solver.h revision 0ae28bd5885b5daa526898fcf7c323dc2c3e1963
1// Ceres Solver - A fast non-linear least squares minimizer
2// Copyright 2010, 2011, 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#ifndef CERES_PUBLIC_SOLVER_H_
32#define CERES_PUBLIC_SOLVER_H_
33
34#include <cmath>
35#include <string>
36#include <vector>
37#include "ceres/crs_matrix.h"
38#include "ceres/internal/macros.h"
39#include "ceres/internal/port.h"
40#include "ceres/iteration_callback.h"
41#include "ceres/ordered_groups.h"
42#include "ceres/types.h"
43
44namespace ceres {
45
46class Problem;
47
48// Interface for non-linear least squares solvers.
49class Solver {
50 public:
51  virtual ~Solver();
52
53  // The options structure contains, not surprisingly, options that control how
54  // the solver operates. The defaults should be suitable for a wide range of
55  // problems; however, better performance is often obtainable with tweaking.
56  //
57  // The constants are defined inside types.h
58  struct Options {
59    // Default constructor that sets up a generic sparse problem.
60    Options() {
61      trust_region_strategy_type = LEVENBERG_MARQUARDT;
62      dogleg_type = TRADITIONAL_DOGLEG;
63      use_nonmonotonic_steps = false;
64      max_consecutive_nonmonotonic_steps = 5;
65      max_num_iterations = 50;
66      max_solver_time_in_seconds = 1e9;
67      num_threads = 1;
68      initial_trust_region_radius = 1e4;
69      max_trust_region_radius = 1e16;
70      min_trust_region_radius = 1e-32;
71      min_relative_decrease = 1e-3;
72      lm_min_diagonal = 1e-6;
73      lm_max_diagonal = 1e32;
74      max_num_consecutive_invalid_steps = 5;
75      function_tolerance = 1e-6;
76      gradient_tolerance = 1e-10;
77      parameter_tolerance = 1e-8;
78
79#if defined(CERES_NO_SUITESPARSE) && defined(CERES_NO_CXSPARSE)
80      linear_solver_type = DENSE_QR;
81#else
82      linear_solver_type = SPARSE_NORMAL_CHOLESKY;
83#endif
84
85      preconditioner_type = JACOBI;
86
87      sparse_linear_algebra_library = SUITE_SPARSE;
88#if defined(CERES_NO_SUITESPARSE) && !defined(CERES_NO_CXSPARSE)
89      sparse_linear_algebra_library = CX_SPARSE;
90#endif
91
92      num_linear_solver_threads = 1;
93
94#if defined(CERES_NO_SUITESPARSE)
95      use_block_amd = false;
96#else
97      use_block_amd = true;
98#endif
99      linear_solver_ordering = NULL;
100      use_inner_iterations = false;
101      inner_iteration_ordering = NULL;
102      linear_solver_min_num_iterations = 1;
103      linear_solver_max_num_iterations = 500;
104      eta = 1e-1;
105      jacobi_scaling = true;
106      logging_type = PER_MINIMIZER_ITERATION;
107      minimizer_progress_to_stdout = false;
108      return_initial_residuals = false;
109      return_initial_gradient = false;
110      return_initial_jacobian = false;
111      return_final_residuals = false;
112      return_final_gradient = false;
113      return_final_jacobian = false;
114      lsqp_dump_directory = "/tmp";
115      lsqp_dump_format_type = TEXTFILE;
116      check_gradients = false;
117      gradient_check_relative_precision = 1e-8;
118      numeric_derivative_relative_step_size = 1e-6;
119      update_state_every_iteration = false;
120    }
121
122    ~Options();
123    // Minimizer options ----------------------------------------
124
125    TrustRegionStrategyType trust_region_strategy_type;
126
127    // Type of dogleg strategy to use.
128    DoglegType dogleg_type;
129
130    // The classical trust region methods are descent methods, in that
131    // they only accept a point if it strictly reduces the value of
132    // the objective function.
133    //
134    // Relaxing this requirement allows the algorithm to be more
135    // efficient in the long term at the cost of some local increase
136    // in the value of the objective function.
