Add IterationSummary::gradient_norm.
Iteration summary now reports the 2-norm of the gradient also.
Change-Id: I1ed7f1456ee4f546c9b42423d7a4ec3079ec078f
diff --git a/docs/source/solving.rst b/docs/source/solving.rst
index f17c695..1ab3dba 100644
--- a/docs/source/solving.rst
+++ b/docs/source/solving.rst
@@ -1588,6 +1588,9 @@
// Infinity norm of the gradient vector.
double gradient_max_norm;
+ // 2-norm of the gradient vector.
+ double gradient_norm;
+
// 2-norm of the size of the step computed by the optimization
// algorithm.
double step_norm;
diff --git a/include/ceres/iteration_callback.h b/include/ceres/iteration_callback.h
index 987c2d9..5689256 100644
--- a/include/ceres/iteration_callback.h
+++ b/include/ceres/iteration_callback.h
@@ -50,6 +50,7 @@
cost(0.0),
cost_change(0.0),
gradient_max_norm(0.0),
+ gradient_norm(0.0),
step_norm(0.0),
eta(0.0),
step_size(0.0),
@@ -100,6 +101,9 @@
// Infinity norm of the gradient vector.
double gradient_max_norm;
+ // 2-norm of the gradient vector.
+ double gradient_norm;
+
// 2-norm of the size of the step computed by the optimization
// algorithm.
double step_norm;
diff --git a/internal/ceres/line_search_minimizer.cc b/internal/ceres/line_search_minimizer.cc
index 6ee514a..b7e96c8 100644
--- a/internal/ceres/line_search_minimizer.cc
+++ b/internal/ceres/line_search_minimizer.cc
@@ -119,6 +119,7 @@
iteration_summary.step_is_successful = false;
iteration_summary.cost_change = 0.0;
iteration_summary.gradient_max_norm = 0.0;
+ iteration_summary.gradient_norm = 0.0;
iteration_summary.step_norm = 0.0;
iteration_summary.linear_solver_iterations = 0;
iteration_summary.step_solver_time_in_seconds = 0;
@@ -135,6 +136,7 @@
iteration_summary.cost = current_state.cost + summary->fixed_cost;
iteration_summary.gradient_max_norm = current_state.gradient_max_norm;
+ iteration_summary.gradient_norm = sqrt(current_state.gradient_squared_norm);
// The initial gradient max_norm is bounded from below so that we do
// not divide by zero.
@@ -331,6 +333,8 @@
}
iteration_summary.gradient_max_norm = current_state.gradient_max_norm;
+ iteration_summary.gradient_norm = sqrt(current_state.gradient_squared_norm);
+
if (iteration_summary.gradient_max_norm <= absolute_gradient_tolerance) {
VLOG_IF(1, is_not_silent)
<< "Terminating: Gradient tolerance reached."
diff --git a/internal/ceres/trust_region_minimizer.cc b/internal/ceres/trust_region_minimizer.cc
index 81dc3e1..ea7ee74 100644
--- a/internal/ceres/trust_region_minimizer.cc
+++ b/internal/ceres/trust_region_minimizer.cc
@@ -113,6 +113,7 @@
iteration_summary.step_is_successful = false;
iteration_summary.cost_change = 0.0;
iteration_summary.gradient_max_norm = 0.0;
+ iteration_summary.gradient_norm = 0.0;
iteration_summary.step_norm = 0.0;
iteration_summary.relative_decrease = 0.0;
iteration_summary.trust_region_radius = strategy->Radius();
@@ -145,6 +146,7 @@
summary->initial_cost = cost + summary->fixed_cost;
iteration_summary.cost = cost + summary->fixed_cost;
iteration_summary.gradient_max_norm = gradient.lpNorm<Eigen::Infinity>();
+ iteration_summary.gradient_norm = gradient.norm();
// The initial gradient max_norm is bounded from below so that we do
// not divide by zero.
@@ -283,6 +285,8 @@
iteration_summary.cost_change = 0.0;
iteration_summary.gradient_max_norm =
summary->iterations.back().gradient_max_norm;
+ iteration_summary.gradient_norm =
+ summary->iterations.back().gradient_norm;
iteration_summary.step_norm = 0.0;
iteration_summary.relative_decrease = 0.0;
iteration_summary.eta = options_.eta;
@@ -478,6 +482,7 @@
}
iteration_summary.gradient_max_norm = gradient.lpNorm<Eigen::Infinity>();
+ iteration_summary.gradient_norm = gradient.norm();
if (iteration_summary.gradient_max_norm <= absolute_gradient_tolerance) {
VLOG_IF(1, is_not_silent) << "Terminating: Gradient tolerance reached."
