internal/ceres/sparse_normal_cholesky_solver_test.cc

   1 // Ceres Solver - A fast non-linear least squares minimizer
   2 // Copyright 2017 Google Inc. All rights reserved.
   3 // http://ceres-solver.org/
   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/block_sparse_matrix.h"
  32 #include "ceres/casts.h"
  33 #include "ceres/internal/scoped_ptr.h"
  34 #include "ceres/linear_least_squares_problems.h"
  35 #include "ceres/linear_solver.h"
  36 #include "ceres/triplet_sparse_matrix.h"
  37 #include "ceres/types.h"
  38 #include "glog/logging.h"
  39 #include "gtest/gtest.h"
  40
  41 #include "Eigen/Cholesky"
  42
  43 namespace ceres {
  44 namespace internal {
  45
  46 // TODO(sameeragarwal): These tests needs to be re-written, since
  47 // SparseNormalCholeskySolver is a composition of two classes now,
  48 // InnerProductComputer and SparseCholesky.
  49 //
  50 // So the test should exercise the composition, rather than the
  51 // numerics of the solver, which are well covered by tests for those
  52 // classes.
  53 class SparseNormalCholeskySolverTest : public ::testing::Test {
  54  protected:
  55   virtual void SetUp() {
  56     scoped_ptr<LinearLeastSquaresProblem> problem(
  57         CreateLinearLeastSquaresProblemFromId(2));
  58
  59     CHECK_NOTNULL(problem.get());
  60     A_.reset(down_cast<BlockSparseMatrix*>(problem->A.release()));
  61     b_.reset(problem->b.release());
  62     D_.reset(problem->D.release());
  63   }
  64
  65   void TestSolver(const LinearSolver::Options& options, double* D) {
  66     Matrix dense_A;
  67     A_->ToDenseMatrix(&dense_A);
  68     Matrix lhs = dense_A.transpose() * dense_A;
  69     if (D != NULL) {
  70       lhs += (ConstVectorRef(D, A_->num_cols()).array() *
  71               ConstVectorRef(D, A_->num_cols()).array())
  72                  .matrix()
  73                  .asDiagonal();
  74     }
  75
  76     Vector rhs(A_->num_cols());
  77     rhs.setZero();
  78     A_->LeftMultiply(b_.get(), rhs.data());
  79     Vector expected_solution = lhs.llt().solve(rhs);
  80
  81     scoped_ptr<LinearSolver> solver(LinearSolver::Create(options));
  82     LinearSolver::PerSolveOptions per_solve_options;
  83     per_solve_options.D = D;
  84     Vector actual_solution(A_->num_cols());
  85     LinearSolver::Summary summary;
  86     summary = solver->Solve(
  87         A_.get(), b_.get(), per_solve_options, actual_solution.data());
  88
  89     EXPECT_EQ(summary.termination_type, LINEAR_SOLVER_SUCCESS);
  90
  91     for (int i = 0; i < A_->num_cols(); ++i) {
  92       EXPECT_NEAR(expected_solution(i), actual_solution(i), 1e-8)
  93           << "\nExpected: " << expected_solution.transpose()
  94           << "\nActual: " << actual_solution.transpose();
  95     }
  96   }
  97
  98   void TestSolver(const LinearSolver::Options& options) {
  99     TestSolver(options, NULL);
 100     TestSolver(options, D_.get());
 101   }
 102
 103   scoped_ptr<BlockSparseMatrix> A_;
 104   scoped_array<double> b_;
 105   scoped_array<double> D_;
 106 };
 107
 108 #ifndef CERES_NO_SUITESPARSE
 109 TEST_F(SparseNormalCholeskySolverTest,
 110        SparseNormalCholeskyUsingSuiteSparsePreOrdering) {
 111   LinearSolver::Options options;
 112   options.sparse_linear_algebra_library_type = SUITE_SPARSE;
 113   options.type = SPARSE_NORMAL_CHOLESKY;
 114   options.use_postordering = false;
 115   TestSolver(options);
 116 }
 117
 118 TEST_F(SparseNormalCholeskySolverTest,
 119        SparseNormalCholeskyUsingSuiteSparsePostOrdering) {
 120   LinearSolver::Options options;
 121   options.sparse_linear_algebra_library_type = SUITE_SPARSE;
 122   options.type = SPARSE_NORMAL_CHOLESKY;
 123   options.use_postordering = true;
 124   TestSolver(options);
 125 }
 126 #endif
 127
 128 #ifndef CERES_NO_CXSPARSE
 129 TEST_F(SparseNormalCholeskySolverTest,
 130        SparseNormalCholeskyUsingCXSparsePreOrdering) {
 131   LinearSolver::Options options;
 132   options.sparse_linear_algebra_library_type = CX_SPARSE;
 133   options.type = SPARSE_NORMAL_CHOLESKY;
 134   options.use_postordering = false;
 135   TestSolver(options);
 136 }
 137
 138 TEST_F(SparseNormalCholeskySolverTest,
 139        SparseNormalCholeskyUsingCXSparsePostOrdering) {
 140   LinearSolver::Options options;
 141   options.sparse_linear_algebra_library_type = CX_SPARSE;
 142   options.type = SPARSE_NORMAL_CHOLESKY;
 143   options.use_postordering = true;
 144   TestSolver(options);
 145 }
 146 #endif
 147
 148 #ifdef CERES_USE_EIGEN_SPARSE
 149 TEST_F(SparseNormalCholeskySolverTest,
 150        SparseNormalCholeskyUsingEigenPreOrdering) {
 151   LinearSolver::Options options;
 152   options.sparse_linear_algebra_library_type = EIGEN_SPARSE;
 153   options.type = SPARSE_NORMAL_CHOLESKY;
 154   options.use_postordering = false;
 155   TestSolver(options);
 156 }
 157
 158 TEST_F(SparseNormalCholeskySolverTest,
 159        SparseNormalCholeskyUsingEigenPostOrdering) {
 160   LinearSolver::Options options;
 161   options.sparse_linear_algebra_library_type = EIGEN_SPARSE;
 162   options.type = SPARSE_NORMAL_CHOLESKY;
 163   options.use_postordering = true;
 164   TestSolver(options);
 165 }
 166 #endif  // CERES_USE_EIGEN_SPARSE
 167
 168 }  // namespace internal
 169 }  // namespace ceres