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 (at) google.com (Sameer Agarwal) 30 31 #include <list> 32 33 #include "ceres/internal/eigen.h" 34 #include "ceres/low_rank_inverse_hessian.h" 35 #include "glog/logging.h" 36 37 namespace ceres { 38 namespace internal { 39 40 // The (L)BFGS algorithm explicitly requires that the secant equation: 41 // 42 // B_{k+1} * s_k = y_k 43 // 44 // Is satisfied at each iteration, where B_{k+1} is the approximated 45 // Hessian at the k+1-th iteration, s_k = (x_{k+1} - x_{k}) and 46 // y_k = (grad_{k+1} - grad_{k}). As the approximated Hessian must be 47 // positive definite, this is equivalent to the condition: 48 // 49 // s_k^T * y_k > 0 [s_k^T * B_{k+1} * s_k = s_k^T * y_k > 0] 50 // 51 // This condition would always be satisfied if the function was strictly 52 // convex, alternatively, it is always satisfied provided that a Wolfe line 53 // search is used (even if the function is not strictly convex). See [1] 54 // (p138) for a proof. 55 // 56 // Although Ceres will always use a Wolfe line search when using (L)BFGS, 57 // practical implementation considerations mean that the line search 58 // may return a point that satisfies only the Armijo condition, and thus 59 // could violate the Secant equation. As such, we will only use a step 60 // to update the Hessian approximation if: 61 // 62 // s_k^T * y_k > tolerance 63 // 64 // It is important that tolerance is very small (and >=0), as otherwise we 65 // might skip the update too often and fail to capture important curvature 66 // information in the Hessian. For example going from 1e-10 -> 1e-14 improves 67 // the NIST benchmark score from 43/54 to 53/54. 68 // 69 // [1] Nocedal J., Wright S., Numerical Optimization, 2nd Ed. Springer, 1999. 70 // 71 // TODO(alexs.mac): Consider using Damped BFGS update instead of 72 // skipping update. 73 const double kLBFGSSecantConditionHessianUpdateTolerance = 1e-14; 74 75 LowRankInverseHessian::LowRankInverseHessian( 76 int num_parameters, 77 int max_num_corrections, 78 bool use_approximate_eigenvalue_scaling) 79 : num_parameters_(num_parameters), 80 max_num_corrections_(max_num_corrections), 81 use_approximate_eigenvalue_scaling_(use_approximate_eigenvalue_scaling), 82 approximate_eigenvalue_scale_(1.0), 83 delta_x_history_(num_parameters, max_num_corrections), 84 delta_gradient_history_(num_parameters, max_num_corrections), 85 delta_x_dot_delta_gradient_(max_num_corrections) { 86 } 87 88 bool LowRankInverseHessian::Update(const Vector& delta_x, 89 const Vector& delta_gradient) { 90 const double delta_x_dot_delta_gradient = delta_x.dot(delta_gradient); 91 if (delta_x_dot_delta_gradient <= 92 kLBFGSSecantConditionHessianUpdateTolerance) { 93 VLOG(2) << "Skipping L-BFGS Update, delta_x_dot_delta_gradient too " 94 << "small: " << delta_x_dot_delta_gradient << ", tolerance: " 95 << kLBFGSSecantConditionHessianUpdateTolerance 96 << " (Secant condition)."; 97 return false; 98 } 99 100 101 int next = indices_.size(); 102 // Once the size of the list reaches max_num_corrections_, simulate 103 // a circular buffer by removing the first element of the list and 104 // making it the next position where the LBFGS history is stored. 105 if (next == max_num_corrections_) { 106 next = indices_.front(); 107 indices_.pop_front(); 108 } 109 110 indices_.push_back(next); 111 delta_x_history_.col(next) = delta_x; 112 delta_gradient_history_.col(next) = delta_gradient; 113 delta_x_dot_delta_gradient_(next) = delta_x_dot_delta_gradient; 114 approximate_eigenvalue_scale_ = 115 delta_x_dot_delta_gradient / delta_gradient.squaredNorm(); 116 return true; 117 } 118 119 void LowRankInverseHessian::RightMultiply(const double* x_ptr, 120 double* y_ptr) const { 121 ConstVectorRef gradient(x_ptr, num_parameters_); 122 VectorRef search_direction(y_ptr, num_parameters_); 123 124 search_direction = gradient; 125 126 const int num_corrections = indices_.size(); 127 Vector alpha(num_corrections); 128 129 for (std::list<int>::const_reverse_iterator it = indices_.rbegin(); 130 it != indices_.rend(); 131 ++it) { 132 const double alpha_i = delta_x_history_.col(*it).dot(search_direction) / 133 delta_x_dot_delta_gradient_(*it); 134 search_direction -= alpha_i * delta_gradient_history_.col(*it); 135 alpha(*it) = alpha_i; 136 } 137 138 if (use_approximate_eigenvalue_scaling_) { 139 // Rescale the initial inverse Hessian approximation (H_0) to be iteratively 140 // updated so that it is of similar 'size' to the true inverse Hessian along 141 // the most recent search direction. As shown in [1]: 142 // 143 // \gamma_k = (delta_gradient_{k-1}' * delta_x_{k-1}) / 144 // (delta_gradient_{k-1}' * delta_gradient_{k-1}) 145 // 146 // Satisfies: 147 // 148 // (1 / \lambda_m) <= \gamma_k <= (1 / \lambda_1) 149 // 150 // Where \lambda_1 & \lambda_m are the smallest and largest eigenvalues of 151 // the true Hessian (not the inverse) along the most recent search direction 152 // respectively. Thus \gamma is an approximate eigenvalue of the true 153 // inverse Hessian, and choosing: H_0 = I * \gamma will yield a starting 154 // point that has a similar scale to the true inverse Hessian. This 155 // technique is widely reported to often improve convergence, however this 156 // is not universally true, particularly if there are errors in the initial 157 // jacobians, or if there are significant differences in the sensitivity 158 // of the problem to the parameters (i.e. the range of the magnitudes of 159 // the components of the gradient is large). 160 // 161 // The original origin of this rescaling trick is somewhat unclear, the 162 // earliest reference appears to be Oren [1], however it is widely discussed 163 // without specific attributation in various texts including [2] (p143/178). 164 // 165 // [1] Oren S.S., Self-scaling variable metric (SSVM) algorithms Part II: 166 // Implementation and experiments, Management Science, 167 // 20(5), 863-874, 1974. 168 // [2] Nocedal J., Wright S., Numerical Optimization, Springer, 1999. 169 search_direction *= approximate_eigenvalue_scale_; 170 171 VLOG(4) << "Applying approximate_eigenvalue_scale: " 172 << approximate_eigenvalue_scale_ << " to initial inverse Hessian " 173 << "approximation."; 174 } 175 176 for (std::list<int>::const_iterator it = indices_.begin(); 177 it != indices_.end(); 178 ++it) { 179 const double beta = delta_gradient_history_.col(*it).dot(search_direction) / 180 delta_x_dot_delta_gradient_(*it); 181 search_direction += delta_x_history_.col(*it) * (alpha(*it) - beta); 182 } 183 } 184 185 } // namespace internal 186 } // namespace ceres 187