OpenVino --- C++ 推理流程

 

• openvino 环境配置相关

• C++案例流程

---

openvino环境匹配

1. openvino 2021.3.394
2. vs2019
3. cmake3.20.2
4. python 3.6-3.9 (用于转换其他框架模型到IR, 也支持python推理)
​

系统环境说明

安装好openvino后主要配置好系统环境变量IE_DIR、ngraph_DIR、TBB_DIR、OPENCV_DIR（openvino自带）以及相关lib库-bin目录；

 

python环境说明

单独创建一个独立的python虚拟环境，直接将openvino自带的一些环境复制到site-packages中，不用按照网上所述的需要pip安装一些openvino包，也不需要添加"PYTHONPATH"系统环境变量。

不用担心相关文件找不到，python 导包的搜索顺序：

1. 当前目录
2. 环境变量PYTHONPATH中的目录
3. Python安装目录第三方库（for Linux OS：/usr/local/lib/python）（win: site-packages下）

---

C++推理流程

将Inter官方提供的person-vehicle-bike-detection-2002为例：

这里将使用官方提供的FP32模型，上图是模型需要的输入以及输出结果。

#include <inference_engine.hpp>
#include <opencv2/opencv.hpp>
#include <opencv2/dnn.hpp>
#include <iostream>
#include "func.h"
​
using namespace std;
using namespace cv;
using namespace cv::dnn;
using namespace InferenceEngine;
​
void test_demo() {
​
    // crerate IE engine, find supported device name
    Core ie;
    vector<string> availableDevices = ie.GetAvailableDevices();
    for (int i = 0; i < availableDevices.size(); i++) {
        printf("supported device name : %s \n", availableDevices[i].c_str());
    }
​
    //  load model
    auto network = ie.ReadNetwork("D:/download/intel/person-vehicle-bike-detection-2002/FP32/person-vehicle-        bike-detection-2002.xml", "D:/download/intel/person-vehicle-bike-detection-2002/FP32/person-vehicle-        bike-detection-2002.bin");
​
    // request network input and output info
    InferenceEngine::InputsDataMap input_info(network.getInputsInfo());
    InferenceEngine::OutputsDataMap output_info(network.getOutputsInfo());
​
    // set input format
    for (auto& item : input_info) {
        auto input_data = item.second;
        input_data->setPrecision(Precision::FP32);
        input_data->setLayout(Layout::NCHW);
        input_data->getPreProcess().setResizeAlgorithm(RESIZE_BILINEAR);
        input_data->getPreProcess().setColorFormat(ColorFormat::RGB);
    }
​
    // set output format
    for (auto& item : output_info) {
        auto output_data = item.second;
        output_data->setPrecision(Precision::FP32);
    }
    auto executable_network = ie.LoadNetwork(network, "GPU");
​
    // output result 
    vector<Rect> boxes;
    vector<int> classIds;
    vector<float> confidences;
​
    // request infer
    auto infer_request = executable_network.CreateInferRequest();
​
    Mat src = imread("F:/Downloads/opencv_tutorial_data/images/objects.jpg");
    int image_height = src.rows;
    int image_width = src.cols;
​
    /** Iterating over all input blobs **/
    for (auto& item : input_info) {
        auto input_name = item.first;
​
        /** Getting input blob **/
        auto input = infer_request.GetBlob(input_name);
        size_t num_channels = input->getTensorDesc().getDims()[1];
        size_t h = input->getTensorDesc().getDims()[2];
        size_t w = input->getTensorDesc().getDims()[3];
        size_t image_size = h * w;
        Mat blob_image;
        resize(src, blob_image, Size(w, h));
        cvtColor(blob_image, blob_image, COLOR_BGR2RGB);
​
        // NCHW
        float* data = static_cast<float*>(input->buffer());
        for (size_t row = 0; row < h; row++) {
            for (size_t col = 0; col < w; col++) {
                for (size_t ch = 0; ch < num_channels; ch++) {
                    data[image_size * ch + row * w + col] = float(blob_image.at<Vec3b>(row, col)[ch]);
                }
            }
        }
    }
​
    int64 start = getTickCount();
​
    // do inference
    infer_request.Infer();
​
    for (auto& item : output_info) {
        auto output_name = item.first;
        printf("output_name : %s \n", output_name.c_str());
        // get output blob
        auto output = infer_request.GetBlob(output_name);
​
        const float* output_blob = static_cast<PrecisionTrait<Precision::FP32>::value_type*>(output-                                            >buffer());
        const SizeVector outputDims = output->getTensorDesc().getDims();
​
        const int out_num = outputDims[2];   // 200
        const int out_info = outputDims[3];  // 7
​
        for (int n = 0; n < out_num; n++) {
            float conf = output_blob[n * 7 + 2];
            if (conf < 0.5) {
                continue;
            }
            int x1 = saturate_cast<int>(output_blob[n * 7 + 3] * image_width);
            int y1 = saturate_cast<int>(output_blob[n * 7 + 4] * image_height);
            int x2 = saturate_cast<int>(output_blob[n * 7 + 5] * image_width);
            int y2 = saturate_cast<int>(output_blob[n * 7 + 6] * image_height);
​
            //label
            int label = saturate_cast<int>(output_blob[n * 7 + 1]);
​
            classIds.push_back(label);
            confidences.push_back(conf);
            boxes.push_back(Rect(x1, y1, x2 - x1, y2 - y1));
        }
    }
​
    vector<int> indices;
    NMSBoxes(boxes, confidences, 0.25, 0.5, indices);
    float time = (getTickCount() - start) / getTickFrequency();
​
    printf("time used：%f", time);
    for (size_t i = 0; i < indices.size(); ++i)
    {
        int idx = indices[i];
        Rect box = boxes[idx];
        rectangle(src, box, Scalar(140, 199, 0), 4, 8, 0);
    }
​
    imshow("OpenVINO-test", src);
    waitKey(0);
}

最终结果如上图所示（效果一般），推理+后处理一般在30多ms, 在Intel(R) Core(TM) i7-9750H CPU @ 2.60GHz 2.59 GHz上