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/coordinate_descent_minimizer.h"
32
33 #ifdef CERES_USE_OPENMP
34 #include <omp.h>
35 #endif
36
37 #include <iterator>
38 #include <numeric>
39 #include <vector>
40 #include "ceres/evaluator.h"
41 #include "ceres/linear_solver.h"
42 #include "ceres/minimizer.h"
43 #include "ceres/parameter_block.h"
44 #include "ceres/parameter_block_ordering.h"
45 #include "ceres/problem_impl.h"
46 #include "ceres/program.h"
47 #include "ceres/residual_block.h"
48 #include "ceres/solver.h"
49 #include "ceres/trust_region_minimizer.h"
50 #include "ceres/trust_region_strategy.h"
51 #include "ceres/parameter_block_ordering.h"
52
53 namespace ceres {
54 namespace internal {
55
~CoordinateDescentMinimizer()56 CoordinateDescentMinimizer::~CoordinateDescentMinimizer() {
57 }
58
Init(const Program & program,const ProblemImpl::ParameterMap & parameter_map,const ParameterBlockOrdering & ordering,string * error)59 bool CoordinateDescentMinimizer::Init(
60 const Program& program,
61 const ProblemImpl::ParameterMap& parameter_map,
62 const ParameterBlockOrdering& ordering,
63 string* error) {
64 parameter_blocks_.clear();
65 independent_set_offsets_.clear();
66 independent_set_offsets_.push_back(0);
67
68 // Serialize the OrderedGroups into a vector of parameter block
69 // offsets for parallel access.
70 map<ParameterBlock*, int> parameter_block_index;
71 map<int, set<double*> > group_to_elements = ordering.group_to_elements();
72 for (map<int, set<double*> >::const_iterator it = group_to_elements.begin();
73 it != group_to_elements.end();
74 ++it) {
75 for (set<double*>::const_iterator ptr_it = it->second.begin();
76 ptr_it != it->second.end();
77 ++ptr_it) {
78 parameter_blocks_.push_back(parameter_map.find(*ptr_it)->second);
79 parameter_block_index[parameter_blocks_.back()] =
80 parameter_blocks_.size() - 1;
81 }
82 independent_set_offsets_.push_back(
83 independent_set_offsets_.back() + it->second.size());
84 }
85
86 // The ordering does not have to contain all parameter blocks, so
87 // assign zero offsets/empty independent sets to these parameter
88 // blocks.
89 const vector<ParameterBlock*>& parameter_blocks = program.parameter_blocks();
90 for (int i = 0; i < parameter_blocks.size(); ++i) {
91 if (!ordering.IsMember(parameter_blocks[i]->mutable_user_state())) {
92 parameter_blocks_.push_back(parameter_blocks[i]);
93 independent_set_offsets_.push_back(independent_set_offsets_.back());
94 }
95 }
96
97 // Compute the set of residual blocks that depend on each parameter
98 // block.
99 residual_blocks_.resize(parameter_block_index.size());
100 const vector<ResidualBlock*>& residual_blocks = program.residual_blocks();
101 for (int i = 0; i < residual_blocks.size(); ++i) {
102 ResidualBlock* residual_block = residual_blocks[i];
103 const int num_parameter_blocks = residual_block->NumParameterBlocks();
104 for (int j = 0; j < num_parameter_blocks; ++j) {
105 ParameterBlock* parameter_block = residual_block->parameter_blocks()[j];
106 const map<ParameterBlock*, int>::const_iterator it =
107 parameter_block_index.find(parameter_block);
108 if (it != parameter_block_index.end()) {
109 residual_blocks_[it->second].push_back(residual_block);
110 }
111 }
112 }
113
114 evaluator_options_.linear_solver_type = DENSE_QR;
115 evaluator_options_.num_eliminate_blocks = 0;
116 evaluator_options_.num_threads = 1;
117
118 return true;
119 }
120
Minimize(const Minimizer::Options & options,double * parameters,Solver::Summary * summary)121 void CoordinateDescentMinimizer::Minimize(
122 const Minimizer::Options& options,
123 double* parameters,
124 Solver::Summary* summary) {
125 // Set the state and mark all parameter blocks constant.
126 for (int i = 0; i < parameter_blocks_.size(); ++i) {
127 ParameterBlock* parameter_block = parameter_blocks_[i];
128 parameter_block->SetState(parameters + parameter_block->state_offset());
129 parameter_block->SetConstant();
130 }
131
132 scoped_array<LinearSolver*> linear_solvers(
133 new LinearSolver*[options.num_threads]);
134
135 LinearSolver::Options linear_solver_options;
136 linear_solver_options.type = DENSE_QR;
137
138 for (int i = 0; i < options.num_threads; ++i) {
139 linear_solvers[i] = LinearSolver::Create(linear_solver_options);
140 }
141
142 for (int i = 0; i < independent_set_offsets_.size() - 1; ++i) {
143 // No point paying the price for an OpemMP call if the set if of
144 // size zero.
