• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Ceres Solver - A fast non-linear least squares minimizer
2 // Copyright 2013 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 //         mierle@gmail.com (Keir Mierle)
31 //
32 // This autodiff implementation differs from the one found in
33 // autodiff_cost_function.h by supporting autodiff on cost functions
34 // with variable numbers of parameters with variable sizes. With the
35 // other implementation, all the sizes (both the number of parameter
36 // blocks and the size of each block) must be fixed at compile time.
37 //
38 // The functor API differs slightly from the API for fixed size
39 // autodiff; the expected interface for the cost functors is:
40 //
41 //   struct MyCostFunctor {
42 //     template<typename T>
43 //     bool operator()(T const* const* parameters, T* residuals) const {
44 //       // Use parameters[i] to access the i'th parameter block.
45 //     }
46 //   }
47 //
48 // Since the sizing of the parameters is done at runtime, you must
49 // also specify the sizes after creating the dynamic autodiff cost
50 // function. For example:
51 //
52 //   DynamicAutoDiffCostFunction<MyCostFunctor, 3> cost_function(
53 //       new MyCostFunctor());
54 //   cost_function.AddParameterBlock(5);
55 //   cost_function.AddParameterBlock(10);
56 //   cost_function.SetNumResiduals(21);
57 //
58 // Under the hood, the implementation evaluates the cost function
59 // multiple times, computing a small set of the derivatives (four by
60 // default, controlled by the Stride template parameter) with each
61 // pass. There is a tradeoff with the size of the passes; you may want
62 // to experiment with the stride.
63 
64 #ifndef CERES_PUBLIC_DYNAMIC_AUTODIFF_COST_FUNCTION_H_
65 #define CERES_PUBLIC_DYNAMIC_AUTODIFF_COST_FUNCTION_H_
66 
67 #include <cmath>
68 #include <numeric>
69 #include <vector>
70 
71 #include "ceres/cost_function.h"
72 #include "ceres/internal/scoped_ptr.h"
73 #include "ceres/jet.h"
74 #include "glog/logging.h"
75 
76 namespace ceres {
77 
78 template <typename CostFunctor, int Stride = 4>
79 class DynamicAutoDiffCostFunction : public CostFunction {
80  public:
DynamicAutoDiffCostFunction(CostFunctor * functor)81   explicit DynamicAutoDiffCostFunction(CostFunctor* functor)
82     : functor_(functor) {}
83 
~DynamicAutoDiffCostFunction()84   virtual ~DynamicAutoDiffCostFunction() {}
85 
AddParameterBlock(int size)86   void AddParameterBlock(int size) {
87     mutable_parameter_block_sizes()->push_back(size);
88   }
89 
SetNumResiduals(int num_residuals)90   void SetNumResiduals(int num_residuals) {
91     set_num_residuals(num_residuals);
92   }
93 
Evaluate(double const * const * parameters,double * residuals,double ** jacobians)94   virtual bool Evaluate(double const* const* parameters,
95                         double* residuals,
96                         double** jacobians) const {
97     CHECK_GT(num_residuals(), 0)
98         << "You must call DynamicAutoDiffCostFunction::SetNumResiduals() "
99         << "before DynamicAutoDiffCostFunction::Evaluate().";
100 
101     if (jacobians == NULL) {
102       return (*functor_)(parameters, residuals);
103     }
104 
105     // The difficulty with Jets, as implemented in Ceres, is that they were
106     // originally designed for strictly compile-sized use. At this point, there
107     // is a large body of code that assumes inside a cost functor it is
108     // acceptable to do e.g. T(1.5) and get an appropriately sized jet back.
109     //
110     // Unfortunately, it is impossible to communicate the expected size of a
111     // dynamically sized jet to the static instantiations that existing code
112     // depends on.
113     //
114     // To work around this issue, the solution here is to evaluate the
115     // jacobians in a series of passes, each one computing Stripe *
116     // num_residuals() derivatives. This is done with small, fixed-size jets.
117     const int num_parameter_blocks = parameter_block_sizes().size();
118     const int num_parameters = std::accumulate(parameter_block_sizes().begin(),
119                                                parameter_block_sizes().end(),
120                                                0);
121 
122     // Allocate scratch space for the strided evaluation.
123     vector<Jet<double, Stride> > input_jets(num_parameters);
124     vector<Jet<double, Stride> > output_jets(num_residuals());
125 
126     // Make the parameter pack that is sent to the functor (reused).
127     vector<Jet<double, Stride>* > jet_parameters(num_parameter_blocks,
128         static_cast<Jet<double, Stride>* >(NULL));
129     int num_active_parameters = 0;
130 
131     // To handle constant parameters between non-constant parameter blocks, the
132     // start position --- a raw parameter index --- of each contiguous block of
133     // non-constant parameters is recorded in start_derivative_section.
