| 1 | /* ---------------------------------------------------------------------------- |
| 2 | |
| 3 | * GTSAM Copyright 2010, Georgia Tech Research Corporation, |
| 4 | * Atlanta, Georgia 30332-0415 |
| 5 | * All Rights Reserved |
| 6 | * Authors: Frank Dellaert, et al. (see THANKS for the full author list) |
| 7 | |
| 8 | * See LICENSE for the license information |
| 9 | |
| 10 | * -------------------------------------------------------------------------- */ |
| 11 | |
| 12 | /** |
| 13 | * @file LinearizedFactor.cpp |
| 14 | * @brief A dummy factor that allows a linear factor to act as a nonlinear factor |
| 15 | * @author Alex Cunningham |
| 16 | */ |
| 17 | |
| 18 | #include <gtsam_unstable/nonlinear/LinearizedFactor.h> |
| 19 | #include <iostream> |
| 20 | #include <cassert> |
| 21 | |
| 22 | namespace gtsam { |
| 23 | |
| 24 | /* ************************************************************************* */ |
| 25 | LinearizedGaussianFactor::LinearizedGaussianFactor( |
| 26 | const GaussianFactor::shared_ptr& gaussian, const Values& lin_points) |
| 27 | : NonlinearFactor(gaussian->keys()) |
| 28 | { |
| 29 | // Extract the keys and linearization points |
| 30 | for(const Key& key: gaussian->keys()) { |
| 31 | // extract linearization point |
| 32 | assert(lin_points.exists(key)); |
| 33 | this->lin_points_.insert(j: key, val: lin_points.at(j: key)); |
| 34 | } |
| 35 | } |
| 36 | |
| 37 | /* ************************************************************************* */ |
| 38 | // LinearizedJacobianFactor |
| 39 | /* ************************************************************************* */ |
| 40 | LinearizedJacobianFactor::LinearizedJacobianFactor() { |
| 41 | } |
| 42 | |
| 43 | /* ************************************************************************* */ |
| 44 | LinearizedJacobianFactor::LinearizedJacobianFactor( |
| 45 | const JacobianFactor::shared_ptr& jacobian, const Values& lin_points) |
| 46 | : Base(jacobian, lin_points) { |
| 47 | |
| 48 | // Create the dims array |
| 49 | size_t *dims = (size_t *)alloca(sizeof(size_t) * (jacobian->size() + 1)); |
| 50 | size_t index = 0; |
| 51 | for(JacobianFactor::const_iterator iter = jacobian->begin(); iter != jacobian->end(); ++iter) { |
| 52 | dims[index++] = jacobian->getDim(variable: iter); |
| 53 | } |
| 54 | dims[index] = 1; |
| 55 | |
| 56 | // Update the BlockInfo accessor |
| 57 | Ab_ = VerticalBlockMatrix(dims, dims+jacobian->size()+1, jacobian->rows()); |
| 58 | // Get the Ab matrix from the Jacobian factor, with any covariance baked in |
| 59 | Ab_.matrix() = jacobian->augmentedJacobian(); |
| 60 | } |
| 61 | |
| 62 | /* ************************************************************************* */ |
| 63 | void LinearizedJacobianFactor::print(const std::string& s, const KeyFormatter& keyFormatter) const { |
| 64 | |
| 65 | std::cout << s << std::endl; |
| 66 | |
| 67 | std::cout << "Nonlinear Keys: " ; |
| 68 | for(const Key& key: this->keys()) |
| 69 | std::cout << keyFormatter(key) << " " ; |
| 70 | std::cout << std::endl; |
| 71 | |
| 72 | for(const_iterator key=begin(); key!=end(); ++key) { |
| 73 | std::cout << "A[" << keyFormatter(*key) << "]=\n" << A(key: *key) << std::endl; |
| 74 | } |
| 75 | std::cout << "b=\n" << b() << std::endl; |
| 76 | |
| 77 | lin_points_.print(str: "Linearization Point: " ); |
| 78 | } |
| 79 | |
| 80 | /* ************************************************************************* */ |
| 81 | bool LinearizedJacobianFactor::equals(const NonlinearFactor& expected, double tol) const { |
| 82 | |
