• 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 // NIST non-linear regression problems solved using Ceres.
32 //
33 // The data was obtained from
34 // http://www.itl.nist.gov/div898/strd/nls/nls_main.shtml, where more
35 // background on these problems can also be found.
36 //
37 // Currently not all problems are solved successfully. Some of the
38 // failures are due to convergence to a local minimum, and some fail
39 // because of numerical issues.
40 //
41 // TODO(sameeragarwal): Fix numerical issues so that all the problems
42 // converge and then look at convergence to the wrong solution issues.
43 
44 #include <iostream>
45 #include <fstream>
46 #include "ceres/ceres.h"
47 #include "ceres/split.h"
48 #include "gflags/gflags.h"
49 #include "glog/logging.h"
50 #include "Eigen/Core"
51 
52 DEFINE_string(nist_data_dir, "", "Directory containing the NIST non-linear"
53               "regression examples");
54 DEFINE_string(trust_region_strategy, "levenberg_marquardt",
55               "Options are: levenberg_marquardt, dogleg");
56 DEFINE_string(dogleg, "traditional_dogleg",
57               "Options are: traditional_dogleg, subspace_dogleg");
58 DEFINE_string(linear_solver, "dense_qr", "Options are: "
59               "sparse_cholesky, dense_qr, dense_normal_cholesky and"
60               "cgnr");
61 DEFINE_string(preconditioner, "jacobi", "Options are: "
62               "identity, jacobi");
63 DEFINE_int32(num_iterations, 10000, "Number of iterations");
64 DEFINE_bool(nonmonotonic_steps, false, "Trust region algorithm can use"
65             " nonmonotic steps");
66 DEFINE_double(initial_trust_region_radius, 1e4, "Initial trust region radius");
67 
68 using Eigen::Dynamic;
69 using Eigen::RowMajor;
70 typedef Eigen::Matrix<double, Dynamic, 1> Vector;
71 typedef Eigen::Matrix<double, Dynamic, Dynamic, RowMajor> Matrix;
72 
GetAndSplitLine(std::ifstream & ifs,std::vector<std::string> * pieces)73 bool GetAndSplitLine(std::ifstream& ifs, std::vector<std::string>* pieces) {
74   pieces->clear();
75   char buf[256];
76   ifs.getline(buf, 256);
77   ceres::SplitStringUsing(std::string(buf), " ", pieces);
78   return true;
79 }
80 
SkipLines(std::ifstream & ifs,int num_lines)81 void SkipLines(std::ifstream& ifs, int num_lines) {
82   char buf[256];
83   for (int i = 0; i < num_lines; ++i) {
84     ifs.getline(buf, 256);
85   }
86 }
87 
IsSuccessfulTermination(ceres::SolverTerminationType status)88 bool IsSuccessfulTermination(ceres::SolverTerminationType status) {
89   return
90       (status == ceres::FUNCTION_TOLERANCE) ||
91       (status == ceres::GRADIENT_TOLERANCE) ||
92       (status == ceres::PARAMETER_TOLERANCE) ||
93       (status == ceres::USER_SUCCESS);
94 }
95 
96 class NISTProblem {
97  public:
NISTProblem(const std::string & filename)98   explicit NISTProblem(const std::string& filename) {
99     std::ifstream ifs(filename.c_str(), std::ifstream::in);
100 
101     std::vector<std::string> pieces;
102     SkipLines(ifs, 24);
103     GetAndSplitLine(ifs, &pieces);
104     const int kNumResponses = std::atoi(pieces[1].c_str());
105 
106     GetAndSplitLine(ifs, &pieces);
107     const int kNumPredictors = std::atoi(pieces[0].c_str());
108 
109     GetAndSplitLine(ifs, &pieces);
110     const int kNumObservations = std::atoi(pieces[0].c_str());
111 
112     SkipLines(ifs, 4);
113     GetAndSplitLine(ifs, &pieces);
114     const int kNumParameters = std::atoi(pieces[0].c_str());
115     SkipLines(ifs, 8);
116 
117     // Get the first line of initial and final parameter values to
118     // determine the number of tries.
119     GetAndSplitLine(ifs, &pieces);
120     const int kNumTries = pieces.size() - 4;
121 
122     predictor_.resize(kNumObservations, kNumPredictors);
123     response_.resize(kNumObservations, kNumResponses);
124     initial_parameters_.resize(kNumTries, kNumParameters);
125     final_parameters_.resize(1, kNumParameters);
126 
127     // Parse the line for parameter b1.
