• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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@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 
LowRankInverseHessian(int num_parameters,int max_num_corrections,bool use_approximate_eigenvalue_scaling)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 
Update(const Vector & delta_x,const Vector & delta_gradient)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 
RightMultiply(const double * x_ptr,double * y_ptr) const119 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