自定义numpy类型导致numpy.mean中的崩溃
Custom numpy type causes crash in numpy.mean
我定义了自己的numpy类型,具有C 中的基本功能。数组创建有效,并且可以添加和划分此类类型的数组。可以使用np.sum
减少数组,但使用np.mean
时,解释器会崩溃。
import numpy as np
from example import Dual
x = np.array([Dual(1, 2), Dual(1, 3), Dual(1, 1)])
print(x.dtype) # dtype(Dual) -- correct type
# Mean 1
m = x.sum() / x.size
print(m.x, m.y) # 1.0 2.0 -- correct result
# Mean 2
x.mean() # -- CRASH
通过在我的C 代码中插入调试输出,我验证了在我的代码中正确计算的总和(实际上是3/3和6/3(。因此,该功能在计算我的结果后崩溃。(Windows告诉我 python停止工作,所以大概发生了引擎盖下的segfault。(
什么是mean
与sum()/n
不同?我如何修改我的类型以使其正常工作?
我试图将我的C 代码煮沸以创建一个最小而完整的示例。不幸的是,在定义工作numpy类型时,最小仍然相当长。这是实现:
#include <Python.h>
#include <numpy/arrayobject.h>
#include <numpy/npy_math.h>
#include <numpy/ufuncobject.h>
#include "structmember.h"
#include <iostream>
#include <math.h>
struct Dual
{
public:
double x;
double y;
Dual(double x, double y) : x(x), y(y) { }
inline static Dual add(const Dual& a, const Dual& b) {
return Dual(a.x + b.x, a.y + b.y); }
inline static Dual div(const Dual& a, const Dual& b) {
return Dual(a.x / b.x, a.y / b.y); }
};
typedef struct {
PyObject_HEAD
Dual ob_val;
} PyDual;
PyArray_Descr* dual_descr;
static PyObject *
PyDual_new(PyTypeObject *type, PyObject *args, PyObject *kwds)
{
PyDual *self;
self = (PyDual *)type->tp_alloc(type, 0);
if (self == NULL) {
return NULL;
}
double x = 0;
double y = 0;
static char *kwlist[] = {"x", "y", NULL};
if(!PyArg_ParseTupleAndKeywords(args, kwds, "d|d", kwlist, &x, &y))
return NULL;
self->ob_val = Dual(x, y);
return (PyObject *)self;
}
static PyTypeObject PyDual_Type = {
PyVarObject_HEAD_INIT(NULL, 0)
"example.Dual", /* tp_name */
sizeof(PyDual), /* tp_basicsize */
0, /* tp_itemsize */
0, /* tp_dealloc */
0, /* tp_print */
0, /* tp_getattr */
0, /* tp_setattr */
0, /* tp_reserved */
0, /* tp_repr */
0, /* tp_as_number */
0, /* tp_as_sequence */
0, /* tp_as_mapping */
0, /* tp_hash */
0, /* tp_call */
0, /* tp_str */
0, /* tp_getattro */
0, /* tp_setattro */
0, /* tp_as_buffer */
Py_TPFLAGS_DEFAULT, /* tp_flags */
"Dual value/derivative", /* tp_doc */
0, /* tp_traverse */
0, /* tp_clear */
0, /* tp_richcompare */
0, /* tp_weaklistoffset */
0, /* tp_iter */
0, /* tp_iternext */
0, /* tp_methods */
0, /* tp_members */
0, /* tp_getset */
0, /* tp_base */
0, /* tp_dict */
0, /* tp_descr_get */
0, /* tp_descr_set */
0, /* tp_dictoffset */
0, /* tp_init */
0, /* tp_alloc */
PyDual_new, /* tp_new */
};
static int PyDual_Check(PyObject* obj)
{
return PyObject_IsInstance(obj, (PyObject*)&PyDual_Type);
}
static PyObject*
PyDual_from_Dual(Dual d)
{
PyDual *obj = (PyDual*)PyDual_Type.tp_alloc(&PyDual_Type, 0);
if(obj)
obj->ob_val = d;
return (PyObject*)obj;
}
static PyMemberDef PyDual_members[] = {
{"x", T_DOUBLE, offsetof(PyDual, ob_val) + offsetof(Dual, x), READONLY, "value"},
{"y", T_DOUBLE, offsetof(PyDual, ob_val) + offsetof(Dual, y), READONLY, "derivative"},
{NULL}
};
//----------------------------------------------------------------------------
static PyArray_ArrFuncs _PyDual_ArrFuncs;
