C++ openvino 模型部署efficientnetv2

#include <time.h>
#include <openvino/openvino.hpp>
#include <opencv2/opencv.hpp>
using namespace cv;
using namespace std;
/*
* 源码训练的时候中心裁剪图片至300*300
*/
std::string path = "acc-76.onnx";//模型文件路径
ov::CompiledModel model;
cv::Mat input_img;

void center_resize(cv::Mat &img, cv::Size new_shape)//中心裁剪方法1
{
    float width = img.cols;
    float height = img.rows;
    int center_x = round(width / 2);
    int center_y = round(height / 2);
    int dh = int(round((new_shape.height - height) / 2));
    int dw = int(round((new_shape.width - width) / 2));
    cv::Scalar color = cv::Scalar(0, 0, 0);//填充色
    if (width >= 300 && height >= 300)
    {
        img = img(cv::Range(center_y - (round(new_shape.height / 2)), center_y + (round(new_shape.height / 2))),
            cv::Range(center_x - (round(new_shape.width / 2)), center_x + (round(new_shape.width / 2))));
    }
    else if (width >= 300)
    {
        cv::copyMakeBorder(img, img, dh, dh, 0, 0, cv::BORDER_CONSTANT, color);
        img = img(cv::Range(0, 300), cv::Range(center_x - (round(new_shape.width / 2)), center_x + (round(new_shape.width / 2))));
    }
    else if (height >= 300)
    {
        cv::copyMakeBorder(img, img, 0, 0, dw, dw, cv::BORDER_CONSTANT, color);
        img = img(cv::Range(center_y - (round(new_shape.height / 2)), center_y + (round(new_shape.height / 2))), cv::Range(0, 300));
    }
    else
    {
        cv::copyMakeBorder(img, img, dh, dh, dw, dw, cv::BORDER_CONSTANT, color);
        img = img(cv::Range(0, 300), cv::Range(0, 300));
    }
     
}

void center_resize(cv::Mat &src)//中心裁剪方法2
{
    int width = src.cols;
    int height = src.rows;
    Point center(width / 2, height / 2); // 指定的中心点 
    Size size(300, 300); // 裁剪出的图片大小 
    getRectSubPix(src, size, center, src); // 裁剪出目标图片
}


void compile_model(std::string xml_path, ov::CompiledModel& compile_model)//加载模型
{
    ov::Core core;                 // core是interference Engine的管理核心，一般使用openvino之前都需要先创建一个core，负责其中的设备管理
    compile_model = core.compile_model(xml_path); // 模型加载并编译
}

void openvino_classifer(cv::Mat image, ov::CompiledModel compile_model)//模型推理
{
    clock_t startTime, endTime;
    /*startTime = clock();*/
    Mat out_image;
    size_t input_batch_size = 1;    // size_t是c和c++标准在stddef.h中定义的，这个类型用于表示对象的大小，其真实类型与操作系统有关
                                    // 在32位架构中被普遍定义为：typedef   unsigned int size_t;
    size_t input_channel_size = 3;  // 而在64位架构中被定义为： typedef  unsigned long size_t;
    size_t input_height_size = 300;
    size_t input_width_size = 300;
    auto iq = compile_model.create_infer_request();  // 创建用于推断已编译模型的推理请求对象  创建的请求分配了输入和输出张量
    auto input = iq.get_input_tensor(0);
    auto output = iq.get_output_tensor(0);
    input.set_shape({ input_batch_size, input_channel_size, input_height_size, input_height_size }); // 指定shape的大小
    float* input_data_host = input.data<float>();  // 获取输入的地址，并传递给指针input_data_host

    
    startTime = clock();
    //auto image = cv::imread("dog.jpg");
    float mean[] = { 0.406, 0.456, 0.485 };
    float std[] = { 0.225, 0.224, 0.229 };

    // 对应于pytorch的代码部分
    //cv::resize(image, image, cv::Size(input_height_size, input_width_size));
    center_resize(image);
    int image_area = image.cols * image.rows;
    unsigned char* pimage = image.data;
    float* phost_b = input_data_host + image_area * 0;   // input_data_host和phost_*进行地址关联
    float* phost_g = input_data_host + image_area * 1;
    float* phost_r = input_data_host + image_area * 2;
    for (int i = 0; i < image_area; ++i, pimage += 3) 
    {
        // 注意这里的顺序rgb调换了
        *phost_r++ = (pimage[0] / 255.0f - mean[0]) / std[0];  // 将图片中的像素点进行减去均值除方差，并赋值给input
        *phost_g++ = (pimage[1] / 255.0f - mean[1]) / std[1];
        *phost_b++ = (pimage[2] / 255.0f - mean[2]) / std[2];
    }
    iq.infer();  // 模型推理
    const int num_classes = 4;
    float* prob = output.data<float>();
    int predict_label = std::max_element(prob, prob + num_classes) - prob;  // 确定预测类别的下标

    float confidence = prob[predict_label];    // 获得预测值的置信度
    printf("confidence = %f, label = %d\n", confidence, predict_label);
    endTime = clock();//计时结束
    std::cout << "total推理时间: " << (double)(endTime - startTime) / CLOCKS_PER_SEC << "s" << std::endl;
}



int main()
{
    compile_model(path, model);
    cv::String pattern = "D:\\vs_project\\effv2_openvino\\effv2_openvino\\wcj\\*.png";//测试图片路径
    vector<string>img_path;
    vector<cv::String>fn;
    glob(pattern, fn, false);
    size_t count = fn.size();
    for (size_t j = 0; j < count; j++)
    {
        img_path.push_back(fn[j]);
    }

    for (int i = 0; i < img_path.size(); i++)
    {
        string a = to_string(i + 1);// 1开始命名存储
        input_img = cv::imread(img_path[i]);
        //start = clock();//开始计时
        openvino_classifer(input_img, model);

    }

    return 0;
}

推理测试结果：

测试了8张图片，结果正确