3 卷积层的训练
同全连接层一样,卷积层的训练也需要从上一层回传的误差矩阵,然后计算:
- 本层的权重矩阵的误差项
- 本层的需要回传到下一层的误差矩阵
在下面的描述中,我们假设已经得到了从上一层回传的误差矩阵,并且已经经过了激活函数的反向传导。
3.1 计算反向传播的梯度矩阵
正向公式:
Z = W ∗ A + b (0) Z = W*A+b \tag{0} Z=W∗A+b(0)
其中,W是卷积核,*表示卷积(互相关)计算,A为当前层的输入项,b是偏移(未在图中画出),Z为当前层的输出项,但尚未经过激活函数处理。
我们举一个具体的例子便于分析。图17-21是正向计算过程。
图17-21 卷积正向运算
分解到每一项就是下列公式:
z
11
=
w
11
⋅
a
11
+
w
12
⋅
a
12
+
w
21
⋅
a
21
+
w
22
⋅
a
22
+
b
(1)
z_{11} = w_{11} \cdot a_{11} + w_{12} \cdot a_{12} + w_{21} \cdot a_{21} + w_{22} \cdot a_{22} + b \tag{1}
z11=w11⋅a11+w12⋅a12+w21⋅a21+w22⋅a22+b(1)
z
12
=
w
11
⋅
a
12
+
w
12
⋅
a
13
+
w
21
⋅
a
22
+
w
22
⋅
a
23
+
b
(2)
z_{12} = w_{11} \cdot a_{12} + w_{12} \cdot a_{13} + w_{21} \cdot a_{22} + w_{22} \cdot a_{23} + b \tag{2}
z12=w11⋅a12+w12⋅a13+w21⋅a22+w22⋅a23+b(2)
z
21
=
w
11
⋅
a
21
+
w
12
⋅
a
22
+
w
21
⋅
a
31
+
w
22
⋅
a
32
+
b
(3)
z_{21} = w_{11} \cdot a_{21} + w_{12} \cdot a_{22} + w_{21} \cdot a_{31} + w_{22} \cdot a_{32} + b \tag{3}
z21=w11⋅a21+w12⋅a22+w21⋅a31+w22⋅a32+b(3)
z
22
=
w
11
⋅
a
22
+
w
12
⋅
a
23
+
w
21
⋅
a
32
+
w
22
⋅
a
33
+
b
(4)
z_{22} = w_{11} \cdot a_{22} + w_{12} \cdot a_{23} + w_{21} \cdot a_{32} + w_{22} \cdot a_{33} + b \tag{4}
z22=w11⋅a22+w12⋅a23+w21⋅a32+w22⋅a33+b(4)
求损失函数J对a11的梯度:
∂ J ∂ a 11 = ∂ J ∂ z 11 ∂ z 11 ∂ a 11 = δ z 11 ⋅ w 11 (5) \frac{\partial J}{\partial a_{11}}=\frac{\partial J}{\partial z_{11}} \frac{\partial z_{11}}{\partial a_{11}}=\delta_{z11}\cdot w_{11} \tag{5} ∂a11∂J=∂z11∂J∂a11∂z11=δz11⋅w11(5)
上式中, δ z 11 \delta_{z11} δz11是从网络后端回传到本层的z11单元的梯度。
求J对a12的梯度时,先看正向公式,发现a12对z11和z12都有贡献,因此需要二者的偏导数相加:
∂ J ∂ a 12 = ∂ J ∂ z 11 ∂ z 11 ∂ a 12 + ∂ J ∂ z 12 ∂ z 12 ∂ a 12 = δ z 11 ⋅ w 12 + δ z 12 ⋅ w 11 (6) \frac{\partial J}{\partial a_{12}}=\frac{\partial J}{\partial z_{11}} \frac{\partial z_{11}}{\partial a_{12}}+\frac{\partial J}{\partial z_{12}} \frac{\partial z_{12}}{\partial a_{12}}=\delta_{z11} \cdot w_{12}+\delta_{z12} \cdot w_{11} \tag{6} ∂a12∂J=∂z11∂J∂a12∂z11+∂z12∂J∂a12∂z12=δz11⋅w12+δz12⋅w11(6)
最复杂的是求a22的梯度,因为从正向公式看,所有的输出都有a22的贡献,所以:
∂
J
∂
a
22
=
∂
J
∂
z
11
∂
z
11
∂
a
22
+
∂
J
∂
z
12
∂
z
12
∂
a
22
+
∂
J
∂
z
21
∂
z
21
∂
a
22
+
∂
J
∂
z
22
∂
z
22
∂
a
22
