为了说清楚问题,画了如下图帮助理解。
如果lossFunction 是 mse,真值为z_true,loss = ( z_true - z ) ** 2
d(loss)/d(z) = 2(z - z_true)
d(z)/d(y) = sigmoid’(z)
d(y)/d(w) = x
d(y)/d(b) = 1
d(loss)/d(w) = d(loss)/d(z) * d(z)/d(y) * d(y)/d(w) = 2(z - z_true) * sigmoid’(z) * x
d(loss)/d(b) = d(loss)/d(z) * d(z)/d(y) * d(y)/d(w) = 2(z - z_true) * sigmoid’(z) * 1
================================================================
如果lossFunction 是 binary_cross_entropy,真值为z_true,
loss = - [ z_true * log(z) + (1 - z_true) * log( 1 - z ) ]
d(loss)/d(z) = - z_true / z + (1 - z_true) / (1-z)
如果 z_true == 1 则 d(loss)/d(z) = -1/z
如果 z_true == 0 则 d(loss)/d(z) = 1/(1-z)
d(z)/d(y) = sigmoid’(z) = z * (1-z) (这个就是sigmoid的导数,下面有推导过程)
d(y)/d(w) = x
d(y)/d(b) = 1
如果 z_true == 1 则
d(loss)/d(w) = d(loss)/d(z) * d(z)/d(y) * d(y)/d(w) = -1/z * z * (1-z) * x = (z-1)*x
d(loss)/d(b) = d(loss)/d(z) * d(z)/d(y) * d(y)/d(b) = -1/z * z * (1-z) * 1 = (z-1)
如果 z_true == 0 则
d(loss)/d(w) = d(loss)/d(z) * d(z)/d(y) * d(y)/d(w) = 1/(1-z) * z * (1-z) * x = (z - 0)*x
d(loss)/d(b) = d(loss)/d(z) * d(z)/d(y) * d(y)/d(b) = 1/(1-z) * z * (1-z) * 1 = ( z - 0 )
合并之后:
d(loss)/d(w) = ( z - z_true ) * x
d(loss)/d(b) = ( z - z_true )
.
.
###sigmoid函数求导过程如下:
如果是多分类,则激励函数为softmax,
那么softmax 又和 sigmoid 有什么关系那?