137    //
138    // This is because allowing for non-decreasing objective function
139    // values in a princpled manner allows the algorithm to "jump over
140    // boulders" as the method is not restricted to move into narrow
141    // valleys while preserving its convergence properties.
142    //
143    // Setting use_nonmonotonic_steps to true enables the
144    // non-monotonic trust region algorithm as described by Conn,
145    // Gould & Toint in "Trust Region Methods", Section 10.1.
146    //
147    // The parameter max_consecutive_nonmonotonic_steps controls the
148    // window size used by the step selection algorithm to accept
149    // non-monotonic steps.
150    //
151    // Even though the value of the objective function may be larger
152    // than the minimum value encountered over the course of the
153    // optimization, the final parameters returned to the user are the
154    // ones corresponding to the minimum cost over all iterations.
155    bool use_nonmonotonic_steps;
156    int max_consecutive_nonmonotonic_steps;
157
158    // Maximum number of iterations for the minimizer to run for.
159    int max_num_iterations;
160
161    // Maximum time for which the minimizer should run for.
162    double max_solver_time_in_seconds;
163
164    // Number of threads used by Ceres for evaluating the cost and
165    // jacobians.
166    int num_threads;
167
168    // Trust region minimizer settings.
169    double initial_trust_region_radius;
170    double max_trust_region_radius;
171
172    // Minimizer terminates when the trust region radius becomes
173    // smaller than this value.
174    double min_trust_region_radius;
175
176    // Lower bound for the relative decrease before a step is
177    // accepted.
178    double min_relative_decrease;
179
180    // For the Levenberg-Marquadt algorithm, the scaled diagonal of
181    // the normal equations J'J is used to control the size of the
182    // trust region. Extremely small and large values along the
183    // diagonal can make this regularization scheme
184    // fail. lm_max_diagonal and lm_min_diagonal, clamp the values of
185    // diag(J'J) from above and below. In the normal course of
186    // operation, the user should not have to modify these parameters.
187    double lm_min_diagonal;
188    double lm_max_diagonal;
189
190    // Sometimes due to numerical conditioning problems or linear
191    // solver flakiness, the trust region strategy may return a
192    // numerically invalid step that can be fixed by reducing the
193    // trust region size. So the TrustRegionMinimizer allows for a few
194    // successive invalid steps before it declares NUMERICAL_FAILURE.
195    int max_num_consecutive_invalid_steps;
196
197    // Minimizer terminates when
198    //
199    //   (new_cost - old_cost) < function_tolerance * old_cost;
200    //
201    double function_tolerance;
202
203    // Minimizer terminates when
204    //
205    //   max_i |gradient_i| < gradient_tolerance * max_i|initial_gradient_i|
206    //
207    // This value should typically be 1e-4 * function_tolerance.
208    double gradient_tolerance;
209
210    // Minimizer terminates when
211    //
212    //   |step|_2 <= parameter_tolerance * ( |x|_2 +  parameter_tolerance)
213    //
214    double parameter_tolerance;
215
216    // Linear least squares solver options -------------------------------------
217
218    LinearSolverType linear_solver_type;
219
220    // Type of preconditioner to use with the iterative linear solvers.
221    PreconditionerType preconditioner_type;
222
223    // Ceres supports using multiple sparse linear algebra libraries
224    // for sparse matrix ordering and factorizations. Currently,
225    // SUITE_SPARSE and CX_SPARSE are the valid choices, depending on
226    // whether they are linked into Ceres at build time.
227    SparseLinearAlgebraLibraryType sparse_linear_algebra_library;
228
229    // Number of threads used by Ceres to solve the Newton
230    // step. Currently only the SPARSE_SCHUR solver is capable of
231    // using this setting.
232    int num_linear_solver_threads;
233
234    // The order in which variables are eliminated in a linear solver
235    // can have a significant of impact on the efficiency and accuracy
236    // of the method. e.g., when doing sparse Cholesky factorization,
237    // there are matrices for which a good ordering will give a
238    // Cholesky factor with O(n) storage, where as a bad ordering will
239    // result in an completely dense factor.
