numpy逆矩阵不适用于满秩矩阵-牛顿法logistic回归中的hessian
我试图用numpy计算一个满秩矩阵的逆,但是当我测试点积时,我发现它没有得到单位矩阵,这意味着它没有正确地反转 我的代码:numpy逆矩阵不适用于满秩矩阵-牛顿法logistic回归中的hessian,numpy,logistic-regression,matrix-inverse,newtons-method,hessian,Numpy,Logistic Regression,Matrix Inverse,Newtons Method,Hessian,我试图用numpy计算一个满秩矩阵的逆,但是当我测试点积时,我发现它没有得到单位矩阵,这意味着它没有正确地反转 我的代码: H = calculateLogisticHessian(theta, X) #returns a 5x5 matrix Hinv = np.linalg.inv(H) print("H = " + str(H)) print("Hinv = " + str(Hinv)) I = np.dot(H, Hinv) isIdentity = np.allclose(I , np
H = calculateLogisticHessian(theta, X) #returns a 5x5 matrix
Hinv = np.linalg.inv(H)
print("H = " + str(H))
print("Hinv = " + str(Hinv))
I = np.dot(H, Hinv)
isIdentity = np.allclose(I , np.eye(5))
print("invdotinv = " + str(isIdentity) + "\n" + str(I))
以及输出:
H = [[ 77.88167948 81.49914902 85.11661855 88.73408809 92.35155763]
[ 81.49914902 85.36097831 89.2228076 93.0846369 96.94646619]
[ 85.11661855 89.2228076 93.32899665 97.4351857 101.54137475]
[ 88.73408809 93.0846369 97.4351857 101.7857345 106.1362833 ]
[ 92.35155763 96.94646619 101.54137475 106.1362833 110.73119186]]
Hinv = [[ 1.41918134e+02 1.00000206e+08 -1.00000632e+08 -9.99999204e+07
1.00000205e+08]
[ 1.00000347e+08 1.00000647e+08 -4.00001421e+08 9.99994941e+07
1.00000932e+08]
[ -1.00000916e+08 -4.00001424e+08 8.00003700e+08 5.68436971e+02
-3.00001928e+08]
[ -9.99997780e+07 1.00000065e+08 -5.72321511e+02 1.00000063e+08
-9.99997769e+07]
[ 1.00000205e+08 1.00000505e+08 -3.00001073e+08 -1.00000205e+08
2.00000567e+08]]
invdotinv = False
[[ 1.00000000e+00 -3.81469727e-06 -7.62939453e-06 3.81469727e-06
3.81469727e-06]
[ 0.00000000e+00 1.00000191e+00 -1.52587891e-05 3.81469727e-06
0.00000000e+00]
[ -3.81469727e-06 1.90734863e-06 9.99992371e-01 3.81469727e-06
3.81469727e-06]
[ 1.90734863e-06 -1.90734863e-06 -7.62939453e-06 1.00000191e+00
3.81469727e-06]
[ 0.00000000e+00 -1.90734863e-06 0.00000000e+00 0.00000000e+00
1.00000000e+00]]
正如您所看到的,np.dot(H,Hinv)
矩阵在计算np.allclose(I,np.eye(5))
时不返回标识并导致False
我做错了什么
以后编辑
这是计算海森曲线的函数:
def calculateLogisticHessian(theta, X):
'''
calculate the hessian matrix based on given function, assuming it is some king of logistic funciton
:param theta: the weights
:param x: 2d array of arguments
:return: the hessian matrix
'''
m, n = X.shape
H = np.zeros((n,n))
for i in range(0,m):
hxi = h(theta, X[i]) #in case of logistic, will return p(y|x)
xiDotxiT = np.outer(X[i], np.transpose(X[i]))
hxiTimesOneMinHxi = hxi*(1-hxi)
currh = np.multiply(hxiTimesOneMinHxi, xiDotxiT)
H = np.add(H, currh)
return np.divide(H, m)
根据andrew ng视频中关于逻辑回归牛顿法的hessian计算公式:
五点零六分
1/m*(从i=1到m的和[h(X[i])*(1-h(X[i])*(X[i]*
(X[i][T])
其中X是数据的2x2矩阵,h()是基于θ的函数(θ是weigts),在本例中,该函数返回逻辑函数
我使用的输入:
theta = np.array([0.001, 0.002, 0.003, 0.004, 0.005])
X = np.array(range(5*7))
X = X.reshape((7,5))
H = calculateLogisticHessian(theta, X)
那么,我实现海森公式的方式是否存在错误,或者输入中是否存在问题,问题是什么
谢谢!Hessian矩阵经常出现。
用于计算条件编号:
由于
H
的条件数很大,因此计算其逆运算存在舍入问题。您确定可以反转此矩阵吗?行列式近似为零:np.linalg.det(H)
给出了-5.94505055443953825E-24
,这是在数值不安全性下面,我会在1e-15
的某个地方命名。另外,试着打印I
并阅读np.allclose
的文档。我不确定矩阵是否可以反转。我添加了一个关于我如何到达该矩阵的解释。我看,我没有学习numerical分析。那么人们通常如何处理hessian矩阵求逆呢?我知道这是逻辑回归牛顿法等方法所必需的。
In [188]: np.linalg.cond(H)
Out[188]: 522295671550.72644