// Ceres Solver - A fast non-linear least squares minimizer // Copyright 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) #include "ceres/dogleg_strategy.h" #include <cmath> #include "Eigen/Dense" #include "ceres/array_utils.h" #include "ceres/internal/eigen.h" #include "ceres/linear_solver.h" #include "ceres/polynomial_solver.h" #include "ceres/sparse_matrix.h" #include "ceres/trust_region_strategy.h" #include "ceres/types.h" #include "glog/logging.h" namespace ceres { namespace internal { namespace { const double kMaxMu = 1.0; const double kMinMu = 1e-8; } DoglegStrategy::DoglegStrategy(const TrustRegionStrategy::Options& options) : linear_solver_(options.linear_solver), radius_(options.initial_radius), max_radius_(options.max_radius), min_diagonal_(options.lm_min_diagonal), max_diagonal_(options.lm_max_diagonal), mu_(kMinMu), min_mu_(kMinMu), max_mu_(kMaxMu), mu_increase_factor_(10.0), increase_threshold_(0.75), decrease_threshold_(0.25), dogleg_step_norm_(0.0), reuse_(false), dogleg_type_(options.dogleg_type) { CHECK_NOTNULL(linear_solver_); CHECK_GT(min_diagonal_, 0.0); CHECK_LE(min_diagonal_, max_diagonal_); CHECK_GT(max_radius_, 0.0); } // If the reuse_ flag is not set, then the Cauchy point (scaled // gradient) and the new Gauss-Newton step are computed from // scratch. The Dogleg step is then computed as interpolation of these // two vectors. TrustRegionStrategy::Summary DoglegStrategy::ComputeStep( const TrustRegionStrategy::PerSolveOptions& per_solve_options, SparseMatrix* jacobian, const double* residuals, double* step) { CHECK_NOTNULL(jacobian); CHECK_NOTNULL(residuals); CHECK_NOTNULL(step); const int n = jacobian->num_cols(); if (reuse_) { // Gauss-Newton and gradient vectors are always available, only a // new interpolant need to be computed. For the subspace case, // the subspace and the two-dimensional model are also still valid. switch(dogleg_type_) { case TRADITIONAL_DOGLEG: ComputeTraditionalDoglegStep(step); break; case SUBSPACE_DOGLEG: ComputeSubspaceDoglegStep(step); break; } TrustRegionStrategy::Summary summary; summary.num_iterations = 0; summary.termination_type = TOLERANCE; return summary; } reuse_ = true; // Check that we have the storage needed to hold the various // temporary vectors. if (diagonal_.rows() != n) { diagonal_.resize(n, 1); gradient_.resize(n, 1); gauss_newton_step_.resize(n, 1); } // Vector used to form the diagonal matrix that is used to // regularize the Gauss-Newton solve and that defines the // elliptical trust region // // || D * step || <= radius_ . // jacobian->SquaredColumnNorm(diagonal_.data()); for (int i = 0; i < n; ++i) { diagonal_[i] = min(max(diagonal_[i], min_diagonal_), max_diagonal_); } diagonal_ = diagonal_.array().sqrt(); ComputeGradient(jacobian, residuals); ComputeCauchyPoint(jacobian); LinearSolver::Summary linear_solver_summary = ComputeGaussNewtonStep(jacobian, residuals); TrustRegionStrategy::Summary summary; summary.residual_norm = linear_solver_summary.residual_norm; summary.num_iterations = linear_solver_summary.num_iterations; summary.termination_type = linear_solver_summary.termination_type; if (linear_solver_summary.termination_type != FAILURE) { switch(dogleg_type_) { // Interpolate the Cauchy point and the Gauss-Newton step. case TRADITIONAL_DOGLEG: ComputeTraditionalDoglegStep(step); break; // Find the minimum in the subspace defined by the // Cauchy point and the (Gauss-)Newton step. case SUBSPACE_DOGLEG: if (!ComputeSubspaceModel(jacobian)) { summary.termination_type = FAILURE; break; } ComputeSubspaceDoglegStep(step); break; } } return summary; } // The trust region is assumed to be elliptical with