| 83 | const This *e = dynamic_cast<const This*> (&expected); |
| 84 | if (e) { |
| 85 | |
| 86 | Matrix thisMatrix = this->Ab_.range(startBlock: 0, endBlock: Ab_.nBlocks()); |
| 87 | Matrix rhsMatrix = e->Ab_.range(startBlock: 0, endBlock: Ab_.nBlocks()); |
| 88 | |
| 89 | return Base::equals(f: expected, tol) |
| 90 | && lin_points_.equals(other: e->lin_points_, tol) |
| 91 | && equal_with_abs_tol(A: thisMatrix, B: rhsMatrix, tol); |
| 92 | } else { |
| 93 | return false; |
| 94 | } |
| 95 | } |
| 96 | |
| 97 | /* ************************************************************************* */ |
| 98 | double LinearizedJacobianFactor::error(const Values& c) const { |
| 99 | Vector errorVector = error_vector(c); |
| 100 | return 0.5 * errorVector.dot(other: errorVector); |
| 101 | } |
| 102 | |
| 103 | /* ************************************************************************* */ |
| 104 | std::shared_ptr<GaussianFactor> |
| 105 | LinearizedJacobianFactor::linearize(const Values& c) const { |
| 106 | |
| 107 | // Create the 'terms' data structure for the Jacobian constructor |
| 108 | std::vector<std::pair<Key, Matrix> > terms; |
| 109 | for(Key key: keys()) { |
| 110 | terms.push_back(x: std::make_pair(x&: key, y: this->A(key))); |
| 111 | } |
| 112 | |
| 113 | // compute rhs |
| 114 | Vector b = -error_vector(c); |
| 115 | |
| 116 | return std::shared_ptr<GaussianFactor>(new JacobianFactor(terms, b, noiseModel::Unit::Create(dim: dim()))); |
| 117 | } |
| 118 | |
| 119 | /* ************************************************************************* */ |
| 120 | Vector LinearizedJacobianFactor::error_vector(const Values& c) const { |
| 121 | |
| 122 | Vector errorVector = -b(); |
| 123 | |
| 124 | for(Key key: this->keys()) { |
| 125 | const Value& newPt = c.at(j: key); |
| 126 | const Value& linPt = lin_points_.at(j: key); |
| 127 | Vector d = linPt.localCoordinates_(value: newPt); |
| 128 | const constABlock A = this->A(key); |
| 129 | errorVector += A * d; |
| 130 | } |
| 131 | |
| 132 | return errorVector; |
| 133 | } |
| 134 | |
| 135 | /* ************************************************************************* */ |
| 136 | // LinearizedHessianFactor |
| 137 | /* ************************************************************************* */ |
| 138 | LinearizedHessianFactor::LinearizedHessianFactor() { |
| 139 | } |
| 140 | |
| 141 | /* ************************************************************************* */ |
| 142 | LinearizedHessianFactor::LinearizedHessianFactor( |
| 143 | const HessianFactor::shared_ptr& hessian, const Values& lin_points) |
| 144 | : Base(hessian, lin_points), info_(hessian->info()) {} |
| 145 | |
| 146 | /* ************************************************************************* */ |
| 147 | void LinearizedHessianFactor::print(const std::string& s, const KeyFormatter& keyFormatter) const { |
| 148 | |
| 149 | std::cout << s << std::endl; |
| 150 | |
| 151 | std::cout << "Nonlinear Keys: " ; |
| 152 | for(const Key& key: this->keys()) |
| 153 | std::cout << keyFormatter(key) << " " ; |
| 154 | std::cout << std::endl; |
| 155 | |
| 156 | gtsam::print(A: Matrix(info_.selfadjointView()), s: "Ab^T * Ab: " ); |
| 157 | |
| 158 | lin_points_.print(str: "Linearization Point: " ); |
| 159 | } |
| 160 | |
| 161 | /* ************************************************************************* */ |
| 162 | bool LinearizedHessianFactor::equals(const NonlinearFactor& expected, double tol) const { |
| 163 | |
| 164 | const This *e = dynamic_cast<const This*> (&expected); |
| 165 | if (e) { |
| 166 | |
| 167 | Matrix thisMatrix = this->info_.selfadjointView(); |
