一、简介

OpenVINO™是英特尔推出的一个用于优化和部署AI推理的开源工具包。常用于 Inter 的集成显卡网络推理使用。

官网地址:https://docs.openvino.ai

二、下载

下载地址:https://docs.openvino.ai/latest/openvino_docs_install_guides_installing_openvino_linux.html
在这里插入图片描述
针对不同平台,在如图红框处选择不同的文档参考,按照官网文档一步步执行就行。

三、使用

注意:笔者当前使用的版本为 openvino_2021。

假设你已经有了模型的 xml 文件和对应的 bin 文件了,基本代码流程如下:

#include <stdio.h>
#include <string>
#include "inference_engine.hpp"

#define LOGD(fmt, ...) printf("[%s][%s][%d]: " fmt "\n", __FILE__, __FUNCTION__, __LINE__, ##__VA_ARGS__)

using namespace InferenceEngine;

int main(int argc, char *argv[]) {
    // 1.查看版本号信息
    const Version* version = GetInferenceEngineVersion();
    LOGD("version description: %s, buildNumber: %s, major.minor: %d.%d",
        version->description, version->buildNumber, version->apiVersion.major, version->apiVersion.minor);
    
    // 2.创建推理引擎
    Core ie;
    std::vector<std::string> devices = ie.GetAvailableDevices(); // 查看可使用的Devices,包含 CPU、GPU 等
    for (std::string device : devices) {
        LOGD("GetAvailableDevices: %s", device.c_str());
    }

    // 3.读取模型文件
    const std::string input_model_xml = "model.xml";
    CNNNetwork network = ie.ReadNetwork(input_model_xml);

    // 4.配置输入输出信息
    InputsDataMap& inputs = network.getInputsInfo();
    for (auto& input : inputs) {
        auto& input_name = input.first; //input是一个键值对类型
        InputInfo::Ptr& input_info = input.second;
        input_info->setLayout(Layout::NCHW); // 设置排列方式
        input_info->setPrecision(Precision::FP32); // 设置精度为float32
        input_info->getPreProcess().setResizeAlgorithm(ResizeAlgorithm::RESIZE_BILINEAR);
        input_info->getPreProcess().setColorFormat(ColorFormat::RAW); // 设置图片格式
    }
    OutputsDataMap& outputs = network.getOutputsInfo();
    for (auto& output : outputs) {
        auto& output_name = output.first; //output也是一个键值对类型
        DataPtr& output_info = output.second;
        output_info->setPrecision(Precision::FP32);
        auto& dims = output_info->getDims();
        LOGD("output shape name: %s, dims: [%d, %d, %d, %d]", output_name.c_str(), dims[0], dims[1], dims[2], dims[3]);
    }

    // 5.根据设备(CPU、GPU 等)加载网络
    std::string device_name = "CPU"; // 可用的device通过ie.GetAvailableDevices查询
    ExecutableNetwork executable_network = ie.LoadNetwork(network, device_name);

    // 6.创建推理请求
    InferRequest infer_request = executable_network.CreateInferRequest();

    /* 如上6步,在多次执行网络推理过程中,可以缓存起来只创建一次,节约耗时*/

    // 7.设置输入数据
    InputsDataMap& inputs = network.getInputsInfo();
    for (auto& input : inputs) {
        auto& input_name = input.first; //input是一个键值对类型
        Blob::Ptr blob = infer_request.GetBlob(name);
        unsigned char* data = static_cast<unsigned char*>(blob->buffer());
        // TODO: 通过memcpy等方式给data赋值
        // readFile(input_path, data);
    }

    // 8.网络推理
    infer_request.Infer();

    // 9.获取输出
    OutputsDataMap& outputs = network.getOutputsInfo();
    for (auto& output : outputs) {
        auto& output_name = output.first; //output也是一个键值对类型
        const Blob::Ptr output_blob = infer_request.GetBlob(name);
        LOGD("size: %d, byte_size: %d", output_blob->size(), output_blob->byteSize());
        const float* output_data = static_cast<PrecisionTrait<Precision::FP32>::value_type*>(output_blob->buffer());
        // writeFile(path, (void *)output_data, output_blob->byteSize());
    }
}

其余更复杂的使用场景,可以参考下载的SDK中的示例,路径是 .\openvino_2021\inference_engine\samples\cpp。

四、ReadNetwork说明

1.通过文件路径读取模型

通常我们的模型文件就是本地的一个文件,通过路径加载即可,对应的接口为:

/**
 * @brief Reads models from IR and ONNX formats
 * @param modelPath path to model
 * @param binPath path to data file
 * For IR format (*.bin):
 *  * if path is empty, will try to read bin file with the same name as xml and
 *  * if bin file with the same name was not found, will load IR without weights.
 * For ONNX format (*.onnx or *.prototxt):
 *  * binPath parameter is not used.
 * @return CNNNetwork
 */
CNNNetwork ReadNetwork(const std::string& modelPath, const std::string& binPath = {}) const;

如果bin文件路径和xml文件路径一致且文件名相同,该参数可以省略,如:CNNNetwork network = ie.ReadNetwork("model.xml")

2.通过内存地址读取模型

假设我们的模型已经在内存中了,可以通过如下接口创建:

/**
 * @brief Reads models from IR and ONNX formats
 * @param model string with model in IR or ONNX format
 * @param weights shared pointer to constant blob with weights
 * Reading ONNX models doesn't support loading weights from data blobs.
 * If you are using an ONNX model with external data files, please use the
 * `InferenceEngine::Core::ReadNetwork(const std::string& model, const Blob::CPtr& weights) const`
 * function overload which takes a filesystem path to the model.
 * For ONNX case the second parameter should contain empty blob.
 * @note Created InferenceEngine::CNNNetwork object shares the weights with `weights` object.
 * So, do not create `weights` on temporary data which can be later freed, since the network
 * constant datas become to point to invalid memory.
 * @return CNNNetwork
 */
CNNNetwork ReadNetwork(const std::string& model, const Blob::CPtr& weights) const;

使用示例:

extern unsigned char __res_model_xml [];
extern unsigned int __res_model_xml_size;
extern unsigned char __res_model_bin [];
extern unsigned int __res_model_bin_size;

std::string model(__res_model_xml, __res_model_xml + __res_model_xml_size);
CNNNetwork network = ie.ReadNetwork(model,
    InferenceEngine::make_shared_blob<uint8_t>({InferenceEngine::Precision::U8,
        {__res_model_bin_size}, InferenceEngine::C}, __res_model_bin));
GitHub 加速计划 / li / linux-dash
10.39 K
1.2 K
下载
A beautiful web dashboard for Linux
最近提交(Master分支:2 个月前 )
186a802e added ecosystem file for PM2 4 年前
5def40a3 Add host customization support for the NodeJS version 4 年前
Logo

旨在为数千万中国开发者提供一个无缝且高效的云端环境,以支持学习、使用和贡献开源项目。

更多推荐