy.get();
246 
247   TrustRegionMinimizer minimizer;
248   Solver::Summary summary;
249   minimizer.Minimize(minimizer_options, parameters, &summary);
250 
251   // The minimum is at x1 = x2 = x3 = x4 = 0.
252   EXPECT_NEAR(0.0, parameters[0], 0.001);
253   EXPECT_NEAR(0.0, parameters[1], 0.001);
254   EXPECT_NEAR(0.0, parameters[2], 0.001);
255   EXPECT_NEAR(0.0, parameters[3], 0.001);
256 };
257 
TEST(TrustRegionMinimizer,PowellsSingularFunctionUsingLevenbergMarquardt)258 TEST(TrustRegionMinimizer, PowellsSingularFunctionUsingLevenbergMarquardt) {
259   // This case is excluded because this has a local minimum and does
260   // not find the optimum. This should not affect the correctness of
261   // this test since we are testing all the other 14 combinations of
262   // column activations.
263   //
264   //   IsSolveSuccessful<true, true, false, true>();
265 
266   const TrustRegionStrategyType kStrategy = LEVENBERG_MARQUARDT;
267   IsTrustRegionSolveSuccessful<true,  true,  true,  true >(kStrategy);
268   IsTrustRegionSolveSuccessful<true,  true,  true,  false>(kStrategy);
269   IsTrustRegionSolveSuccessful<true,  false, true,  true >(kStrategy);
270   IsTrustRegionSolveSuccessful<false, true,  true,  true >(kStrategy);
271   IsTrustRegionSolveSuccessful<true,  true,  false, false>(kStrategy);
272   IsTrustRegionSolveSuccessful<true,  false, true,  false>(kStrategy);
273   IsTrustRegionSolveSuccessful<false, true,  true,  false>(kStrategy);
274   IsTrustRegionSolveSuccessful<true,  false, false, true >(kStrategy);
275   IsTrustRegionSolveSuccessful<false, true,  false, true >(kStrategy);
276   IsTrustRegionSolveSuccessful<false, false, true,  true >(kStrategy);
277   IsTrustRegionSolveSuccessful<true,  false, false, false>(kStrategy);
278   IsTrustRegionSolveSuccessful<false, true,  false, false>(kStrategy);
279   IsTrustRegionSolveSuccessful<false, false, true,  false>(kStrategy);
280   IsTrustRegionSolveSuccessful<false, false, false, true >(kStrategy);
281 }
282 
TEST(TrustRegionMinimizer,PowellsSingularFunctionUsingDogleg)283 TEST(TrustRegionMinimizer, PowellsSingularFunctionUsingDogleg) {
284   // The following two cases are excluded because they encounter a
285   // local minimum.
286   //
287   //  IsTrustRegionSolveSuccessful<true, true, false, true >(kStrategy);
288   //  IsTrustRegionSolveSuccessful<true,  true,  true,  true >(kStrategy);
289 
290   const TrustRegionStrategyType kStrategy = DOGLEG;
291   IsTrustRegionSolveSuccessful<true,  true,  true,  false>(kStrategy);
292   IsTrustRegionSolveSuccessful<true,  false, true,  true >(kStrategy);
293   IsTrustRegionSolveSuccessful<false, true,  true,  true >(kStrategy);
294   IsTrustRegionSolveSuccessful<true,  true,  false, false>(kStrategy);
295   IsTrustRegionSolveSuccessful<true,  false, true,  false>(kStrategy);
296   IsTrustRegionSolveSuccessful<false, true,  true,  false>(kStrategy);
297   IsTrustRegionSolveSuccessful<true,  false, false, true >(kStrategy);
298   IsTrustRegionSolveSuccessful<false, true,  false, true >(kStrategy);
299   IsTrustRegionSolveSuccessful<false, false, true,  true >(kStrategy);
300   IsTrustRegionSolveSuccessful<true,  false, false, false>(kStrategy);
301   IsTrustRegionSolveSuccessful<false, true,  false, false>(kStrategy);
302   IsTrustRegionSolveSuccessful<false, false, true,  false>(kStrategy);
303   IsTrustRegionSolveSuccessful<false, false, false, true >(kStrategy);
304 }
305 
306 
307 class CurveCostFunction : public CostFunction {
308  public:
CurveCostFunction(int num_vertices,double target_length)309   CurveCostFunction(int num_vertices, double target_length)
310       : num_vertices_(num_vertices), target_length_(target_length) {
311     set_num_residuals(1);
312     for (int i = 0; i < num_vertices_; ++i) {
313       mutable_parameter_block_sizes()->push_back(2);
314     }
315   }
316 
Evaluate(double const * const * parameters,double * residuals,double ** jacobians) const317   bool Evaluate(double const* const* parameters,
318                 double* residuals,
319                 double** jacobians) const {
320     residuals[0] = target_length_;
321 
322     for (int i = 0; i < num_vertices_; ++i) {
323       int prev = (num_vertices_ + i - 1) % num_vertices_;
324       double length = 0.0;
325       for (int dim = 0; dim < 2; dim++) {
326         const double diff = parameters[prev][dim] - parameters[i][dim];
327         length += diff * diff;
328       }
329       residuals[0] -= sqrt(length);
330     }
331 
332     if (jacobians == NULL) {
333       return true;
334     }
335 
336     for (int i = 0; i < num_vertices_; ++i) {
337       if (jacobians[i] != NULL) {
338         int prev = (num_vertices_ + i - 1) % num_vertices_;
339         int next = (i + 1) % num_vertices_;
340 
341         double u[2], v[2];
342         double norm_u = 0., norm_v = 0.;
343         for (int dim = 0; dim < 2; dim++) {
344           u[dim] = parameters[i][dim] - parameters[prev][dim];
345           norm_u += u[dim] * u[dim];
346           v[dim] = parameters[next][dim] - parameters[i][dim];
347           norm_v += v[dim] * v[dim];
348         }
349 
350         norm_u = sqrt(norm_u);
351         norm_v = sqrt(norm_v);
352 
353         for (int dim = 0; dim < 2; dim++) {
354           jacobians[i][dim] = 0.;
355 
356           if (norm_u > std::numeric_limits< double >::min()) {
357             jacobians[i][dim] -= u[dim] / norm_u;
358           }
359 
360           if (norm_v > std::numeric_limits< double >::min()) {
361             jacobians[i][dim] += v[dim] / norm_v;
362           }
363         }
364       }
365     }
366 
367     return true;
368   }
369 
370  private:
371   int     num_vertices_;
372   double  target_length_;
373 };
374 
TEST(TrustRegionMinimizer,JacobiScalingTest)375 TEST(TrustRegionMinimizer, JacobiScalingTest) {
376   int N = 6;
377   std::vector< double* > y(N);
378   const double pi = 3.1415926535897932384626433;
379   for (int i = 0; i < N; i++) {
380     double theta = i * 2. * pi/ static_cast< double >(N);
381     y[i] = new double[2];
382     y[i][0] = cos(theta);
383     y[i][1] = sin(theta);
384   }
385 
386   Problem problem;
387   problem.AddResidualBlock(new CurveCostFunction(N, 10.), NULL, y);
388   Solver::Options options;
389   options.linear_solver_type = ceres::DENSE_QR;
390   Solver::Summary summary;
391   Solve(options, &problem, &summary);
392   EXPECT_LE(summary.final_cost, 1e-10);
393 
394   for (int i = 0; i < N; i++) {
395     delete []y[i];
396   }
397 }
398 
399 }  // namespace internal
400 }  // namespace ceres
401