gtsam/gtsam/nonlinear/factorTesting.h

118 lines
4.4 KiB
C++

/* ----------------------------------------------------------------------------
* GTSAM Copyright 2010, Georgia Tech Research Corporation,
* Atlanta, Georgia 30332-0415
* All Rights Reserved
* Authors: Frank Dellaert, et al. (see THANKS for the full author list)
* See LICENSE for the license information
* -------------------------------------------------------------------------- */
/**
* @file factorTesting.h
* @date September 18, 2014
* @author Frank Dellaert
* @author Paul Furgale
* @brief Evaluate derivatives of a nonlinear factor numerically
*/
#pragma once
#include <gtsam/nonlinear/NonlinearFactor.h>
#include <gtsam/base/numericalDerivative.h>
#include <string>
#include <vector>
namespace gtsam {
/**
* Linearize a nonlinear factor using numerical differentiation
* The benefit of this method is that it does not need to know what types are
* involved to evaluate the factor. If all the machinery of gtsam is working
* correctly, we should get the correct numerical derivatives out the other side.
* NOTE(frank): factors that have non vector-space measurements use between or LocalCoordinates
* to evaluate the error, and their derivatives will only be correct for near-zero errors.
* This is fixable but expensive, and does not matter in practice as most factors will sit near
* zero errors anyway. However, it means that below will only be exact for the correct measurement.
*/
inline JacobianFactor linearizeNumerically(const NoiseModelFactor& factor,
const Values& values,
double delta = 1e-5) {
// We will fill a vector of key/Jacobians pairs (a map would sort)
std::vector<std::pair<Key, Matrix> > jacobians;
// Get size
const Vector e = factor.whitenedError(values);
const size_t rows = e.size();
// Loop over all variables
const double one_over_2delta = 1.0 / (2.0 * delta);
for (Key key : factor) {
// Compute central differences using the values struct.
VectorValues dX = values.zeroVectors();
const size_t cols = dX.dim(key);
Matrix J = Matrix::Zero(rows, cols);
for (size_t col = 0; col < cols; ++col) {
Vector dx = Vector::Zero(cols);
dx(col) = delta;
dX[key] = dx;
Values eval_values = values.retract(dX);
const Vector left = factor.whitenedError(eval_values);
dx(col) = -delta;
dX[key] = dx;
eval_values = values.retract(dX);
const Vector right = factor.whitenedError(eval_values);
J.col(col) = (left - right) * one_over_2delta;
}
jacobians.emplace_back(key, J);
}
// Next step...return JacobianFactor
return JacobianFactor(jacobians, -e);
}
namespace internal {
// CPPUnitLite-style test for linearization of a factor
inline bool testFactorJacobians(const std::string& name_,
const NoiseModelFactor& factor,
const gtsam::Values& values, double delta,
double tolerance) {
// Create expected value by numerical differentiation
JacobianFactor expected = linearizeNumerically(factor, values, delta);
// Create actual value by linearize
auto actual =
boost::dynamic_pointer_cast<JacobianFactor>(factor.linearize(values));
if (!actual) return false;
// Check cast result and then equality
bool equal = assert_equal(expected, *actual, tolerance);
// if not equal, test individual jacobians:
if (!equal) {
for (size_t i = 0; i < actual->size(); i++) {
bool i_good =
assert_equal((Matrix)(expected.getA(expected.begin() + i)),
(Matrix)(actual->getA(actual->begin() + i)), tolerance);
if (!i_good) {
std::cout << "Mismatch in Jacobian " << i + 1
<< " (base 1), as shown above" << std::endl;
}
}
}
return equal;
}
} // namespace internal
/// \brief Check the Jacobians produced by a factor against finite differences.
/// \param factor The factor to test.
/// \param values Values filled in for testing the Jacobians.
/// \param numerical_derivative_step The step to use when computing the numerical derivative Jacobians
/// \param tolerance The numerical tolerance to use when comparing Jacobians.
#define EXPECT_CORRECT_FACTOR_JACOBIANS(factor, values, numerical_derivative_step, tolerance) \
{ EXPECT(gtsam::internal::testFactorJacobians(name_, factor, values, numerical_derivative_step, tolerance)); }
} // namespace gtsam