240    //
241    // Ceres allows the user to provide varying amounts of hints to
242    // the solver about the variable elimination ordering to use. This
243    // can range from no hints, where the solver is free to decide the
244    // best possible ordering based on the user's choices like the
245    // linear solver being used, to an exact order in which the
246    // variables should be eliminated, and a variety of possibilities
247    // in between.
248    //
249    // Instances of the ParameterBlockOrdering class are used to
250    // communicate this information to Ceres.
251    //
252    // Formally an ordering is an ordered partitioning of the
253    // parameter blocks, i.e, each parameter block belongs to exactly
254    // one group, and each group has a unique non-negative integer
255    // associated with it, that determines its order in the set of
256    // groups.
257    //
258    // Given such an ordering, Ceres ensures that the parameter blocks in
259    // the lowest numbered group are eliminated first, and then the
260    // parmeter blocks in the next lowest numbered group and so on. Within
261    // each group, Ceres is free to order the parameter blocks as it
262    // chooses.
263    //
264    // If NULL, then all parameter blocks are assumed to be in the
265    // same group and the solver is free to decide the best
266    // ordering.
267    //
268    // e.g. Consider the linear system
269    //
270    //   x + y = 3
271    //   2x + 3y = 7
272    //
273    // There are two ways in which it can be solved. First eliminating x
274    // from the two equations, solving for y and then back substituting
275    // for x, or first eliminating y, solving for x and back substituting
276    // for y. The user can construct three orderings here.
277    //
278    //   {0: x}, {1: y} - eliminate x first.
279    //   {0: y}, {1: x} - eliminate y first.
280    //   {0: x, y}      - Solver gets to decide the elimination order.
281    //
282    // Thus, to have Ceres determine the ordering automatically using
283    // heuristics, put all the variables in group 0 and to control the
284    // ordering for every variable, create groups 0..N-1, one per
285    // variable, in the desired order.
286    //
287    // Bundle Adjustment
288    // -----------------
289    //
290    // A particular case of interest is bundle adjustment, where the user
291    // has two options. The default is to not specify an ordering at all,
292    // the solver will see that the user wants to use a Schur type solver
293    // and figure out the right elimination ordering.
294    //
295    // But if the user already knows what parameter blocks are points and
296    // what are cameras, they can save preprocessing time by partitioning
297    // the parameter blocks into two groups, one for the points and one
298    // for the cameras, where the group containing the points has an id
299    // smaller than the group containing cameras.
300    //
301    // Once assigned, Solver::Options owns this pointer and will
302    // deallocate the memory when destroyed.
303    ParameterBlockOrdering* linear_solver_ordering;
304
305    // By virtue of the modeling layer in Ceres being block oriented,
306    // all the matrices used by Ceres are also block oriented. When
307    // doing sparse direct factorization of these matrices (for
308    // SPARSE_NORMAL_CHOLESKY, SPARSE_SCHUR and ITERATIVE in
309    // conjunction with CLUSTER_TRIDIAGONAL AND CLUSTER_JACOBI
310    // preconditioners), the fill-reducing ordering algorithms can
311    // either be run on the block or the scalar form of these matrices.
312    // Running it on the block form exposes more of the super-nodal
313    // structure of the matrix to the factorization routines. Setting
314    // this parameter to true runs the ordering algorithms in block
315    // form. Currently this option only makes sense with
316    // sparse_linear_algebra_library = SUITE_SPARSE.
317    bool use_block_amd;
318
319    // Some non-linear least squares problems have additional
320    // structure in the way the parameter blocks interact that it is
321    // beneficial to modify the way the trust region step is computed.
322    //
323    // e.g., consider the following regression problem
324    //
325    //   y = a_1 exp(b_1 x) + a_2 exp(b_3 x^2 + c_1)
326    //
327    // Given a set of pairs{(x_i, y_i)}, the user wishes to estimate
328    // a_1, a_2, b_1, b_2, and c_1.
329    //
330    // Notice here that the expression on the left is linear in a_1
331    // and a_2, and given any value for b_1, b_2 and c_1, it is
332    // possible to use linear regression to estimate the optimal
333    // values of a_1 and a_2. Indeed, its possible to analytically
334    // eliminate the variables a_1 and a_2 from the problem all
335    // together. Problems like these are known as separable least
336    // squares problem and the most famous algorithm for solving them
337    // is the Variable Projection algorithm invented by Golub &
338    // Pereyra.
