Numpy转Cython：编译错误：无法将“npy_intp *”转换为Python对象

Question

我有以下代码需要正确转换为Cython：

from numpy import *

## returns winning players or [] if undecided. 
def score(board):
    scores = []
    checked = zeros(board.shape)
    for i in xrange(len(board)):
        for j in xrange(len(board)):
            if checked[i,j] == 0 and board[i,j] !=0:
                ... do stuf

我尝试将代码转换为Cython：

import numpy as np
cimport numpy as np

@cython.boundscheck(False)
@cython.wraparound(False)
@cython.nonecheck(False)

## returns winning players or [] if undecided. 
def score(np.ndarray[int, ndim=2] board):
    scores = []
    cdef np.ndarray[int, ndim = 2 ] checked
    checked = np.zeros(board.shape)
    for i in xrange(len(board)):
        for j in xrange(len(board)):
            if checked[i,j] == 0 and board[i,j] !=0:
                ... do stuf

但当我编译时，我得到了：

$ python setup.py build_ext --inplace
running build_ext
cythoning newgox.pyx to newgox.c


## returns winning players or [] if undecided. 
def score(np.ndarray[int, ndim=2] board):
    scores = []
    cdef np.ndarray[int, ndim = 2 ] checked
    checked = np.zeros(board.shape)
                       ^
------------------------------------------------------------

newgox.pyx:58:28: Cannot convert 'npy_intp *' to Python object 
building 'newgox' extension

另外，我不确定这是否是在cython中处理列表的正确方式：

scores = []
if some_stuff_is_true:
    scores.append(some_integer)

编辑：

谢谢，代码现在可以编译了，但是当我运行它时，我得到了以下错误：

File "newgox.pyx", line 63, in newgox.score (newgox.c:1710)
    def score(np.ndarray[np.int, ndim=2] board):
ValueError: Buffer dtype mismatch, expected 'int object' but got 'long'

我尝试了这两个选项：

ctypedef np.int_t DTYPE_t
DTYPE = np.int

然后继续：

board = zeros((5,5), dtype = DTYPE)
def score(np.ndarray[DTYPE, ndim=2] board):

或者在两个声明中都使用：

    ...  np.int ...

这将导致相同的错误，但是签名如下：

 ValueError: Buffer dtype mismatch, expected 'int' but got 'long'

感谢bellamyj提供的建议，但是你所提供的代码无法编译。

$ python setup.py build_ext --inplace
running build_ext
cythoning newgox.pyx to newgox.c
building 'newgox' extension
gcc-4.2 -fno-strict-aliasing -fno-common -dynamic -arch i386 -arch x86_64 -g -O2 -DNDEBUG -g -O3 -I/Library/Frameworks/Python.framework/Versions/2.7/include/python2.7 -c newgox.c -o   build/temp.macosx-10.6-intel-2.7/newgox.o

newgox.c:238:31: error: numpy/arrayobject.h: No such file or directory
newgox.c:239:31: error: numpy/ufuncobject.h: No such file or directory
newgox.c:356: error: expected ‘=’, ‘,’, ‘;’, ‘asm’ or ‘__attribute__’ before ‘__pyx_t_5numpy_int8_t’
newgox.c:365: error: expected ‘=’, ‘,’, ‘;’, ‘asm’ or ‘__attribute__’ before ‘__pyx_t_5numpy_int16_t’
newgox.c:374: error: expected ‘=’, ‘,’, ‘;’, ‘asm’ or ‘__attribute__’ before ‘__pyx_t_5numpy_int32_t’

它继续列举所有其他类型。

然后告诉我这个：

newgox.c:978: error: expected ‘)’ before ‘*’ token
newgox.c:979: error: expected ‘)’ before ‘*’ token
newgox.c:980: error: expected ‘)’ before ‘*’ token
newgox.c:983: error: expected declaration specifiers or ‘...’ before ‘PyArrayObject’
newgox.c:984: error: expected declaration specifiers or ‘...’ before ‘PyArrayObject’
newgox.c:985: error: expected ‘)’ before ‘*’ token
newgox.c:987: error: ‘__pyx_t_5numpy_int32_t’ undeclared here (not in a function)
newgox.c: In function ‘__pyx_pf_6newgox_7score’:
newgox.c:1638: error: ‘PyArrayObject’ undeclared (first use in this function)
newgox.c:1638: error: (Each undeclared identifier is reported only once
newgox.c:1638: error: for each function it appears in.)
newgox.c:1638: error: ‘__pyx_v_checked’ undeclared (first use in this function)
newgox.c:1659: error: ‘__pyx_t_5’ undeclared (first use in this function)
newgox.c:1721: error: expected expression before ‘)’ token
newgox.c:1723: error: expected expression before ‘)’ token
newgox.c:1747: error: expected expression before ‘)’ token
newgox.c:1766: error: expected expression before ‘)’ token
newgox.c:1813: error: expected expression before ‘)’ token
newgox.c:1846: error: expected expression before ‘)’ token
newgox.c:1846: error: too many arguments to function ‘__pyx_f_6newgox_check_life’
newgox.c:2009: error: expected expression before ‘)’ token
newgox.c:2009: warning: assignment makes pointer from integer without a cast
newgox.c:2012: error: expected expression before ‘)’ token
newgox.c:2032: error: expected expression before ‘)’ token
newgox.c: At top level:
newgox.c:2088: error: expected declaration specifiers or ‘...’ before ‘PyArrayObject’
newgox.c: In function ‘__pyx_f_6newgox_check_life’:
newgox.c:2124: error: ‘__pyx_v_board’ undeclared (first use in this function)
newgox.c:2160: error: ‘PyArrayObject’ undeclared (first use in this function)
newgox.c:2160: error: expected expression before ‘)’ token
newgox.c:2160: error: too many arguments to function ‘__pyx_f_6newgox_liberty’
newgox.c:2420: error: too many arguments to function ‘__pyx_f_6newgox_check_life’
newgox.c: At top level:
newgox.c:2583: error: expected declaration specifiers or ‘...’ before ‘PyArrayObject’
newgox.c: In function ‘__pyx_f_6newgox_liberty’:
newgox.c:2610: error: ‘__pyx_v_board’ undeclared (first use in this function)
newgox.c: At top level:
newgox.c:2859: error: expected ‘)’ before ‘*’ token
newgox.c: In function ‘__pyx_pf_5numpy_7ndarray___getbuffer__’:
newgox.c:2999: error: ‘PyArray_Descr’ undeclared (first use in this function)
newgox.c:2999: error: ‘__pyx_v_descr’ undeclared (first use in this function)
newgox.c:3062: error: ‘PyArrayObject’ undeclared (first use in this function)
newgox.c:3062: error: expected expression before ‘)’ token
newgox.c:3071: error: ‘npy_intp’ undeclared (first use in this function)
newgox.c:3114: error: expected expression before ‘)’ token
newgox.c:3114: error: ‘NPY_C_CONTIGUOUS’ undeclared (first use in this function)
newgox.c:3154: error: expected expression before ‘)’ token
newgox.c:3154: error: ‘NPY_F_CONTIGUOUS’ undeclared (first use in this function)
newgox.c:3184: error: expected expression before ‘)’ token
newgox.c:3184: warning: assignment makes pointer from integer without a cast
newgox.c:3240: error: expected expression before ‘)’ token
newgox.c:3240: error: subscripted value is neither array nor pointer
newgox.c:3249: error: expected expression before ‘)’ token
newgox.c:3249: error: subscripted value is neither array nor pointer
newgox.c:3262: error: expected expression before ‘)’ token
newgox.c:3271: error: expected expression before ‘)’ token
newgox.c:3291: error: expected expression before ‘)’ token

