这里求导可以参看上一节这张图,
∂
<
x
,
w
>
∂
w
=
x
T
∂
w
∂
w
+
w
T
∂
x
∂
w
=
x
T
∂
w
∂
w
=
x
T
\frac {\partial <x,w>}{\partial w}=x^T\frac{\partial w}{\partial w}+w^T\frac{\partial x}{\partial w}=x^T\frac{\partial w}{\partial w}=x^T
∂w∂<x,w>=xT∂w∂w+wT∂w∂x=xT∂w∂w=xT,因为
∂
w
∂
w
=
1
\frac{\partial w}{\partial w}=1
∂w∂w=1,
∂
x
∂
w
=
0
\frac{\partial x}{\partial w}=0
∂w∂x=0
这里的
∂
X
w
∂
w
\frac{\partial Xw}{\partial w}
∂w∂Xw的X与w无关,可以提出来,因为
∂
w
∂
w
=
1
\frac{\partial w}{\partial w}=1
∂w∂w=1,所以最后
∂
X
w
∂
w
=
X
×
1
=
X
\frac{\partial Xw}{\partial w}=X \times 1=X
∂w∂Xw=X×1=X
MXNet和TensorFlow都可以显示构造计算图,数学上都是这样显示的构造,先有公式再代入值
MXNet和Pytorch可以隐式的构造