在Windows下安裝tensorflow-gpu,可參考windows+tensorflow-gpu+anaconda3+cuda8.0+cudnn安裝指南https://blog.csdn.net/hdd0411/article/details/71305931?locationNum=8&fps=1。

2.1具體呼叫指南：

新建vs2013工程，複製anaconda3資料夾下的libs和include檔案拷貝到.sln目錄下，並新增到工程屬性包含目錄和庫目錄下；複製libs資料夾下的python36.lib，命名為python36_d.lib，新增到屬性-》聯結器中；複製anaconda3資料夾下的python36.dll檔案放在debug檔案下，使.exe,python36.dll在同一個debug資料夾中。

C++程式碼：(呼叫表情識別模型)

DWORD WINAPI testImage(LPVOID lParam)

{

char msg[256] = "11111 ";

PyObject* pFunc = NULL;

PyObject* pArg = NULL;

PyObject* module = NULL;

Py_Initialize();

module = PyImport_ImportModule("predictd");//myModel:Python檔名  

if (!module) {

printf("cannot open module!");

//Py_Finalize();  

}

pFunc = PyObject_GetAttrString(module, "test_one_image");//test_one_image:Python檔案中的函式名  

if (!pFunc) {

printf("cannot open FUNC!");

//Py_Finalize();  

}

//開始呼叫model  

pArg = Py_BuildValue("(s)", "00028.jpg");

if (module != NULL) {

PyGILState_STATE gstate;

gstate = PyGILState_Ensure();

PyEval_CallObject(pFunc, pArg);

PyGILState_Release(gstate);

}

return 0;

}

int main()

{

cout << "建立執行緒" << endl;

CreateThread(NULL, 0, testImage, 0, 0, NULL);

cout << "建立成功" << endl;

system("pause");

return 0;

}

Python程式碼：建立predictd.py檔案，將predictd.py檔案和python36.dll檔案放在同一個資料夾中。

import tensorflow as tf

import numpy as np

import os,glob,cv2

import sys,argparse

#os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'

# First, pass the path of the image

def test_one_image(test_dir):

    stdout_backup = sys.stdout

    log_file = open("log.txt", "w")

    sys.stdout = log_file

    print(test_dir)

    image_size=96#跟face-expression-model.meta中的大小一致。

    num_channels=3

    images = []

# Reading the image using OpenCV

    image = cv2.imread(test_dir)

    print(image)

# Resizing the image to our desired size and preprocessing will be done exactly as done during training

    image = cv2.resize(image, (image_size, image_size), 0, 0, cv2.INTER_LINEAR)

    images.append(image)

    images = np.array(images, dtype=np.uint8)

    images = images.astype('float32')

    images = np.multiply(images, 1.0 / 255.0)

    # The input to the network is of shape [None image_size image_size num_channels]. Hence we reshape.

    x_batch = images.reshape(1, image_size, image_size, num_channels)

    ## Let us restore the saved model

    sess = tf.Session()

    # Step-1: Recreate the network graph. At this step only graph is created.

    saver = tf.train.import_meta_graph('face-expression-model.meta')

    # Step-2: Now let's load the weights saved using the restore method.

    saver.restore(sess, tf.train.latest_checkpoint('./'))

    # Accessing the default graph which we have restored

    graph = tf.get_default_graph()

    # Now, let's get hold of the op that we can be processed to get the output.

    # In the original network y_pred is the tensor that is the prediction of the network

    y_pred = graph.get_tensor_by_name('y_pred:0')

    ## Let's feed the images to the input placeholders

    x = graph.get_tensor_by_name('x:0')

    y_true = graph.get_tensor_by_name("y_true:0")

    y_test_images = np.zeros((1, len(os.listdir('training_data'))))

    ### Creating the feed_dict that is required to be fed to calculate y_pred

    feed_dict_testing = {x: x_batch, y_true: y_test_images}

    result = sess.run(y_pred, feed_dict=feed_dict_testing)

    # result is of this format [probabiliy_of_rose probability_of_sunflower]

    print(result)

    #print(feed_dict_testing)

    log_file.close()

    sys.stdout = stdout_backup

sess.close()

將result儲存到log.txt檔案中；

2、2讀取log.txt中的結果值實現程式碼：

#include "stdafx.h"

#include "string.h"

#include "stdlib.h"

#include <stdio.h>

int main(){

FILE *fp;

fp = fopen("data.txt","r");

float x;

char a[100] = "0";

char str1[100] = "0";

char ch;

int i,j,len;

double b[100] = {0};

for (i =j= 0; (ch = fgetc(fp)) != EOF;  i++)

{

if (ch=='['||ch==']')

{

continue;

}

a[j] = ch;

j++;

}

printf("%s\n", a);

len = strlen(a);

int n = 0;

for ( i=j = 0; i <= len; i++)

{

if (a[i] != ' ')

{

str1[j++] = a[i];

}

else

{

str1[j] = 0;

if (j>0)

{

b[n++] = atof(str1);

}

j = 0;

}

}

if (j>0)

{

str1[j] = 0;

b[n++] = atof(str1);

}

for ( i = 0; i <n; i++)

{

printf("%lf\n", b[i]);

}

return 0;

}