128     int parameter_id = 0;
129     for (int i = 0; i < kNumTries; ++i) {
130       initial_parameters_(i, parameter_id) = std::atof(pieces[i + 2].c_str());
131     }
132     final_parameters_(0, parameter_id) = std::atof(pieces[2 + kNumTries].c_str());
133 
134     // Parse the remaining parameter lines.
135     for (int parameter_id = 1; parameter_id < kNumParameters; ++parameter_id) {
136      GetAndSplitLine(ifs, &pieces);
137      // b2, b3, ....
138      for (int i = 0; i < kNumTries; ++i) {
139        initial_parameters_(i, parameter_id) = std::atof(pieces[i + 2].c_str());
140      }
141      final_parameters_(0, parameter_id) = std::atof(pieces[2 + kNumTries].c_str());
142     }
143 
144     // Certfied cost
145     SkipLines(ifs, 1);
146     GetAndSplitLine(ifs, &pieces);
147     certified_cost_ = std::atof(pieces[4].c_str()) / 2.0;
148 
149     // Read the observations.
150     SkipLines(ifs, 18 - kNumParameters);
151     for (int i = 0; i < kNumObservations; ++i) {
152       GetAndSplitLine(ifs, &pieces);
153       // Response.
154       for (int j = 0; j < kNumResponses; ++j) {
155         response_(i, j) =  std::atof(pieces[j].c_str());
156       }
157 
158       // Predictor variables.
159       for (int j = 0; j < kNumPredictors; ++j) {
160         predictor_(i, j) =  std::atof(pieces[j + kNumResponses].c_str());
161       }
162     }
163   }
164 
initial_parameters(int start) const165   Matrix initial_parameters(int start) const { return initial_parameters_.row(start); }
final_parameters() const166   Matrix final_parameters() const  { return final_parameters_; }
predictor() const167   Matrix predictor()        const { return predictor_;         }
response() const168   Matrix response()         const { return response_;          }
predictor_size() const169   int predictor_size()      const { return predictor_.cols();  }
num_observations() const170   int num_observations()    const { return predictor_.rows();  }
response_size() const171   int response_size()       const { return response_.cols();   }
num_parameters() const172   int num_parameters()      const { return initial_parameters_.cols(); }
num_starts() const173   int num_starts()          const { return initial_parameters_.rows(); }
certified_cost() const174   double certified_cost()   const { return certified_cost_; }
175 
176  private:
177   Matrix predictor_;
178   Matrix response_;
179   Matrix initial_parameters_;
180   Matrix final_parameters_;
181   double certified_cost_;
182 };
183 
184 #define NIST_BEGIN(CostFunctionName) \
185   struct CostFunctionName { \
186     CostFunctionName(const double* const x, \
187                      const double* const y) \
188         : x_(*x), y_(*y) {} \
189     double x_; \
190     double y_; \
191     template <typename T> \
192     bool operator()(const T* const b, T* residual) const { \
193     const T y(y_); \
194     const T x(x_); \
195       residual[0] = y - (
196 
197 #define NIST_END ); return true; }};
198 
199 // y = b1 * (b2+x)**(-1/b3)  +  e
200 NIST_BEGIN(Bennet5)
201   b[0] * pow(b[1] + x, T(-1.0) / b[2])
202 NIST_END
203 
204 // y = b1*(1-exp[-b2*x])  +  e
205 NIST_BEGIN(BoxBOD)
206   b[0] * (T(1.0) - exp(-b[1] * x))
207 NIST_END
208 
209 // y = exp[-b1*x]/(b2+b3*x)  +  e
210 NIST_BEGIN(Chwirut)
211   exp(-b[0] * x) / (b[1] + b[2] * x)
212 NIST_END
213 
214 // y  = b1*x**b2  +  e
215 NIST_BEGIN(DanWood)
216   b[0] * pow(x, b[1])
217 NIST_END
218 
219 // y = b1*exp( -b2*x ) + b3*exp( -(x-b4)**2 / b5**2 )
220 //     + b6*exp( -(x-b7)**2 / b8**2 ) + e