\frac{\partial J}{\partial a_{22}}=\frac{\partial J}{\partial z_{11}} \frac{\partial z_{11}}{\partial a_{22}}+\frac{\partial J}{\partial z_{12}} \frac{\partial z_{12}}{\partial a_{22}}+\frac{\partial J}{\partial z_{21}} \frac{\partial z_{21}}{\partial a_{22}}+\frac{\partial J}{\partial z_{22}} \frac{\partial z_{22}}{\partial a_{22}}
∂a22∂J=∂z11∂J∂a22∂z11+∂z12∂J∂a22∂z12+∂z21∂J∂a22∂z21+∂z22∂J∂a22∂z22
=
δ
z
11
⋅
w
22
+
δ
z
12
⋅
w
21
+
δ
z
21
⋅
w
12
+
δ
z
22
⋅
w
11
(7)
=\delta_{z11} \cdot w_{22} + \delta_{z12} \cdot w_{21} + \delta_{z21} \cdot w_{12} + \delta_{z22} \cdot w_{11} \tag{7}
=δz11⋅w22+δz12⋅w21+δz21⋅w12+δz22⋅w11(7)
同理可得所有a的梯度。
观察公式7中的w的顺序,貌似是把原始的卷积核旋转了180度,再与传入误差项做卷积操作,即可得到所有元素的误差项。而公式5和公式6并不完备,是因为二者处于角落,这和卷积正向计算中的padding是相同的现象。因此,我们把传入的误差矩阵Delta-In做一个zero padding,再乘以旋转180度的卷积核,就是要传出的误差矩阵Delta-Out,如图17-22所示。
图17-22 卷积运算中的误差反向传播
最后可以统一成为一个简洁的公式:
δ o u t = δ i n ∗ W r o t 180 (8) \delta_{out} = \delta_{in} * W^{rot180} \tag{8} δout=δin∗Wrot180(8)
这个误差矩阵可以继续回传到下一层。
- 当Weights是3x3时, δ i n \delta_{in} δin需要padding=2,即加2圈0,才能和Weights卷积后,得到正确尺寸的 δ o u t \delta_{out} δout
- 当Weights是5x5时, δ i n \delta_{in} δin需要padding=4,即加4圈0,才能和Weights卷积后,得到正确尺寸的 δ o u t \delta_{out} δout
- 以此类推:当Weights是NxN时, δ i n \delta_{in} δin需要padding=N-1,即加N-1圈0
举例:
正向时stride=1: A ( 10 × 8 ) ∗ W ( 5 × 5 ) = Z ( 6 × 4 ) A^{(10 \times 8)}*W^{(5 \times 5)}=Z^{(6 \times 4)} A(10×8)∗W(5×5)=Z(6×4)
反向时, δ z ( 6 × 4 ) + 4 p a d d i n g = δ z ( 14 × 12 ) \delta_z^{(6 \times 4)} + 4 padding = \delta_z^{(14 \times 12)} δz(6×4)+4padding=δz(14×12)
然后: δ z ( 14 × 12 ) ∗ W r o t 180 ( 5 × 5 ) = δ a ( 10 × 8 ) \delta_z^{(14 \times 12)} * W^{rot180(5 \times 5)}= \delta_a^{(10 \times 8)} δz(14×12)∗Wrot180(5×5)=δa(10×8)
3.2 步长不为1时的梯度矩阵还原
我们先观察一下stride=1和2时,卷积结果的差异如图17-23。
图17-23 步长为1和步长为2的卷积结果的比较
二者的差别就是中间那个结果图的灰色部分。如果反向传播时,传入的误差矩阵是stride=2时的2x2的形状,那么我们只需要把它补上一个十字,变成3x3的误差矩阵,就可以用步长为1的算法了。
以此类推,如果步长为3时,需要补一个双线的十字。所以,当知道当前的卷积层步长为S(S>1)时:
- 得到从上层回传的误差矩阵形状,假设为 M × N M \times N M×N
- 初始化一个 ( M ⋅ S ) × ( N ⋅ S ) (M \cdot S) \times (N \cdot S) (M⋅S)×(N⋅S)的零矩阵
- 把传入的误差矩阵的第一行值放到零矩阵第0行的0,S,2S,3S…位置
- 然后把误差矩阵的第二行的值放到零矩阵第S行的0,S,2S,3S…位置
- …
步长为2时,用实例表示就是这样:
