我正在探索一些numpy / scipy函数,我注意到与直接函数调用相比,scipy.optimize.fmin_bfgs要求对正在调用的函数进行更改以提供正确的结果。我对fnRSS
函数的第一个定义在调用函数时会返回正确的值,但拒绝在优化中工作;我的第二个定义在调用函数时给出了错误的结果,但在运行优化时给出了正确的结果。有人可以告诉我,为优化转置vY
参数至关重要吗?它应该已经是164x1。
import numpy as np
import scipy as sp
import pandas as pd
from scipy import optimize
if __name__ == "__main__":
urlSheatherData = "http://www.stat.tamu.edu/~sheather/book/docs/datasets/MichelinNY.csv"
data = pd.read_csv(urlSheatherData)
Xs = np.vstack(data[['Service','Decor', 'Food', 'Price']].values)
Xs = np.concatenate((np.vstack(np.ones(Xs.shape[0])),Xs), axis=1)
Ys = np.vstack(data[['InMichelin']].values)
# optimal solution (given)
vBeta = np.array([-1.49209249, -0.01117662, 0.044193, 0.05773374, 0.00179794]).reshape(5,1)
print Ys.shape, Xs.shape, vBeta.shape
# first definition of function
def fnRSS(vBeta, vY, mX):
return np.sum((vY - np.dot(mX, vBeta))**2)
print fnRSS(vBeta, Ys, Xs) # correct value
print np.linalg.lstsq(Xs, Ys)[1] # confirm correct value
print sp.optimize.fmin_bfgs(fnRSS, x0=vBeta, args=(Ys,Xs)) # wrong value
# second definition
def fnRSS(vBeta, vY, mX):
return np.sum((vY.T - np.dot(mX, vBeta))**2)
print fnRSS(vBeta, Ys, Xs) # incorrect value
print sp.optimize.fmin_bfgs(fnRSS, x0=vBeta, args=(Ys,Xs)) # correct convergence but simple call gives different value
我的输出:
(164, 1) (164, 5) (5, 1)
26.3239061505
[ 26.32390615]
Warning: Desired error not necessarily achieved due to precision loss.
Current function value: 6660.000000
Iterations: 39
Function evaluations: 3558
Gradient evaluations: 480
[ 4.51220111e-01 1.32711255e-07 8.09143368e-08 -1.06633003e-07
-5.18448332e-08]
9002.87916028
Warning: Desired error not necessarily achieved due to precision loss.
Current function value: 26.323906
Iterations: 29
Function evaluations: 1954
Gradient evaluations: 260
[-1.49209095 -0.0111764 0.04419313 0.05773347 0.00179789]
答案 0 :(得分:0)
它不是关于vY.T
,而是vBeta
,即x
被fmin_bfgs
传递给fnRSS
,而不是2d矢量,而是1d阵列。因此,尽管您试图将x0=vBeta
指定为形状数组(5,1),但它会在内部转换为1d数组形状(5,),最后会返回。< / p>