使用 ctypes 从 C 结构构造 numpy 数组。 值错误：' <p' is not a valid pep 3118 buffer format string< div><div id="text_translate"><p> 考虑以下代码：</p><p> by_ref.h</p><pre> typedef struct OutPutImage{ double *** output_img; int nb_images; int nb_cols; int nb_rows; }opi_; void test_output_images(struct OutPutImage * out, int nb_images, int nb_cols, int nb_rows);</pre><p> by_ref.c</p><pre> #include &lt;stdlib.h&gt; #include "by_ref.h" void test_output_images(struct OutPutImage* out, int nb_images, int nb_cols, int nb_rows){ out-&gt;nb_images = nb_images; out-&gt;nb_cols = nb_cols; out-&gt;nb_rows = nb_rows; out-&gt;output_img = (double***)malloc((nb_images) * sizeof(double**)); for(int i = 0; i &lt; nb_images; i++){ out-&gt;output_img[i] = (double**)malloc((nb_cols) * sizeof(double*)); for(int j = 0; j &lt; nb_cols; j++){ out-&gt;output_img[i][j] = (double*)malloc((nb_rows) * sizeof(double)); for(int k = 0; k &lt; nb_rows; k++){ out-&gt;output_img[i][j][k] = 0; } } } }</pre><p> 和</p><p> by_ref.py</p><pre> import ctypes import numpy.ctypeslib as npct import numpy as np class OutPutImage(ctypes.Structure): _fields_ = [('output_img', npct.ndpointer(dtype=np.double, ndim=3)), ('nb_images', ctypes.c_int), ('nb_cols', ctypes.c_int), ('nb_rows', ctypes.c_int)] _libc = ctypes.CDLL("./by_ref.so") def __init__(self, nb_images=None, nb_cols=None, nb_rows=None): self.nb_images = nb_images self.nb_cols = nb_cols self.nb_rows = nb_rows if __name__ == '__main__': libc_adm = ctypes.CDLL("./by_ref.so") libc_adm.test_output_images.restype = ctypes.c_int libc_adm.test_output_images.argtypes = [ctypes.POINTER(OutPutImage), ctypes.c_int, ctypes.c_int, ctypes.c_int] output_image = OutPutImage(1, 2, 3) libc_adm.test_output_images(ctypes.byref(output_image), 4, 5, 6) print(np.array(output_image.output_img, dtype=np.float)) # error ocures here</pre><p> 当我使用以下 Makefile 制作并运行此代码时</p><pre>by_ref: by_ref.so python by_ref.py by_ref.so: by_ref.o gcc -shared -o by_ref.so by_ref.o by_ref.o: by_ref.c gcc -c -Wall -fpic by_ref.c -o by_ref.o</pre><p> 我得到错误：</p><pre> Traceback (most recent call last): File "by_ref.py", line 46, in &lt;module&gt; print(np.array(output_image.output_img, dtype=np.float)) ValueError: '&lt;P' is not a valid PEP 3118 buffer format string make: *** [Makefile:2: by_ref] Error 1</pre><p> 我确信test_output_images正在做它应该做的事情。 但是我无法从结构中的数据构建 numpy 数组。 我该如何做到这一点？ 另外，我什么时候释放 memory？ 谢谢。</p><p> <strong>编辑：</strong></p><p> 如果我使用np.ctypeslib.as_array(output_image.output_img)我得到同样的错误：</p><pre> ValueError: '&lt;P' is not a valid PEP 3118 buffer format string</pre><p> <strong>更新：</strong></p><p> 如果我使用x = np.ctypeslib.as_array(( ctypes.c_double*array_length ).from_address( ctypes.addressof(output_image.output_img) )) ，其中array_length=nb_images*nb_cols*nb_rows ，那么我避免了上述错误，但是新数组 x 包含垃圾，无法重塑。</p></div></p'>

Question

Consider the following code:

by_ref.h

typedef struct OutPutImage{
    double *** output_img;
    int nb_images;
    int nb_cols;
    int nb_rows;
}opi_;

void test_output_images(struct OutPutImage * out, int nb_images, int nb_cols, int nb_rows);

by_ref.c

#include <stdlib.h>
#include "by_ref.h"

void test_output_images(struct OutPutImage* out, int nb_images, int nb_cols, int nb_rows){
    out->nb_images = nb_images;
    out->nb_cols = nb_cols;
    out->nb_rows = nb_rows;

    out->output_img = (double***)malloc((nb_images) * sizeof(double**));
    for(int i = 0; i < nb_images; i++){
        out->output_img[i] = (double**)malloc((nb_cols) * sizeof(double*));
        for(int j = 0; j < nb_cols; j++){
            out->output_img[i][j] = (double*)malloc((nb_rows) * sizeof(double));
            for(int k = 0; k < nb_rows; k++){
                out->output_img[i][j][k] = 0;
            } 
        }
    }
}

and

by_ref.py

import ctypes
import numpy.ctypeslib as npct
import numpy as np

class OutPutImage(ctypes.Structure):

    _fields_ = [('output_img', npct.ndpointer(dtype=np.double, ndim=3)), 
                ('nb_images', ctypes.c_int), 
                ('nb_cols', ctypes.c_int),
                ('nb_rows', ctypes.c_int)]

    _libc = ctypes.CDLL("./by_ref.so")

    def __init__(self, nb_images=None, nb_cols=None, nb_rows=None):
        self.nb_images = nb_images
        self.nb_cols = nb_cols
        self.nb_rows = nb_rows

if __name__ == '__main__':
    libc_adm = ctypes.CDLL("./by_ref.so")
    libc_adm.test_output_images.restype = ctypes.c_int
    libc_adm.test_output_images.argtypes = [ctypes.POINTER(OutPutImage), ctypes.c_int,
                                            ctypes.c_int, ctypes.c_int]

    output_image = OutPutImage(1, 2, 3)
    libc_adm.test_output_images(ctypes.byref(output_image), 4, 5, 6)  
    print(np.array(output_image.output_img, dtype=np.float))    # error ocures here

When I make and run this code using the following Makefile

by_ref: by_ref.so
    python by_ref.py

by_ref.so: by_ref.o
    gcc -shared -o by_ref.so by_ref.o

by_ref.o: by_ref.c
    gcc -c -Wall -fpic by_ref.c -o by_ref.o

I get the error:

Traceback (most recent call last):
  File "by_ref.py", line 46, in <module>
    print(np.array(output_image.output_img, dtype=np.float))
ValueError: '<P' is not a valid PEP 3118 buffer format string
make: *** [Makefile:2: by_ref] Error 1

I am sure test_output_images is doing what it is supposed to do. However I can't build the numpy array from the data in the structure. How do I accomplish this? Also, when do I free the memory? Thanks.

Edit:

If I use np.ctypeslib.as_array(output_image.output_img) I get the same error:

ValueError: '<P' is not a valid PEP 3118 buffer format string

Update:

If I use x = np.ctypeslib.as_array(( ctypes.c_double*array_length ).from_address( ctypes.addressof(output_image.output_img) )) , where array_length=nb_images*nb_cols*nb_rows , then I avoid the error described above, but the new array x contains garbage and it can't be reshaped.

Answer 1

It's a bug in NumPy 1.15.x. Either downgrade to 1.14.6 or older, or upgrade to 1.16.0 or newer.