static void
DUAL_copyswap(Dual *dst, Dual *src, int swap, void *NPY_UNUSED(arr))
{
PyArray_Descr *descr;
descr = PyArray_DescrFromType(NPY_DOUBLE);
descr->f->copyswapn(dst, sizeof(double), src, sizeof(double), 2, swap, NULL);
Py_DECREF(descr);
}
static int DUAL_setitem(PyObject* item, void* data, void* ap)
{
Dual d(0, 0);
if(PyDual_Check(item)) {
memcpy(data, &(((PyDual *)item)->ob_val),sizeof(Dual));
} else if(PySequence_Check(item) && PySequence_Length(item)==4) {
auto element = PySequence_GetItem(item, 0);
if(element == NULL) { return -1; } /* Not a sequence, or other failure */
d.x = PyFloat_AsDouble(element);
Py_DECREF(element);
element = PySequence_GetItem(item, 1);
if(element == NULL) { return -1; } /* Not a sequence, or other failure */
d.y = PyFloat_AsDouble(element);
Py_DECREF(element);
} else {
PyErr_SetString(PyExc_TypeError,
"Unknown input to DUAL_setitem");
return -1;
}
return 0;
}
static PyObject *DUAL_getitem(void* data, void* arr)
{
Dual d(0, 0);
memcpy(&d, data, sizeof(Dual));
return PyDual_from_Dual(d);
}
template<typename T>
scalar_to_dual(T* ip, Dual* op, npy_intp n,
PyArrayObject *NPY_UNUSED(aip), PyArrayObject *NPY_UNUSED(aop))
{
while(n--)
{
op->x = *ip++;
op->y = op->x;
}
}
static void register_cast_function(int sourceType, int destType, PyArray_VectorUnaryFunc *castfunc)
{
PyArray_Descr *descr = PyArray_DescrFromType(sourceType);
PyArray_RegisterCastFunc(descr, destType, castfunc);
PyArray_RegisterCanCast(descr, destType, NPY_NOSCALAR);
Py_DECREF(descr);
}
static void sum_ufunc(char** args, npy_intp* dimensions, npy_intp* steps, void* data)
{
char *ip1 = args[0], *ip2 = args[1], *op1 = args[2];
npy_intp is1 = steps[0], is2 = steps[1], os1 = steps[2];
npy_intp n = dimensions[0];
for(npy_intp i=0; i<n; i++, ip1+=is1, ip2+=is2, op1+=os1)
{
const Dual in1 = *(Dual*)ip1;
const Dual in2 = *(Dual*)ip2;
*((Dual*)op1) = Dual::add(in1, in2);
}
}
static void div_ufunc(char** args, npy_intp* dimensions, npy_intp* steps, void* data)
{
char *ip1 = args[0], *ip2 = args[1], *op1 = args[2];
npy_intp is1 = steps[0], is2 = steps[1], os1 = steps[2];
npy_intp n = dimensions[0];
for(npy_intp i=0; i<n; i++, ip1+=is1, ip2+=is2, op1+=os1)
{
const Dual in1 = *(Dual*)ip1;
const Dual in2 = *(Dual*)ip2;
*((Dual*)op1) = Dual::div(in1, in2);
std::cout << in1.x << "/" << in2.x << std::endl;
std::cout << in1.y << "/" << in2.y << std::endl;
}
}
//----------------------------------------------------------------------------
static struct PyModuleDef examplemodule = {
PyModuleDef_HEAD_INIT,
"example", /* name of module */
NULL, /* module documentation, may be NULL */
-1, /* size of per-interpreter state of the module,
or -1 if the module keeps state in global variables. */
NULL, NULL, NULL, NULL, NULL
};
PyMODINIT_FUNC PyInit_example(void)
{
// initialize numpy
import_array(); if(PyErr_Occurred()) return NULL;
import_umath(); if(PyErr_Occurred()) return NULL;
auto numpy = PyImport_ImportModule("numpy"); if(!numpy) return NULL;