[ δ 11 0 δ 12 0 δ 13 0 0 0 0 0 δ 21 0 δ 22 0 δ 23 ] \begin{bmatrix} \delta_{11} & 0 & \delta_{12} & 0 & \delta_{13}\\ 0 & 0 & 0 & 0 & 0\\ \delta_{21} & 0 & \delta_{22} & 0 & \delta_{23}\\ \end{bmatrix} ⎣⎡δ110δ21000δ120δ22000δ130δ23⎦⎤
步长为3时,用实例表示就是这样:
[ δ 11 0 0 δ 12 0 0 δ 13 0 0 0 0 0 0 0 0 0 0 0 0 0 0 δ 21 0 0 δ 22 0 0 δ 23 ] \begin{bmatrix} \delta_{11} & 0 & 0 & \delta_{12} & 0 & 0 & \delta_{13}\\ 0 & 0 & 0 & 0 & 0 & 0 & 0\\ 0 & 0 & 0 & 0 & 0 & 0 & 0\\ \delta_{21} & 0 & 0 & \delta_{22} & 0 & 0 & \delta_{23}\\ \end{bmatrix} ⎣⎢⎢⎡δ1100δ2100000000δ1200δ2200000000δ1300δ23⎦⎥⎥⎤
3.3 有多个卷积核时的梯度计算
有多个卷积核也就意味着有多个输出通道。
也就是14.1中的升维卷积,如图17-24。
图17-24 升维卷积
正向公式:
z
111
=
w
111
⋅
a
11
+
w
112
⋅
a
12
+
w
121
⋅
a
21
+
w
122
⋅
a
22
z111 = w111 \cdot a11 + w112 \cdot a12 + w121 \cdot a21 + w122 \cdot a22
z111=w111⋅a11+w112⋅a12+w121⋅a21+w122⋅a22
z
112
=
w
111
⋅
a
12
+
w
112
⋅
a
13
+
w
121
⋅
a
22
+
w
122
⋅
a
23
z112 = w111 \cdot a12 + w112 \cdot a13 + w121 \cdot a22 + w122 \cdot a23
z112=w111⋅a12+w112⋅a13+w121⋅a22+w122⋅a23
z
121
=
w
111
⋅
a
21
+
w
112
⋅
a
22
+
w
121
⋅
a
31
+
w
122
⋅
a
32
z121 = w111 \cdot a21 + w112 \cdot a22 + w121 \cdot a31 + w122 \cdot a32
z121=w111⋅a21+w112⋅a22+w121⋅a31+w122⋅a32
z
122
=
w
111
⋅
a
22
+
w
112
⋅
a
23
+
w
121
⋅
a
32
+
w
122
⋅
a
33
z122 = w111 \cdot a22 + w112 \cdot a23 + w121 \cdot a32 + w122 \cdot a33
z122=w111⋅a22+w112⋅a23+w121⋅a32+w122⋅a33
z
211
=
w
211
⋅
a
11
+
w
212
⋅
a
12
+
w
221
⋅
a
21
+
w
222
⋅
a
22
z211 = w211 \cdot a11 + w212 \cdot a12 + w221 \cdot a21 + w222 \cdot a22
z211=w211⋅a11+w212⋅a12+w221⋅a21+w222⋅a22
z
212
=
w
211
⋅
a
12
+
w
212
⋅
a
13
+
w
221
⋅
a
22
+
w
222
⋅
a
23
z212 = w211 \cdot a12 + w212 \cdot a13 + w221 \cdot a22 + w222 \cdot a23
z212=w211⋅a12+w212⋅a13+w221⋅a22+w222⋅a23
z
221
=
w
211
⋅
a
21
+
w
212
⋅
a
22
+
w
221
⋅
a
31
+
w
222
⋅
a
32
z221 = w211 \cdot a21 + w212 \cdot a22 + w221 \cdot a31 + w222 \cdot a32
z221=w211⋅a21+w212⋅a22+w221⋅a31+w222⋅a32
z
222
=
w
211
⋅
a
22
+
w
212
⋅
a
23
+
w
221
⋅
a
32
+
w
222
⋅
a
33
z222 = w211 \cdot a22 + w212 \cdot a23 + w221 \cdot a32 + w222 \cdot a33
z222=w211⋅a22+w212⋅a23+w221⋅a32+w222⋅a33
求J对a22的梯度:
∂
J
∂
a
22
=
∂
J
∂
Z
1
∂
Z
1
∂
a
22
+
∂
J
∂
Z
2
∂
Z
2
∂
a
22
