search

Alibaba-MIIL / ASL

Official Pytorch Implementation of: "Asymmetric Loss For Multi-Label Classification"(ICCV, 2021) paper
MIT License
722 stars 101 forks source link

What is the value of FocalLoss alpha hyperprameter in ASL experiment? #57

Closed sunlanchang closed 3 years ago

sunlanchang commented 3 years ago

In original FocalLoss paper, there are two hyperparameter in FocalLoss which are alpha and gamma. But I can't find alpha hyperparameter in your paper, so I want to know what is the value of alpha. Thank a lot!

mrT23 commented 3 years ago

The article discusses it, read thoroughly (chapter 2.2)

sunlanchang commented 3 years ago

In chaper 2.2 of ASL paper, it said

In practice, [19] even suggested a weighting factor which favors background samples (α = 0.25). In section 3 we will show that simple linear weighting is insufficient to tackle the negative-positive imbalance issue in multi-label classification properly. For those reasons, we chose to avoid adding static weighting factors to our focusing formulation.

Is that equally mean the hyperparameter alpha of focal loss in figure5 is being set to 1.0? Thanks a lot!

mrT23 commented 3 years ago

since the weighting is (α,1-α), equal factor is α=0.5

(or just remove alpha from focal loss equations, we are not using it to favor one over the other)

sunlanchang commented 3 years ago

since the weighting is (α,1-α), equal factor is α=0.5

(or just remove alpha from focal loss equations, we are not using it to favor one over the other)

OK, I get you, thans a lot.