auto numpy_dict = PyModule_GetDict(numpy); if(!numpy_dict) return NULL;
PyDual_Type.tp_members = PyDual_members;
PyDual_Type.tp_base = &PyGenericArrType_Type;
if( PyType_Ready(&PyDual_Type) < 0)
return NULL;
PyArray_InitArrFuncs(&_PyDual_ArrFuncs);
_PyDual_ArrFuncs.copyswap = (PyArray_CopySwapFunc*)DUAL_copyswap;
_PyDual_ArrFuncs.setitem = (PyArray_SetItemFunc*)DUAL_setitem;
_PyDual_ArrFuncs.getitem = (PyArray_GetItemFunc*)DUAL_getitem;
dual_descr = PyObject_New(PyArray_Descr, &PyArrayDescr_Type);
dual_descr->typeobj = &PyDual_Type;
dual_descr->kind = 'x';
dual_descr->type = 'x';
dual_descr->byteorder = '=';
dual_descr->flags = 0;
dual_descr->type_num = 0; // assigned at registration
dual_descr->elsize = 8*2;
dual_descr->alignment = 8;
dual_descr->subarray = NULL;
dual_descr->fields = NULL;
dual_descr->names = NULL;
dual_descr->f = &_PyDual_ArrFuncs;
dual_descr->metadata = NULL;
dual_descr->c_metadata = NULL;
Py_INCREF(&PyDual_Type);
auto dualNum = PyArray_RegisterDataType(dual_descr);
if(dualNum < 0) return NULL;
int AD = dual_descr->type_num;
register_cast_function(NPY_BOOL, dualNum, (PyArray_VectorUnaryFunc*)scalar_to_dual<npy_bool>);
register_cast_function(NPY_LONG, dualNum, (PyArray_VectorUnaryFunc*)scalar_to_dual<npy_long>);
register_cast_function(NPY_DOUBLE, dualNum, (PyArray_VectorUnaryFunc*)scalar_to_dual<npy_double>);
int arg_types[] = {AD, AD, AD};
PyUFuncObject* ufunc = (PyUFuncObject*)PyObject_GetAttrString(numpy, "add");
PyUFunc_RegisterLoopForType(ufunc, AD, sum_ufunc, arg_types, NULL);
ufunc = (PyUFuncObject*)PyObject_GetAttrString(numpy, "true_divide");
PyUFunc_RegisterLoopForType(ufunc, AD, div_ufunc, arg_types, NULL);
auto module = PyModule_Create(&examplemodule);
if( module == NULL )
return NULL;
Py_INCREF(&PyDual_Type);
PyModule_AddObject(module, "Dual", (PyObject*)&PyDual_Type);
return module;
}
回想起来,错误几乎是显而易见的,并且与实现的数字部分无关,而与python对象的初始化有关。荣誉转到 @hpaulj的评论,这使我走上了正确的轨道。
x.mean
和 np.mean
call np.core._methods.mean
,崩溃发生在该行中
ret = ret.dtype.type(ret / rcount)
这是一种冗余类型的转换(至少在我的情况下(,等效于
Dual(dual_object)
用文字,将Dual
构造函数称为Dual
的实例作为参数。但是,构造函数的实施期望两倍作为参数:
static char *kwlist[] = {"x", "y", NULL};
if(!PyArg_ParseTupleAndKeywords(args, kwds, "d|d", kwlist, &x, &y))
return NULL;
PyArg_ParseTupleAndKeywords
试图将Dual
施加到double
时崩溃。这是要研究的另一个问题,但是原始问题是通过在构造函数中实现相关转换来解决的。
也许这是一个愚蠢的观察,但是您尝试过:
import numpy as np
###code
np.mean(x) # instead of x.mean?
根据文档:
相关文章:
- 当回溯以零开始时,如何调试崩溃
- 内联映射初始化的动态atexit析构函数崩溃
- 执行函数时导致崩溃的变量
- 程序崩溃并显示"std::out_of_range"错误
- CoInitialize()在单独的线程上崩溃而不返回
- 使用调试/崩溃报告将应用程序部署到客户端
- 为什么所有C++编译器都会崩溃或挂起此代码
- 为什么lambda在clang上崩溃而不是在gcc上崩溃
- 为什么我的多线程作业队列崩溃
- ExtractIconEx:可以工作,但偶尔会崩溃
- 为什么引用传递会导致此崩溃(C++)
- 试图创建流或fopen时程序崩溃
- 类对象数组的问题会导致崩溃
- 排序时无法执行交换操作.我做的时候它会崩溃.为什么
- 为什么要增加导致崩溃的指针
- 在虚幻引擎中删除NXOpen对象时崩溃
- 为什么它只打印双链接列表的第一个值,而我的程序却崩溃了
- 应用程序崩溃并显示"symbol _ZdlPvm, version Qt_5 not defined in file libQt5Core.so.5 with link time reference"
- 自定义numpy类型导致numpy.mean中的崩溃
- 使用C 扩展Python/numpy,模块在初始化时崩溃