\frac{\partial J}{\partial a_{22}}=\frac{\partial J}{\partial Z_{1}} \frac{\partial Z_{1}}{\partial a_{22}}+\frac{\partial J}{\partial Z_{2}} \frac{\partial Z_{2}}{\partial a_{22}}
∂a22∂J=∂Z1∂J∂a22∂Z1+∂Z2∂J∂a22∂Z2
=
∂
J
∂
z
111
∂
z
111
∂
a
22
+
∂
J
∂
z
112
∂
z
112
∂
a
22
+
∂
J
∂
z
121
∂
z
121
∂
a
22
+
∂
J
∂
z
122
∂
z
122
∂
a
22
=\frac{\partial J}{\partial z_{111}} \frac{\partial z_{111}}{\partial a_{22}}+\frac{\partial J}{\partial z_{112}} \frac{\partial z_{112}}{\partial a_{22}}+\frac{\partial J}{\partial z_{121}} \frac{\partial z_{121}}{\partial a_{22}}+\frac{\partial J}{\partial z_{122}} \frac{\partial z_{122}}{\partial a_{22}}
=∂z111∂J∂a22∂z111+∂z112∂J∂a22∂z112+∂z121∂J∂a22∂z121+∂z122∂J∂a22∂z122
+
∂
J
∂
z
211
∂
z
211
∂
a
22
+
∂
J
∂
z
212
∂
z
212
∂
a
22
+
∂
J
∂
z
221
∂
z
221
∂
a
22
+
∂
J
∂
z
222
∂
z
222
∂
a
22
+\frac{\partial J}{\partial z_{211}} \frac{\partial z_{211}}{\partial a_{22}}+\frac{\partial J}{\partial z_{212}} \frac{\partial z_{212}}{\partial a_{22}}+\frac{\partial J}{\partial z_{221}} \frac{\partial z_{221}}{\partial a_{22}}+\frac{\partial J}{\partial z_{222}} \frac{\partial z_{222}}{\partial a_{22}}
+∂z211∂J∂a22∂z211+∂z212∂J∂a22∂z212+∂z221∂J∂a22∂z221+∂z222∂J∂a22∂z222
=
(
δ
z
111
⋅
w
122
+
δ
z
112
⋅
w
121
+
δ
z
121
⋅
w
112
+
δ
z
122
⋅
w
111
)
=(\delta_{z111} \cdot w_{122} + \delta_{z112} \cdot w_{121} + \delta_{z121} \cdot w_{112} + \delta_{z122} \cdot w_{111})
=(δz111⋅w122+δz112⋅w121+δz121⋅w112+δz122⋅w111)
+
(
δ
z
211
⋅
w
222
+
δ
z
212
⋅
w
221
+
δ
z
221
⋅
w
212
+
δ
z
222
⋅
w
211
)
+(\delta_{z211} \cdot w_{222} + \delta_{z212} \cdot w_{221} + \delta_{z221} \cdot w_{212} + \delta_{z222} \cdot w_{211})
+(δz211⋅w222+δz212⋅w221+δz221⋅w212+δz222⋅w211)
=
δ
z
1
∗
W
1
r
o
t
180
+
δ
z
2
∗
W
2
r
o
t
180
=\delta_{z1} * W_1^{rot180} + \delta_{z2} * W_2^{rot180}
=δz1∗W1rot180+δz2∗W2rot180
因此和公式8相似,先在 δ i n \delta_{in} δin外面加padding,然后和对应的旋转后的卷积核相乘,再把几个结果相加,就得到了需要前传的梯度矩阵:
δ o u t = ∑ m δ i n _ m ∗ W m r o t 180 (9) \delta_{out} = \sum_m \delta_{in\_m} * W^{rot180}_m \tag{9} δout=m∑δin_m∗Wmrot180(9)
3.4 有多个输入时的梯度计算
当输入层是多个图层时,每个图层必须对应一个卷积核,如图17-25。
图17-25 多个图层的卷积必须有一一对应的卷积核
所以有前向公式:
z
11
=
w
111
⋅
a
