// Ceres Solver - A fast non-linear least squares minimizer // Copyright 2010, 2011, 2012 Google Inc. All rights reserved. // http://code.google.com/p/ceres-solver/ // // Redistribution and use in source and binary forms, with or without // modification, are permitted provided that the following conditions are met: // // * Redistributions of source code must retain the above copyright notice, // this list of conditions and the following disclaimer. // * Redistributions in binary form must reproduce the above copyright notice, // this list of conditions and the following disclaimer in the documentation // and/or other materials provided with the distribution. // * Neither the name of Google Inc. nor the names of its contributors may be // used to endorse or promote products derived from this software without // specific prior written permission. // // THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" // AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE // IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE // ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE // LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR // CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF // SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS // INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN // CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) // ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE // POSSIBILITY OF SUCH DAMAGE. // // Author: sameeragarwal@google.com (Sameer Agarwal) #ifndef CERES_PUBLIC_SOLVER_H_ #define CERES_PUBLIC_SOLVER_H_ #include <cmath> #include <string> #include <vector> #include "ceres/crs_matrix.h" #include "ceres/internal/macros.h" #include "ceres/internal/port.h" #include "ceres/iteration_callback.h" #include "ceres/ordered_groups.h" #include "ceres/types.h" namespace ceres { class Problem; // Interface for non-linear least squares solvers. class Solver { public: virtual ~Solver(); // The options structure contains, not surprisingly, options that control how // the solver operates. The defaults should be suitable for a wide range of // problems; however, better performance is often obtainable with tweaking. // // The constants are defined inside types.h struct Options { // Default constructor that sets up a generic sparse problem. Options() { trust_region_strategy_type = LEVENBERG_MARQUARDT; dogleg_type = TRADITIONAL_DOGLEG; use_nonmonotonic_steps = false; max_consecutive_nonmonotonic_steps = 5; max_num_iterations = 50; max_solver_time_in_seconds = 1e9; num_threads = 1; initial_trust_region_radius = 1e4; max_trust_region_radius = 1e16; min_trust_region_radius = 1e-32; min_relative_decrease = 1e-3; lm_min_diagonal = 1e-6; lm_max_diagonal = 1e32; max_num_consecutive_invalid_steps = 5; function_tolerance = 1e-6; gradient_tolerance = 1e-10; parameter_tolerance = 1e-8; #if defined(CERES_NO_SUITESPARSE) && defined(CERES_NO_CXSPARSE) linear_solver_type = DENSE_QR; #else linear_solver_type = SPARSE_NORMAL_CHOLESKY; #endif preconditioner_type = JACOBI; sparse_linear_algebra_library = SUITE_SPARSE; #if defined(CERES_NO_SUITESPARSE) && !defined(CERES_NO_CXSPARSE) sparse_linear_algebra_library = CX_SPARSE; #endif num_linear_solver_threads = 1; #if defined(CERES_NO_SUITESPARSE) use_block_amd = false; #else use_block_amd = true; #endif linear_solver_ordering = NULL; use_inner_iterations = false; inner_iteration_ordering = NULL; linear_solver_min_num_iterations = 1; linear_solver_max_num_iterations = 500; eta = 1e-1; jacobi_scaling = true; logging_type = PER_MINIMIZER_ITERATION; minimizer_progress_to_stdout = false; return_initial_residuals = false; return_initial_gradient = false; return_initial_jacobian = false; return_final_residuals = false; return_final_gradient = false; return_final_jacobian = false; lsqp_dump_directory = "/tmp"; lsqp_dump_format_type = TEXTFILE; check_gradients = false; gradient_check_relative_precision = 1e-8; numeric_derivative_relative_step_size = 1e-6; update_state_every_iteration = false; } ~Options(); // Minimizer options ---------------------------------------- TrustRegionStrategyType trust_region_strategy_type; // Type of dogleg strategy to use. DoglegType dogleg_type; // The classical trust region methods are descent methods, in that // they only accept a point if it strictly reduces the value of // the objective function. // // Relaxing this requirement allows the algorithm to be more // efficient in the long term at the cost of some local increase // in the value of the objective function. // // This is because allowing for non-decreasing objective function // values in a princpled manner allows the algorithm to "jump over // boulders" as the method is not restricted to move into narrow // valleys while preserving its convergence properties. // // Setting use_nonmonotonic_steps to true enables the // non-monotonic trust region algorithm as described by Conn, // Gould & Toint in "Trust Region Methods", Section 10.1. // // The parameter max_consecutive_nonmonotonic_steps controls the // window size used by the step selection algorithm to accept // non-monotonic steps. // // Even though the value of the objective function may be larger // than the minimum value encountered over the course of the // optimization, the final parameters returned to the user are the // ones corresponding to the minimum cost over all iterations. bool use_nonmonotonic_steps; int max_consecutive_nonmonotonic_steps; // Maximum number of iterations for the minimizer to run for. int max_num_iterations; // Maximum time for which the minimizer should run for. double max_solver_time_in_seconds; // Number of threads used by Ceres for evaluating the cost and // jacobians. int num_threads; // Trust region minimizer settings. double initial_trust_region_radius; double max_trust_region_radius; // Minimizer terminates when the trust region radius becomes // smaller than this value. double min_trust_region_radius; // Lower bound for the relative decrease before a step is // accepted. double min_relative_decrease; // For the Levenberg-Marquadt algorithm, the scaled diagonal of // the normal equations J'J is used to control the size of the // trust region. Extremely small and large values along the // diagonal can make this regularization scheme // fail. lm_max_diagonal and lm_min_diagonal, clamp the values of // diag(J'J) from above and below. In the normal course of // operation, the user should not have to modify these parameters. double lm_min_diagonal; double lm_max_diagonal; // Sometimes due to numerical conditioning problems or linear // solver flakiness, the trust region strategy may return a // numerically invalid step that can be fixed by reducing the // trust region size. So the TrustRegionMinimizer allows for a few // successive invalid steps before it declares NUMERICAL_FAILURE. int max_num_consecutive_invalid_steps; // Minimizer terminates when // // (new_cost - old_cost) < function_tolerance * old_cost; // double function_tolerance; // Minimizer terminates when // // max_i |gradient_i| < gradient_tolerance * max_i|initial_gradient_i| // // This value should typically be 1e-4 * function_tolerance. double gradient_tolerance; // Minimizer terminates when // // |step|_2 <= parameter_tolerance * ( |x|_2 + parameter_tolerance) // double parameter_tolerance; // Linear least squares solver options ------------------------------------- LinearSolverType linear_solver_type; // Type of preconditioner to use with the iterative linear solvers. PreconditionerType preconditioner_type; // Ceres supports using multiple sparse linear algebra libraries // for sparse matrix ordering and factorizations. Currently, // SUITE_SPARSE and CX_SPARSE are the valid choices, depending on // whether they are linked into Ceres at build time. SparseLinearAlgebraLibraryType sparse_linear_algebra_library; // Number of threads used by Ceres to solve the Newton // step. Currently only the SPARSE_SCHUR solver is capable of // using this setting. int num_linear_solver_threads; // The order in which variables are eliminated in a linear solver // can have a significant of impact on the efficiency and accuracy // of the method. e.g., when doing sparse Cholesky factorization, // there are matrices for which a good ordering will give a // Cholesky factor with O(n) storage, where as a bad ordering will // result in an completely dense factor. // // Ceres allows the user to provide varying amounts of hints to // the solver about the variable elimination ordering to use. This // can range from no hints, where the solver is free to decide the // best possible ordering based on the user's choices like the // linear solver being used, to an exact order in which the // variables should be eliminated, and a variety of possibilities // in between. // // Instances of the ParameterBlockOrdering class are used to // communicate this information to Ceres. // // Formally an ordering is an ordered partitioning of the // parameter blocks, i.e, each parameter block belongs to exactly // one group, and each group has a unique non-negative integer // associated with it, that determines its order in the set of // groups. // // Given such an ordering, Ceres ensures that the parameter blocks in // the lowest numbered group are eliminated first, and then the // parmeter blocks in the next lowest numbered group and so on. Within // each group, Ceres is free to order the parameter blocks as it // chooses. // // If NULL, then all parameter blocks are assumed to be in the // same group and the solver is free to decide the best // ordering. // // e.g. Consider the linear system // // x + y = 3 // 2x + 3y = 7 // // There are two ways in which it can be solved. First eliminating x // from the two equations, solving for y and then back substituting // for x, or first eliminating y, solving for x and back substituting // for y. The