[C++][onnxruntime]攻克onnxruntime之模型热启动warmup

当我们使用onnxruntime加载模型后，需要热启动。热启动过程就是随机输入一段数据然后再模型推理一遍，一般用做激活神经元。下面以yolov8s.onnx为例我们加载模型热启动一下。热启动过程其实就是模拟模拟模型推理全过程，有助于我们简单了解推理过程是怎么样的。

测试环境：

vs2019

onnxruntime==1.12.0

代码部分：

#include <iostream>
#include<onnxruntime_cxx_api.h>
#include <numeric>
using namespace std;
using namespace Ort;

template <typename T>
T VectorProduct(const std::vector<T>& v)
{
    return std::accumulate(v.begin(), v.end(), 1, std::multiplies<T>());
};


int main()
{

    const wchar_t* model_path = L"D:\\yolov8s.onnx";//模型路径
    Ort::Env env;//创建env
    Ort::Session session(nullptr);//创建一个空会话
    Ort::SessionOptions sessionOptions{ nullptr };//创建会话配置
    session = Ort::Session(env, model_path, sessionOptions);

    //获取输入节点数量，名称和shape

    size_t inputNodeCount= session.GetInputCount();
    std::cout << "输入节点数量:" << inputNodeCount << "\n";

    Ort::AllocatorWithDefaultOptions allocator;

    std::shared_ptr<char> inputName = std::move(session.GetInputNameAllocated(0, allocator));
    std::vector<char*> inputNodeNames;
    inputNodeNames.push_back(inputName.get());
    std::cout << "输入节点名称:" << inputName << "\n";


    Ort::TypeInfo inputTypeInfo = session.GetInputTypeInfo(0);
    auto input_tensor_info = inputTypeInfo.GetTensorTypeAndShapeInfo();
    ONNXTensorElementDataType inputNodeDataType = input_tensor_info.GetElementType();
    std::vector<int64_t> inputTensorShape = input_tensor_info.GetShape();
    std::cout << "输入节点shape:";
    for (int i = 0; i<inputTensorShape.size(); i++)
    {
        std::cout << inputTensorShape[i]<<" ";
    }
    std::cout << "\n";


    //获取输出节点数量、名称和shape
    size_t outputNodeCount = session.GetInputCount();
    std::cout << "输出节点数量:" << outputNodeCount << "\n";

    std::shared_ptr<char> outputName = std::move(session.GetOutputNameAllocated(0, allocator));
    std::vector<char*> outputNodeNames;
    outputNodeNames.push_back(outputName.get());
    std::cout << "输出节点名称:" << outputName << "\n";


    Ort::TypeInfo type_info_output0(nullptr);
    type_info_output0 = session.GetOutputTypeInfo(0);  //output0

    auto tensor_info_output0 = type_info_output0.GetTensorTypeAndShapeInfo();
    ONNXTensorElementDataType outputNodeDataType = tensor_info_output0.GetElementType();
    std::vector<int64_t> outputTensorShape = tensor_info_output0.GetShape();
    std::cout << "输出节点shape:";
    for (int i = 0; i<outputTensorShape.size(); i++)
    {
        std::cout << outputTensorShape[i]<<" ";
    }
    std::cout << "\n";


    size_t input_tensor_length = VectorProduct(inputTensorShape);
    std::cout << "输入tensor的长度:" << input_tensor_length << "\n";





    Ort::MemoryInfo OrtMemoryInfo= Ort::MemoryInfo::CreateCpu(OrtAllocatorType::OrtDeviceAllocator, OrtMemType::OrtMemTypeCPUOutput);
    std::cout << "Start warming up" << endl;
    float* temp = new float[input_tensor_length];
    std::vector<Ort::Value> input_tensors;
    std::vector<Ort::Value> output_tensors;
    input_tensors.push_back(Ort::Value::CreateTensor<float>(
        OrtMemoryInfo, temp, input_tensor_length, inputTensorShape.data(),
        inputTensorShape.size()));
    for (int i = 0; i < 3; ++i) {
        output_tensors = session.Run(Ort::RunOptions{ nullptr },
            inputNodeNames.data(),
            input_tensors.data(),
            inputNodeNames.size(),
            outputNodeNames.data(),
            outputNodeNames.size());
    }
    delete[]temp;
    std::cout << "warming up over" << endl;
	getchar();
}

输出结果：