221 NIST_BEGIN(Gauss)
222   b[0] * exp(-b[1] * x) +
223   b[2] * exp(-pow((x - b[3])/b[4], 2)) +
224   b[5] * exp(-pow((x - b[6])/b[7],2))
225 NIST_END
226 
227 // y = b1*exp(-b2*x) + b3*exp(-b4*x) + b5*exp(-b6*x)  +  e
228 NIST_BEGIN(Lanczos)
229   b[0] * exp(-b[1] * x) + b[2] * exp(-b[3] * x) + b[4] * exp(-b[5] * x)
230 NIST_END
231 
232 // y = (b1+b2*x+b3*x**2+b4*x**3) /
233 //     (1+b5*x+b6*x**2+b7*x**3)  +  e
234 NIST_BEGIN(Hahn1)
235   (b[0] + b[1] * x + b[2] * x * x + b[3] * x * x * x) /
236   (T(1.0) + b[4] * x + b[5] * x * x + b[6] * x * x * x)
237 NIST_END
238 
239 // y = (b1 + b2*x + b3*x**2) /
240 //    (1 + b4*x + b5*x**2)  +  e
241 NIST_BEGIN(Kirby2)
242   (b[0] + b[1] * x + b[2] * x * x) /
243   (T(1.0) + b[3] * x + b[4] * x * x)
244 NIST_END
245 
246 // y = b1*(x**2+x*b2) / (x**2+x*b3+b4)  +  e
247 NIST_BEGIN(MGH09)
248   b[0] * (x * x + x * b[1]) / (x * x + x * b[2] + b[3])
249 NIST_END
250 
251 // y = b1 * exp[b2/(x+b3)]  +  e
252 NIST_BEGIN(MGH10)
253   b[0] * exp(b[1] / (x + b[2]))
254 NIST_END
255 
256 // y = b1 + b2*exp[-x*b4] + b3*exp[-x*b5]
257 NIST_BEGIN(MGH17)
258   b[0] + b[1] * exp(-x * b[3]) + b[2] * exp(-x * b[4])
259 NIST_END
260 
261 // y = b1*(1-exp[-b2*x])  +  e
262 NIST_BEGIN(Misra1a)
263   b[0] * (T(1.0) - exp(-b[1] * x))
264 NIST_END
265 
266 // y = b1 * (1-(1+b2*x/2)**(-2))  +  e
267 NIST_BEGIN(Misra1b)
268   b[0] * (T(1.0) - T(1.0)/ ((T(1.0) + b[1] * x / 2.0) * (T(1.0) + b[1] * x / 2.0)))
269 NIST_END
270 
271 // y = b1 * (1-(1+2*b2*x)**(-.5))  +  e
272 NIST_BEGIN(Misra1c)
273   b[0] * (T(1.0) - pow(T(1.0) + T(2.0) * b[1] * x, -0.5))
274 NIST_END
275 
276 // y = b1*b2*x*((1+b2*x)**(-1))  +  e
277 NIST_BEGIN(Misra1d)
278   b[0] * b[1] * x / (T(1.0) + b[1] * x)
279 NIST_END
280 
281 const double kPi = 3.141592653589793238462643383279;
282 // pi = 3.141592653589793238462643383279E0
283 // y =  b1 - b2*x - arctan[b3/(x-b4)]/pi  +  e
284 NIST_BEGIN(Roszman1)
285   b[0] - b[1] * x - atan2(b[2], (x - b[3]))/T(kPi)
286 NIST_END
287 
288 // y = b1 / (1+exp[b2-b3*x])  +  e
289 NIST_BEGIN(Rat42)
290   b[0] / (T(1.0) + exp(b[1] - b[2] * x))
291 NIST_END
292 
293 // y = b1 / ((1+exp[b2-b3*x])**(1/b4))  +  e
294 NIST_BEGIN(Rat43)
295   b[0] / pow(T(1.0) + exp(b[1] - b[2] * x), T(1.0) / b[3])
296 NIST_END
297 
298 // y = (b1 + b2*x + b3*x**2 + b4*x**3) /
299 //    (1 + b5*x + b6*x**2 + b7*x**3)  +  e
300 NIST_BEGIN(Thurber)
301   (b[0] + b[1] * x + b[2] * x * x  + b[3] * x * x * x) /
302   (T(1.0) + b[4] * x + b[5] * x * x + b[6] * x * x * x)
303 NIST_END
304 
305 // y = b1 + b2*cos( 2*pi*x/12 ) + b3*sin( 2*pi*x/12 )
306 //        + b5*cos( 2*pi*x/b4 ) + b6*sin( 2*pi*x/b4 )
307 //        + b8*cos( 2*pi*x/b7 ) + b9*sin( 2*pi*x/b7 )  + e
308 NIST_BEGIN(ENSO)
309   b[0] + b[1] * cos(T(2.0 * kPi) * x / T(12.0)) +
310          b[2] * sin(T(2.0 * kPi) * x / T(12.0)) +
311          b[4] * cos(T(2.0 * kPi) * x / b[3]) +
312          b[5] * sin(T(2.0 * kPi) * x / b[3]) +
313          b[7] * cos(T(2.0 * kPi) * x / b[6]) +
314          b[8] * sin(T(2.0 * kPi) * x / b[6])
315 NIST_END
316 
317 // y = (b1/b2) * exp[-0.5*((x-b3)/b2)**2]  +  e
318 NIST_BEGIN(Eckerle4)
319   b[0] / b[1] * exp(T(-0.5) * pow((x - b[2])/b[1], 2))
320 NIST_END
321 
322 struct Nelson {
323  public:
NelsonNelson324   Nelson(const double* const x, const double* const y)
325       : x1_(x[0]), x2_(x[1]), y_(y[0]) {}
326 
327   template <typename T>
operator ()Nelson328   bool operator()(const T* const b, T* residual) const {
329     // log[y] = b1 - b2*x1 * exp[-b3*x2]  +  e