user can construct three orderings here. // // {0: x}, {1: y} - eliminate x first. // {0: y}, {1: x} - eliminate y first. // {0: x, y} - Solver gets to decide the elimination order. // // Thus, to have Ceres determine the ordering automatically using // heuristics, put all the variables in group 0 and to control the // ordering for every variable, create groups 0..N-1, one per // variable, in the desired order. // // Bundle Adjustment // ----------------- // // A particular case of interest is bundle adjustment, where the user // has two options. The default is to not specify an ordering at all, // the solver will see that the user wants to use a Schur type solver // and figure out the right elimination ordering. // // But if the user already knows what parameter blocks are points and // what are cameras, they can save preprocessing time by partitioning // the parameter blocks into two groups, one for the points and one // for the cameras, where the group containing the points has an id // smaller than the group containing cameras. // // Once assigned, Solver::Options owns this pointer and will // deallocate the memory when destroyed. ParameterBlockOrdering* linear_solver_ordering; // By virtue of the modeling layer in Ceres being block oriented, // all the matrices used by Ceres are also block oriented. When // doing sparse direct factorization of these matrices (for // SPARSE_NORMAL_CHOLESKY, SPARSE_SCHUR and ITERATIVE in // conjunction with CLUSTER_TRIDIAGONAL AND CLUSTER_JACOBI // preconditioners), the fill-reducing ordering algorithms can // either be run on the block or the scalar form of these matrices. // Running it on the block form exposes more of the super-nodal // structure of the matrix to the factorization routines. Setting // this parameter to true runs the ordering algorithms in block // form. Currently this option only makes sense with // sparse_linear_algebra_library = SUITE_SPARSE. bool use_block_amd; // Some non-linear least squares problems have additional // structure in the way the parameter blocks interact that it is // beneficial to modify the way the trust region step is computed. // // e.g., consider the following regression problem // // y = a_1 exp(b_1 x) + a_2 exp(b_3 x^2 + c_1) // // Given a set of pairs{(x_i, y_i)}, the user wishes to estimate // a_1, a_2, b_1, b_2, and c_1. // // Notice here that the expression on the left is linear in a_1 // and a_2, and given any value for b_1, b_2 and c_1, it is // possible to use linear regression to estimate the optimal // values of a_1 and a_2. Indeed, its possible to analytically // eliminate the variables a_1 and a_2 from the problem all // together. Problems like these are known as separable least // squares problem and the most famous algorithm for solving them // is the Variable Projection algorithm invented by Golub & // Pereyra. // // Similar structure can be found in the matrix factorization with // missing data problem. There the corresponding algorithm is // known as Wiberg's algorithm. // // Ruhe & Wedin (Algorithms for Separable Nonlinear Least Squares // Problems, SIAM Reviews, 22(3), 1980) present an analyis of // various algorithms for solving separable non-linear least // squares problems and refer to "Variable Projection" as // Algorithm I in their paper. // // Implementing Variable Projection is tedious and expensive, and // they present a simpler algorithm, which they refer to as // Algorithm II, where once the Newton/Trust Region step has been // computed for the whole problem (a_1, a_2, b_1, b_2, c_1) and // additional optimization step is performed to estimate a_1 and // a_2 exactly. // // This idea can be generalized to cases where the residual is not // linear in a_1 and a_2, i.e., Solve for the trust region step // for the full problem, and then use it as the starting point to // further optimize just a_1 and a_2. For the linear case, this // amounts to doing a single linear least squares solve. For // non-linear problems, any method for solving the a_1 and a_2 // optimization problems will do. The only constraint on a_1 and // a_2 is that they do not co-occur in any residual block. // // This idea can be further generalized, by not just optimizing // (a_1, a_2), but decomposing the graph corresponding to the // Hessian matrix's sparsity structure in a collection of // non-overlapping independent sets and optimizing each of them. // // Setting "use_inner_iterations" to true enables the use of this // non-linear generalization of Ruhe & Wedin's Algorithm II. This // version of Ceres has a higher iteration complexity, but also // displays better convergence behaviour per iteration. Setting // Solver::Options::num_threads to the maximum number possible is // highly recommended. bool use_inner_iterations; // If inner_iterations is true, then the user has two choices. // // 1. Let the solver heuristically decide which parameter blocks // to optimize in each inner