cython 中没有 python 个对象的高效矩阵运算

Efficient matrix operations in cython with no python objects

我正在编写一个函数,我想使用 Cython 尽可能多地将其转换为 C。为此,我需要使用线性代数运算。这是我的功能。 编辑:我学到的教训是尝试处理循环之外的线性代数,这在很大程度上是我能够做到的。否则,求助于包装 LAPACK/BLAS 或编写我自己的函数。

import numpy as np
from scipy.stats import multivariate_normal as mv
import itertools

def llf(data, rho, mu, sigma, A, V, n):
    '''evaluate likelihood by guass-hermite quadrature

    Parameters
    ----------
    data : array
        N x J matrix, columns are measurements
    rho : array
        length L vector of weights for mixture of normals
    mu : array
        L x K vector of means of mixture of normals
    sigma : array
        K x L*K matrix of variance matrices for mixture of normals
    A : array
        J x (K + 1) matrix of loadings
    V : array
        J x J variance matrix of measurement errors
    n : int
        number of sample points for quadrature
    '''

    N = data.shape[0]
    L, K = mu.shape

    # getting weights and sample points for approximating integral
    v, w = np.polynomial.hermite.hermgauss(n)

    totllf = 0
    for i in range(N):
        M_i = data[i, :]
        totllf_i = 0
        for l in range(L):
            rho_l = rho[l]
            sigma_l = sigma[:, K*l:K*(l+1)]
            mu_l = mu[l, :]
            chol_l = np.linalg.cholesky(sigma_l)
            for ix in itertools.product(*(list(range(n)) for k in range(K))):
               wt =  np.prod(w[list(ix)])
               pt = np.sqrt(2)*chol_l.dot(v[list(ix)]) + mu_l
               totllf_i += wt*rho_l*mv.pdf(M_i, A[:, 0] + A[:, 1:].dot(pt), V)

        totllf += np.log(totllf_i)

    return totllf

为了完成这个,我需要有矩阵乘法、转置、行列式、矩阵求逆和 cholesky 分解的函数。我看过 some posts 使用 BLAS 函数,但我真的不清楚如何使用这些函数。

编辑 2018 年 4 月 29 日

按照建议,我采用了内存视图方法并在循环之前初始化了所有内容。我的新函数写成

def llf_c(double[:, ::1] data, double[::1] rho, double[:, ::1] mu, 
                double[:, ::1] sigma, double[:, ::1] A, double[:, ::1] V, int n):
    '''evaluate likelihood by guass-hermite quadrature

    Parameters
    ----------
    data : array
        N x J matrix, columns are measurements
    rho : array
        length L vector of weights for mixture of normals
    mu : array
        L x K vector of means of mixture of normals
    sigma : array
        K x L*K matrix of variance matrices for mixture of normals
    A : array
        J x (K + 1) matrix of loadings
    V : array
        J x J variance matrix of measurement errors
    n : int
        number of sample points for quadrature
    '''

    cdef Py_ssize_t N = data.shape[0], J = data.shape[1], L = mu.shape[0], K = mu.shape[1]

    # initializing indexing variables
    cdef Py_ssize_t i, l, j, k

    # getting weights and sample points for approximating integral
    v_a, w_a = np.polynomial.hermite.hermgauss(n)
    cdef double[::1] v = v_a
    cdef double[::1] w = w_a
    cdef double[::1] v_ix = np.zeros(K, dtype=np.float)

    # initializing memory views for cholesky decomposition of sigma matrices
    sigma_chol_a = np.zeros((K, L*K), dtype=np.float)
    for l in range(L):
        sigma_chol_a[:, K*l:K*(l+1)] = np.linalg.cholesky(sigma[:, K*l:K*(l+1)])

    cdef double[:, ::1] sigma_chol = sigma_chol_a

    # intializing V inverse and determinant
    cdef double[:, ::1] V_inv = np.linalg.inv(V)
    cdef double V_det = np.linalg.det(V)

    # initializing memoryviews for work matrices
    cdef double[::1] work_K = np.zeros(K, dtype=np.float)
    cdef double[::1] work_J = np.zeros(J, dtype=np.float)

    # initializing memoryview for quadrature points
    cdef double[::1] pt = np.zeros(K, dtype=np.float)

    # initializing memorview for means for multivariate normal
    cdef double[::1] loc = np.zeros(J, dtype=np.float)

    # initializing values for loop
    cdef double[::1] totllf = np.zeros(N, dtype=np.float)
    cdef double wt, pdf_init = 1./sqrt(((2*pi)**J)*V_det)
    cdef int[:, ::1] ix_vals = np.vstack(itertools.product(*(list(range(n)) for k in range(K)))).astype(np.int32)
    cdef Py_ssize_t ix_len = ix_vals.shape[0]

    for ix_row in range(ix_len):

        ix = ix_vals[ix_row]

        # weights and points for quadrature
        wt = 1.
        for k in range(K):
            wt *= w[ix[k]]
            v_ix[k] = v[ix[k]]

        for l in range(L):

            # change of variables
            dotmv_c(sigma_chol[:, K*l:K*(l+1)], v_ix, work_K)
            for k in range(K):
                pt[k] = sqrt(2)*work_K[k]
            addvv_c(pt, mu[l, :], pt)

            for i in range(N):

                # generating demeaned vector for multivariate normal pdf
                dotmv_c(A[:, 1:], pt, work_J)
                addvv_c(A[:, 0], work_J, work_J)
                for j in range(J):
                    loc[j] = data[i, j] - work_J[j]

                # performing matrix products in exponential
                # print(wt, rho[l], np.asarray(work_J))
                dotvm_c(loc, V_inv, work_J)
                totllf[i] += wt*rho[l]*pdf_init*exp(-0.5*dotvv_c(work_J, loc))

    return np.log(np.asarray(totllf)).sum()

dotvm_cdotmv_caddvv_c是执行向量与矩阵的矩阵乘法、矩阵与向量的矩阵乘法以及两个向量的元素相加的函数。我也在 Cython 中编写了这些,但为简洁起见不包括在内。我不再像在使用 numpy 循环之前处理所有其他线性代数那样包装任何 LAPACK 函数。我还有几个问题。为什么我的循环中仍然有黄色? (见下面的简介)。我认为现在一切都应该在 C 中。另外,如果您对新的实施有任何其他建议,请告诉我。

例如,在第 221 行,我在编译时收到此消息:"Index should be typed for more efficient access." 但我认为我输入了索引 k。另外,由于 addvv_c 显示为黄色,我将在下面向您展示它的定义。

cpdef void addvv_c(double[:] a, double[:] b, double[:] out):
    '''add two vectors elementwise
    '''
    cdef Py_ssize_t i, n = a.shape[0]

    for i in range(n):
        out[i] = a[i] + b[i]

关于你优化的Cython/BLAS函数的几点小意见:

ipiv_a = np.zeros(n).astype(np.int32)
cdef int[::1] ipiv = ipiv_a

可以有两个简单的改进:它不必通过临时变量,你可以直接创建一个类型 np.int32 的数组,而不是创建一个不同类型的数组然后转换:

 cdef int[::1] ipiv = np.zeros(n,dtype=np.int32)

同样,在这两个函数中,您可以通过执行

以更少的步骤初始化 B
cdef double[:, ::1] B = A.copy()

而不是创建零数组并复制


第二个(更重要的)变化是 to use C arrays 用于临时变量,例如 Fortran 工作区。我仍然将 return 值之类的东西保留为 numpy 数组,因为引用计数和将它们发送回 Python 的能力 非常 有用。

 cdef double* work = <double*>malloc(n*n*sizeof(double))
 try:
     # rest of function
 finally:
     free(work)

您需要从 libc.stdlib 导入 mallocfreetry: ... finally: 确保内存被正确释放。不要太过头了 - 例如,如果在何处释放 C 数组并不明显,那么只需使用 numpy。


要查看的最后一个选项是没有 return 值但修改输入:

cdef void inv_c(double[:,::1] A, double[:,::1] B):
    # check that A and B are the right size, then just write into B
    # ...

这样做的好处是,如果您需要在具有相同大小输入的循环中调用它,那么您只需为整个循环进行一次分配。您可以将其扩展为也包含工作数组,尽管这可能会稍微复杂一些。