111
+
w
112
⋅
a
112
+
w
121
⋅
a
121
+
w
122
⋅
a
122
+
w
211
⋅
a
211
+
w
212
⋅
a
212
+
w
221
⋅
a
221
+
w
222
⋅
a
222
(10)
\begin{aligned} z11 &= w111 \cdot a111 + w112 \cdot a112 + w121 \cdot a121 + w122 \cdot a122 \\ &+ w211 \cdot a211 + w212 \cdot a212 + w221 \cdot a221 + w222 \cdot a222 \end{aligned} \tag{10}
z11=w111⋅a111+w112⋅a112+w121⋅a121+w122⋅a122+w211⋅a211+w212⋅a212+w221⋅a221+w222⋅a222(10)
z
12
=
w
111
⋅
a
112
+
w
112
⋅
a
113
+
w
121
⋅
a
122
+
w
122
⋅
a
123
+
w
211
⋅
a
212
+
w
212
⋅
a
213
+
w
221
⋅
a
222
+
w
222
⋅
a
223
(11)
\begin{aligned} z12 &= w111 \cdot a112 + w112 \cdot a113 + w121 \cdot a122 + w122 \cdot a123 \\ &+ w211 \cdot a212 + w212 \cdot a213 + w221 \cdot a222 + w222 \cdot a223 \end{aligned}\tag{11}
z12=w111⋅a112+w112⋅a113+w121⋅a122+w122⋅a123+w211⋅a212+w212⋅a213+w221⋅a222+w222⋅a223(11)
z
21
=
w
111
⋅
a
121
+
w
112
⋅
a
122
+
w
121
⋅
a
131
+
w
122
⋅
a
132
+
w
211
⋅
a
221
+
w
212
⋅
a
222
+
w
221
⋅
a
231
+
w
222
⋅
a
232
(12)
\begin{aligned} z21 &= w111 \cdot a121 + w112 \cdot a122 + w121 \cdot a131 + w122 \cdot a132 \\ &+ w211 \cdot a221 + w212 \cdot a222 + w221 \cdot a231 + w222 \cdot a232 \end{aligned}\tag{12}
z21=w111⋅a121+w112⋅a122+w121⋅a131+w122⋅a132+w211⋅a221+w212⋅a222+w221⋅a231+w222⋅a232(12)
z
22
=
w
111
⋅
a
122
+
w
112
⋅
a
123
+
w
121
⋅
a
132
+
w
122
⋅
a
133
+
w
211
⋅
a
222
+
w
212
⋅
a
223
+
w
221
⋅
a
232
+
w
222
⋅
a
233
(13)
\begin{aligned} z22 &= w111 \cdot a122 + w112 \cdot a123 + w121 \cdot a132 + w122 \cdot a133 \\ &+ w211 \cdot a222 + w212 \cdot a223 + w221 \cdot a232 + w222 \cdot a233 \end{aligned}\tag{13}
z22=w111⋅a122+w112⋅a123+w121⋅a132+w122⋅a133+w211⋅a222+w212⋅a223+w221⋅a232+w222⋅a233(13)
最复杂的情况,求J对a122的梯度:
∂ J ∂ a 111 = ∂ J ∂ z 11 ∂ z 11 ∂ a 122 + ∂ J ∂ z 12 ∂ z 12 ∂ a 122 + ∂ J ∂ z 21 ∂ z 21 ∂ a 122 + ∂ J ∂ z 22 ∂ z 22 ∂ a 122 = δ z 11 ⋅ w 122 + δ z 12 ⋅ w 121 + δ z 21 ⋅ w 112 + δ z 22 ⋅ w 111 \begin{aligned} \frac{\partial J}{\partial a111}&=\frac{\partial J}{\partial z11}\frac{\partial z11}{\partial a122} + \frac{\partial J}{\partial z12}\frac{\partial z12}{\partial a122} + \frac{\partial J}{\partial z21}\frac{\partial z21}{\partial a122} + \frac{\partial J}{\partial z22}\frac{\partial z22}{\partial a122} \\ &=\delta_{z11} \cdot w122 + \delta_{z12} \cdot w121 + \delta_{z21} \cdot w112 + \delta_{z22} \cdot w111 \end{aligned} ∂a111∂J=∂z11∂J∂a122∂z11+∂z12∂J∂a122∂z12+∂z21∂J∂a122∂z21+∂z22∂J∂a122∂z22=δz11⋅w122+δz12⋅w121+δz21⋅w112+δz22⋅w111