the // diagonal scaling matrix D defined by sqrt(diagonal_). // It is implemented by substituting step' = D * step. // The trust region for step' is spherical. // The gradient, the Gauss-Newton step, the Cauchy point, // and all calculations involving the Jacobian have to // be adjusted accordingly. void DoglegStrategy::ComputeGradient( SparseMatrix* jacobian, const double* residuals) { gradient_.setZero(); jacobian->LeftMultiply(residuals, gradient_.data()); gradient_.array() /= diagonal_.array(); } // The Cauchy point is the global minimizer of the quadratic model // along the one-dimensional subspace spanned by the gradient. void DoglegStrategy::ComputeCauchyPoint(SparseMatrix* jacobian) { // alpha * -gradient is the Cauchy point. Vector Jg(jacobian->num_rows()); Jg.setZero(); // The Jacobian is scaled implicitly by computing J * (D^-1 * (D^-1 * g)) // instead of (J * D^-1) * (D^-1 * g). Vector scaled_gradient = (gradient_.array() / diagonal_.array()).matrix(); jacobian->RightMultiply(scaled_gradient.data(), Jg.data()); alpha_ = gradient_.squaredNorm() / Jg.squaredNorm(); } // The dogleg step is defined as the intersection of the trust region // boundary with the piecewise linear path from the origin to the Cauchy // point and then from there to the Gauss-Newton point (global minimizer // of the model function). The Gauss-Newton point is taken if it lies // within the trust region. void DoglegStrategy::ComputeTraditionalDoglegStep(double* dogleg) { VectorRef dogleg_step(dogleg, gradient_.rows()); // Case 1. The Gauss-Newton step lies inside the trust region, and // is therefore the optimal solution to the trust-region problem. const double gradient_norm = gradient_.norm(); const double gauss_newton_norm = gauss_newton_step_.norm(); if (gauss_newton_norm <= radius_) { dogleg_step = gauss_newton_step_; dogleg_step_norm_ = gauss_newton_norm; dogleg_step.array() /= diagonal_.array(); VLOG(3) << "GaussNewton step size: " << dogleg_step_norm_ << " radius: " << radius_; return; } // Case 2. The Cauchy point and the Gauss-Newton steps lie outside // the trust region. Rescale the Cauchy point to the trust region // and return. if (gradient_norm * alpha_ >= radius_) { dogleg_step = -(radius_ / gradient_norm) * gradient_; dogleg_step_norm_ = radius_; dogleg_step.array() /= diagonal_.array(); VLOG(3) << "Cauchy step size: " << dogleg_step_norm_ << " radius: " << radius_; return; } // Case 3. The Cauchy point is inside the trust region and the // Gauss-Newton step is outside. Compute the line joining the two // points and the point on it which intersects the trust region // boundary. // a = alpha * -gradient // b = gauss_newton_step const double b_dot_a = -alpha_ * gradient_.dot(gauss_newton_step_); const double a_squared_norm = pow(alpha_ * gradient_norm, 2.0); const double b_minus_a_squared_norm = a_squared_norm - 2 * b_dot_a + pow(gauss_newton_norm, 2); // c = a' (b - a) // = alpha * -gradient' gauss_newton_step - alpha^2 |gradient|^2 const double c = b_dot_a - a_squared_norm; const double d = sqrt(c * c + b_minus_a_squared_norm * (pow(radius_, 2.0) - a_squared_norm)); double beta = (c <= 0) ? (d - c) / b_minus_a_squared_norm : (radius_ * radius_ - a_squared_norm) / (d + c); dogleg_step = (-alpha_ * (1.0 - beta)) * gradient_ + beta * gauss_newton_step_; dogleg_step_norm_ = dogleg_step.norm(); dogleg_step.array() /= diagonal_.array(); VLOG(3) << "Dogleg step size: " << dogleg_step_norm_ << " radius: " << radius_; } // The subspace method finds the minimum of the two-dimensional problem // // min. 1/2 x' B' H B x + g' B x // s.t. || B x ||^2 <= r^2 // // where r is the trust region radius and B is