loss

Hello, excuse me. I would like to ask if the corresponding loss of weak label and strong label in your paper is Lc and Ls respectively. However, there is an item in the appendix that compares the corresponding results of Lc, Ls and Lc + Ls, how does Lc + Ls implement dene in the code？ Looking forward to hearing from you, Thanks.

Hello!

if the corresponding loss of weak label and strong label in your paper is Lc and Ls

Yes.

how does Lc + Ls implement dene in the code

ifsl/fs-cs/model/ifsl.py

Lines 57 to 60 in eb2cf96

    
           if self.weak: 
        
               loss = self.compute_cls_objective(shared_masks, batch['query_class_presence']) 
        
           else: 
        
               loss = self.compute_seg_objective(logit_seg, batch['query_mask'])

I simply hardcoded by adding the above two losses with 0.1 multiplied to the weak label loss (Lc).

Have a great day! 😃

Dahyun.

Thank you very much for your reply！

Regarding the hardcode you mentioned, can I understand the test indicator of ASNet Lc + Ls (as shown below) in the third line of the article, and directly train it by changing loss to loss = Ls + 0.1 * Lc?

Yes, exactly.

Ok, thank you very much for your reply, very helpful！

Hope to have the opportunity to communicate more.

Haha my pleasure, thank you for your attention on the work.
You are more than welcome to chat here or reach out to me.

	if self.weak:
	loss = self.compute_cls_objective(shared_masks, batch['query_class_presence'])
	else:
	loss = self.compute_seg_objective(logit_seg, batch['query_mask'])