泛化以后得到:
δ o u t 1 = δ i n ∗ W 1 r o t 180 (14) \delta_{out1} = \delta_{in} * W_1^{rot180} \tag{14} δout1=δin∗W1rot180(14)
最复杂的情况,求J对a222的梯度:
∂ J ∂ a 211 = ∂ J ∂ z 11 ∂ z 11 ∂ a 222 + ∂ J ∂ z 12 ∂ z 12 ∂ a 222 + ∂ J ∂ z 21 ∂ z 21 ∂ a 222 + ∂ J ∂ z 22 ∂ z 22 ∂ a 222 = δ z 11 ⋅ w 222 + δ z 12 ⋅ w 221 + δ z 21 ⋅ w 212 + δ z 22 ⋅ w 211 \begin{aligned} \frac{\partial J}{\partial a211}&=\frac{\partial J}{\partial z11}\frac{\partial z11}{\partial a222} + \frac{\partial J}{\partial z12}\frac{\partial z12}{\partial a222} + \frac{\partial J}{\partial z21}\frac{\partial z21}{\partial a222} + \frac{\partial J}{\partial z22}\frac{\partial z22}{\partial a222} \\ &=\delta_{z11} \cdot w222 + \delta_{z12} \cdot w221 + \delta_{z21} \cdot w212 + \delta_{z22} \cdot w211 \end{aligned} ∂a211∂J=∂z11∂J∂a222∂z11+∂z12∂J∂a222∂z12+∂z21∂J∂a222∂z21+∂z22∂J∂a222∂z22=δz11⋅w222+δz12⋅w221+δz21⋅w212+δz22⋅w211
泛化以后得到:
δ o u t 2 = δ i n ∗ W 2 r o t 180 (15) \delta_{out2} = \delta_{in} * W_2^{rot180} \tag{15} δout2=δin∗W2rot180(15)
3.5 权重(卷积核)梯度计算
图17-26展示了我们已经熟悉的卷积正向运算。
图17-26 卷积正向计算
要求J对w11的梯度,从正向公式可以看到,w11对所有的z都有贡献,所以:
∂ J ∂ w 11 = ∂ J ∂ z 11 ∂ z 11 ∂ w 11 + ∂ J ∂ z 12 ∂ z 12 ∂ w 11 + ∂ J ∂ z 21 ∂ z 21 ∂ w 11 + ∂ J ∂ z 22 ∂ z 22 ∂ w 11 = δ z 11 ⋅ a 11 + δ z 12 ⋅ a 12 + δ z 21 ⋅ a 21 + δ z 22 ⋅ a 22 (9) \begin{aligned} \frac{\partial J}{\partial w_{11}} &= \frac{\partial J}{\partial z_{11}}\frac{\partial z_{11}}{\partial w_{11}} + \frac{\partial J}{\partial z_{12}}\frac{\partial z_{12}}{\partial w_{11}} + \frac{\partial J}{\partial z_{21}}\frac{\partial z_{21}}{\partial w_{11}} + \frac{\partial J}{\partial z_{22}}\frac{\partial z_{22}}{\partial w_{11}} \\ &=\delta_{z11} \cdot a_{11} + \delta_{z12} \cdot a_{12} + \delta_{z21} \cdot a_{21} + \delta_{z22} \cdot a_{22} \end{aligned} \tag{9} ∂w11∂J=∂z11∂J∂w11∂z11+∂z12∂J∂w11∂z12+∂z21∂J∂w11∂z21+∂z22∂J∂w11∂z22=δz11⋅a11+δz12⋅a12+δz21⋅a21+δz22⋅a22(9)
