[Work Log] TULIPS: index issue

April 02, 2014

Tested gradient and hessian -- is it correct with new model? ... Yes.

Is index optimization working? Quick test: ran this code twice, once using analytical hessian, once with hessian-free algorithm:

[mu_, Tracks_] = get_wacv_result(5, params_trained.model_type, 'deleteme_new2');
out = optimize_ml_wrt_indices(Tracks_(14), [], data, params_trained);

Analytical method halts early -- can't improve. Final gradient is noticibly non-zero

Hessian-free method seems very effective:

Perhaps the optimization is failing; let's run end-to-end with hessian-free algorithm. It will be super slow, but if it gives better results, we may have our solution.

[mu_, Tracks_] = get_wacv_result(5, params_trained.model_type, 'deleteme_new3');
Posted by Kyle Simek
blog comments powered by Disqus