iteration. To do this leave // Solver::Options::inner_iteration_ordering untouched. // // 2. Specify a collection of of ordered independent sets. Where // the lower numbered groups are optimized before the higher // number groups. Each group must be an independent set. ParameterBlockOrdering* inner_iteration_ordering; // Minimum number of iterations for which the linear solver should // run, even if the convergence criterion is satisfied. int linear_solver_min_num_iterations; // Maximum number of iterations for which the linear solver should // run. If the solver does not converge in less than // linear_solver_max_num_iterations, then it returns // MAX_ITERATIONS, as its termination type. int linear_solver_max_num_iterations; // Forcing sequence parameter. The truncated Newton solver uses // this number to control the relative accuracy with which the // Newton step is computed. // // This constant is passed to ConjugateGradientsSolver which uses // it to terminate the iterations when // // (Q_i - Q_{i-1})/Q_i < eta/i double eta; // Normalize the jacobian using Jacobi scaling before calling // the linear least squares solver. bool jacobi_scaling; // Logging options --------------------------------------------------------- LoggingType logging_type; // By default the Minimizer progress is logged to VLOG(1), which // is sent to STDERR depending on the vlog level. If this flag is // set to true, and logging_type is not SILENT, the logging output // is sent to STDOUT. bool minimizer_progress_to_stdout; bool return_initial_residuals; bool return_initial_gradient; bool return_initial_jacobian; bool return_final_residuals; bool return_final_gradient; bool return_final_jacobian; // List of iterations at which the optimizer should dump the // linear least squares problem to disk. Useful for testing and // benchmarking. If empty (default), no problems are dumped. // // This is ignored if protocol buffers are disabled. vector<int> lsqp_iterations_to_dump; string lsqp_dump_directory; DumpFormatType lsqp_dump_format_type; // Finite differences options ---------------------------------------------- // Check all jacobians computed by each residual block with finite // differences. This is expensive since it involves computing the // derivative by normal means (e.g. user specified, autodiff, // etc), then also computing it using finite differences. The // results are compared, and if they differ substantially, details // are printed to the log. bool check_gradients; // Relative precision to check for in the gradient checker. If the // relative difference between an element in a jacobian exceeds // this number, then the jacobian for that cost term is dumped. double gradient_check_relative_precision; // Relative shift used for taking numeric derivatives. For finite // differencing, each dimension is evaluated at slightly shifted // values; for the case of central difference, this is what gets // evaluated: // // delta = numeric_derivative_relative_step_size; // f_initial = f(x) // f_forward = f((1 + delta) * x) // f_backward = f((1 - delta) * x) // // The finite differencing is done along each dimension. The // reason to use a relative (rather than absolute) step size is // that this way, numeric differentation works for functions where // the arguments are typically large (e.g. 1e9) and when the // values are small (e.g. 1e-5). It is possible to construct // "torture cases" which break this finite difference heuristic, // but they do not come up often in practice. // // TODO(keir): Pick a smarter number than the default above! In // theory a good choice is sqrt(eps) * x, which for doubles means // about 1e-8 * x. However, I have found this number too // optimistic. This number should be exposed for users to change. double numeric_derivative_relative_step_size; // If true, the user's parameter blocks are updated at the end of // every Minimizer iteration, otherwise they are updated when the // Minimizer terminates. This is useful if, for example, the user // wishes to visualize the state of the optimization every // iteration. bool update_state_every_iteration; // Callbacks that are executed at the end of each iteration of the // Minimizer. An iteration may terminate midway, either due to // numerical failures or because one of the convergence tests has // been satisfied. In this case none of the callbacks are // executed. // Callbacks are executed in the order that they are specified in // this vector. By default, parameter blocks are updated only at // the end of the optimization, i.e when the Minimizer // terminates. This behaviour is controlled by // update_state_every_variable. If the user wishes to have access // to the update parameter blocks when his/her callbacks are // executed, then set update_state_every_iteration to true. // // The solver does NOT take ownership of these pointers. vector<IterationCallback*> callbacks; // If non-empty, a summary of the execution