• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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 // A preconditioned conjugate gradients solver
32 // (ConjugateGradientsSolver) for positive semidefinite linear
33 // systems.
34 //
35 // We have also augmented the termination criterion used by this
36 // solver to support not just residual based termination but also
37 // termination based on decrease in the value of the quadratic model
38 // that CG optimizes.
39 
40 #include "ceres/conjugate_gradients_solver.h"
41 
42 #include <cmath>
43 #include <cstddef>
44 #include "ceres/fpclassify.h"
45 #include "ceres/internal/eigen.h"
46 #include "ceres/linear_operator.h"
47 #include "ceres/types.h"
48 #include "glog/logging.h"
49 
50 namespace ceres {
51 namespace internal {
52 namespace {
53 
IsZeroOrInfinity(double x)54 bool IsZeroOrInfinity(double x) {
55   return ((x == 0.0) || (IsInfinite(x)));
56 }
57 
58 // Constant used in the MATLAB implementation ~ 2 * eps.
59 const double kEpsilon = 2.2204e-16;
60 
61 }  // namespace
62 
ConjugateGradientsSolver(const LinearSolver::Options & options)63 ConjugateGradientsSolver::ConjugateGradientsSolver(
64     const LinearSolver::Options& options)
65     : options_(options) {
66 }
67 
Solve(LinearOperator * A,const double * b,const LinearSolver::PerSolveOptions & per_solve_options,double * x)68 LinearSolver::Summary ConjugateGradientsSolver::Solve(
69     LinearOperator* A,
70     const double* b,
71     const LinearSolver::PerSolveOptions& per_solve_options,
72     double* x) {
73   CHECK_NOTNULL(A);
74   CHECK_NOTNULL(x);
75   CHECK_NOTNULL(b);
76   CHECK_EQ(A->num_rows(), A->num_cols());
77 
78   LinearSolver::Summary summary;
79   summary.termination_type = MAX_ITERATIONS;
80   summary.num_iterations = 0;
81 
82   int num_cols = A->num_cols();
83   VectorRef xref(x, num_cols);
84   ConstVectorRef bref(b, num_cols);
85 
86   double norm_b = bref.norm();
87   if (norm_b == 0.0) {
88     xref.setZero();
89     summary.termination_type = TOLERANCE;
90     return summary;
91   }
92 
93   Vector r(num_cols);
94   Vector p(num_cols);
95   Vector z(num_cols);
96   Vector tmp(num_cols);
97 
98   double tol_r = per_solve_options.r_tolerance * norm_b;
99 
100   tmp.setZero();
101   A->RightMultiply(x, tmp.data());
102   r = bref - tmp;
103   double norm_r = r.norm();
104 
105   if (norm_r <= tol_r) {
106     summary.termination_type = TOLERANCE;
107     return summary;
108   }
109 
110   double rho = 1.0;
111 
112   // Initial value of the quadratic model Q = x'Ax - 2 * b'x.
113   double Q0 = -1.0 * xref.dot(bref + r);
114 
115   for (summary.num_iterations = 1;
116        summary.num_iterations < options_.max_num_iterations;
117        ++summary.num_iterations) {
118     VLOG(3) << "cg iteration " << summary.num_iterations;
119 
120     // Apply preconditioner
121     if (per_solve_options.preconditioner != NULL) {
122       z.setZero();
123       per_solve_options.preconditioner->RightMultiply(r.data(), z.data());
124     } else {
125       z = r;
126     }
127 
128     double last_rho = rho;
129     rho = r.dot(z);
130 
131     if (IsZeroOrInfinity(rho)) {
132       LOG(ERROR) << "Numerical failure. rho = " << rho;
133       summary.termination_type = FAILURE;
134       break;
135     };
136 
137     if (summary.num_iterations == 1) {
138       p = z;
139     } else {
140       double beta = rho / last_rho;
141       if (IsZeroOrInfinity(beta)) {
142         LOG(ERROR) << "Numerical failure. beta = " << beta;
143         summary.termination_type = FAILURE;
144         break;
145       }
146       p = z + beta * p;
147     }
148 
149     Vector& q = z;
150     q.setZero();
151     A->RightMultiply(p.data(), q.data());
152     double pq = p.dot(q);
153 
154     if ((pq <= 0) || IsInfinite(pq))  {
155       LOG(ERROR) << "Numerical failure. pq = " << pq;
156       summary.termination_type = FAILURE;
157       break;
158     }
159 
160     double alpha = rho / pq;
161     if (IsInfinite(alpha)) {
162       LOG(ERROR) << "Numerical failure. alpha " << alpha;
163       summary.termination_type = FAILURE;
164       break;
165     }
166 
167     xref = xref + alpha * p;
168 
169     // Ideally we would just use the update r = r - alpha*q to keep
170     // track of the residual vector. However this estimate tends to
171     // drift over time due to round off errors. Thus every
172     // residual_reset_period iterations, we calculate the residual as
173     // r = b - Ax. We do not do this every iteration because this
174     // requires an additional matrix vector multiply which would
175     // double the complexity of the CG algorithm.
176     if (summary.num_iterations % options_.residual_reset_period == 0) {
177       tmp.setZero();
178       A->RightMultiply(x, tmp.data());
179       r = bref - tmp;
180     } else {
181       r = r - alpha * q;
182     }
183 
184     // Quadratic model based termination.
185     //   Q1 = x'Ax - 2 * b' x.
186     double Q1 = -1.0 * xref.dot(bref + r);
187 
188     // For PSD matrices A, let
189     //
190     //   Q(x) = x'Ax - 2b'x
191     //
192     // be the cost of the quadratic function defined by A and b. Then,
193     // the solver terminates at iteration i if
194     //
195     //   i * (Q(x_i) - Q(x_i-1)) / Q(x_i) < q_tolerance.
196     //
197     // This termination criterion is more useful when using CG to
198     // solve the Newton step. This particular convergence test comes
199     // from Stephen Nash's work on truncated Newton
200     // methods. References:
201     //
202     //   1. Stephen G. Nash & Ariela Sofer, Assessing A Search
203     //   Direction Within A Truncated Newton Method, Operation
204     //   Research Letters 9(1990) 219-221.
205     //
206     //   2. Stephen G. Nash, A Survey of Truncated Newton Methods,
207     //   Journal of Computational and Applied Mathematics,
208     //   124(1-2), 45-59, 2000.
209     //
210     double zeta = summary.num_iterations * (Q1 - Q0) / Q1;
211     VLOG(3) << "Q termination: zeta " << zeta
212             << " " << per_solve_options.q_tolerance;
213     if (zeta < per_solve_options.q_tolerance) {
214       summary.termination_type = TOLERANCE;
215       break;
216     }
217     Q0 = Q1;
218 
219     // Residual based termination.
220     norm_r = r. norm();
221     VLOG(3) << "R termination: norm_r " << norm_r
222             << " " << tol_r;
223     if (norm_r <= tol_r) {
224       summary.termination_type = TOLERANCE;
225       break;
226     }
227   }
228 
229   return summary;
230 };
231 
232 }  // namespace internal
233 }  // namespace ceres
234