internal/ceres/dense_linear_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/casts.h"
  32 #include "ceres/internal/scoped_ptr.h"
  33 #include "ceres/linear_least_squares_problems.h"
  34 #include "ceres/linear_solver.h"
  35 #include "ceres/triplet_sparse_matrix.h"
  36 #include "ceres/types.h"
  37 #include "glog/logging.h"
  38 #include "gtest/gtest.h"
  39
  40 namespace ceres {
  41 namespace internal {
  42
  43 typedef ::testing::
  44     tuple<LinearSolverType, DenseLinearAlgebraLibraryType, bool, int>
  45         Param;
  46
  47 std::string ParamInfoToString(testing::TestParamInfo<Param> info) {
  48   Param param = info.param;
  49   std::stringstream ss;
  50   ss << LinearSolverTypeToString(::testing::get<0>(param)) << "_"
  51      << DenseLinearAlgebraLibraryTypeToString(::testing::get<1>(param)) << "_"
  52      << (::testing::get<2>(param) ? "Regularized" : "Unregularized") << "_"
  53      << ::testing::get<3>(param);
  54   return ss.str();
  55 }
  56
  57 class DenseLinearSolverTest : public ::testing::TestWithParam<Param> {};
  58
  59 TEST_P(DenseLinearSolverTest, _) {
  60   Param param = GetParam();
  61   const bool regularized = testing::get<2>(param);
  62
  63   scoped_ptr<LinearLeastSquaresProblem> problem(
  64       CreateLinearLeastSquaresProblemFromId(testing::get<3>(param)));
  65   DenseSparseMatrix lhs(*down_cast<TripletSparseMatrix*>(problem->A.get()));
  66
  67   const int num_cols = lhs.num_cols();
  68   const int num_rows = lhs.num_rows();
  69
  70   Vector rhs = Vector::Zero(num_rows + num_cols);
  71   rhs.head(num_rows) = ConstVectorRef(problem->b.get(), num_rows);
  72
  73   LinearSolver::Options options;
  74   options.type = ::testing::get<0>(param);
  75   options.dense_linear_algebra_library_type = ::testing::get<1>(param);
  76   scoped_ptr<LinearSolver> solver(LinearSolver::Create(options));
  77
  78   LinearSolver::PerSolveOptions per_solve_options;
  79   if (regularized) {
  80     per_solve_options.D = problem->D.get();
  81   }
  82
  83   Vector solution(num_cols);
  84   LinearSolver::Summary summary =
  85       solver->Solve(&lhs, rhs.data(), per_solve_options, solution.data());
  86   EXPECT_EQ(summary.termination_type, LINEAR_SOLVER_SUCCESS);
  87
  88   // If solving for the regularized solution, add the diagonal to the
  89   // matrix. This makes subsequent computations simpler.
  90   if (testing::get<2>(param)) {
  91     lhs.AppendDiagonal(problem->D.get());
  92   };
  93
  94   Vector tmp = Vector::Zero(num_rows + num_cols);
  95   lhs.RightMultiply(solution.data(), tmp.data());
  96   Vector actual_normal_rhs = Vector::Zero(num_cols);
  97   lhs.LeftMultiply(tmp.data(), actual_normal_rhs.data());
  98
  99   Vector expected_normal_rhs = Vector::Zero(num_cols);
 100   lhs.LeftMultiply(rhs.data(), expected_normal_rhs.data());
 101   const double residual = (expected_normal_rhs - actual_normal_rhs).norm() /
 102                           expected_normal_rhs.norm();
 103
 104   EXPECT_NEAR(residual, 0.0, 10 * std::numeric_limits<double>::epsilon());
 105 }
 106
 107 // TODO(sameeragarwal): Should we move away from hard coded linear
 108 // least squares problem to randomly generated ones?
 109 #ifndef CERES_NO_LAPACK
 110
 111 INSTANTIATE_TEST_CASE_P(
 112     DenseLinearSolver,
 113     DenseLinearSolverTest,
 114     ::testing::Combine(::testing::Values(DENSE_QR, DENSE_NORMAL_CHOLESKY),
 115                        ::testing::Values(EIGEN, LAPACK),
 116                        ::testing::Values(true, false),
 117                        ::testing::Values(0, 1)),
 118     ParamInfoToString);
 119
 120 #else
 121
 122 INSTANTIATE_TEST_CASE_P(
 123     DenseLinearSolver,
 124     DenseLinearSolverTest,
 125     ::testing::Combine(::testing::Values(DENSE_QR, DENSE_NORMAL_CHOLESKY),
 126                        ::testing::Values(EIGEN),
 127                        ::testing::Values(true, false),
 128                        ::testing::Values(0, 1)),
 129     ParamInfoToString);
 130
 131 #endif
 132
 133 }  // namespace internal
 134 }  // namespace ceres