对W22也是一样的:
∂ J ∂ w 12 = ∂ J ∂ z 11 ∂ z 11 ∂ w 12 + ∂ J ∂ z 12 ∂ z 12 ∂ w 12 + ∂ J ∂ z 21 ∂ z 21 ∂ w 12 + ∂ J ∂ z 22 ∂ z 22 ∂ w 12 = δ z 11 ⋅ a 12 + δ z 12 ⋅ a 13 + δ z 21 ⋅ a 22 + δ z 22 ⋅ a 23 (10) \begin{aligned} \frac{\partial J}{\partial w_{12}} &= \frac{\partial J}{\partial z_{11}}\frac{\partial z_{11}}{\partial w_{12}} + \frac{\partial J}{\partial z_{12}}\frac{\partial z_{12}}{\partial w_{12}} + \frac{\partial J}{\partial z_{21}}\frac{\partial z_{21}}{\partial w_{12}} + \frac{\partial J}{\partial z_{22}}\frac{\partial z_{22}}{\partial w_{12}} \\ &=\delta_{z11} \cdot a_{12} + \delta_{z12} \cdot a_{13} + \delta_{z21} \cdot a_{22} + \delta_{z22} \cdot a_{23} \end{aligned} \tag{10} ∂w12∂J=∂z11∂J∂w12∂z11+∂z12∂J∂w12∂z12+∂z21∂J∂w12∂z21+∂z22∂J∂w12∂z22=δz11⋅a12+δz12⋅a13+δz21⋅a22+δz22⋅a23(10)
观察公式8和公式9,其实也是一个标准的卷积(互相关)操作过程,因此,可以把这个过程看成图17-27。
图17-27 卷积核的梯度计算
总结成一个公式:
δ w = A ∗ δ i n (11) \delta_w = A * \delta_{in} \tag{11} δw=A∗δin(11)
3.6 偏移的梯度计算
根据前向计算公式1,2,3,4,可以得到:
∂ J ∂ b = ∂ J ∂ z 11 ∂ z 11 ∂ b + ∂ J ∂ z 12 ∂ z 12 ∂ b + ∂ J ∂ z 21 ∂ z 21 ∂ b + ∂ J ∂ z 22 ∂ z 22 ∂ b = δ z 11 + δ z 12 + δ z 21 + δ z 22 (12) \begin{aligned} \frac{\partial J}{\partial b} &= \frac{\partial J}{\partial z_{11}}\frac{\partial z_{11}}{\partial b} + \frac{\partial J}{\partial z_{12}}\frac{\partial z_{12}}{\partial b} + \frac{\partial J}{\partial z_{21}}\frac{\partial z_{21}}{\partial b} + \frac{\partial J}{\partial z_{22}}\frac{\partial z_{22}}{\partial b} \\ &=\delta_{z11} + \delta_{z12} + \delta_{z21} + \delta_{z22} \end{aligned} \tag{12} ∂b∂J=∂z11∂J∂b∂z11+∂z12∂J∂b∂z12+∂z21∂J∂b∂z21+∂z22∂J∂b∂z22=δz11+δz12+δz21+δz22(12)
所以:
δ b = δ i n (13) \delta_b = \delta_{in} \tag{13} δb=δin(13)
每个卷积核W可能会有多个filter,或者叫子核,但是一个卷积核只有一个偏移,无论有多少子核。
3.7 计算卷积核梯度的实例说明
下面我们会用一个简单的例子来说明卷积核的训练过程。我们先制作一张样本图片,然后使用“横边检测”算子做为卷积核对该样本进行卷积,得到对比如图17-28。
图17-28 原图和经过横边检测算子的卷积结果
左侧为原始图片(80x80的灰度图),右侧为经过3x3的卷积后的结果图片(78x78的灰度图)。由于算子是横边检测,所以只保留了原始图片中的横边。
卷积核矩阵:
w = ( 0 − 1 0 0 2 0 0 − 1 0 ) w=\begin{pmatrix} 0 & -1 & 0 \\ 0 & 2 & 0 \\ 0 & -1 & 0 \end{pmatrix} w=⎝⎛000−12−1000⎠⎞
现在我们转换一下问题:假设我们有一张原始图片(如左侧)和一张目标图片(如右侧),我们如何得到对应的卷积核呢?