the matrix with unit columns // spanning the subspace defined by the steepest descent and Newton direction. // This subspace by definition includes the Gauss-Newton point, which is // therefore taken if it lies within the trust region. void DoglegStrategy::ComputeSubspaceDoglegStep(double* dogleg) { VectorRef dogleg_step(dogleg, gradient_.rows()); // The Gauss-Newton point is inside the trust region if |GN| <= radius_. // This test is valid even though radius_ is a length in the two-dimensional // subspace while gauss_newton_step_ is expressed in the (scaled) // higher dimensional original space. This is because // // 1. gauss_newton_step_ by definition lies in the subspace, and // 2. the subspace basis is orthonormal. // // As a consequence, the norm of the gauss_newton_step_ in the subspace is // the same as its norm in the original space. const double gauss_newton_norm = gauss_newton_step_.norm(); if (gauss_newton_norm <= radius_) { dogleg_step = gauss_newton_step_; dogleg_step_norm_ = gauss_newton_norm; dogleg_step.array() /= diagonal_.array(); VLOG(3) << "GaussNewton step size: " << dogleg_step_norm_ << " radius: " << radius_; return; } // The optimum lies on the boundary of the trust region. The above problem // therefore becomes // // min. 1/2 x^T B^T H B x + g^T B x // s.t. || B x ||^2 = r^2 // // Notice the equality in the constraint. // // This can be solved by forming the Lagrangian, solving for x(y), where // y is the Lagrange multiplier, using the gradient of the objective, and // putting x(y) back into the constraint. This results in a fourth order // polynomial in y, which can be solved using e.g. the companion matrix. // See the description of MakePolynomialForBoundaryConstrainedProblem for // details. The result is up to four real roots y*, not all of which // correspond to feasible points. The feasible points x(y*) have to be // tested for optimality. if (subspace_is_one_dimensional_) { // The subspace is one-dimensional, so both the gradient and // the Gauss-Newton step point towards the same direction. // In this case, we move along the gradient until we reach the trust // region boundary. dogleg_step = -(radius_ / gradient_.norm()) * gradient_; dogleg_step_norm_ = radius_; dogleg_step.array() /= diagonal_.array(); VLOG(3) << "Dogleg subspace step size (1D): " << dogleg_step_norm_ << " radius: " << radius_; return; } Vector2d minimum(0.0, 0.0); if (!FindMinimumOnTrustRegionBoundary(&minimum)) { // For the positive semi-definite case, a traditional dogleg step // is taken in this case. LOG(WARNING) << "Failed to compute polynomial roots. " << "Taking traditional dogleg step instead."; ComputeTraditionalDoglegStep(dogleg); return; } // Test first order optimality at the minimum. // The first order KKT conditions state that the minimum x* // has to satisfy either || x* ||^2 < r^2 (i.e. has to lie within // the trust region), or // // (B x* + g) + y x* = 0 // // for some positive scalar y. // Here, as it is already known that the minimum lies on the boundary, the // latter condition is tested. To allow for small imprecisions, we test if // the angle between (B x* + g) and -x* is smaller than acos(0.99). // The exact value of the cosine is arbitrary but should be close to 1. // // This condition should not be violated. If it is, the minimum was not // correctly determined. const double kCosineThreshold = 0.99; const Vector2d grad_minimum = subspace_B_ * minimum + subspace_g_; const double cosine_angle = -minimum.dot(grad_minimum) / (minimum.norm() * grad_minimum.norm()); if (cosine_angle < kCosineThreshold) { LOG(WARNING) << "First order optimality seems to be violated " << "in the subspace method!