Can't see asymmetric probability margin implementation in loss function.

[sunlanchang]

After reading ASL paper and source code, I am trying to find source code about asymmetric probability margin. But I can't see any code with respect to asymmetric probability margin in losses.py. Maybe bellow source code is about asymmetric probability margin?

# Asymmetric Clipping
if self.clip is not None and self.clip > 0:
    xs_neg = (xs_neg + self.clip).clamp(max=1)

But in ASL paper, asymmetric probability margin formula for negative sample is pm=max(p−m, 0), the source code does not make sense. I think the source code should be bellow:

# Asymmetric Clipping
if self.clip is not None and self.clip > 0:
    xs_neg = (xs_neg - self.clip).clamp(min=0)

I am not certain my understanding is right. Maybe I missed something, Please help me. Thanks!!!

[mrT23]

see:
#10

[sunlanchang]

I can see define variable xs_pos and xs_neg to prevent calculating again and again (1-p) along the code. But I still can't understand why xs_neg = (xs_neg + self.clip).clamp(max=1) is equal to p_m = max(p-m, 0). Could you please describe the reason in details. Thanks a million.

[mrT23]

put actual numbers instead of a formula:

a sample got, for some label, probability 0.3, meaning p_xs_neg=1-0.3=0.7
if the ground-truth for the label is indeed negative, our target is: p_target_neg=1

with margin of 0.3 and above, we shouldnt have a loss. how do it with math:
xs_neg = (xs_neg + self.clip).clamp(max=1)
(don't forget - if the target is indeed negative, we want xs_neg to be 1, not 0, for not propagating a loss)

[sunlanchang]

I can see, I'll close this issue. Thank you most sincerely.