No need to write robust_pow function?

[thedjofsis]

Hello.
Firstly, the two variables used as g2 and g5 are the same. So I think the mathematical equations can be represented in 4 variables instead of 5 and so can be used in the code. As per the code here
g2 = label + ((-1) ** label) * sigmoid_pred
g5 = label + ((-1) ** label) * sigmoid_pred
Also, the only term being used when calling the function robust_pow to calculate the exponent for, is the g2 (or g5) and that value can never be negative. Here there are two cases, either the label is 0 or 1.
In case of 0, g2 becomes sigmoid_pred with range 0 to 1.
In case of 1, g2 becomes 1-sigmoid_pred with the range 1 to 0.
So I don't see the point of having a separate function than using the normal exponent operator of **. In case I'm wrong, please correct me.

[thedjofsis]

Also robust_pow function is not exactly correct. I am wondering how does it change the calculations because as per this code, (-16)0.25 will be -2. But -24 is 16. And in case this is needed in fact, how about using the answer here ?

[jhwjhw0123]

Hi, thanks for the comments. After checking the codes, I believe the robust_pow is the legacy of a previous (scrap) version with the wrong math derivation. That results in a (wrong) procedure which will incur (a negative number) ** (a fraction number).
In terms of the implementation of the focal loss, this shouldn't have any effect -- as you said g_{2} can only be [0,1]. I'll fix this in a later version. As for now, the performance of the software should be ok.