The RMI loss provides negative loss

Question

The RMI loss provides negative loss

koda12344505 opened this issue 5 years ago · comments

koda12344505 commented 5 years ago

Hello,
Firstly thanks for this work

I'm currently working with this RMI loss on my own segmentation toolbox,
but i found the RMI loss provide negative loss

I just copied all codes from rmi.py and rmi_utils.py, then use this RMI loss instead of cross entropy loss

is it normal issue that RMI loss provide negative loss at the beginning of training?

thanks you

xiong224 · Answer 1 · Wed Dec 18 2019 11:52:57 GMT+0800 (China Standard Time)

I have met the same problems

Shuai Zhao · Answer 2 · Fri Dec 20 2019 16:27:42 GMT+0800 (China Standard Time)

This phenomenon is normal. -log det (A) can definitely get a negative value when A is a positive definite matrix. You can also check the value of the det(A) by printing it.

The sign of the loss does not matter, the gradient matters.
You can add -100 to the normal cross entropy loss, then your loss will be a negative value, but it will not change the training process and results.

Yu Xianwang · Answer 3 · Sun Apr 17 2022 16:52:23 GMT+0800 (China Standard Time)

“-log det (A) can definitely get a negative value when A is a positive definite matrix.” I can't understand this sentence. For example, A=Diag(0.1,0.1,0.1)，A is a positive definite matrix, but det(A)=0.001, -log(det(A)) is a positive value.