\n" << "Cosine of angle between x and B x + g is " << cosine_angle << ".\n" << "Taking a regular dogleg step instead.\n" << "Please consider filing a bug report if this " << "happens frequently or consistently.\n"; ComputeTraditionalDoglegStep(dogleg); return; } // Create the full step from the optimal 2d solution. dogleg_step = subspace_basis_ * minimum; dogleg_step_norm_ = radius_; dogleg_step.array() /= diagonal_.array(); VLOG(3) << "Dogleg subspace step size: " << dogleg_step_norm_ << " radius: " << radius_; } // Build the polynomial that defines the optimal Lagrange multipliers. // Let the Lagrangian be // // L(x, y) = 0.5 x^T B x + x^T g + y (0.5 x^T x - 0.5 r^2). (1) // // Stationary points of the Lagrangian are given by // // 0 = d L(x, y) / dx = Bx + g + y x (2) // 0 = d L(x, y) / dy = 0.5 x^T x - 0.5 r^2 (3) // // For any given y, we can solve (2) for x as // // x(y) = -(B + y I)^-1 g . (4) // // As B + y I is 2x2, we form the inverse explicitly: // // (B + y I)^-1 = (1 / det(B + y I)) adj(B + y I) (5) // // where adj() denotes adjugation. This should be safe, as B is positive // semi-definite and y is necessarily positive, so (B + y I) is indeed // invertible. // Plugging (5) into (4) and the result into (3), then dividing by 0.5 we // obtain // // 0 = (1 / det(B + y I))^2 g^T adj(B + y I)^T adj(B + y I) g - r^2 // (6) // // or // // det(B + y I)^2 r^2 = g^T adj(B + y I)^T adj(B + y I) g (7a) // = g^T adj(B)^T adj(B) g // + 2 y g^T adj(B)^T g + y^2 g^T g (7b) // // as // // adj(B + y I) = adj(B) + y I = adj(B)^T + y I . (8) // // The left hand side can be expressed explicitly using // // det(B + y I) = det(B) + y tr(B) + y^2 . (9) // // So (7) is a polynomial in y of degree four. // Bringing everything back to the left hand side, the coefficients can // be read off as // // y^4 r^2 // + y^3 2 r^2 tr(B) // + y^2 (r^2 tr(B)^2 + 2 r^2 det(B) - g^T g) // + y^1 (2 r^2 det(B) tr(B) - 2 g^T adj(B)^T g) // + y^0 (r^2 det(B)^2 - g^T adj(B)^T adj(B) g) // Vector DoglegStrategy::MakePolynomialForBoundaryConstrainedProblem() const { const double detB = subspace_B_.determinant(); const double trB = subspace_B_.trace(); const double r2 = radius_ * radius_; Matrix2d B_adj; B_adj << subspace_B_(1,1) , -subspace_B_(0,1), -subspace_B_(1,0) , subspace_B_(0,0); Vector polynomial(5); polynomial(0) = r2; polynomial(1) = 2.0 * r2 * trB; polynomial(2) = r2 * ( trB * trB + 2.0 * detB ) - subspace_g_.squaredNorm(); polynomial(3) = -2.0 * ( subspace_g_.transpose() * B_adj * subspace_g_ - r2 * detB * trB ); polynomial(4) = r2 * detB * detB - (B_adj * subspace_g_).squaredNorm(); return polynomial; } // Given a Lagrange multiplier y that corresponds to a stationary point // of the Lagrangian L(x, y), compute the corresponding x from the // equation // // 0 = d L(x, y) / dx // = B * x + g + y * x // = (B + y * I) * x + g // DoglegStrategy::Vector2d DoglegStrategy::ComputeSubspaceStepFromRoot( double y) const { const Matrix2d B_i = subspace_B_ + y * Matrix2d::Identity(); return -B_i.partialPivLu().solve(subspace_g_); } // This function evaluates the quadratic model at a point x in the // subspace spanned by subspace_basis_. double DoglegStrategy::EvaluateSubspaceModel(const Vector2d& x) const { return 0.5 * x.dot(subspace_B_ * x) + subspace_g_.dot(x); } // This function attempts to solve the boundary-constrained subspace problem // // min. 1/2 x^T B^T H B x + g^T B x // s.t. || B x ||^2 = r^2 // // where B is an orthonormal subspace basis and r is the trust-region radius. // // This is done by finding the roots of a fourth degree polynomial. If the // root finding fails, the