Adding VLOG output to line search.
- Previously line search was sparse in terms of debug orientated VLOG
output which made debugging failure cases difficult.
Change-Id: Idfabf74d2b3f7b8256f79dff8c6b7fcdc2fcf4d3
diff --git a/internal/ceres/line_search.cc b/internal/ceres/line_search.cc
index 2601398..c62cda7 100644
--- a/internal/ceres/line_search.cc
+++ b/internal/ceres/line_search.cc
@@ -384,6 +384,13 @@
return;
}
+ VLOG(3) << std::scientific << std::setprecision(kErrorMessageNumericPrecision)
+ << "Starting line search zoom phase with bracket_low: "
+ << bracket_low << ", bracket_high: " << bracket_high
+ << ", bracket width: " << fabs(bracket_low.x - bracket_high.x)
+ << ", bracket abs delta cost: "
+ << fabs(bracket_low.value - bracket_high.value);
+
// Wolfe Zoom phase: Called when the Bracketing phase finds an interval of
// non-zero, finite width that should bracket step sizes which satisfy the
// (strong) Wolfe conditions (before finding a step size that satisfies the
@@ -493,6 +500,10 @@
*do_zoom_search = true;
*bracket_low = previous;
*bracket_high = current;
+ VLOG(3) << std::scientific << std::setprecision(kErrorMessageNumericPrecision)
+ << "Bracket found: current step (" << current.x
+ << ") violates Armijo sufficient condition, or has passed an "
+ << "inflection point of f() based on value.";
break;
}
@@ -503,6 +514,10 @@
// valid termination point, therefore a Zoom not required.
*bracket_low = current;
*bracket_high = current;
+ VLOG(3) << std::scientific << std::setprecision(kErrorMessageNumericPrecision)
+ << "Bracketing phase found step size: " << current.x
+ << ", satisfying strong Wolfe conditions, initial_position: "
+ << initial_position << ", current: " << current;
break;
} else if (current.value_is_valid && current.gradient >= 0) {
@@ -515,6 +530,9 @@
// Note inverse ordering from first bracket case.
*bracket_low = current;
*bracket_high = previous;
+ VLOG(3) << "Bracket found: current step (" << current.x
+ << ") satisfies Armijo, but has gradient >= 0, thus have passed "
+ << "an inflection point of f().";
break;
} else if (current.value_is_valid &&
@@ -757,6 +775,12 @@
return false;
}
+ VLOG(3) << "Zoom iteration: "
+ << summary->num_iterations - num_bracketing_iterations
+ << ", bracket_low: " << bracket_low
+ << ", bracket_high: " << bracket_high
+ << ", minimizing solution: " << *solution;
+
if ((solution->value > (initial_position.value
+ options().sufficient_decrease
* initial_position.gradient
@@ -772,6 +796,9 @@
if (fabs(solution->gradient) <=
-options().sufficient_curvature_decrease * initial_position.gradient) {
// Found a valid termination point satisfying strong Wolfe conditions.
+ VLOG(3) << std::scientific << std::setprecision(kErrorMessageNumericPrecision)
+ << "Zoom phase found step size: " << solution->x
+ << ", satisfying strong Wolfe conditions.";
break;
} else if (solution->gradient * (bracket_high.x - bracket_low.x) >= 0) {
diff --git a/internal/ceres/line_search_direction.cc b/internal/ceres/line_search_direction.cc
index ced8da8..0230e8d 100644
--- a/internal/ceres/line_search_direction.cc
+++ b/internal/ceres/line_search_direction.cc
@@ -251,8 +251,13 @@
// Part II: Implementation and experiments, Management Science,
// 20(5), 863-874, 1974.
// [2] Nocedal J., Wright S., Numerical Optimization, Springer, 1999.
- inverse_hessian_ *=
+ const double approximate_eigenvalue_scale =
delta_x_dot_delta_gradient / delta_gradient.dot(delta_gradient);
+ inverse_hessian_ *= approximate_eigenvalue_scale;
+
+ VLOG(4) << "Applying approximate_eigenvalue_scale: "
+ << approximate_eigenvalue_scale << " to initial inverse "
+ << "Hessian approximation.";
}
initialized_ = true;
diff --git a/internal/ceres/low_rank_inverse_hessian.cc b/internal/ceres/low_rank_inverse_hessian.cc
index d675395..16d84c6 100644
--- a/internal/ceres/low_rank_inverse_hessian.cc
+++ b/internal/ceres/low_rank_inverse_hessian.cc
@@ -170,6 +170,10 @@
// 20(5), 863-874, 1974.
// [2] Nocedal J., Wright S., Numerical Optimization, Springer, 1999.
search_direction *= approximate_eigenvalue_scale_;
+
+ VLOG(4) << "Applying approximate_eigenvalue_scale: "
+ << approximate_eigenvalue_scale_ << " to initial inverse Hessian "
+ << "approximation.";
}
for (int i = 0; i < num_corrections_; ++i) {
