Hello: This work is very interesting. The article is well-founded. I manually derived the formula in the paper and found that the expression of δ²_t in the paper seems to be problematic.

ZjjConan / SimAM

The official pytorch implemention of our ICML-2021 paper "SimAM: A Simple, Parameter-Free Attention Module for Convolutional Neural Networks".

409 stars 36 forks source link

Closed Suke0 closed 3 years ago

Suke0 commented 3 years ago

δ²_t is not equal to 1/(M-1) ∑(x_i - u_t)², it is equal to 1/(M-1) ∑(x²_i - u²_t).

ZjjConan commented 3 years ago

@Suke0, Thanks for your attention. Actually, you can expand (*) in the first equation.

Suke0 commented 3 years ago

ok，I’ve got it。think you。

Ema1997 commented 3 years ago

@Suke0, Thanks for your attention. Actually, you can expand (*) in the first equation.

There will still be a 2u_t x_i when expanding the first equation, so the two equations are still unequal? @ZjjConan @Suke0

Suke0 commented 3 years ago

They are equal. You should calculate more carefully.

Ema1997 commented 3 years ago

They are equal. You should calculate more carefully.

Ok, I got it. Thanks.