330     residual[0] = T(log(y_)) - (b[0] - b[1] * T(x1_) * exp(-b[2] * T(x2_)));
331     return true;
332   }
333 
334  private:
335   double x1_;
336   double x2_;
337   double y_;
338 };
339 
340 template <typename Model, int num_residuals, int num_parameters>
RegressionDriver(const std::string & filename,const ceres::Solver::Options & options)341 int RegressionDriver(const std::string& filename,
342                       const ceres::Solver::Options& options) {
343   NISTProblem nist_problem(FLAGS_nist_data_dir + filename);
344   CHECK_EQ(num_residuals, nist_problem.response_size());
345   CHECK_EQ(num_parameters, nist_problem.num_parameters());
346 
347   Matrix predictor = nist_problem.predictor();
348   Matrix response = nist_problem.response();
349   Matrix final_parameters = nist_problem.final_parameters();
350   std::vector<ceres::Solver::Summary> summaries(nist_problem.num_starts() + 1);
351   std::cerr << filename << std::endl;
352 
353   // Each NIST problem comes with multiple starting points, so we
354   // construct the problem from scratch for each case and solve it.
355   for (int start = 0; start < nist_problem.num_starts(); ++start) {
356     Matrix initial_parameters = nist_problem.initial_parameters(start);
357 
358     ceres::Problem problem;
359     for (int i = 0; i < nist_problem.num_observations(); ++i) {
360       problem.AddResidualBlock(
361           new ceres::AutoDiffCostFunction<Model, num_residuals, num_parameters>(
362               new Model(predictor.data() + nist_problem.predictor_size() * i,
363                         response.data() + nist_problem.response_size() * i)),
364           NULL,
365           initial_parameters.data());
366     }
367 
368     Solve(options, &problem, &summaries[start]);
369   }
370 
371   const double certified_cost = nist_problem.certified_cost();
372 
373   int num_success = 0;
374   const int kMinNumMatchingDigits = 4;
375   for (int start = 0; start < nist_problem.num_starts(); ++start) {
376     const ceres::Solver::Summary& summary = summaries[start];
377 
378     int num_matching_digits = 0;
379     if (IsSuccessfulTermination(summary.termination_type)
380         && summary.final_cost < certified_cost) {
381       num_matching_digits = kMinNumMatchingDigits + 1;
382     } else {
383       num_matching_digits =
384           -std::log10(fabs(summary.final_cost - certified_cost) / certified_cost);
385     }
386 
387     std::cerr << "start " << start + 1 << " " ;
388     if (num_matching_digits <= kMinNumMatchingDigits) {
389       std::cerr <<  "FAILURE";
390     } else {
391       std::cerr <<  "SUCCESS";
392       ++num_success;
393     }
394     std::cerr << " summary: "
395               << summary.BriefReport()
396               << " Certified cost: " << certified_cost
397               << std::endl;
398 
399   }
400 
401   return num_success;
402 }
403 
SetMinimizerOptions(ceres::Solver::Options * options)404 void SetMinimizerOptions(ceres::Solver::Options* options) {
405   CHECK(ceres::StringToLinearSolverType(FLAGS_linear_solver,
406                                         &options->linear_solver_type));
407   CHECK(ceres::StringToPreconditionerType(FLAGS_preconditioner,
408                                           &options->preconditioner_type));
409   CHECK(ceres::StringToTrustRegionStrategyType(
410             FLAGS_trust_region_strategy,
411             &options->trust_region_strategy_type));
412   CHECK(ceres::StringToDoglegType(FLAGS_dogleg, &options->dogleg_type));
413 
414   options->max_num_iterations = FLAGS_num_iterations;
