1 // Ceres Solver - A fast non-linear least squares minimizer
2 // Copyright 2019 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/schur_eliminator.h"
32 
33 #include <memory>
34 
35 #include "Eigen/Dense"
36 #include "ceres/block_random_access_dense_matrix.h"
37 #include "ceres/block_sparse_matrix.h"
38 #include "ceres/block_structure.h"
39 #include "ceres/casts.h"
40 #include "ceres/context_impl.h"
41 #include "ceres/detect_structure.h"
42 #include "ceres/internal/eigen.h"
43 #include "ceres/linear_least_squares_problems.h"
44 #include "ceres/random.h"
45 #include "ceres/test_util.h"
46 #include "ceres/triplet_sparse_matrix.h"
47 #include "ceres/types.h"
48 #include "glog/logging.h"
49 #include "gtest/gtest.h"
50 
51 // TODO(sameeragarwal): Reduce the size of these tests and redo the
52 // parameterization to be more efficient.
53 
54 namespace ceres {
55 namespace internal {
56 
57 class SchurEliminatorTest : public ::testing::Test {
58  protected:
SetUpFromId(int id)59   void SetUpFromId(int id) {
60     std::unique_ptr<LinearLeastSquaresProblem> problem(
61         CreateLinearLeastSquaresProblemFromId(id));
62     CHECK(problem != nullptr);
63     SetupHelper(problem.get());
64   }
65 
SetupHelper(LinearLeastSquaresProblem * problem)66   void SetupHelper(LinearLeastSquaresProblem* problem) {
67     A.reset(down_cast<BlockSparseMatrix*>(problem->A.release()));
68     b.reset(problem->b.release());
69     D.reset(problem->D.release());
70 
71     num_eliminate_blocks = problem->num_eliminate_blocks;
72     num_eliminate_cols = 0;
73     const CompressedRowBlockStructure* bs = A->block_structure();
74 
75     for (int i = 0; i < num_eliminate_blocks; ++i) {
76       num_eliminate_cols += bs->cols[i].size;
77     }
78   }
79 
80   // Compute the golden values for the reduced linear system and the
81   // solution to the linear least squares problem using dense linear
82   // algebra.
ComputeReferenceSolution(const Vector & D)83   void ComputeReferenceSolution(const Vector& D) {
84     Matrix J;
85     A->ToDenseMatrix(&J);
86     VectorRef f(b.get(), J.rows());
87 
88     Matrix H = (D.cwiseProduct(D)).asDiagonal();
89     H.noalias() += J.transpose() * J;
90 
91     const Vector g = J.transpose() * f;
92     const int schur_size = J.cols() - num_eliminate_cols;
93 
94     lhs_expected.resize(schur_size, schur_size);
95     lhs_expected.setZero();
96 
97     rhs_expected.resize(schur_size);
98     rhs_expected.setZero();
99 
100     sol_expected.resize(J.cols());
101     sol_expected.setZero();
102 
103     Matrix P = H.block(0, 0, num_eliminate_cols, num_eliminate_cols);
104     Matrix Q = H.block(0, num_eliminate_cols, num_eliminate_cols, schur_size);
105     Matrix R =
106         H.block(num_eliminate_cols, num_eliminate_cols, schur_size, schur_size);
107     int row = 0;
108     const CompressedRowBlockStructure* bs = A->block_structure();
109     for (int i = 0; i < num_eliminate_blocks; ++i) {
110       const int block_size = bs->cols[i].size;
111       P.block(row, row, block_size, block_size) =
112           P.block(row, row, block_size, block_size)
113               .llt()
114               .solve(Matrix::Identity(block_size, block_size));
115       row += block_size;
116     }
117 
118     lhs_expected.triangularView<Eigen::Upper>() = R - Q.transpose() * P * Q;
119     rhs_expected =
120         g.tail(schur_size) - Q.transpose() * P * g.head(num_eliminate_cols);
121     sol_expected = H.llt().solve(g);
122   }
123 
EliminateSolveAndCompare(const VectorRef & diagonal,bool use_static_structure,const double relative_tolerance)124   void EliminateSolveAndCompare(const VectorRef& diagonal,
125                                 bool use_static_structure,
126                                 const double relative_tolerance) {
127     const CompressedRowBlockStructure* bs = A->block_structure();
128     const int num_col_blocks = bs->cols.size();
129     std::vector<int> blocks(num_col_blocks - num_eliminate_blocks, 0);
130     for (int i = num_eliminate_blocks; i < num_col_blocks; ++i) {
131       blocks[i - num_eliminate_blocks] = bs->cols[i].size;
132     }
133 
134     BlockRandomAccessDenseMatrix lhs(blocks);
135 
136     const int num_cols = A->num_cols();
137     const int schur_size = lhs.num_rows();
138 
139     Vector rhs(schur_size);
140 
141     LinearSolver::Options options;
142     ContextImpl context;
143     options.context = &context;
144     options.elimination_groups.push_back(num_eliminate_blocks);
145     if (use_static_structure) {
146       DetectStructure(*bs,
147                       num_eliminate_blocks,
148                       &options.row_block_size,
149                       &options.e_block_size,
150                       &options.f_block_size);
151     }
152 
153     std::unique_ptr<SchurEliminatorBase> eliminator;
154     eliminator.reset(SchurEliminatorBase::Create(options));
155     const bool kFullRankETE = true;
156     eliminator->Init(num_eliminate_blocks, kFullRankETE, A->block_structure());
157     eliminator->Eliminate(
158         BlockSparseMatrixData(*A), b.get(), diagonal.data(), &lhs, rhs.data());
159 
160     MatrixRef lhs_ref(lhs.mutable_values(), lhs.num_rows(), lhs.num_cols());
161     Vector reduced_sol =
162         lhs_ref.selfadjointView<Eigen::Upper>().llt().solve(rhs);
163 
164     // Solution to the linear least squares problem.
