detach during salient mode

[fdengmark]

Hi:

Nice Work.

I notice that you detach the backprop after block layers when training your joint_poolnet with mode == 1, which is the salient mode. So the deeppool layer and your backbone only gets update during the edge mode. Can you explain the reason behind it?

Thank you.

[backseason]

During jointly training with edge, we try to use the edge features extracted from the edge branch to help refine the details around the edge areas of the salient features. By detaching the edge features, we can avoid the influence caused by the supervision information from the salient branch, as the edge branch aims to learn all possible edge areas while the salient branch only cares the salient objects. fdengmark <notifications@github.com> 于2020年11月27日周五 下午10:15写道：

…

Hi: Nice Work. I notice that you detach the backprop after block layers when training your joint_poolnet with mode == 1, which is the salient mode. So the deeppool layer and your backbone only gets update during the edge mode. Can you explain the reason behind it? Thank you. — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#74>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ABYO2OBDKBKXAWYHIGT6RHDSR6YAXANCNFSM4UE6V2TQ> .

-- From BackSeason-刘姜江 Major in Computer Science and Technology Nankai University, Tianjin,300350 P.R.China J04.liu@gmail.com <j04.liu@gmail.com>

[fdengmark]

I understand your idea, but all the parameters of backbone/convert_layers/deep_pool_layers will not be updated from salinet branch right? The objective of this network is salient object segmentation, and the majority of the parameters of the network will not be update from salinet obj segmentation task. I'm still confused by this

[backseason]

It will be updated by this line https://github.com/backseason/PoolNet/blob/16b77d9e9f85bbd862d2f9be43d9d6f2eb9c8e95/networks/joint_poolnet.py#L193 . fdengmark <notifications@github.com> 于2020年11月30日周一 下午4:53写道：

…

I understand your idea, but all the parameters of backbone/convert_layers/deep_pool_layers will not be updated from salinet branch right? The objective of this network is salient object segmentation, and the majority of the parameters of the network will not be update from salinet obj segmentation task. I'm still confused by this — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#74 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ABYO2OC7PW5JRGK26VQNPC3SSNMPVANCNFSM4UE6V2TQ> .

-- From BackSeason-刘姜江 Major in Computer Science and Technology Nankai University, Tianjin,300350 P.R.China J04.liu@gmail.com <j04.liu@gmail.com>