我们在前面学习了线性拟合的解决方案,实际上这个问题是同一种性质的,只不过把直线拟合点阵的问题,变成了图像拟合图像的问题,如表17-3所示。
表17-3 直线拟合与图像拟合的比较
样本数据 | 标签数据 | 预测数据 | 公式 | 损失函数 | |
---|---|---|---|---|---|
直线拟合 | 样本点x | 标签值y | 预测直线z | z = x ⋅ w + b z=x \cdot w+b z=x⋅w+b | 均方差 |
图片拟合 | 原始图片x | 目标图片y | 预测图片z | z = x ∗ w + b z=x * w+b z=x∗w+b | 均方差 |
直线拟合中的均方差,是计算预测值与样本点之间的距离;图片拟合中的均方差,可以直接计算两张图片对应的像素点之间的差值。
为了简化问题,我们令b=0,只求卷积核w的值,则前向公式为:
z
=
x
∗
w
z = x * w
z=x∗w
l
o
s
s
=
1
2
(
z
−
y
)
2
loss = {1 \over 2}(z-y)^2
loss=21(z−y)2
反向求解w的梯度公式(从公式11得到):
∂ l o s s ∂ w = ∂ l o s s ∂ z ∂ z ∂ w = x ∗ ( z − y ) {\partial loss \over \partial w}={\partial loss \over \partial z}{\partial z \over \partial w}=x * (z-y) ∂w∂loss=∂z∂loss∂w∂z=x∗(z−y)
即w的梯度为预测图片z减去目标图片y的结果,再与原始图片x做卷积,其中x为被卷积图片,z-y为卷积核。
训练部分的代码实现如下:
def train(x, w, b, y):
output = create_zero_array(x, w)
for i in range(10000):
# forward
jit_conv_2d(x, w, b, output)
# loss
t1 = (output - y)
m = t1.shape[0]*t1.shape[1]
LOSS = np.multiply(t1, t1)
loss = np.sum(LOSS)/2/m
print(i,loss)
if loss < 1e-7:
break
# delta
delta = output - y
# backward
dw = np.zeros(w.shape)
jit_conv_2d(x, delta, b, dw)
w = w - 0.5 * dw/m
#end for
return w
一共迭代10000次:
- 用jit_conv_2d(x,w…)做一次前向计算
- 计算loss值以便检测停止条件,当loss值小于1e-7时停止迭代
- 然后计算delta值
- 再用jit_conv_2d(x,delta)做一次反向计算,得到w的梯度
- 最后更新卷积核w的值
运行结果:
......
3458 1.0063169744079507e-07
3459 1.0031151142628902e-07
3460 9.999234418532805e-08
w_true:
[[ 0 -1 0]
[ 0 2 0]
[ 0 -1 0]]
w_result:
[[-1.86879237e-03 -9.97261724e-01 -1.01212359e-03]
[ 2.58961697e-03 1.99494606e+00 2.74435794e-03]
[-8.67754199e-04 -9.97404263e-01 -1.87580756e-03]]
w allclose: True
y allclose: True
当迭代到3460次的时候,loss值小于1e-7,迭代停止。比较w_true和w_result的值,两者非常接近。用numpy.allclose()方法比较真实卷积核和训练出来的卷积核的值,结果为True。比如-1.86879237e-03,接近于0;-9.97261724e-01,接近于-1。
再比较卷积结果,当然也会非常接近,误差很小,allclose结果为True。用图示方法显示卷积结果比较如图17-29。
图17-29 真实值和训练值的卷积结果区别
人眼是看不出什么差异来的。由此我们可以直观地理解到卷积核的训练过程并不复杂。
代码位置
ch17, Level3