VTM中直接调用python(pytorch模型)

参考：

​​​​​​C++调用python文件（包含第三方库） - 知乎 (zhihu.com)

奇巧：C++ 调用python方法 - 知乎 (zhihu.com)

(31条消息) C调用Python（传递数字、字符串、list数组（一维、二维），结构体）_sunshine_9990的博客-CSDN博客

一、环境配置：

一、VC++目录：

1、包含目录：python所在文件夹下include文件夹“C:\Program Files (x86)\Microsoft Visual Studio\Shared\Python37_64\include”

2、库目录：python所在文件夹下libs文件夹“C:\Program Files (x86)\Microsoft Visual Studio\Shared\Python37_64\libs”

 二、链接器输入

附加依赖项：python37.lib

三、release文件夹（同EncoderApp.exe文件夹）

复制python文件夹下DLLs、libs、Lib文件夹、所调用的python文件、python37.lib到该文件夹下

测试所用的.bat、.cfg也放在该文件夹下

二、C++代码，传递一维list[[list0],[list1],[list2]],返回值一维list

int main(int argc, char* argv[])
{//addPy_Initialize();//end// print informationfprintf( stdout, "\n" );fprintf( stdout, "VVCSoftware: VTM Encoder Version %s ", VTM_VERSION );fprintf( stdout, NVM_ONOS );fprintf( stdout, NVM_COMPILEDBY );fprintf( stdout, NVM_BITS );............return 0;//addPy_Finalize();//end
}

void EncCu::python()
{//初始化使用的变量PyObject* pModule = NULL;PyObject* pFunc = NULL;PyObject* pName = NULL;PyObject* pReturn = NULL;int wh = width * height;//2、初始化python系统文件路径，保证可以访问到 .py文件PyRun_SimpleString("import sys");PyRun_SimpleString("sys.path.append('./')");//3、导入所调用的python文件pModule = PyImport_ImportModule("GEAPI");//4、调用python文件函数pFunc = PyObject_GetAttrString(pModule, "out");//5、定义三个Pylist列表PyObject* PyListP0 = PyList_New(wh);PyObject* PyListP1 = PyList_New(wh);PyObject* PyListP2 = PyList_New(wh);int num = 0;for (unsigned y = 0; y < height; y++){for (unsigned x = 0; x < width; x++){Pel pixel = (pcYuvSrc0.bufs[0].at(x, y) + IF_INTERNAL_OFFS) >> IF_INTERNAL_FRAC_BITS(ClipBD);float pixelL0 = float(round(float(pixel) / 4.0)); //向列表中添加元素PyList_SetItem(PyListP0, num, PyFloat_FromDouble(pixelL0));num++;}//cout << endl;}//PylistP1、PylistP2同理//...//6、三个list添加到一个新list中PyObject* PyList = PyList_New(0);PyList_Append(PyList, PyListP0);PyList_Append(PyList,PyListP1);PyList_Append(PyList, PyListP2);//7、定义一个Tuple对象，Tuple对象的长度与Python函数参数个数一致PyObject* ArgList = PyTuple_New(1);//8、传递list到python函数中PyTuple_SetItem(ArgList, 0, PyList);//9、接收返回值listpReturn = PyObject_CallObject(pFunc, ArgList);for (unsigned y = 0; y < height; y++){for (unsigned x = 0; x < width; x++){PyObject* Item = PyList_GetItem(pReturn, num);//获取List对象中的每一个元素num++;float result;PyArg_Parse(Item, "f", &result);//f表示转换成float型变量Pel pixel = round(result);pcYuvdst.bufs[0].at(x, y) = Pel(pixel * 4);Py_DECREF(Item);}//cout << endl;}
}

三、c++调用python完整代码，神经网络pytorch模型

import sys
import torch
import torch.nn as nn
import torch.nn.functional as F
import os
import numpy as np
from pure_network import DetailEnhance
def out(inputorg):#inputorg = [[list0],[list1],[list2]]inputP0 = inputorg[0]inputP1 = inputorg[1]inputP2 = inputorg[2]#模型加载model = DetailEnhance()model = torch.load("",map_location='cpu')model.eval()#输入转tensor，转数组P0 = torch.as_tensor(inputP0,dtype=torch.float32)P1 = torch.as_tensor(inputP1,dtype=torch.float32)P2 = torch.as_tensor(inputP2,dtype=torch.float32)refP0 = np.array(P0).reshape(height,width)refP1 = np.array(P1).reshape(height, width)refP2 = np.array(P2).reshape(height, width)input0 = np.array([[refP0], [refP1], [refP2]])#增减维度，确保与模型输入维度一致input = (torch.from_numpy(input0).unsqueeze(0) / 255)output = model(input)*255#输出转list，返回一维list[]out = output.detach().numpy().reshape(1,width*height).tolist()dst = out[0]return dst

踩坑记录：

1、c++中调用python

//PyObject* Item 写在循环外，否则无法调用大块
PyObject* Item = NULL;  
for (unsigned y = 0; y < height; y++){for (unsigned x = 0; x < width; x++){Item = PyList_GetItem(pReturn, num);//获取List对象中的每一个元素num++;float result;PyArg_Parse(Item, "f", &result);//f表示转换成float型变量Pel pixel = round(result);pcYuvdst.bufs[0].at(x, y) = Pel(pixel * 4);}//cout << endl;}
Py_DECREF(Item);

2、换设备测试，复制release文件，其中Lib文件夹使用新设备环境下Lib文件夹