| 168 | thisMatrix(thisMatrix.rows()-1, thisMatrix.cols()-1) = 0.0; |
| 169 | Matrix rhsMatrix = e->info_.selfadjointView(); |
| 170 | rhsMatrix(rhsMatrix.rows()-1, rhsMatrix.cols()-1) = 0.0; |
| 171 | |
| 172 | return Base::equals(f: expected, tol) |
| 173 | && lin_points_.equals(other: e->lin_points_, tol) |
| 174 | && equal_with_abs_tol(A: thisMatrix, B: rhsMatrix, tol); |
| 175 | } else { |
| 176 | return false; |
| 177 | } |
| 178 | } |
| 179 | |
| 180 | /* ************************************************************************* */ |
| 181 | double LinearizedHessianFactor::error(const Values& c) const { |
| 182 | |
| 183 | // Construct an error vector in key-order from the Values |
| 184 | Vector dx = Vector::Zero(size: dim()); |
| 185 | size_t index = 0; |
| 186 | for(unsigned int i = 0; i < this->size(); ++i){ |
| 187 | Key key = this->keys()[i]; |
| 188 | const Value& newPt = c.at(j: key); |
| 189 | const Value& linPt = lin_points_.at(j: key); |
| 190 | dx.segment(start: index, n: linPt.dim()) = linPt.localCoordinates_(value: newPt); |
| 191 | index += linPt.dim(); |
| 192 | } |
| 193 | |
| 194 | // error 0.5*(f - 2*x'*g + x'*G*x) |
| 195 | double f = constantTerm(); |
| 196 | double xtg = dx.dot(other: linearTerm()); |
| 197 | double xGx = dx.transpose() * squaredTerm() * dx; |
| 198 | |
| 199 | return 0.5 * (f - 2.0 * xtg + xGx); |
| 200 | } |
| 201 | |
| 202 | /* ************************************************************************* */ |
| 203 | std::shared_ptr<GaussianFactor> |
| 204 | LinearizedHessianFactor::linearize(const Values& c) const { |
| 205 | |
| 206 | // Construct an error vector in key-order from the Values |
| 207 | Vector dx = Vector::Zero(size: dim()); |
| 208 | size_t index = 0; |
| 209 | for(unsigned int i = 0; i < this->size(); ++i){ |
| 210 | Key key = this->keys()[i]; |
| 211 | const Value& newPt = c.at(j: key); |
| 212 | const Value& linPt = lin_points_.at(j: key); |
| 213 | dx.segment(start: index, n: linPt.dim()) = linPt.localCoordinates_(value: newPt); |
| 214 | index += linPt.dim(); |
| 215 | } |
| 216 | |
| 217 | // f2 = f1 - 2*dx'*g1 + dx'*G1*dx |
| 218 | //newInfo(this->size(), this->size())(0,0) += -2*dx.dot(linearTerm()) + dx.transpose() * squaredTerm().selfadjointView<Eigen::Upper>() * dx; |
| 219 | double f = constantTerm() - 2*dx.dot(other: linearTerm()) + dx.transpose() * squaredTerm() * dx; |
| 220 | |
| 221 | // g2 = g1 - G1*dx |
| 222 | //newInfo.rangeColumn(0, this->size(), this->size(), 0) -= squaredTerm().selfadjointView<Eigen::Upper>() * dx; |
| 223 | Vector g = linearTerm() - squaredTerm() * dx; |
| 224 | std::vector<Vector> gs; |
| 225 | std::size_t offset = 0; |
| 226 | for(DenseIndex i = 0; i < info_.nBlocks()-1; ++i) { |
| 227 | const std::size_t dim = info_.getDim(block: i); |
| 228 | gs.push_back(x: g.segment(start: offset, n: dim)); |
| 229 | offset += dim; |
| 230 | } |
| 231 | |
| 232 | // G2 = G1 |
| 233 | // Do Nothing |
| 234 | std::vector<Matrix> Gs; |
| 235 | for(DenseIndex i = 0; i < info_.nBlocks()-1; ++i) { |
| 236 | Gs.push_back(x: info_.diagonalBlock(J: i)); |
| 237 | for(DenseIndex j = i + 1; j < info_.nBlocks()-1; ++j) { |
| 238 | Gs.push_back(x: info_.aboveDiagonalBlock(I: i, J: j)); |
| 239 | } |
| 240 | } |
| 241 | |
| 242 | // Create a Hessian Factor from the modified info matrix |
| 243 | //return std::shared_ptr<GaussianFactor>(new HessianFactor(js, newInfo)); |
| 244 | return std::shared_ptr<GaussianFactor>(new HessianFactor(keys(), Gs, gs, f)); |
| 245 | } |
| 246 | |
| 247 | } // \namespace aspn |
| 248 | |