我正在使用数值稳定的softmax版本: -
def softmax(arr):
print(arr)
expArr=np.exp(arr-np.max(arr))
print(expArr)
return expArr/np.sum(expArr)
它被用作: -
def feedforward(x_i,W):
...
outputLayer = softmax(np.dot(network[-1],W[-1]))
...
并且迭代地调用此函数: -
for j in range(len(x)):
....
network = feedforward(x[j],weights)
....
不过,对于某些数组序列,我收到警告: -
RuntimeWarning: invalid value encountered in subtract
expArr=np.exp(arr-np.max(arr))
在警告之前进入功能的输入(和输出)是: -
input
[-1.36678160e+211 -1.97916134e+206 -5.44472726e+204 -5.47948095e+276
-6.30134248e+251 -4.04707279e+210 7.72371508e+204 1.34861349e+268
5.47948093e+276 1.06699784e+206]
output
[0. 0. 0. 0. 0. 0. 0. 0. 1. 0.]
input
[-7.06701455e+257 1.47067222e+250 inf -inf
-1.13669521e+298 -6.54589076e+254 8.22221348e+250 inf
-inf -5.44761594e+251]
digit.py:22: RuntimeWarning: invalid value encountered in subtract
expArr=np.exp(arr-np.max(arr))
output
[ 0. 0. nan 0. 0. 0. 0. nan 0. 0.]
input
[nan nan nan nan nan nan nan nan nan nan]
output
[nan nan nan nan nan nan nan nan nan nan]
我想知道即使我通过引入np.max(arr)
术语来稳定softmax函数,为什么我仍然会收到此错误,我该如何解决?谢谢!
另外,我还使用了scipy.special
中给出的softmax函数,但仍然得到了同样的警告。
输入数组中的“inf”使得除法成为
"<non inf number>/inf"
给出“0”和
"inf/inf"
这给了“南”
你应该从输入数组中消除“inf”。