165     Vector sol(num_cols);
166     sol.setZero();
167     sol.tail(schur_size) = reduced_sol;
168     eliminator->BackSubstitute(BlockSparseMatrixData(*A),
169                                b.get(),
170                                diagonal.data(),
171                                reduced_sol.data(),
172                                sol.data());
173 
174     Matrix delta = (lhs_ref - lhs_expected).selfadjointView<Eigen::Upper>();
175     double diff = delta.norm();
176     EXPECT_NEAR(diff / lhs_expected.norm(), 0.0, relative_tolerance);
177     EXPECT_NEAR((rhs - rhs_expected).norm() / rhs_expected.norm(),
178                 0.0,
179                 relative_tolerance);
180     EXPECT_NEAR((sol - sol_expected).norm() / sol_expected.norm(),
181                 0.0,
182                 relative_tolerance);
183   }
184 
185   std::unique_ptr<BlockSparseMatrix> A;
186   std::unique_ptr<double[]> b;
187   std::unique_ptr<double[]> D;
188   int num_eliminate_blocks;
189   int num_eliminate_cols;
190 
191   Matrix lhs_expected;
192   Vector rhs_expected;
193   Vector sol_expected;
194 };
195 
TEST_F(SchurEliminatorTest,ScalarProblemNoRegularization)196 TEST_F(SchurEliminatorTest, ScalarProblemNoRegularization) {
197   SetUpFromId(2);
198   Vector zero(A->num_cols());
199   zero.setZero();
200 
201   ComputeReferenceSolution(VectorRef(zero.data(), A->num_cols()));
202   EliminateSolveAndCompare(VectorRef(zero.data(), A->num_cols()), true, 1e-14);
203   EliminateSolveAndCompare(VectorRef(zero.data(), A->num_cols()), false, 1e-14);
204 }
205 
TEST_F(SchurEliminatorTest,ScalarProblemWithRegularization)206 TEST_F(SchurEliminatorTest, ScalarProblemWithRegularization) {
207   SetUpFromId(2);
208   ComputeReferenceSolution(VectorRef(D.get(), A->num_cols()));
209   EliminateSolveAndCompare(VectorRef(D.get(), A->num_cols()), true, 1e-14);
210   EliminateSolveAndCompare(VectorRef(D.get(), A->num_cols()), false, 1e-14);
211 }
212 
TEST_F(SchurEliminatorTest,VaryingFBlockSizeWithStaticStructure)213 TEST_F(SchurEliminatorTest, VaryingFBlockSizeWithStaticStructure) {
214   SetUpFromId(4);
215   ComputeReferenceSolution(VectorRef(D.get(), A->num_cols()));
216   EliminateSolveAndCompare(VectorRef(D.get(), A->num_cols()), true, 1e-14);
217 }
218 
TEST_F(SchurEliminatorTest,VaryingFBlockSizeWithoutStaticStructure)219 TEST_F(SchurEliminatorTest, VaryingFBlockSizeWithoutStaticStructure) {
220   SetUpFromId(4);
221   ComputeReferenceSolution(VectorRef(D.get(), A->num_cols()));
222   EliminateSolveAndCompare(VectorRef(D.get(), A->num_cols()), false, 1e-14);
223 }
224 
TEST(SchurEliminatorForOneFBlock,MatchesSchurEliminator)225 TEST(SchurEliminatorForOneFBlock, MatchesSchurEliminator) {
226   constexpr int kRowBlockSize = 2;
227   constexpr int kEBlockSize = 3;
228   constexpr int kFBlockSize = 6;
229   constexpr int num_e_blocks = 5;
230 
231   CompressedRowBlockStructure* bs = new CompressedRowBlockStructure;
232   bs->cols.resize(num_e_blocks + 1);
233   int col_pos = 0;
234   for (int i = 0; i < num_e_blocks; ++i) {
235     bs->cols[i].position = col_pos;
236     bs->cols[i].size = kEBlockSize;
237     col_pos += kEBlockSize;
238   }
239   bs->cols.back().position = col_pos;
240   bs->cols.back().size = kFBlockSize;
241 
242   bs->rows.resize(2 * num_e_blocks + 1);
243   int row_pos = 0;
244   int cell_pos = 0;
245   for (int i = 0; i < num_e_blocks; ++i) {
246     {
247       auto& row = bs->rows[2 * i];
248       row.block.position = row_pos;
249       row.block.size = kRowBlockSize;
250       row_pos += kRowBlockSize;
251       auto& cells = row.cells;
252       cells.resize(2);
253       cells[0].block_id = i;
254       cells[0].position = cell_pos;
255       cell_pos += kRowBlockSize * kEBlockSize;
256       cells[1].block_id = num_e_blocks;
257       cells[1].position = cell_pos;
258       cell_pos += kRowBlockSize * kFBlockSize;
259     }
260     {
261       auto& row = bs->rows[2 * i + 1];
262       row.block.position = row_pos;
263       row.block.size = kRowBlockSize;
