commit | eafeca5dcb7af8688d40a9c14b0d2fcb856c96fc | [log] [tgz] |
---|---|---|
author | Sameer Agarwal <sameeragarwal@google.com> | Mon Sep 28 11:12:59 2020 -0700 |
committer | Sameer Agarwal <sameeragarwal@google.com> | Mon Sep 28 12:19:45 2020 -0700 |
tree | 5efbf3f94dd6e77ff6a2b03161670ae1b7d1ec92 | |
parent | 1fd0be916dd4ff4241bd52264b9e9170bc7e4339 [diff] |
Fix a logging bug in TrustRegionMinimizer. Upon encountering an unsuccessful step (one where the cost goes up) the the trust region minimizer failed to populate the gradient norm in the IterationSummary. This would cause the gradient norm to be logged as zero which is incorrect. Instead it should be the gradient norm at the current point. This CL fixes this issue. Before: iter cost cost_change |gradient| |step| tr_ratio tr_radius ls_iter iter_time total_time 0 1.115206e+07 0.00e+00 1.90e+07 0.00e+00 0.00e+00 1.00e+04 0 2.72e-01 1.33e+00 1 3.687552e+06 7.46e+06 1.84e+08 2.86e+03 6.91e-01 1.06e+04 1 1.32e+00 2.65e+00 2 3.670266e+10 -3.67e+10 0.00e+00 3.27e+03 -1.07e+04 5.30e+03 1 7.52e-01 3.40e+00 3 4.335397e+07 -3.97e+07 0.00e+00 2.74e+03 -1.16e+01 1.32e+03 1 7.28e-01 4.13e+00 4 1.345488e+06 2.34e+06 4.12e+07 1.55e+03 6.87e-01 1.40e+03 1 9.31e-01 5.06e+00 5 5.376653e+05 8.08e+05 9.99e+06 6.64e+02 7.46e-01 1.59e+03 1 9.64e-01 6.03e+00 After: iter cost cost_change |gradient| |step| tr_ratio tr_radius ls_iter iter_time total_time 0 1.115206e+07 0.00e+00 1.90e+07 0.00e+00 0.00e+00 1.00e+04 0 2.37e-01 1.13e+00 1 3.687552e+06 7.46e+06 1.84e+08 2.86e+03 6.91e-01 1.06e+04 1 1.08e+00 2.21e+00 2 3.670266e+10 -3.67e+10 1.84e+08 3.27e+03 -1.07e+04 5.30e+03 1 7.50e-01 2.96e+00 3 4.335397e+07 -3.97e+07 1.84e+08 2.74e+03 -1.16e+01 1.32e+03 1 7.13e-01 3.67e+00 4 1.345488e+06 2.34e+06 4.12e+07 1.55e+03 6.87e-01 1.40e+03 1 9.01e-01 4.57e+00 5 5.376653e+05 8.08e+05 9.99e+06 6.64e+02 7.46e-01 1.59e+03 1 9.36e-01 5.51e+00 Change-Id: Iae538fe089be07c7bb219337a6f1392f7213acfe
Ceres Solver is an open source C++ library for modeling and solving large, complicated optimization problems. It is a feature rich, mature and performant library which has been used in production at Google since 2010. Ceres Solver can solve two kinds of problems.
Please see ceres-solver.org for more information.