function returns false and minimum will be set // to (0, 0). If it succeeds, true is returned. // // In the failure case, another step should be taken, such as the traditional // dogleg step. bool DoglegStrategy::FindMinimumOnTrustRegionBoundary(Vector2d* minimum) const { CHECK_NOTNULL(minimum); // Return (0, 0) in all error cases. minimum->setZero(); // Create the fourth-degree polynomial that is a necessary condition for // optimality. const Vector polynomial = MakePolynomialForBoundaryConstrainedProblem(); // Find the real parts y_i of its roots (not only the real roots). Vector roots_real; if (!FindPolynomialRoots(polynomial, &roots_real, NULL)) { // Failed to find the roots of the polynomial, i.e. the candidate // solutions of the constrained problem. Report this back to the caller. return false; } // For each root y, compute B x(y) and check for feasibility. // Notice that there should always be four roots, as the leading term of // the polynomial is r^2 and therefore non-zero. However, as some roots // may be complex, the real parts are not necessarily unique. double minimum_value = std::numeric_limits<double>::max(); bool valid_root_found = false; for (int i = 0; i < roots_real.size(); ++i) { const Vector2d x_i = ComputeSubspaceStepFromRoot(roots_real(i)); // Not all roots correspond to points on the trust region boundary. // There are at most four candidate solutions. As we are interested // in the minimum, it is safe to consider all of them after projecting // them onto the trust region boundary. if (x_i.norm() > 0) { const double f_i = EvaluateSubspaceModel((radius_ / x_i.norm()) * x_i); valid_root_found = true; if (f_i < minimum_value) { minimum_value = f_i; *minimum = x_i; } } } return valid_root_found; } LinearSolver::Summary DoglegStrategy::ComputeGaussNewtonStep( SparseMatrix* jacobian, const double* residuals) { const int n = jacobian->num_cols(); LinearSolver::Summary linear_solver_summary; linear_solver_summary.termination_type = FAILURE; // The Jacobian matrix is often quite poorly conditioned. Thus it is // necessary to add a diagonal matrix at the bottom to prevent the // linear solver from failing. // // We do this by computing the same diagonal matrix as the one used // by Levenberg-Marquardt (other choices are possible), and scaling // it by a small constant (independent of the trust region radius). // // If the solve fails, the multiplier to the diagonal is increased // up to max_mu_ by a factor of mu_increase_factor_ every time. If // the linear solver is still not successful, the strategy returns // with FAILURE. // // Next time when a new Gauss-Newton step is requested, the // multiplier starts out from the last successful solve. // // When a step is declared successful, the multiplier is decreased // by half of mu_increase_factor_. while (mu_ < max_mu_) { // Dogleg, as far as I (sameeragarwal) understand it, requires a // reasonably good estimate of the Gauss-Newton step. This means // that we need to solve the normal equations more or less // exactly. This is reflected in the values of the tolerances set // below. // // For now, this strategy should only be used with exact // factorization based solvers, for which these tolerances are // automatically satisfied. // // The right way to combine inexact solves with trust region // methods is to use Stiehaug's method. LinearSolver::PerSolveOptions solve_options; solve_options.q_tolerance = 0.0; solve_options.r_tolerance = 0.0; lm_diagonal_ = diagonal_ * std::sqrt(mu_); solve_options.D = lm_diagonal_.data(); // As in the LevenbergMarquardtStrategy, solve Jy = r instead // of Jx = -r and later set x = -y to avoid