Cpp DenseNet OpenVino部署主体

Cpp OpenVino DenseNet部署主体

Python版本导引

• Python DenseNet OpenVino导出
• Python DenseNet OpenVino推理
• Python DenseNet OpenVino打包

CMake工程示例

测试性质的工程结构如下:

Project:

• Network: 存放网络推理相关
  • DebugTools: 打印网络结构
  • DenseGradeWrapper: 推理主结构
  • NetworkTools: 部分CV辅助函数
• Script: 存放前面博客中的导出以及验证Py文件
  • Cpp DenseNet OpenVino导出
  • Cpp DenseNet OpenVino测试
• CMakeLists.txt: cmake工程文件
• main.cpp: 简单的执行文件

DenseGradeWrapper

这里给出了我代码风格的封装，进行模型读取前注意中英文路径，我下载的OpenVINO不会报一些无法读取的错误，可能日志level设置错误，导致我困惑了一段时间

这里的推理使用了直接await, 官方写法中推荐使用回调函数的形式，提高程序的运行效率，可以自行阅读官方源码进行修改

#include "DenseGradeWrapper.h"

DenseGradeWrapper::DenseGradeWrapper(QString xmlPath, QObject *parent)
    : QObject{parent}
{
    qDebug() << ov::get_openvino_version().buildNumber << " " << ov::get_openvino_version().description;
    qDebug() << "Start To Compile CPU Model";
    ov::Core core;
    this->xmlModel = core.read_model(xmlPath.toStdString());
    printInputAndOutputsInfo(*this->xmlModel);
    double tS, tE;
    tS = (double) clock();
    this->compiledModel = core.compile_model(this->xmlModel, "CPU");
    tE = (double) clock();
    qDebug() << "Compile Done, cost " << (tE - tS) / (CLOCKS_PER_SEC) << " s";
}

DenseInferResult DenseGradeWrapper::infer(QString imageSrc){
    double tS, tE;

    tS = (double) clock();
    cv::Mat transDim = this->preprocess(imageSrc);
    tE = (double) clock();
    qDebug() << "Preprocess Done, cost " << (tE - tS) / (CLOCKS_PER_SEC) << " s";

    tS = (double) clock();
    QVector<float> buffer = this->_infer(transDim);
    tE = (double) clock();
    qDebug() << "Infer Done, cost " << (tE - tS) / (CLOCKS_PER_SEC) << " s";

    tS = (double) clock();
    this->softmax(buffer);
    DenseInferResult res = this->postProcess(buffer);
    tE = (double) clock();
    qDebug() << "Postprocess Done, cost " << (tE - tS) / (CLOCKS_PER_SEC) << " s";

    return res;
}

QVector<float> DenseGradeWrapper::_infer(cv::Mat transDim){
    ov::InferRequest inferRequest = this->compiledModel.create_infer_request();
    ov::Shape inputShape = {1, 3, 224, 224};
    ov::Tensor inputTensor = ov::Tensor(ov::element::f32, inputShape, transDim.data);
    inferRequest.set_input_tensor(inputTensor);

    // async
    inferRequest.start_async();
    inferRequest.wait();

    const ov::Tensor& outputTensor = inferRequest.get_output_tensor();
    const float *outputBuffer = outputTensor.data<const float>();
    return QVector<float>(outputBuffer, (outputBuffer + outputTensor.get_size()));
}

void DenseGradeWrapper::softmax(QVector<float> &buffer){
    float denominator = 0.0;

    for(size_t i = 0; i < buffer.size(); i++){
        buffer[i] = std::exp(buffer[i]);
        denominator += buffer[i];
    }

    for(size_t i = 0; i < buffer.size(); i++){
        buffer[i] /= denominator;
    }
}

DenseInferResult DenseGradeWrapper::postProcess(const QVector<float> &buffer){
    int index = buffer.indexOf(*std::max_element(buffer.begin(), buffer.end()));
    DenseInferResult result;
    result.index = index;
    std::copy(buffer.begin(), buffer.end(), result.pie);
    return result;
}

cv::Mat DenseGradeWrapper::preprocess(QString imageSrc){
    cv::Mat cropMat = this->circleCrop(imageSrc);
    std::vector<cv::Mat> batch_mat;
    batch_mat.push_back(cropMat);
    return cv::dnn::blobFromImages(batch_mat, 1.0 / 255.0, cv::Size(),
                                   cv::Scalar(), true);
}


cv::Mat DenseGradeWrapper::circleCrop(QString imageSrc){
    cv::Mat cropMask = NetworkTool::ToolCropImageFromMask(imageSrc);
    return NetworkTool::ToolCropWithGaussian(cropMask);
}

main

这里给出简单的调用方式，至此OpenVINO的Cpp集成方式已经完成，后面的博客将会介绍如何集成Tensorrt版本的代码

// intel official instruction:
// https://docs.openvino.ai/2024/openvino-workflow/running-inference/integrate-openvino-with-your-application.html

#include <iostream>
#include <QCoreApplication>
#include "Network/DenseGradeWrapper.h"

int main(){
    QString model_path = "../openvino-cpu-win-msvc2019/Script/export_dense121_cpu.xml";
    QString image_path = "../openvino-cpu-win-msvc2019/Script/1.jpg";

    DenseGradeWrapper denseWrapper(model_path);
    DenseInferResult res = denseWrapper.infer(image_path);

    qDebug() << "Level is: " << res.index;
    qDebug() << "Each possible is: ";
    for(int i = 0; i < 5; i++){
        qDebug() << "\t" << res.pie[i];
    }
}