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: keir (at) google.com (Keir Mierle) 30 31 #include "ceres/gradient_checking_cost_function.h" 32 33 #include <algorithm> 34 #include <cmath> 35 #include <numeric> 36 #include <string> 37 #include <vector> 38 39 #include "ceres/cost_function.h" 40 #include "ceres/internal/eigen.h" 41 #include "ceres/internal/scoped_ptr.h" 42 #include "ceres/parameter_block.h" 43 #include "ceres/problem.h" 44 #include "ceres/problem_impl.h" 45 #include "ceres/program.h" 46 #include "ceres/residual_block.h" 47 #include "ceres/runtime_numeric_diff_cost_function.h" 48 #include "ceres/stringprintf.h" 49 #include "ceres/types.h" 50 #include "glog/logging.h" 51 52 namespace ceres { 53 namespace internal { 54 namespace { 55 56 // True if x and y have an absolute relative difference less than 57 // relative_precision and false otherwise. Stores the relative and absolute 58 // difference in relative/absolute_error if non-NULL. 59 bool IsClose(double x, double y, double relative_precision, 60 double *relative_error, 61 double *absolute_error) { 62 double local_absolute_error; 63 double local_relative_error; 64 if (!absolute_error) { 65 absolute_error = &local_absolute_error; 66 } 67 if (!relative_error) { 68 relative_error = &local_relative_error; 69 } 70 *absolute_error = fabs(x - y); 71 *relative_error = *absolute_error / max(fabs(x), fabs(y)); 72 if (x == 0 || y == 0) { 73 // If x or y is exactly zero, then relative difference doesn't have any 74 // meaning. Take the absolute difference instead. 75 *relative_error = *absolute_error; 76 } 77 return fabs(*relative_error) < fabs(relative_precision); 78 } 79 80 class GradientCheckingCostFunction : public CostFunction { 81 public: 82 GradientCheckingCostFunction(const CostFunction* function, 83 double relative_step_size, 84 double relative_precision, 85 const string& extra_info) 86 : function_(function), 87 finite_diff_cost_function_( 88 CreateRuntimeNumericDiffCostFunction(function, 89 CENTRAL, 90 relative_step_size)), 91 relative_precision_(relative_precision), 92 extra_info_(extra_info) { 93 *mutable_parameter_block_sizes() = function->parameter_block_sizes(); 94 set_num_residuals(function->num_residuals()); 95 } 96 97 virtual ~GradientCheckingCostFunction() { } 98 99 virtual bool Evaluate(double const* const* parameters, 100 double* residuals, 101 double** jacobians) const { 102 if (!jacobians) { 103 // Nothing to check in this case; just forward. 104 return function_->Evaluate(parameters, residuals, NULL); 105 } 106 107 int num_residuals = function_->num_residuals(); 108 109 // Make space for the jacobians of the two methods. 110 const vector<int16>& block_sizes = function_->parameter_block_sizes(); 111 vector<Matrix> term_jacobians(block_sizes.size()); 112 vector<Matrix> finite_difference_jacobians(block_sizes.size()); 113 vector<double*> term_jacobian_pointers(block_sizes.size()); 114 vector<double*> finite_difference_jacobian_pointers(block_sizes.size()); 115 for (int i = 0; i < block_sizes.size(); i++) { 116 term_jacobians[i].resize(num_residuals, block_sizes[i]); 117 term_jacobian_pointers[i] = term_jacobians[i].data(); 118 finite_difference_jacobians[i].resize(num_residuals, block_sizes[i]); 119 finite_difference_jacobian_pointers[i] = 120 finite_difference_jacobians[i].data(); 121 } 122 123 // Evaluate the derivative using the user supplied code. 124 if (!function_->Evaluate(parameters, 125 residuals, 126 &term_jacobian_pointers[0])) { 127 LOG(WARNING) << "Function evaluation failed."; 128 return false; 129 } 130 131 // Evaluate the derivative using numeric derivatives. 132 finite_diff_cost_function_->Evaluate( 133 parameters, 134 residuals, 135 &finite_difference_jacobian_pointers[0]); 136 137 // See if any elements have relative error larger than the threshold. 138 int num_bad_jacobian_components = 0; 139 double worst_relative_error = 0; 140 141 // Accumulate the error message for all the jacobians, since it won't get 142 // output if there are no bad jacobian components. 143 string m; 144 for (int k = 0; k < block_sizes.size(); k++) { 145 // Copy the original jacobian blocks into the jacobians array. 146 if (jacobians[k] != NULL) { 147 MatrixRef(jacobians[k], 148 term_jacobians[k].rows(), 149 term_jacobians[k].cols()) = term_jacobians[k]; 150 } 151 152 StringAppendF(&m, 153 "========== " 154 "Jacobian for " "block %d: (%ld by %ld)) " 155 "==========\n", 156 k, 157 static_cast<long>(term_jacobians[k].rows()), 158 static_cast<long>(term_jacobians[k].cols())); 159 // The funny spacing creates appropriately aligned column headers. 