264       row_pos += kRowBlockSize;
265       auto& cells = row.cells;
266       cells.resize(1);
267       cells[0].block_id = i;
268       cells[0].position = cell_pos;
269       cell_pos += kRowBlockSize * kEBlockSize;
270     }
271   }
272 
273   {
274     auto& row = bs->rows.back();
275     row.block.position = row_pos;
276     row.block.size = kEBlockSize;
277     row_pos += kRowBlockSize;
278     auto& cells = row.cells;
279     cells.resize(1);
280     cells[0].block_id = num_e_blocks;
281     cells[0].position = cell_pos;
282     cell_pos += kEBlockSize * kEBlockSize;
283   }
284 
285   BlockSparseMatrix matrix(bs);
286   double* values = matrix.mutable_values();
287   for (int i = 0; i < matrix.num_nonzeros(); ++i) {
288     values[i] = RandNormal();
289   }
290 
291   Vector b(matrix.num_rows());
292   b.setRandom();
293 
294   Vector diagonal(matrix.num_cols());
295   diagonal.setOnes();
296 
297   std::vector<int> blocks(1, kFBlockSize);
298   BlockRandomAccessDenseMatrix actual_lhs(blocks);
299   BlockRandomAccessDenseMatrix expected_lhs(blocks);
300   Vector actual_rhs(kFBlockSize);
301   Vector expected_rhs(kFBlockSize);
302 
303   Vector f_sol(kFBlockSize);
304   f_sol.setRandom();
305   Vector actual_e_sol(num_e_blocks * kEBlockSize);
306   actual_e_sol.setZero();
307   Vector expected_e_sol(num_e_blocks * kEBlockSize);
308   expected_e_sol.setZero();
309 
310   {
311     ContextImpl context;
312     LinearSolver::Options linear_solver_options;
313     linear_solver_options.e_block_size = kEBlockSize;
314     linear_solver_options.row_block_size = kRowBlockSize;
315     linear_solver_options.f_block_size = kFBlockSize;
316     linear_solver_options.context = &context;
317     std::unique_ptr<SchurEliminatorBase> eliminator(
318         SchurEliminatorBase::Create(linear_solver_options));
319     eliminator->Init(num_e_blocks, true, matrix.block_structure());
320     eliminator->Eliminate(BlockSparseMatrixData(matrix),
321                           b.data(),
322                           diagonal.data(),
323                           &expected_lhs,
324                           expected_rhs.data());
325     eliminator->BackSubstitute(BlockSparseMatrixData(matrix),
326                                b.data(),
327                                diagonal.data(),
328                                f_sol.data(),
329                                actual_e_sol.data());
330   }
331 
332   {
333     SchurEliminatorForOneFBlock<2, 3, 6> eliminator;
334     eliminator.Init(num_e_blocks, true, matrix.block_structure());
335     eliminator.Eliminate(BlockSparseMatrixData(matrix),
336                          b.data(),
337                          diagonal.data(),
338                          &actual_lhs,
339                          actual_rhs.data());
340     eliminator.BackSubstitute(BlockSparseMatrixData(matrix),
341                               b.data(),
342                               diagonal.data(),
343                               f_sol.data(),
344                               expected_e_sol.data());
345   }
346   ConstMatrixRef actual_lhsref(
347       actual_lhs.values(), actual_lhs.num_cols(), actual_lhs.num_cols());
348   ConstMatrixRef expected_lhsref(
349       expected_lhs.values(), actual_lhs.num_cols(), actual_lhs.num_cols());
350 
351   EXPECT_NEAR((actual_lhsref - expected_lhsref).norm() / expected_lhsref.norm(),
352               0.0,
353               1e-12)
354       << "expected: \n"
355       << expected_lhsref << "\nactual: \n"
356       << actual_lhsref;
357 
358   EXPECT_NEAR(
359       (actual_rhs - expected_rhs).norm() / expected_rhs.norm(), 0.0, 1e-12)
360       << "expected: \n"
361       << expected_rhs << "\nactual: \n"
362       << actual_rhs;
363 
364   EXPECT_NEAR((actual_e_sol - expected_e_sol).norm() / expected_e_sol.norm(),
365               0.0,
366               1e-12)
367       << "expected: \n"
368       << expected_e_sol << "\nactual: \n"
369       << actual_e_sol;
370 }
371 
372 }  // namespace internal
373 }  // namespace ceres
374