W
m
e
a
n
=
∑
i
n
W
i
n
,
o
u
t
/
N
i
n
W_{mean} = \sum_{in} W_{in,out}/N_{in}
Wmean=in∑Win,out/Nin
W
′
=
W
−
W
m
e
a
n
W^{'}= W - W_{mean}
W′=W−Wmean
f
a
k
e
o
u
t
=
X
∗
W
′
fakeout = X*W^{'}
fakeout=X∗W′
W
′
′
=
W
′
/
σ
(
f
a
k
e
o
u
t
)
W^{''} = W^{'}/\sigma(fakeout)
W′′=W′/σ(fakeout)
l
o
s
s
w
e
i
g
h
t
=
∣
∣
X
∗
W
′
′
−
X
∗
W
′
∣
∣
2
loss_{weight} = ||X*W^{''} - X*W^{'}||_{2}
lossweight=∣∣X∗W′′−X∗W′∣∣2
r
e
a
l
o
u
t
=
X
∗
W
′
′
realout = X*W^{''}
realout=X∗W′′
L
o
s
s
t
o
t
a
l
=
l
o
s
s
c
l
a
s
s
f
i
c
a
t
i
o
n
+
λ
∗
l
o
s
s
w
e
i
g
h
t
Loss_{total}=loss_{classfication}+\lambda *loss_{weight}
Losstotal=lossclassfication+λ∗lossweight
目前做法整理
最新推荐文章于 2021-06-02 21:41:33 发布