160 m += " block row col user dx/dy num diff dx/dy " 161 "abs error relative error parameter residual\n"; 162 163 for (int i = 0; i < term_jacobians[k].rows(); i++) { 164 for (int j = 0; j < term_jacobians[k].cols(); j++) { 165 double term_jacobian = term_jacobians[k](i, j); 166 double finite_jacobian = finite_difference_jacobians[k](i, j); 167 double relative_error, absolute_error; 168 bool bad_jacobian_entry = 169 !IsClose(term_jacobian, 170 finite_jacobian, 171 relative_precision_, 172 &relative_error, 173 &absolute_error); 174 worst_relative_error = std::max(worst_relative_error, 175 relative_error); 176 177 StringAppendF(&m, "%6d %4d %4d %17g %17g %17g %17g %17g %17g", 178 k, i, j, 179 term_jacobian, finite_jacobian, 180 absolute_error, relative_error, 181 parameters[k][j], 182 residuals[i]); 183 184 if (bad_jacobian_entry) { 185 num_bad_jacobian_components++; 186 StringAppendF( 187 &m, " ------ (%d,%d,%d) Relative error worse than %g", 188 k, i, j, relative_precision_); 189 } 190 m += "\n"; 191 } 192 } 193 } 194 195 // Since there were some bad errors, dump comprehensive debug info. 196 if (num_bad_jacobian_components) { 197 string header = StringPrintf("Detected %d bad jacobian component(s). " 198 "Worst relative error was %g.\n", 199 num_bad_jacobian_components, 200 worst_relative_error); 201 if (!extra_info_.empty()) { 202 header += "Extra info for this residual: " + extra_info_ + "\n"; 203 } 204 LOG(WARNING) << "\n" << header << m; 205 } 206 return true; 207 } 208 209 private: 210 const CostFunction* function_; 211 internal::scoped_ptr<CostFunction> finite_diff_cost_function_; 212 double relative_precision_; 213 string extra_info_; 214 }; 215 216 } // namespace 217 218 CostFunction *CreateGradientCheckingCostFunction( 219 const CostFunction *cost_function, 220 double relative_step_size, 221 double relative_precision, 222 const string& extra_info) { 223 return new GradientCheckingCostFunction(cost_function, 224 relative_step_size, 225 relative_precision, 226 extra_info); 227 } 228 229 ProblemImpl* CreateGradientCheckingProblemImpl(ProblemImpl* problem_impl, 230 double relative_step_size, 231 double relative_precision) { 232 // We create new CostFunctions by wrapping the original CostFunction 233 // in a gradient checking CostFunction. So its okay for the 234 // ProblemImpl to take ownership of it and destroy it. The 235 // LossFunctions and LocalParameterizations are reused and since 236 // they are owned by problem_impl, gradient_checking_problem_impl 237 // should not take ownership of it. 238 Problem::Options gradient_checking_problem_options; 239 gradient_checking_problem_options.cost_function_ownership = TAKE_OWNERSHIP; 240 gradient_checking_problem_options.loss_function_ownership = 241 DO_NOT_TAKE_OWNERSHIP; 242 gradient_checking_problem_options.local_parameterization_ownership = 243 DO_NOT_TAKE_OWNERSHIP; 244 245 ProblemImpl* gradient_checking_problem_impl = new ProblemImpl( 246 gradient_checking_problem_options); 247 248 Program* program = problem_impl->mutable_program(); 249 250 // For every ParameterBlock in problem_impl, create a new parameter 251 // block with the same local parameterization and constancy. 252 const vector<ParameterBlock*>& parameter_blocks = program->parameter_blocks(); 253 for (int i = 0; i < parameter_blocks.size(); ++i) { 254 ParameterBlock* parameter_block = parameter_blocks[i]; 255 gradient_checking_problem_impl->AddParameterBlock( 256 parameter_block->mutable_user_state(), 257 parameter_block->Size(), 258 parameter_block->mutable_local_parameterization()); 259 260 if (parameter_block->IsConstant()) { 261 gradient_checking_problem_impl->SetParameterBlockConstant( 262 parameter_block->mutable_user_state()); 263 } 264 } 265 266 // For every ResidualBlock in problem_impl, create a new 267 // ResidualBlock by wrapping its CostFunction inside a 268 // GradientCheckingCostFunction. 269 const vector<ResidualBlock*>& residual_blocks = program->residual_blocks(); 270 for (int i = 0; i < residual_blocks.size(); ++i) { 271 ResidualBlock* residual_block = residual_blocks[i]; 272 273 // Build a human readable string which identifies the 274 // ResidualBlock. This is used by the GradientCheckingCostFunction 275 // when logging debugging information. 276 string extra_info = StringPrintf( 277 "Residual block id %d; depends on parameters [", i); 278 vector<double*> parameter_blocks; 279 for (int j = 0; j < residual_block->NumParameterBlocks(); ++j) { 280 ParameterBlock* parameter_block = residual_block->parameter_blocks()[j]; 281 parameter_blocks.push_back(parameter_block->mutable_user_state()); 282 StringAppendF(&extra_info, "%p", parameter_block->mutable_user_state()); 283 extra_info += (j < residual_block->NumParameterBlocks() - 1) ? ", " : "]"; 284 } 285 286 // Wrap the original CostFunction in a GradientCheckingCostFunction. 287 CostFunction* gradient_checking_cost_function = 288 CreateGradientCheckingCostFunction(residual_block->cost_function(), 289 relative_step_size, 290 relative_precision, 291 extra_info); 292 293 // The const_cast is necessary because 294 // ProblemImpl::AddResidualBlock can potentially take ownership of 295 // the LossFunction, but in this case we are guaranteed that this 296 // will not be the case, so this const_cast is harmless. 297 gradient_checking_problem_impl->AddResidualBlock( 298 gradient_checking_cost_function, 299 const_cast<LossFunction*>(residual_block->loss_function()), 300 parameter_blocks); 301 } 302 303 return gradient_checking_problem_impl; 304 } 305 306 307 } // namespace internal 308 } // namespace ceres 309