还有一些。

请注意，它说：

newgox.c:2160: error: too many arguments to function ‘__pyx_f_6newgox_liberty’
newgox.c:2420: error: too many arguments to function ‘__pyx_f_6newgox_check_life’

但是我从score/1调用check_life的方法：

life, newly_checked = check_life(i,j,board,[])

接收数据的方法如下：

# helper functions of score/1
cdef check_life(int i, int j, np.ndarray[int, ndim=2] board, checked):
    ...
    return life, checked

最后，你使用的“i4”是什么数据类型？

- Ihmahr

Answer 1

我认为在 Cython 中，元组拆包的用法和 Python 不同。你需要使用 (board.shape[0], board.shape[1]) 来指定 checked 的大小。

你还需要指定 checked 的数据类型。默认情况下，np.zeros 返回一个 float64 数据类型。你需要将其改为 int32 数据类型以与声明的类型兼容。

这是一个经过修改后应该可以运行的版本。

def score(np.ndarray[int, ndim=2] board):

    cdef np.ndarray[np.int32_t, ndim = 2] checked
    checked = np.zeros((board.shape[0], board.shape[1]), dtype='i4')

    for i in xrange(len(board)):
        for j in xrange(len(board)):
            if checked[i,j] == 0 and board[i,j] !=0:
                ... do stuff

回答你问题的第二部分：是的，你可以在Cython函数中添加到列表中。然而，据我所知，这涉及Python函数调用，因此它不会比Python内部的相同操作更快。

更新1：

我认为那些错误是因为编译器找不到NumPy头文件。请更新你的setup.py以包括这些文件的路径。该路径应由np.get_include()函数输出。

import numpy as np
from distutils.core import setup
from distutils.extension import Extension
from Cython.Distutils import build_ext

setup(
      cmdclass={'build_ext': build_ext},
      ext_modules = [
              Extension("new_gox",
                        ["new_gox.pyx"],
                        include_dirs=[np.get_include()])
                    ]
     )

dtype i4 表示一个32位整数，它等同于 np.int32。

对于 'too many arguments' 错误，我不太确定具体情况。如果你在Python中调用该函数，它需要被声明为def或cpdef。使用cdef声明的函数只能从Cython中调用。 在这里的Cython文档中有描述。

- joshayers

Answer 2

np.intp与(void *)相同

定义其类型如下：

ctypedef np.npy_intp SIZE_t              # Type for indices and counters

翻译：

NPY_INTP == 有符号整数类型的枚举值，其大小与（void *）指针相同。这是所有索引数组使用的类型。

因此，例如在Cython中...

头文件.h文件中的结构体：

struct Cudamat_Struct1 {
    float * dataHost;
    float * dataDevice;
    int onDevice;
    int onHost;
    (void *) shape;
    int size;
    int isTrans; // 0 or 1
    int ownsData;

};

在Cython文件.pxd中：

cdef extern from "file.h":
    cdef cppclass c_Cudamat_Struct "Cudamat_Struct":
        float *dataHost
        float *dataDevice
        bint onDevice
        bint onHost
        SIZE_t * shape
        int size
        bint isTrans
        bint ownsData

最后在Cython函数中：

cdef pycublas_thrust_2d():
    cdef c_Cudamat_Struct1* mat = new c_Cudamat_Struct1()

    cdef int N = 10

    cdef np.ndarray[ndim=2, dtype=F32t] arrHost = np.ones((1,10),dtype=F32)

    cdef int size = 10

    cdef F32t[:,:] arrHost_ = arrHost

    cdef F32t * arrHostPtr = &arrHost[0,0]

    mat.dataHost = arrHostPtr
    mat.shape = arrHost.shape
    mat.size = size

- yourstruly