415   options->use_nonmonotonic_steps = FLAGS_nonmonotonic_steps;
416   options->initial_trust_region_radius = FLAGS_initial_trust_region_radius;
417   options->function_tolerance = 1e-18;
418   options->gradient_tolerance = 1e-18;
419   options->parameter_tolerance = 1e-18;
420 }
421 
SolveNISTProblems()422 void SolveNISTProblems() {
423   if (FLAGS_nist_data_dir.empty()) {
424     LOG(FATAL) << "Must specify the directory containing the NIST problems";
425   }
426 
427   ceres::Solver::Options options;
428   SetMinimizerOptions(&options);
429 
430   std::cerr << "Lower Difficulty\n";
431   int easy_success = 0;
432   easy_success += RegressionDriver<Misra1a,  1, 2>("Misra1a.dat",  options);
433   easy_success += RegressionDriver<Chwirut,  1, 3>("Chwirut1.dat", options);
434   easy_success += RegressionDriver<Chwirut,  1, 3>("Chwirut2.dat", options);
435   easy_success += RegressionDriver<Lanczos,  1, 6>("Lanczos3.dat", options);
436   easy_success += RegressionDriver<Gauss,    1, 8>("Gauss1.dat",   options);
437   easy_success += RegressionDriver<Gauss,    1, 8>("Gauss2.dat",   options);
438   easy_success += RegressionDriver<DanWood,  1, 2>("DanWood.dat",  options);
439   easy_success += RegressionDriver<Misra1b,  1, 2>("Misra1b.dat",  options);
440 
441   std::cerr << "\nMedium Difficulty\n";
442   int medium_success = 0;
443   medium_success += RegressionDriver<Kirby2,   1, 5>("Kirby2.dat",   options);
444   medium_success += RegressionDriver<Hahn1,    1, 7>("Hahn1.dat",    options);
445   medium_success += RegressionDriver<Nelson,   1, 3>("Nelson.dat",   options);
446   medium_success += RegressionDriver<MGH17,    1, 5>("MGH17.dat",    options);
447   medium_success += RegressionDriver<Lanczos,  1, 6>("Lanczos1.dat", options);
448   medium_success += RegressionDriver<Lanczos,  1, 6>("Lanczos2.dat", options);
449   medium_success += RegressionDriver<Gauss,    1, 8>("Gauss3.dat",   options);
450   medium_success += RegressionDriver<Misra1c,  1, 2>("Misra1c.dat",  options);
451   medium_success += RegressionDriver<Misra1d,  1, 2>("Misra1d.dat",  options);
452   medium_success += RegressionDriver<Roszman1, 1, 4>("Roszman1.dat", options);
453   medium_success += RegressionDriver<ENSO,     1, 9>("ENSO.dat",     options);
454 
455   std::cerr << "\nHigher Difficulty\n";
456   int hard_success = 0;
457   hard_success += RegressionDriver<MGH09,    1, 4>("MGH09.dat",    options);
458   hard_success += RegressionDriver<Thurber,  1, 7>("Thurber.dat",  options);
459   hard_success += RegressionDriver<BoxBOD,   1, 2>("BoxBOD.dat",   options);
460   hard_success += RegressionDriver<Rat42,    1, 3>("Rat42.dat",    options);
461   hard_success += RegressionDriver<MGH10,    1, 3>("MGH10.dat",    options);
462 
463   hard_success += RegressionDriver<Eckerle4, 1, 3>("Eckerle4.dat", options);
464   hard_success += RegressionDriver<Rat43,    1, 4>("Rat43.dat",    options);
465   hard_success += RegressionDriver<Bennet5,  1, 3>("Bennett5.dat", options);
466 
467   std::cerr << "\n";
468   std::cerr << "Easy    : " << easy_success << "/16\n";
469   std::cerr << "Medium  : " << medium_success << "/22\n";
470   std::cerr << "Hard    : " << hard_success << "/16\n";
471   std::cerr << "Total   : " << easy_success + medium_success + hard_success << "/54\n";
472 }
473 
main(int argc,char ** argv)474 int main(int argc, char** argv) {
475   google::ParseCommandLineFlags(&argc, &argv, true);
476   google::InitGoogleLogging(argv[0]);
477   SolveNISTProblems();
478   return 0;
479 };
480