having to modify // either jacobian or residuals. InvalidateArray(n, gauss_newton_step_.data()); linear_solver_summary = linear_solver_->Solve(jacobian, residuals, solve_options, gauss_newton_step_.data()); if (linear_solver_summary.termination_type == FAILURE || !IsArrayValid(n, gauss_newton_step_.data())) { mu_ *= mu_increase_factor_; VLOG(2) << "Increasing mu " << mu_; linear_solver_summary.termination_type = FAILURE; continue; } break; } if (linear_solver_summary.termination_type != FAILURE) { // The scaled Gauss-Newton step is D * GN: // // - (D^-1 J^T J D^-1)^-1 (D^-1 g) // = - D (J^T J)^-1 D D^-1 g // = D -(J^T J)^-1 g // gauss_newton_step_.array() *= -diagonal_.array(); } return linear_solver_summary; } void DoglegStrategy::StepAccepted(double step_quality) { CHECK_GT(step_quality, 0.0); if (step_quality < decrease_threshold_) { radius_ *= 0.5; } if (step_quality > increase_threshold_) { radius_ = max(radius_, 3.0 * dogleg_step_norm_); } // Reduce the regularization multiplier, in the hope that whatever // was causing the rank deficiency has gone away and we can return // to doing a pure Gauss-Newton solve. mu_ = max(min_mu_, 2.0 * mu_ / mu_increase_factor_ ); reuse_ = false; } void DoglegStrategy::StepRejected(double step_quality) { radius_ *= 0.5; reuse_ = true; } void DoglegStrategy::StepIsInvalid() { mu_ *= mu_increase_factor_; reuse_ = false; } double DoglegStrategy::Radius() const { return radius_; } bool DoglegStrategy::ComputeSubspaceModel(SparseMatrix* jacobian) { // Compute an orthogonal basis for the subspace using QR decomposition. Matrix basis_vectors(jacobian->num_cols(), 2); basis_vectors.col(0) = gradient_; basis_vectors.col(1) = gauss_newton_step_; Eigen::ColPivHouseholderQR<Matrix> basis_qr(basis_vectors); switch (basis_qr.rank()) { case 0: // This should never happen, as it implies that both the gradient // and the Gauss-Newton step are zero. In this case, the minimizer should // have stopped due to the gradient being too small. LOG(ERROR) << "Rank of subspace basis is 0. " << "This means that the gradient at the current iterate is " << "zero but the optimization has not been terminated. " << "You may have found a bug in Ceres."; return false; case 1: // Gradient and Gauss-Newton step coincide, so we lie on one of the // major axes of the quadratic problem. In this case, we simply move // along the gradient until we reach the trust region boundary. subspace_is_one_dimensional_ = true; return true; case 2: subspace_is_one_dimensional_ = false; break; default: LOG(ERROR) << "Rank of the subspace basis matrix is reported to be " << "greater than 2. As the matrix contains only two " << "columns this cannot be true and is indicative of " << "a bug."; return false; } // The subspace is two-dimensional, so compute the subspace model. // Given the basis U, this is // // subspace_g_ = g_scaled^T U // // and // // subspace_B_ = U^T (J_scaled^T J_scaled) U // // As J_scaled = J * D^-1, the latter becomes // // subspace_B_ = ((U^T D^-1) J^T) (J (D^-1 U)) // = (J (D^-1 U))^T (J (D^-1 U)) subspace_basis_ = basis_qr.householderQ() * Matrix::Identity(jacobian->num_cols(), 2); subspace_g_ = subspace_basis_.transpose() * gradient_; Eigen::Matrix<double, 2, Eigen::Dynamic, Eigen::RowMajor> Jb(2, jacobian->num_rows()); Jb.setZero(); Vector tmp; tmp = (subspace_basis_.col(0).array() / diagonal_.array()).matrix(); jacobian->RightMultiply(tmp.data(), Jb.row(0).data()); tmp = (subspace_basis_.col(1).array() / diagonal_.array()).matrix(); jacobian->RightMultiply(tmp.data(), Jb.row(1).data()); subspace_B_ = Jb * Jb.transpose(); return true; } } // namespace internal } // namespace ceres