C\C++ 提升Python效率

gapore

有没有大佬懂cython的啊？我有段程序想用cython加速一下，但是自己不懂c语言，胡乱编的效果有限

数据来源代码（Python）如下

import numpy as np
import scipy.linalg as linalg

N = 2328
Y = np.random.rand(N)
L = 200   # int in range(2,N/2)
K = N - L + 1

# Embedding
X = linalg.hankel(Y,np.zeros(K))
X = X[:-K+1,:]

# Singular value decomposition (SVD)
Xt = X.T
_,sig,Vt = linalg.svd(X)
Sig = linalg.diagsvd(sig,L,L)
r = np.linalg.matrix_rank(X)
Vt1 = Vt[:r,:]
Sig1 = Sig[:r,:r]
B1 = Sig1.dot(Vt1)
B1t = B1.T

我想用最小一乘法求解矩阵 A ，使得 B1t * A = Xt ，编的python代码如下

# coding = utf-8
import numpy as np
cimport cython

def FitLadReg(double[:] endog,double[:,:] exog):
    
    exog_rank = np.linalg.matrix_rank(exog)
    beta = np.ones(exog_rank)
    xstar = exog

    n_iter = 0
    diff = 10
    cycle = False
    history = []
    while n_iter < 1000 and diff > 1e-6 and not cycle:
        n_iter += 1
        beta0 = beta
        xtx = np.dot(xstar.T, exog)    
        xty = np.dot(xstar.T, endog)
        beta = np.dot(np.linalg.pinv(xtx), xty)
        resid = endog - np.dot(exog, beta)

        mask = np.abs(resid) < .000001
        resid[mask] = ((resid[mask] >= 0) * 2 - 1) * .000001
        resid = np.where(resid < 0, 0.5 * resid, 0.5 * resid)
        resid = np.abs(resid)
        xstar = exog / resid[:, np.newaxis]
        diff = np.max(np.abs(beta - beta0))
        history.append(beta)

        if (n_iter >= 300) and (n_iter % 100 == 0):
            # check for convergence circle, shouldn't happen
            for ii in range(2, 10):
                if np.all(beta == history[-ii]):
                    cycle = True
                    break

    return history

编译后调用代码如下

from FitLadReg import FitLadReg
A = np.zeros((r,L))
for i in range(L):
    A[:,i] = FitLadReg(Xt[:,i],B1t)[0]

FitLadReg函数cython编译前后效率差不多，请问要怎么改进函数代码呢？谢谢！！！

TCY

你用numba吧……