145 if (independent_set_offsets_[i] == independent_set_offsets_[i + 1]) {
146 continue;
147 }
148
149 // The parameter blocks in each independent set can be optimized
150 // in parallel, since they do not co-occur in any residual block.
151 #pragma omp parallel for num_threads(options.num_threads)
152 for (int j = independent_set_offsets_[i];
153 j < independent_set_offsets_[i + 1];
154 ++j) {
155 #ifdef CERES_USE_OPENMP
156 int thread_id = omp_get_thread_num();
157 #else
158 int thread_id = 0;
159 #endif
160
161 ParameterBlock* parameter_block = parameter_blocks_[j];
162 const int old_index = parameter_block->index();
163 const int old_delta_offset = parameter_block->delta_offset();
164 parameter_block->SetVarying();
165 parameter_block->set_index(0);
166 parameter_block->set_delta_offset(0);
167
168 Program inner_program;
169 inner_program.mutable_parameter_blocks()->push_back(parameter_block);
170 *inner_program.mutable_residual_blocks() = residual_blocks_[j];
171
172 // TODO(sameeragarwal): Better error handling. Right now we
173 // assume that this is not going to lead to problems of any
174 // sort. Basically we should be checking for numerical failure
175 // of some sort.
176 //
177 // On the other hand, if the optimization is a failure, that in
178 // some ways is fine, since it won't change the parameters and
179 // we are fine.
180 Solver::Summary inner_summary;
181 Solve(&inner_program,
182 linear_solvers[thread_id],
183 parameters + parameter_block->state_offset(),
184 &inner_summary);
185
186 parameter_block->set_index(old_index);
187 parameter_block->set_delta_offset(old_delta_offset);
188 parameter_block->SetState(parameters + parameter_block->state_offset());
189 parameter_block->SetConstant();
190 }
191 }
192
193 for (int i = 0; i < parameter_blocks_.size(); ++i) {
194 parameter_blocks_[i]->SetVarying();
195 }
196
197 for (int i = 0; i < options.num_threads; ++i) {
198 delete linear_solvers[i];
199 }
200 }
201
202 // Solve the optimization problem for one parameter block.
Solve(Program * program,LinearSolver * linear_solver,double * parameter,Solver::Summary * summary)203 void CoordinateDescentMinimizer::Solve(Program* program,
204 LinearSolver* linear_solver,
205 double* parameter,
206 Solver::Summary* summary) {
207 *summary = Solver::Summary();
208 summary->initial_cost = 0.0;
209 summary->fixed_cost = 0.0;
210 summary->final_cost = 0.0;
211 string error;
212
213 scoped_ptr<Evaluator> evaluator(
214 Evaluator::Create(evaluator_options_, program, &error));
215 CHECK_NOTNULL(evaluator.get());
216
217 scoped_ptr<SparseMatrix> jacobian(evaluator->CreateJacobian());
218 CHECK_NOTNULL(jacobian.get());
219
220 TrustRegionStrategy::Options trs_options;
221 trs_options.linear_solver = linear_solver;
222
223 scoped_ptr<TrustRegionStrategy>trust_region_strategy(
224 CHECK_NOTNULL(TrustRegionStrategy::Create(trs_options)));
225
226 Minimizer::Options minimizer_options;
227 minimizer_options.evaluator = evaluator.get();
228 minimizer_options.jacobian = jacobian.get();
229 minimizer_options.trust_region_strategy = trust_region_strategy.get();
230 minimizer_options.is_silent = true;
231
232 TrustRegionMinimizer minimizer;
233 minimizer.Minimize(minimizer_options, parameter, summary);
234 }
235
IsOrderingValid(const Program & program,const ParameterBlockOrdering & ordering,string * message)236 bool CoordinateDescentMinimizer::IsOrderingValid(
237 const Program& program,
238 const ParameterBlockOrdering& ordering,
239 string* message) {
240 const map<int, set<double*> >& group_to_elements =
241 ordering.group_to_elements();
242
243 // Verify that each group is an independent set
244 map<int, set<double*> >::const_iterator it = group_to_elements.begin();
245 for ( ; it != group_to_elements.end(); ++it) {
246 if (!program.IsParameterBlockSetIndependent(it->second)) {
247 *message =
248 StringPrintf("The user-provided "
249 "parameter_blocks_for_inner_iterations does not "
250 "form an independent set. Group Id: %d", it->first);
251 return false;
252 }
253 }
254 return true;
255 }
256
257 // Find a recursive decomposition of the Hessian matrix as a set
258 // of independent sets of decreasing size and invert it. This
259 // seems to work better in practice, i.e., Cameras before
260 // points.
CreateOrdering(const Program & program)261 ParameterBlockOrdering* CoordinateDescentMinimizer::CreateOrdering(
262 const Program& program) {
263 scoped_ptr<ParameterBlockOrdering> ordering(new ParameterBlockOrdering);
264 ComputeRecursiveIndependentSetOrdering(program, ordering.get());
265 ordering->Reverse();
266 return ordering.release();
267 }
268
269 } // namespace internal
270 } // namespace ceres
271