339    //
340    // Similar structure can be found in the matrix factorization with
341    // missing data problem. There the corresponding algorithm is
342    // known as Wiberg's algorithm.
343    //
344    // Ruhe & Wedin (Algorithms for Separable Nonlinear Least Squares
345    // Problems, SIAM Reviews, 22(3), 1980) present an analyis of
346    // various algorithms for solving separable non-linear least
347    // squares problems and refer to "Variable Projection" as
348    // Algorithm I in their paper.
349    //
350    // Implementing Variable Projection is tedious and expensive, and
351    // they present a simpler algorithm, which they refer to as
352    // Algorithm II, where once the Newton/Trust Region step has been
353    // computed for the whole problem (a_1, a_2, b_1, b_2, c_1) and
354    // additional optimization step is performed to estimate a_1 and
355    // a_2 exactly.
356    //
357    // This idea can be generalized to cases where the residual is not
358    // linear in a_1 and a_2, i.e., Solve for the trust region step
359    // for the full problem, and then use it as the starting point to
360    // further optimize just a_1 and a_2. For the linear case, this
361    // amounts to doing a single linear least squares solve. For
362    // non-linear problems, any method for solving the a_1 and a_2
363    // optimization problems will do. The only constraint on a_1 and
364    // a_2 is that they do not co-occur in any residual block.
365    //
366    // This idea can be further generalized, by not just optimizing
367    // (a_1, a_2), but decomposing the graph corresponding to the
368    // Hessian matrix's sparsity structure in a collection of
369    // non-overlapping independent sets and optimizing each of them.
370    //
371    // Setting "use_inner_iterations" to true enables the use of this
372    // non-linear generalization of Ruhe & Wedin's Algorithm II.  This
373    // version of Ceres has a higher iteration complexity, but also
374    // displays better convergence behaviour per iteration. Setting
375    // Solver::Options::num_threads to the maximum number possible is
376    // highly recommended.
377    bool use_inner_iterations;
378
379    // If inner_iterations is true, then the user has two choices.
380    //
381    // 1. Let the solver heuristically decide which parameter blocks
382    //    to optimize in each inner iteration. To do this leave
383    //    Solver::Options::inner_iteration_ordering untouched.
384    //
385    // 2. Specify a collection of of ordered independent sets. Where
386    //    the lower numbered groups are optimized before the higher
387    //    number groups. Each group must be an independent set.
388    ParameterBlockOrdering* inner_iteration_ordering;
389
390    // Minimum number of iterations for which the linear solver should
391    // run, even if the convergence criterion is satisfied.
392    int linear_solver_min_num_iterations;
393
394    // Maximum number of iterations for which the linear solver should
395    // run. If the solver does not converge in less than
396    // linear_solver_max_num_iterations, then it returns
397    // MAX_ITERATIONS, as its termination type.
398    int linear_solver_max_num_iterations;
399
400    // Forcing sequence parameter. The truncated Newton solver uses
401    // this number to control the relative accuracy with which the
402    // Newton step is computed.
403    //
404    // This constant is passed to ConjugateGradientsSolver which uses
405    // it to terminate the iterations when
406    //
407    //  (Q_i - Q_{i-1})/Q_i < eta/i
408    double eta;
409
410    // Normalize the jacobian using Jacobi scaling before calling
411    // the linear least squares solver.
412    bool jacobi_scaling;
413
414    // Logging options ---------------------------------------------------------
415
416    LoggingType logging_type;
417
418    // By default the Minimizer progress is logged to VLOG(1), which
419    // is sent to STDERR depending on the vlog level. If this flag is
420    // set to true, and logging_type is not SILENT, the logging output
421    // is sent to STDOUT.
422    bool minimizer_progress_to_stdout;
423
424    bool return_initial_residuals;
425    bool return_initial_gradient;
426    bool return_initial_jacobian;
427
428    bool return_final_residuals;
429    bool return_final_gradient;
430    bool return_final_jacobian;
431
432    // List of iterations at which the optimizer should dump the
433    // linear least squares problem to disk. Useful for testing and
434    // benchmarking. If empty (default), no problems are dumped.
