bottom-up-attention icon indicating copy to clipboard operation
bottom-up-attention copied to clipboard

Published 20 hours ago • verified publisher icon peteanderson80

Why choose "SoftmaxWithLoss" for "loss_attr"

Open YuanEZhou opened this issue 6 years ago • 4 comments

Hi @peteanderson80 , Thanks for your sharing.
There are some objects has multiple attributes, why choose "SoftmaxWithLoss" for "loss_attr"?

YuanEZhou avatar Nov 10 '18 02:11 YuanEZhou

Hi, did anybody figure out the reason about the loss function?

zhaohengz avatar Mar 25 '20 01:03 zhaohengz

Actually, the operation of this loss function is what we want. “SoftmaxWithLoss” was rewrited. You can check the souce code at this linkhttps://github.com/peteanderson80/bottom-up-attention/blob/master/caffe/src/caffe/layers/softmax_loss_layer.cpp

YuanEZhou avatar Mar 25 '20 02:03 YuanEZhou

Thanks! So what they implemented is something like softmaxBinaryCrossEntorpy. Did I get it right?

zhaohengz avatar Mar 25 '20 03:03 zhaohengz

yeah, this rewrited function try to maximize the sum of all attribures' score .

YuanEZhou avatar Mar 25 '20 08:03 YuanEZhou