Incorrect Focal-R mse loss?

[5uperpalo]

Hi authors,

page 6 from your paper:
Precisely, Focal-R loss based on L1 distance can be written as 1/n∑n i=1 σ(|βei|)γ ei, where ei is the L1 error for i-th sample, σ(·)

• QUESTION 1 : in the focal_mse loss:
  

  imbalanced-regression/agedb-dir/loss.py

  Line 24 in 055a7b3

  (2 * torch.sigmoid(beta * torch.abs(inputs - targets)) - 1) ** gamma

  
  should be torch.abs((inputs - targets)**2) and not only torch.abs(inputs - targets), am I correct?
• QUESTION 2 : why is there 2*torch.abs(...)-1 ? you do not have and -1 or 2* in the function in your paper?

[YyzHarry]

QUESTION 1 : in the focal_mse loss...

This is a good catch. Since there's already a hyper-parameter beta to control the "scaling" effect, we did not use the squared error -- but one could always try that (and tune the value of beta correspondingly).

QUESTION 2 : why is there 2*torch.abs(...)-1

This is because sigmoid(x) is in [0.5, 1] for all x>=0, so we simple scale it back to [0, 1]. If you choose activation as 'tanh', you do not need to rescale it. Since it's an engineering issue, for cleanness, we did not expand the detail in paper.

[5uperpalo]

thank you! now it's clear to me :)