Keir Mierle | 8ebb073 | 2012-04-30 23:09:08 -0700 | [diff] [blame] | 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@google.com (Keir Mierle) |
| 30 | // |
| 31 | // Create CostFunctions as needed by the least squares framework with jacobians |
| 32 | // computed via numeric (a.k.a. finite) differentiation. For more details see |
| 33 | // http://en.wikipedia.org/wiki/Numerical_differentiation. |
| 34 | // |
| 35 | // To get a numerically differentiated cost function, define a subclass of |
| 36 | // CostFunction such that the Evaluate() function ignores the jacobian |
| 37 | // parameter. The numeric differentiation wrapper will fill in the jacobian |
| 38 | // parameter if nececssary by repeatedly calling the Evaluate() function with |
| 39 | // small changes to the appropriate parameters, and computing the slope. For |
| 40 | // performance, the numeric differentiation wrapper class is templated on the |
| 41 | // concrete cost function, even though it could be implemented only in terms of |
| 42 | // the virtual CostFunction interface. |
| 43 | // |
| 44 | // The numerically differentiated version of a cost function for a cost function |
| 45 | // can be constructed as follows: |
| 46 | // |
| 47 | // CostFunction* cost_function |
| 48 | // = new NumericDiffCostFunction<MyCostFunction, CENTRAL, 1, 4, 8>( |
| 49 | // new MyCostFunction(...), TAKE_OWNERSHIP); |
| 50 | // |
| 51 | // where MyCostFunction has 1 residual and 2 parameter blocks with sizes 4 and 8 |
| 52 | // respectively. Look at the tests for a more detailed example. |
| 53 | // |
| 54 | // The central difference method is considerably more accurate at the cost of |
| 55 | // twice as many function evaluations than forward difference. Consider using |
| 56 | // central differences begin with, and only after that works, trying forward |
| 57 | // difference to improve performance. |
| 58 | // |
| 59 | // TODO(keir): Characterize accuracy; mention pitfalls; provide alternatives. |
| 60 | |
| 61 | #ifndef CERES_PUBLIC_NUMERIC_DIFF_COST_FUNCTION_H_ |
| 62 | #define CERES_PUBLIC_NUMERIC_DIFF_COST_FUNCTION_H_ |
| 63 | |
| 64 | #include <cstring> |
| 65 | #include <glog/logging.h> |
| 66 | #include "Eigen/Dense" |
| 67 | #include "ceres/internal/scoped_ptr.h" |
| 68 | #include "ceres/sized_cost_function.h" |
| 69 | #include "ceres/types.h" |
| 70 | |
| 71 | namespace ceres { |
| 72 | |
| 73 | enum NumericDiffMethod { |
| 74 | CENTRAL, |
Sameer Agarwal | 6447219 | 2012-05-03 21:53:07 -0700 | [diff] [blame] | 75 | FORWARD |
Keir Mierle | 8ebb073 | 2012-04-30 23:09:08 -0700 | [diff] [blame] | 76 | }; |
| 77 | |
| 78 | // This is split from the main class because C++ doesn't allow partial template |
| 79 | // specializations for member functions. The alternative is to repeat the main |
| 80 | // class for differing numbers of parameters, which is also unfortunate. |
| 81 | template <typename CostFunctionNoJacobian, |
| 82 | int num_residuals, |
| 83 | int parameter_block_size, |
| 84 | int parameter_block, |
| 85 | NumericDiffMethod method> |
| 86 | struct Differencer { |
| 87 | // Mutates parameters but must restore them before return. |
| 88 | static bool EvaluateJacobianForParameterBlock( |
| 89 | const CostFunctionNoJacobian *function, |
| 90 | double const* residuals_at_eval_point, |
| 91 | double **parameters, |
| 92 | double **jacobians) { |
| 93 | using Eigen::Map; |
| 94 | using Eigen::Matrix; |
| 95 | using Eigen::RowMajor; |
Sameer Agarwal | 295ade1 | 2012-08-22 06:51:22 -0700 | [diff] [blame] | 96 | using Eigen::ColMajor; |
Keir Mierle | 8ebb073 | 2012-04-30 23:09:08 -0700 | [diff] [blame] | 97 | |
| 98 | typedef Matrix<double, num_residuals, 1> ResidualVector; |
| 99 | typedef Matrix<double, parameter_block_size, 1> ParameterVector; |