of the solver is // recorded to this file. string solver_log; }; struct Summary { Summary(); // A brief one line description of the state of the solver after // termination. string BriefReport() const; // A full multiline description of the state of the solver after // termination. string FullReport() const; // Minimizer summary ------------------------------------------------- SolverTerminationType termination_type; // If the solver did not run, or there was a failure, a // description of the error. string error; // Cost of the problem before and after the optimization. See // problem.h for definition of the cost of a problem. double initial_cost; double final_cost; // The part of the total cost that comes from residual blocks that // were held fixed by the preprocessor because all the parameter // blocks that they depend on were fixed. double fixed_cost; // Vectors of residuals before and after the optimization. The // entries of these vectors are in the order in which // ResidualBlocks were added to the Problem object. // // Whether the residual vectors are populated with values is // controlled by Solver::Options::return_initial_residuals and // Solver::Options::return_final_residuals respectively. vector<double> initial_residuals; vector<double> final_residuals; // Gradient vectors, before and after the optimization. The rows // are in the same order in which the ParameterBlocks were added // to the Problem object. // // NOTE: Since AddResidualBlock adds ParameterBlocks to the // Problem automatically if they do not already exist, if you wish // to have explicit control over the ordering of the vectors, then // use Problem::AddParameterBlock to explicitly add the // ParameterBlocks in the order desired. // // Whether the vectors are populated with values is controlled by // Solver::Options::return_initial_gradient and // Solver::Options::return_final_gradient respectively. vector<double> initial_gradient; vector<double> final_gradient; // Jacobian matrices before and after the optimization. The rows // of these matrices are in the same order in which the // ResidualBlocks were added to the Problem object. The columns // are in the same order in which the ParameterBlocks were added // to the Problem object. // // NOTE: Since AddResidualBlock adds ParameterBlocks to the // Problem automatically if they do not already exist, if you wish // to have explicit control over the column ordering of the // matrix, then use Problem::AddParameterBlock to explicitly add // the ParameterBlocks in the order desired. // // The Jacobian matrices are stored as compressed row sparse // matrices. Please see ceres/crs_matrix.h for more details of the // format. // // Whether the Jacboan matrices are populated with values is // controlled by Solver::Options::return_initial_jacobian and // Solver::Options::return_final_jacobian respectively. CRSMatrix initial_jacobian; CRSMatrix final_jacobian; vector<IterationSummary> iterations; int num_successful_steps; int num_unsuccessful_steps; // When the user calls Solve, before the actual optimization // occurs, Ceres performs a number of preprocessing steps. These // include error checks, memory allocations, and reorderings. This // time is accounted for as preprocessing time. double preprocessor_time_in_seconds; // Time spent in the TrustRegionMinimizer. double minimizer_time_in_seconds; // After the Minimizer is finished, some time is spent in // re-evaluating residuals etc. This time is accounted for in the // postprocessor time. double postprocessor_time_in_seconds; // Some total of all time spent inside Ceres when Solve is called. double total_time_in_seconds; // Preprocessor summary. int num_parameter_blocks; int num_parameters; int num_residual_blocks; int num_residuals; int num_parameter_blocks_reduced; int num_parameters_reduced; int num_residual_blocks_reduced; int num_residuals_reduced; int num_eliminate_blocks_given; int num_eliminate_blocks_used; int num_threads_given; int num_threads_used; int num_linear_solver_threads_given; int num_linear_solver_threads_used; LinearSolverType linear_solver_type_given; LinearSolverType linear_solver_type_used; PreconditionerType preconditioner_type; TrustRegionStrategyType trust_region_strategy_type; DoglegType dogleg_type; SparseLinearAlgebraLibraryType sparse_linear_algebra_library; }; // Once a least squares problem has been built, this function takes // the problem and optimizes it based on the values of the options // parameters. Upon return, a detailed summary of the work performed // by the preprocessor, the non-linear minmizer and the linear // solver are reported in the summary object. virtual void Solve(const Options& options, Problem* problem, Solver::Summary* summary); }; // Helper function which avoids going through the interface. void Solve(const Solver::Options& options, Problem* problem, Solver::Summary* summary); } // namespace ceres #endif // CERES_PUBLIC_SOLVER_H_