134     vector<int> start_derivative_section;
135     bool in_derivative_section = false;
136     int parameter_cursor = 0;
137 
138     // Discover the derivative sections and set the parameter values.
139     for (int i = 0; i < num_parameter_blocks; ++i) {
140       jet_parameters[i] = &input_jets[parameter_cursor];
141 
142       const int parameter_block_size = parameter_block_sizes()[i];
143       if (jacobians[i] != NULL) {
144         if (!in_derivative_section) {
145           start_derivative_section.push_back(parameter_cursor);
146           in_derivative_section = true;
147         }
148 
149         num_active_parameters += parameter_block_size;
150       } else {
151         in_derivative_section = false;
152       }
153 
154       for (int j = 0; j < parameter_block_size; ++j, parameter_cursor++) {
155         input_jets[parameter_cursor].a = parameters[i][j];
156       }
157     }
158 
159     // When `num_active_parameters % Stride != 0` then it can be the case
160     // that `active_parameter_count < Stride` while parameter_cursor is less
161     // than the total number of parameters and with no remaining non-constant
162     // parameter blocks. Pushing parameter_cursor (the total number of
163     // parameters) as a final entry to start_derivative_section is required
164     // because if a constant parameter block is encountered after the
165     // last non-constant block then current_derivative_section is incremented
166     // and would otherwise index an invalid position in
167     // start_derivative_section. Setting the final element to the total number
168     // of parameters means that this can only happen at most once in the loop
169     // below.
170     start_derivative_section.push_back(parameter_cursor);
171 
172     // Evaluate all of the strides. Each stride is a chunk of the derivative to
173     // evaluate, typically some size proportional to the size of the SIMD
174     // registers of the CPU.
175     int num_strides = static_cast<int>(ceil(num_active_parameters /
176                                             static_cast<float>(Stride)));
177 
178     int current_derivative_section = 0;
179     int current_derivative_section_cursor = 0;
180 
181     for (int pass = 0; pass < num_strides; ++pass) {
182       // Set most of the jet components to zero, except for
183       // non-constant #Stride parameters.
184       const int initial_derivative_section = current_derivative_section;
185       const int initial_derivative_section_cursor =
186         current_derivative_section_cursor;
187 
188       int active_parameter_count = 0;
189       parameter_cursor = 0;
190 
191       for (int i = 0; i < num_parameter_blocks; ++i) {
192         for (int j = 0; j < parameter_block_sizes()[i];
193              ++j, parameter_cursor++) {
194           input_jets[parameter_cursor].v.setZero();
195           if (active_parameter_count < Stride &&
196               parameter_cursor >= (
197                 start_derivative_section[current_derivative_section] +
198                 current_derivative_section_cursor)) {
199             if (jacobians[i] != NULL) {
200               input_jets[parameter_cursor].v[active_parameter_count] = 1.0;
201               ++active_parameter_count;
202               ++current_derivative_section_cursor;
203             } else {
204               ++current_derivative_section;
205               current_derivative_section_cursor = 0;
206             }
207           }
208         }
209       }
210 
211       if (!(*functor_)(&jet_parameters[0], &output_jets[0])) {
212         return false;
213       }
214 
215       // Copy the pieces of the jacobians into their final place.
216       active_parameter_count = 0;
217 
218       current_derivative_section = initial_derivative_section;
219       current_derivative_section_cursor = initial_derivative_section_cursor;
220 
221       for (int i = 0, parameter_cursor = 0; i < num_parameter_blocks; ++i) {
222         for (int j = 0; j < parameter_block_sizes()[i];
223              ++j, parameter_cursor++) {
224           if (active_parameter_count < Stride &&
225               parameter_cursor >= (
226                 start_derivative_section[current_derivative_section] +
227                 current_derivative_section_cursor)) {
228             if (jacobians[i] != NULL) {
229               for (int k = 0; k < num_residuals(); ++k) {
230                 jacobians[i][k * parameter_block_sizes()[i] + j] =
231                     output_jets[k].v[active_parameter_count];
232               }
233               ++active_parameter_count;
234               ++current_derivative_section_cursor;
235             } else {
236               ++current_derivative_section;
237               current_derivative_section_cursor = 0;
238             }
239           }
240         }
241       }
242 
243       // Only copy the residuals over once (even though we compute them on
244       // every loop).
245       if (pass == num_strides - 1) {
246         for (int k = 0; k < num_residuals(); ++k) {
247           residuals[k] = output_jets[k].a;
248         }
249       }
250     }
251     return true;
252   }
253 
254  private:
255   internal::scoped_ptr<CostFunctor> functor_;
256 };
257 
258 }  // namespace ceres
259 
260 #endif  // CERES_PUBLIC_DYNAMIC_AUTODIFF_COST_FUNCTION_H_
261