Sameer Agarwal | 295ade1 | 2012-08-22 06:51:22 -0700 | [diff] [blame] | 100 | typedef Matrix<double, num_residuals, parameter_block_size, |
| 101 | (parameter_block_size == 1 && |
| 102 | num_residuals > 1) ? ColMajor : RowMajor> JacobianMatrix; |
Keir Mierle | 8ebb073 | 2012-04-30 23:09:08 -0700 | [diff] [blame] | 103 | |
| 104 | Map<JacobianMatrix> parameter_jacobian(jacobians[parameter_block], |
| 105 | num_residuals, |
| 106 | parameter_block_size); |
| 107 | |
| 108 | // Mutate 1 element at a time and then restore. |
| 109 | Map<ParameterVector> x_plus_delta(parameters[parameter_block], |
| 110 | parameter_block_size); |
| 111 | ParameterVector x(x_plus_delta); |
| 112 | |
| 113 | // TODO(keir): Pick a smarter number! In theory a good choice is sqrt(eps) * |
| 114 | // x, which for doubles means about 1e-8 * x. However, I have found this |
| 115 | // number too optimistic. This number should be exposed for users to change. |
| 116 | const double kRelativeStepSize = 1e-6; |
| 117 | |
| 118 | ParameterVector step_size = x.array().abs() * kRelativeStepSize; |
| 119 | |
| 120 | // To handle cases where a parameter is exactly zero, instead use the mean |
| 121 | // step_size for the other dimensions. |
| 122 | double fallback_step_size = step_size.sum() / step_size.rows(); |
| 123 | if (fallback_step_size == 0.0) { |
| 124 | // If all the parameters are zero, there's no good answer. Take |
| 125 | // kRelativeStepSize as a guess and hope for the best. |
| 126 | fallback_step_size = kRelativeStepSize; |
| 127 | } |
| 128 | |
| 129 | // For each parameter in the parameter block, use finite differences to |
| 130 | // compute the derivative for that parameter. |
| 131 | for (int j = 0; j < parameter_block_size; ++j) { |
| 132 | if (step_size(j) == 0.0) { |
| 133 | // The parameter is exactly zero, so compromise and use the mean |
| 134 | // step_size from the other parameters. This can break in many cases, |
| 135 | // but it's hard to pick a good number without problem specific |
| 136 | // knowledge. |
| 137 | step_size(j) = fallback_step_size; |
| 138 | } |
| 139 | x_plus_delta(j) = x(j) + step_size(j); |
| 140 | |
| 141 | double residuals[num_residuals]; // NOLINT |
| 142 | if (!function->Evaluate(parameters, residuals, NULL)) { |
| 143 | // Something went wrong; bail. |
| 144 | return false; |
| 145 | } |
| 146 | |
| 147 | // Compute this column of the jacobian in 3 steps: |
| 148 | // 1. Store residuals for the forward part. |
| 149 | // 2. Subtract residuals for the backward (or 0) part. |
| 150 | // 3. Divide out the run. |
| 151 | parameter_jacobian.col(j) = |
| 152 | Map<const ResidualVector>(residuals, num_residuals); |
| 153 | |
| 154 | double one_over_h = 1 / step_size(j); |
| 155 | if (method == CENTRAL) { |
| 156 | // Compute the function on the other side of x(j). |
| 157 | x_plus_delta(j) = x(j) - step_size(j); |
| 158 | |
| 159 | if (!function->Evaluate(parameters, residuals, NULL)) { |
| 160 | // Something went wrong; bail. |
| 161 | return false; |
| 162 | } |
| 163 | parameter_jacobian.col(j) -= |
| 164 | Map<ResidualVector>(residuals, num_residuals, 1); |
| 165 | one_over_h /= 2; |
| 166 | } else { |
| 167 | // Forward difference only; reuse existing residuals evaluation. |
| 168 | parameter_jacobian.col(j) -= |
| 169 | Map<const ResidualVector>(residuals_at_eval_point, num_residuals); |
| 170 | } |
| 171 | x_plus_delta(j) = x(j); // Restore x_plus_delta. |
| 172 | |
| 173 | // Divide out the run to get slope. |
| 174 | parameter_jacobian.col(j) *= one_over_h; |
| 175 | } |
| 176 | return true; |
| 177 | } |
| 178 | }; |
| 179 | |
| 180 | // Prevent invalid instantiations. |
| 181 | template <typename CostFunctionNoJacobian, |
| 182 | int num_residuals, |
| 183 | int parameter_block, |
| 184 | NumericDiffMethod method> |
