Open Taruio opened 5 years ago
Good question, moreover shouldn't we use gamma=0 since the author advised not to use focal loss? Things We tried that didn't work: Focal loss. We tried using focal loss. It dropped our mAP about 2 points. YOLOv3 may already be robust to the problem focal loss is trying to solve because it has separate objectness predictions and conditional class predictions. Thus for most examples there is no loss from the class predictions? Or something? We aren’t totally sure. (https://arxiv.org/abs/1804.02767)
@RubenS02 I saw that in YOLOv3 article too. And alpha=1 means it just use a half of focal loss. I'm traing both gamma=0 and gamma=2 model. If I get something I'll write it here
@Taruio Thanks I would appreciate that a lot
In function
loss_layer()
, I find out that you use a coefficientconf_focal
. I would like to know is this a Focal Loss? Whyalpha
in focal loss is always 1? Thanks