negative losses

Open hhjjaaa opened this issue 1 year ago • 3 comments

Hi , I am finding negative losses while training with my own dataset. I would like to ask you what causes this. Snipaste_2024-10-31_16-21-27

Oct 31 '24 08:10 hhjjaaa

Hi, sorry for the late reply, I hope you have solved the issue, mostly this is due to incorrect data processing, e.g., incorrect GT mask value. Carefully checking the data would be a possible solution.

Dec 31 '24 09:12 McGregorWwww

hi, I conducted experiments on my other CT dataset, I got the negative loss too, but the results is good. Is there anything wrong? can you give me some advice to fix this? thanks

Jan 06 '25 02:01 Flash98-cc

Hi, my previous answer was wrong; since we use -dice as a loss rather than 1-dice, a negative loss is correct; if you prefer a positive loss, you can change it to 1-dice in the loss.py file.

Jan 06 '25 02:01 McGregorWwww