435    //
436    // This is ignored if protocol buffers are disabled.
437    vector<int> lsqp_iterations_to_dump;
438    string lsqp_dump_directory;
439    DumpFormatType lsqp_dump_format_type;
440
441    // Finite differences options ----------------------------------------------
442
443    // Check all jacobians computed by each residual block with finite
444    // differences. This is expensive since it involves computing the
445    // derivative by normal means (e.g. user specified, autodiff,
446    // etc), then also computing it using finite differences. The
447    // results are compared, and if they differ substantially, details
448    // are printed to the log.
449    bool check_gradients;
450
451    // Relative precision to check for in the gradient checker. If the
452    // relative difference between an element in a jacobian exceeds
453    // this number, then the jacobian for that cost term is dumped.
454    double gradient_check_relative_precision;
455
456    // Relative shift used for taking numeric derivatives. For finite
457    // differencing, each dimension is evaluated at slightly shifted
458    // values; for the case of central difference, this is what gets
459    // evaluated:
460    //
461    //   delta = numeric_derivative_relative_step_size;
462    //   f_initial  = f(x)
463    //   f_forward  = f((1 + delta) * x)
464    //   f_backward = f((1 - delta) * x)
465    //
466    // The finite differencing is done along each dimension. The
467    // reason to use a relative (rather than absolute) step size is
468    // that this way, numeric differentation works for functions where
469    // the arguments are typically large (e.g. 1e9) and when the
470    // values are small (e.g. 1e-5). It is possible to construct
471    // "torture cases" which break this finite difference heuristic,
472    // but they do not come up often in practice.
473    //
474    // TODO(keir): Pick a smarter number than the default above! In
475    // theory a good choice is sqrt(eps) * x, which for doubles means
476    // about 1e-8 * x. However, I have found this number too
477    // optimistic. This number should be exposed for users to change.
478    double numeric_derivative_relative_step_size;
479
480    // If true, the user's parameter blocks are updated at the end of
481    // every Minimizer iteration, otherwise they are updated when the
482    // Minimizer terminates. This is useful if, for example, the user
483    // wishes to visualize the state of the optimization every
484    // iteration.
485    bool update_state_every_iteration;
486
487    // Callbacks that are executed at the end of each iteration of the
488    // Minimizer. An iteration may terminate midway, either due to
489    // numerical failures or because one of the convergence tests has
490    // been satisfied. In this case none of the callbacks are
491    // executed.
492
493    // Callbacks are executed in the order that they are specified in
494    // this vector. By default, parameter blocks are updated only at
495    // the end of the optimization, i.e when the Minimizer
496    // terminates. This behaviour is controlled by
497    // update_state_every_variable. If the user wishes to have access
498    // to the update parameter blocks when his/her callbacks are
499    // executed, then set update_state_every_iteration to true.
500    //
501    // The solver does NOT take ownership of these pointers.
502    vector<IterationCallback*> callbacks;
503
504    // If non-empty, a summary of the execution of the solver is
505    // recorded to this file.
506    string solver_log;
507  };
508
509  struct Summary {
510    Summary();
511
512    // A brief one line description of the state of the solver after
513    // termination.
514    string BriefReport() const;
515
516    // A full multiline description of the state of the solver after
517    // termination.
518    string FullReport() const;
519
520    // Minimizer summary -------------------------------------------------
521    SolverTerminationType termination_type;
522
523    // If the solver did not run, or there was a failure, a
524    // description of the error.
525    string error;
526
527    // Cost of the problem before and after the optimization. See
528    // problem.h for definition of the cost of a problem.
529    double initial_cost;
530    double final_cost;
531
532    // The part of the total cost that comes from residual blocks that
533    // were held fixed by the preprocessor because all the parameter
534    // blocks that they depend on were fixed.
535    double fixed_cost;
536
537    // Vectors of residuals before and after the optimization. The
538    // entries of these vectors are in the order in which
539    // ResidualBlocks were added to the Problem object.
540    //
541    // Whether the residual vectors are populated with values is
542    // controlled by Solver::Options::return_initial_residuals and
543    // Solver::Options::return_final_residuals respectively.
