Distribution-Balanced Loss
P
I
(
x
k
)
P^I(x^k)
PI(xk)=
1
C
∑
y
i
k
=
1
1
n
i
{1\over C}\sum\limits_{y_i^k=1}{1\over{n_i}}
C1yik=1∑ni1，
P
i
C
(
x
k
)
P^C_i(x^k)
PiC(xk)=
1
C
1
n
i
{1\over C}{1\over{n_i}}
C1ni1
r
i
k
_i^k
ik=
P
i
C
(
x
k
)
P
I
(
x
k
)
{P^C_i(x^k)}\over{P^I(x^k)}
PI(xk)PiC(xk)
NTR means ‘negative-tolerant regularization’，R-BCE performs a re-balanced weighting strategy
辅助信息
作者使用的’MMDataParallel’ has been removed，不建议复现
