Adaptive_Affinity_Fields icon indicating copy to clipboard operation
Adaptive_Affinity_Fields copied to clipboard

Published 20 hours ago • verified publisher icon twke18

if "aff_loss_edge" and "aff_loss_nedge" do not have a coefficient of "dec" that decreases with "step", can "aff_loss" converge?

Open ef-ever opened this issue 6 years ago • 1 comments

Excuse me, if "aff_loss_edge" and "aff_loss_nedge" do not have a coefficient of "dec" that decreases with "step", can "aff_loss" converge? When I run "train_affinity.py" with the parameters mentioned in "train_pspnet_affinity.sh", "aff_loss_edge" will not fall until 2.88.

ef-ever avatar Oct 17 '18 11:10 ef-ever

Yes, according to our empirical experience, it is necessary to exponentially decrease the affinity field loss during training.

twke18 avatar Oct 17 '18 21:10 twke18