544    vector<double> initial_residuals;
545    vector<double> final_residuals;
546
547    // Gradient vectors, before and after the optimization.  The rows
548    // are in the same order in which the ParameterBlocks were added
549    // to the Problem object.
550    //
551    // NOTE: Since AddResidualBlock adds ParameterBlocks to the
552    // Problem automatically if they do not already exist, if you wish
553    // to have explicit control over the ordering of the vectors, then
554    // use Problem::AddParameterBlock to explicitly add the
555    // ParameterBlocks in the order desired.
556    //
557    // Whether the vectors are populated with values is controlled by
558    // Solver::Options::return_initial_gradient and
559    // Solver::Options::return_final_gradient respectively.
560    vector<double> initial_gradient;
561    vector<double> final_gradient;
562
563    // Jacobian matrices before and after the optimization. The rows
564    // of these matrices are in the same order in which the
565    // ResidualBlocks were added to the Problem object. The columns
566    // are in the same order in which the ParameterBlocks were added
567    // to the Problem object.
568    //
569    // NOTE: Since AddResidualBlock adds ParameterBlocks to the
570    // Problem automatically if they do not already exist, if you wish
571    // to have explicit control over the column ordering of the
572    // matrix, then use Problem::AddParameterBlock to explicitly add
573    // the ParameterBlocks in the order desired.
574    //
575    // The Jacobian matrices are stored as compressed row sparse
576    // matrices. Please see ceres/crs_matrix.h for more details of the
577    // format.
578    //
579    // Whether the Jacboan matrices are populated with values is
580    // controlled by Solver::Options::return_initial_jacobian and
581    // Solver::Options::return_final_jacobian respectively.
582    CRSMatrix initial_jacobian;
583    CRSMatrix final_jacobian;
584
585    vector<IterationSummary> iterations;
586
587    int num_successful_steps;
588    int num_unsuccessful_steps;
589
590    // When the user calls Solve, before the actual optimization
591    // occurs, Ceres performs a number of preprocessing steps. These
592    // include error checks, memory allocations, and reorderings. This
593    // time is accounted for as preprocessing time.
594    double preprocessor_time_in_seconds;
595
596    // Time spent in the TrustRegionMinimizer.
597    double minimizer_time_in_seconds;
598
599    // After the Minimizer is finished, some time is spent in
600    // re-evaluating residuals etc. This time is accounted for in the
601    // postprocessor time.
602    double postprocessor_time_in_seconds;
603
604    // Some total of all time spent inside Ceres when Solve is called.
605    double total_time_in_seconds;
606
607    // Preprocessor summary.
608    int num_parameter_blocks;
609    int num_parameters;
610    int num_residual_blocks;
611    int num_residuals;
612
613    int num_parameter_blocks_reduced;
614    int num_parameters_reduced;
615    int num_residual_blocks_reduced;
616    int num_residuals_reduced;
617
618    int num_eliminate_blocks_given;
619    int num_eliminate_blocks_used;
620
621    int num_threads_given;
622    int num_threads_used;
623
624    int num_linear_solver_threads_given;
625    int num_linear_solver_threads_used;
626
627    LinearSolverType linear_solver_type_given;
628    LinearSolverType linear_solver_type_used;
629
630    PreconditionerType preconditioner_type;
631
632    TrustRegionStrategyType trust_region_strategy_type;
633    DoglegType dogleg_type;
634    SparseLinearAlgebraLibraryType sparse_linear_algebra_library;
635  };
636
637  // Once a least squares problem has been built, this function takes
638  // the problem and optimizes it based on the values of the options
639  // parameters. Upon return, a detailed summary of the work performed
640  // by the preprocessor, the non-linear minmizer and the linear
641  // solver are reported in the summary object.
642  virtual void Solve(const Options& options,
643                     Problem* problem,
644                     Solver::Summary* summary);
645};
646
647// Helper function which avoids going through the interface.
648void Solve(const Solver::Options& options,
649           Problem* problem,
650           Solver::Summary* summary);
651
652}  // namespace ceres
653
654#endif  // CERES_PUBLIC_SOLVER_H_
655