| 185 | struct Differencer<CostFunctionNoJacobian, |
| 186 | num_residuals, |
| 187 | 0 /* parameter_block_size */, |
| 188 | parameter_block, |
| 189 | method> { |
| 190 | static bool EvaluateJacobianForParameterBlock( |
| 191 | const CostFunctionNoJacobian *function, |
| 192 | double const* residuals_at_eval_point, |
| 193 | double **parameters, |
| 194 | double **jacobians) { |
| 195 | LOG(FATAL) << "Shouldn't get here."; |
| 196 | return true; |
| 197 | } |
| 198 | }; |
| 199 | |
| 200 | template <typename CostFunctionNoJacobian, |
| 201 | NumericDiffMethod method = CENTRAL, int M = 0, |
| 202 | int N0 = 0, int N1 = 0, int N2 = 0, int N3 = 0, int N4 = 0, int N5 = 0> |
| 203 | class NumericDiffCostFunction |
| 204 | : public SizedCostFunction<M, N0, N1, N2, N3, N4, N5> { |
| 205 | public: |
| 206 | NumericDiffCostFunction(CostFunctionNoJacobian* function, |
| 207 | Ownership ownership) |
| 208 | : function_(function), ownership_(ownership) {} |
| 209 | |
| 210 | virtual ~NumericDiffCostFunction() { |
| 211 | if (ownership_ != TAKE_OWNERSHIP) { |
| 212 | function_.release(); |
| 213 | } |
| 214 | } |
| 215 | |
| 216 | virtual bool Evaluate(double const* const* parameters, |
| 217 | double* residuals, |
| 218 | double** jacobians) const { |
| 219 | // Get the function value (residuals) at the the point to evaluate. |
| 220 | bool success = function_->Evaluate(parameters, residuals, NULL); |
| 221 | if (!success) { |
| 222 | // Something went wrong; ignore the jacobian. |
| 223 | return false; |
| 224 | } |
| 225 | if (!jacobians) { |
| 226 | // Nothing to do; just forward. |
| 227 | return true; |
| 228 | } |
| 229 | |
| 230 | // Create a copy of the parameters which will get mutated. |
| 231 | const int kParametersSize = N0 + N1 + N2 + N3 + N4 + N5; |
| 232 | double parameters_copy[kParametersSize]; |
| 233 | double *parameters_references_copy[6]; |
| 234 | parameters_references_copy[0] = ¶meters_copy[0]; |
| 235 | parameters_references_copy[1] = ¶meters_copy[0] + N0; |
| 236 | parameters_references_copy[2] = ¶meters_copy[0] + N0 + N1; |
| 237 | parameters_references_copy[3] = ¶meters_copy[0] + N0 + N1 + N2; |
| 238 | parameters_references_copy[4] = ¶meters_copy[0] + N0 + N1 + N2 + N3; |
| 239 | parameters_references_copy[5] = |
| 240 | ¶meters_copy[0] + N0 + N1 + N2 + N3 + N4; |
| 241 | |
| 242 | #define COPY_PARAMETER_BLOCK(block) \ |
| 243 | if (N ## block) memcpy(parameters_references_copy[block], \ |
| 244 | parameters[block], \ |
| 245 | sizeof(double) * N ## block); // NOLINT |
| 246 | COPY_PARAMETER_BLOCK(0); |
| 247 | COPY_PARAMETER_BLOCK(1); |
| 248 | COPY_PARAMETER_BLOCK(2); |
| 249 | COPY_PARAMETER_BLOCK(3); |
| 250 | COPY_PARAMETER_BLOCK(4); |
| 251 | COPY_PARAMETER_BLOCK(5); |
| 252 | #undef COPY_PARAMETER_BLOCK |
| 253 | |
| 254 | #define EVALUATE_JACOBIAN_FOR_BLOCK(block) \ |
| 255 | if (N ## block && jacobians[block]) { \ |
| 256 | if (!Differencer<CostFunctionNoJacobian, /* NOLINT */ \ |
| 257 | M, \ |
| 258 | N ## block, \ |
| 259 | block, \ |
| 260 | method>::EvaluateJacobianForParameterBlock( \ |
| 261 | function_.get(), \ |
| 262 | residuals, \ |
| 263 | parameters_references_copy, \ |
| 264 | jacobians)) { \ |
| 265 | return false; \ |
| 266 | } \ |
| 267 | } |
| 268 | EVALUATE_JACOBIAN_FOR_BLOCK(0); |
| 269 | EVALUATE_JACOBIAN_FOR_BLOCK(1); |
| 270 | EVALUATE_JACOBIAN_FOR_BLOCK(2); |
| 271 | EVALUATE_JACOBIAN_FOR_BLOCK(3); |
| 272 | EVALUATE_JACOBIAN_FOR_BLOCK(4); |
| 273 | EVALUATE_JACOBIAN_FOR_BLOCK(5); |
| 274 | #undef EVALUATE_JACOBIAN_FOR_BLOCK |
| 275 | return true; |
| 276 | } |
| 277 | |
| 278 | private: |
| 279 | internal::scoped_ptr<CostFunctionNoJacobian> function_; |
| 280 | Ownership ownership_; |
| 281 | }; |
| 282 | |
| 283 | } // namespace ceres |
| 284 | |
| 285 | #endif // CERES_PUBLIC_NUMERIC_DIFF_COST_FUNCTION_H_ |