Question

我有一些代码需要帮助矢量化。我想将以下内容转换为矢量形式，我该怎么办？我想摆脱内循环 - 显然，它可以这样做。 X是NxD矩阵。 y是1xD向量。

def foo(X, y, mylambda, N, D, epsilon): 
... 
    for j in xrange(D): 
        aj = 0 
        cj = 0 
        for i in xrange(N): 
            aj += 2 * (X[i,j] ** 2) 
            cj += 2 * (X[i,j] * (y[i] - w.transpose()*X[i].transpose() + w[j]*X[i,j])) 

...

如果我在函数上调用numpy.vectorize（），它会在运行时抛出错误。

完整代码：

import scipy
import scipy.io
import numpy
from numpy import linalg
from scipy import *

def data(N, d, k, sigma, seed=12231):
    random.seed(seed)
    X = randn(N, d)
    wg = zeros(1 + d)
    wg[1:k + 1] = 10 * sign(randn(k))
    eps = randn(N) * sigma
    y = X.dot(wg[1:]) + wg[0] + eps
    return (y, X)


def foo(X, y, mylambda, n, D, epsilon):
    identityMatrix = numpy.matrix(numpy.identity(D))

    w = (X.transpose() * X + mylambda * identityMatrix).getI() * X.transpose() * y
    newweight = (X.transpose() * X + mylambda * identityMatrix).getI() * X.transpose() * y

    iterate = 1
    iteration = 0

    while iterate > 0 and iteration < 10000:
        iteration += 1
        iterate = 0
        maxerror = 0
        for j in xrange(D):
            aj = 0
            cj = 0
            for i in xrange(n):
                aj += 2 * (X[i,j] ** 2)
                cj += 2 * (X[i,j] * (y[i] - w.transpose()*X[i].transpose() + w[j]*X[i,j]))

            if cj < -mylambda:
                newweight[j,0] = (cj + mylambda)/ aj
            elif cj > mylambda:
                newweight[j,0] = (cj - mylambda)/ aj
            else:
                newweight[j,0] = 0

            if abs(newweight[j,0] - w[j,0]) > epsilon:
                iterate += 1
            if abs(newweight[j,0] - w[j,0]) > maxerror:
                maxerror = abs(newweight[j,0] - w[j,0])            
            w[j,0] = newweight[j,0]

N, D, k = 50, 75, 5
(y, X) = data(N, D, k, 1, 123)
X = numpy.matrix(X)
y = numpy.matrix(y).transpose()
foo(X, y, 1, N, D, 0.1)

Answer 1

您可以替换：

aj = 0
cj = 0
for i in xrange(n):
    aj += 2 * (X[i,j] ** 2)
    cj += 2 * (X[i,j] * (y[i] - w.transpose()*X[i].transpose() + w[j]*X[i,j]))

使用：

aj = 2*np.sum(X[:,j].T*X[:,j])
cj = 2*np.sum(np.multiply(X[:, j].T, (y.T - w.T*X.T + w[j] * X[:, j].T)))

需要帮助矢量化一些Python代码

1 个答案: