【模型部署 01】C++实现GoogLeNet在OpenCV DNN、ONNXRuntime、TensorRT、OpenVINO上的推理部署

2023-05-26 23:13 由最菜程序员Sxx 发表于 #其他

深度学习领域常用的基于CPU/GPU的推理方式有OpenCV DNN、ONNXRuntime、TensorRT以及OpenVINO。这几种方式的推理过程可以统一用下图来概述。整体可分为模型初始化部分和推理部分，后者包括步骤2-5。

以GoogLeNet模型为例，测得几种推理方式在推理部分的耗时如下：

结论：

GPU加速首选TensorRT；
CPU加速，单图推理首选OpenVINO，多图并行推理可选择ONNXRuntime;
如果需要兼具CPU和GPU推理功能，可选择ONNXRuntime。

下一篇内容：【模型部署 02】Python实现GoogLeNet在OpenCV DNN、ONNXRuntime、TensorRT、OpenVINO上的推理部署

1. 环境配置

2. PyTorch模型文件（pt/pth/pkl）转ONNX

2.1 pt/pth/pkl互转

PyTorch中支持导出三种后缀格式的模型文件：pt、pth和pkl，这三种格式在存储方式上并无区别，只是后缀不同。三种格式之间的转换比较简单，只需要创建模型并加载模型参数，然后再保存为其他格式即可。

以pth转pt为例：

import torch
import torchvision

# 构建模型
model = torchvision.models.googlenet(num_classes=2, init_weights=True)
# 加载模型参数，pt/pth/pkl三种格式均可
model.load_state_dict(torch.load("googlenet_catdog.pth"))
model.eval()
# 重新保存为所需要转换的格式
torch.save(model.state_dict(), 'googlenet_catdog.pt')

2.2 pt/pth/pkl转ONNX

PyTorch中提供了现成的函数torch.onnx.export()，可将模型文件转换成onnx格式。该函数原型如下：

export(model, args, f, export_params=True, verbose=False, training=TrainingMode.EVAL,
           input_names=None, output_names=None, operator_export_type=None,
           opset_version=None, do_constant_folding=True, dynamic_axes=None,
           keep_initializers_as_inputs=None, custom_opsets=None,
           export_modules_as_functions=False)

主要参数含义：

model (torch.nn.Module, torch.jit.ScriptModule or torch.jit.ScriptFunction) ：需要转换的模型。
args (tuple or torch.Tensor) ：args可以被设置为三种形式：
- 一个tuple，这个tuple应该与模型的输入相对应，任何非Tensor的输入都会被硬编码入onnx模型，所有Tensor类型的参数会被当做onnx模型的输入。
```
args = (x, y, z)
```
- 一个Tensor，一般这种情况下模型只有一个输入。
```
args = torch.Tensor([1, 2, 3])
```
- 一个带有字典的tuple，这种情况下，所有字典之前的参数会被当做“非关键字”参数传入网络，字典中的键值对会被当做关键字参数传入网络。如果网络中的关键字参数未出现在此字典中，将会使用默认值，如果没有设定默认值，则会被指定为None。
```
args = (x,
        {'y': input_y,
         'z': input_z})
```
  NOTE：一个特殊情况，当网络本身最后一个参数为字典时，直接在tuple最后写一个字典则会被误认为关键字传参。所以，可以通过在tuple最后添加一个空字典来解决。
```
# 错误写法：
torch.onnx.export(
    model,
    (x,
     # WRONG: will be interpreted as named arguments
     {y: z}),
    "test.onnx.pb")
 
# 纠正
torch.onnx.export(
    model,
    (x,
     {y: z},
     {}),
    "test.onnx.pb") 
```
f：一个文件类对象或一个路径字符串，二进制的protocol buffer将被写入此文件，即onnx文件。
export_params (bool, default False) ：如果为True则导出模型的参数。如果想导出一个未训练的模型，则设为False。
verbose (bool, default False) ：如果为True，则打印一些转换日志，并且onnx模型中会包含doc_string信息。
training (enum, default TrainingMode.EVAL) ：枚举类型包括：
- TrainingMode.EVAL - 以推理模式导出模型。
- TrainingMode.PRESERVE - 如果model.training为False，则以推理模式导出；否则以训练模式导出。
- TrainingMode.TRAINING - 以训练模式导出，此模式将禁止一些影响训练的优化操作。
input_names (list of str, default empty list) ：按顺序分配给onnx图的输入节点的名称列表。
output_names (list of str, default empty list) ：按顺序分配给onnx图的输出节点的名称列表。
operator_export_type (enum, default None) ：默认为OperatorExportTypes.ONNX，如果Pytorch built with DPYTORCH_ONNX_CAFFE2_BUNDLE，则默认为OperatorExportTypes.ONNX_ATEN_FALLBACK。枚举类型包括：
- OperatorExportTypes.ONNX - 将所有操作导出为ONNX操作。
- OperatorExportTypes.ONNX_FALLTHROUGH - 试图将所有操作导出为ONNX操作，但碰到无法转换的操作（如onnx未实现的操作），则将操作导出为“自定义操作”，为了使导出的模型可用，运行时必须支持这些自定义操作。支持自定义操作方法见链接。
- OperatorExportTypes.ONNX_ATEN - 所有ATen操作导出为ATen操作，ATen是Pytorch的内建tensor库，所以这将使得模型直接使用Pytorch实现。（此方法转换的模型只能被Caffe2直接使用）
- OperatorExportTypes.ONNX_ATEN_FALLBACK - 试图将所有的ATen操作也转换为ONNX操作，如果无法转换则转换为ATen操作（此方法转换的模型只能被Caffe2直接使用）。例如：
```
# 转换前：
graph(%0 : Float):
  %3 : int = prim::Constant[value=0]()
  # conversion unsupported
  %4 : Float = aten::triu(%0, %3)
  # conversion supported
  %5 : Float = aten::mul(%4, %0)
  return (%5)


# 转换后：
graph(%0 : Float):
  %1 : Long() = onnx::Constant[value={0}]()
  # not converted
  %2 : Float = aten::ATen[operator="triu"](%0, %1)
  # converted
  %3 : Float = onnx::Mul(%2, %0)
  return (%3)
```
opset_version (int, default 9) ：取值必须等于_onnx_main_opset或在_onnx_stable_opsets之内。具体可在torch/onnx/symbolic_helper.py中找到。例如：
```
_default_onnx_opset_version = 9
_onnx_main_opset = 13
_onnx_stable_opsets = [7, 8, 9, 10, 11, 12]
_export_onnx_opset_version = _default_onnx_opset_version
```
do_constant_folding (bool, default False) ：是否使用“常量折叠”优化。常量折叠将使用一些算好的常量来优化一些输入全为常量的节点。
example_outputs (T or a tuple of T, where T is Tensor or convertible to Tensor, default None) ：当需输入模型为ScriptModule 或 ScriptFunction时必须提供。此参数用于确定输出的类型和形状，而不跟踪（tracing）模型的执行。
dynamic_axes (dict<string, dict<python:int, string>> or dict<string, list(int)>, default empty dict) ：通过以下规则设置动态的维度：
- KEY(str) - 必须是input_names或output_names指定的名称，用来指定哪个变量需要使用到动态尺寸。
- VALUE(dict or list) - 如果是一个dict，dict中的key是变量的某个维度，dict中的value是我们给这个维度取的名称。如果是一个list，则list中的元素都表示此变量的某个维度。

代码实现：

import torch
import torchvision

weight_file = 'googlenet_catdog.pt'
onnx_file = 'googlenet_catdog.onnx'

model = torchvision.models.googlenet(num_classes=2, init_weights=True)
model.load_state_dict(torch.load(weight_file, map_location=torch.device('cpu')))

model.eval()

# 单输入单输出，固定batch
input = torch.randn(1, 3, 224, 224)
input_names = ["input"]
output_names = ["output"]
torch.onnx.export(model=model,
                  args=input,
                  f=onnx_file,
                  input_names=input_names,
                  output_names=output_names,
                  opset_version=11,
                  verbose=True)

通过netron.app可视化onnx的输入输出：　

如果需要多张图片同时进行推理，可以通过设置export的dynamic_axes参数，将模型输入输出的指定维度设置为变量。

import torch
import torchvision

weight_file = 'googlenet_catdog.pt'
onne_file = 'googlenet_catdog.onnx'

model = torchvision.models.googlenet(num_classes=2, init_weights=True)
model.load_state_dict(torch.load(weight_file, map_location=torch.device('cpu')))

model.eval()

# 单输入单输出，动态batch
input = torch.randn(1, 3, 224, 224)
input_names = ["input"]
output_names = ["output"]
torch.onnx.export(model=model,
                  args=input,
                  f=onnx_file,
                  input_names=input_names,
                  output_names=output_names,
                  opset_version=11,
                  verbose=True,
                  dynamic_axes={'input': {0: 'batch'}, 'output': {0: 'batch'}})

动态batch的onnx文件输入输出在netron.app可视化如下，其中batch维度是变量的形式，可以根据自己需要设置为大于0的任意整数。

如果模型有多个输入和输出，按照以下形式导出：

# 模型有两个输入和两个输出，动态batch
input1 = torch.randn(1, 3, 256, 192).to(opt.device)
input2 = torch.randn(1, 3, 256, 192).to(opt.device)
input_names = ["input1", "input2"]
output_names = ["output1", "output2"]
torch.onnx.export(model=model,
                  args=(input1, input2),
                  f=opt.onnx_path,
                  input_names=input_names,
                  output_names=output_names,
                  opset_version=16,
                  verbose=True,
                  dynamic_axes={'input1': {0: 'batch'},
                                'input2': {0: 'batch'},
                                'output1': {0: 'batch'},
                                'output2': {0: 'batch'}})

3. OpenCV DNN部署GoogLeNet

3.1 推理过程及代码实现

整个推理过程可分为前处理、推理、后处理三部分。具体细节请阅读代码，包括单图推理、动态batch推理的实现。

#include <opencv2/opencv.hpp>
#include <opencv2/dnn.hpp>
#include <chrono>
#include <fstream>

using namespace std;
using namespace cv;
using namespace cv::dnn;

std::string onnxPath = "E:/inference-master/models/engine/googlenet-pretrained_batch.onnx";
std::string imagePath = "E:/inference-master/images/catdog";
std::string classNamesPath = "E:/inference-master/imagenet-classes.txt";			// 标签名称列表（类名）

cv::dnn::Net net;
std::vector<std::string> classNameList;		// 标签名，可以从文件读取
int batchSize = 32;

int softmax(const cv::Mat& src, cv::Mat& dst)
{
	float max = 0.0;
	float sum = 0.0;

	max = *max_element(src.begin<float>(), src.end<float>());
	cv::exp((src - max), dst);
	sum = cv::sum(dst)[0];
	dst /= sum;

	return 0;
}

// GoogLeNet模型初始化
void ModelInit(string onnxPath)
{
	net = cv::dnn::readNetFromONNX(onnxPath);
	// net = cv::dnn::readNetFromCaffe("E:/inference-master/2/deploy.prototxt", "E:/inference-master/2/default.caffemodel");

	// 设置计算后台和计算设备
	// CPU（默认）
	// net.setPreferableBackend(cv::dnn::DNN_BACKEND_OPENCV);
	// net.setPreferableTarget(cv::dnn::DNN_TARGET_CPU);
	// CUDA
	net.setPreferableBackend(cv::dnn::DNN_BACKEND_CUDA);
	net.setPreferableTarget(cv::dnn::DNN_TARGET_CUDA);

	// 读取标签名称
	ifstream fin(classNamesPath.c_str());
	string strLine;
	classNameList.clear();
	while (getline(fin, strLine))
		classNameList.push_back(strLine);
	fin.close();
}

// 单图推理
bool ModelInference(cv::Mat srcImage, std::string& className, float& confidence)
{
	auto start = chrono::high_resolution_clock::now();

	cv::Mat image = srcImage.clone();

	// 预处理（尺寸变换、通道变换、归一化）
	cv::cvtColor(image, image, cv::COLOR_BGR2RGB);
	cv::resize(image, image, cv::Size(224, 224));
	image.convertTo(image, CV_32FC3, 1.0 / 255.0);
	cv::Scalar mean(0.485, 0.456, 0.406);
	cv::Scalar std(0.229, 0.224, 0.225);
	cv::subtract(image, mean, image);
	cv::divide(image, std, image);

	// blobFromImage操作顺序：swapRB交换通道 -> scalefactor比例缩放 -> mean求减 -> size进行resize;
	// mean操作时，ddepth不能选取CV_8U;
	// crop=True时，先等比缩放，直到宽高之一率先达到对应的size尺寸，另一个大于或等于对应的size尺寸，然后从中心裁剪;
	// 返回4-D Mat维度顺序：NCHW
	// cv::Mat blob = cv::dnn::blobFromImage(image, 1., cv::Size(224, 224), cv::Scalar(0, 0, 0), false, false);
	cv::Mat blob = cv::dnn::blobFromImage(image);

	// 设置输入
	net.setInput(blob);

	auto end1 = std::chrono::high_resolution_clock::now();
	auto ms1 = std::chrono::duration_cast<std::chrono::microseconds>(end1 - start);
	std::cout << "PreProcess time: " << (ms1 / 1000.0).count() << "ms" << std::endl;

	// 前向推理
	cv::Mat preds = net.forward();

	auto end2 = std::chrono::high_resolution_clock::now();
	auto ms2 = std::chrono::duration_cast<std::chrono::microseconds>(end2 - end1);
	std::cout << "Inference time: " << (ms2 / 1000.0).count() << "ms" << std::endl;

	// 结果归一化（每个batch分别求softmax）
	softmax(preds, preds);

	Point minLoc, maxLoc;
	double minValue = 0, maxValue = 0;
	cv::minMaxLoc(preds, &minValue, &maxValue, &minLoc, &maxLoc);
	int labelIndex = maxLoc.x;
	double probability = maxValue;

	className = classNameList[labelIndex];
	confidence = probability;

	// std::cout << "class：" << className << endl << "confidence：" << confidence << endl;

	auto end3 = std::chrono::high_resolution_clock::now();
	auto ms3 = std::chrono::duration_cast<std::chrono::microseconds>(end3 - end2);
	std::cout << "PostProcess time: " << (ms3 / 1000.0).count() << "ms" << std::endl;

	auto ms = chrono::duration_cast<std::chrono::microseconds>(end3 - start);
	std::cout << "opencv_dnn 推理时间：" << (ms / 1000.0).count() << "ms" << std::endl;
}

// 多图并行推理（动态batch）
bool ModelInference_Batch(std::vector<cv::Mat> srcImages, std::vector<string>& classNames, std::vector<float>& confidences)
{
	auto start = chrono::high_resolution_clock::now();

	// 预处理（尺寸变换、通道变换、归一化）
	std::vector<cv::Mat> images;
	for (size_t i = 0; i < srcImages.size(); i++)
	{
		cv::Mat image = srcImages[i].clone();
		cv::cvtColor(image, image, cv::COLOR_BGR2RGB);
		cv::resize(image, image, cv::Size(224, 224));
		image.convertTo(image, CV_32FC3, 1.0 / 255.0);
		cv::Scalar mean(0.485, 0.456, 0.406);
		cv::Scalar std(0.229, 0.224, 0.225);
		cv::subtract(image, mean, image);
		cv::divide(image, std, image);

		images.push_back(image);
	}

	cv::Mat blob = cv::dnn::blobFromImages(images);

	auto end1 = std::chrono::high_resolution_clock::now();
	auto ms1 = std::chrono::duration_cast<std::chrono::microseconds>(end1 - start);
	std::cout << "PreProcess time: " << (ms1 / 1000.0).count() << "ms" << std::endl;

	// 设置输入
	net.setInput(blob);

	// 前向推理
	cv::Mat	preds =  net.forward();

	auto end2 = std::chrono::high_resolution_clock::now();
	auto ms2 = std::chrono::duration_cast<std::chrono::microseconds>(end2 - end1) / 100.0;
	std::cout << "Inference time: " << (ms2 / 1000.0).count() << "ms" << std::endl;

	int rows = preds.size[0];	// batch
	int cols = preds.size[1];	// 类别数（每一个类别的得分）
	for (int row = 0; row < rows; row++)
	{
		cv::Mat scores(1, cols, CV_32FC1, preds.ptr<float>(row));
		softmax(scores, scores);	// 结果归一化

		Point minLoc, maxLoc;
		double minValue = 0, maxValue = 0;
		cv::minMaxLoc(scores, &minValue, &maxValue, &minLoc, &maxLoc);
		int labelIndex = maxLoc.x;
		double probability = maxValue;

		classNames.push_back(classNameList[labelIndex]);
		confidences.push_back(probability);
	}

	auto end3 = std::chrono::high_resolution_clock::now();
	auto ms3 = std::chrono::duration_cast<std::chrono::microseconds>(end3 - end2);
	std::cout << "PostProcess time: " << (ms3 / 1000.0).count() << "ms" << std::endl;

	auto ms = chrono::duration_cast<std::chrono::microseconds>(end3 - start);
	std::cout << "opencv_dnn batch" << rows << " 推理时间：" << (ms / 1000.0).count() << "ms" << std::endl;
}

int main(int argc, char** argv)
{
	// 模型初始化
	ModelInit(onnxPath);

	// 读取图像
	vector<string> filenames;
	glob(imagePath, filenames);

	// 单图推理测试
	for (int n = 0; n < filenames.size(); n++)
	{
		// 重复100次，计算平均时间
		auto start = chrono::high_resolution_clock::now();
		cv::Mat src = imread(filenames[n]);
		std::string classname;
		float confidence;
		for (int i = 0; i < 101; i++) {
			if (i==1)
				start = chrono::high_resolution_clock::now();
			ModelInference(src, classname, confidence);
		}
		auto end = chrono::high_resolution_clock::now();
		auto ms = chrono::duration_cast<std::chrono::microseconds>(end - start) / 100;
		std::cout << "opencv_dnn 平均推理时间：---------------------" << (ms / 1000.0).count() << "ms" << std::endl;
	}

	// 批量（动态batch）推理测试
	std::vector<cv::Mat> srcImages;
	for (int n = 0; n < filenames.size(); n++)
	{
		cv::Mat image = imread(filenames[n]);
		srcImages.push_back(image);
		if ((n + 1) % batchSize == 0 || n == filenames.size() - 1)
		{
			// 重复100次，计算平均时间
			auto start = chrono::high_resolution_clock::now();
			for (int i = 0; i < 101; i++) {
				if (i == 1)
					start = chrono::high_resolution_clock::now();
				std::vector<std::string> classNames;
				std::vector<float> confidences;
				ModelInference_Batch(srcImages, classNames, confidences);	
			}	
			srcImages.clear();
			auto end = chrono::high_resolution_clock::now();
			auto ms = chrono::duration_cast<std::chrono::microseconds>(end - start) / 100;
			std::cout << "opencv_dnn batch" << batchSize << " 平均推理时间：---------------------" << (ms / 1000.0).count() << "ms" << std::endl;	
		}	
	}

	return 0;
}

3.2 选择CPU/GPU

OpenCV DNN切换CPU和GPU推理，只需要通过下边两行代码设置计算后台和计算设备。

CPU推理

net.setPreferableBackend(cv::dnn::DNN_BACKEND_OPENCV);
net.setPreferableTarget(cv::dnn::DNN_TARGET_CPU);

GPU推理

net.setPreferableBackend(cv::dnn::DNN_BACKEND_CUDA);
net.setPreferableTarget(cv::dnn::DNN_TARGET_CUDA);

以下两点需要注意：

在不做任何设置的情况下，默认使用CPU进行推理。
在设置为GPU推理时，如果电脑没有搜索到CUDA环境，则会自动转换成CPU进行推理。

3.3 多输出模型推理

当模型有多个输出时，使用forward的重载方法，返回Mat类型的数组：

// 模型多输出
std::vector<cv::Mat> preds;
net.forward(preds);

cv::Mat pred1 = preds[0];
cv::Mat pred2 = preds[1];

4. ONNXRuntime部署GoogLeNet

4.1 推理过程及代码实现

代码：

#include <opencv2/opencv.hpp>
#include <opencv2/dnn.hpp>
#include <onnxruntime_cxx_api.h>
#include <vector>
#include <fstream>
#include <chrono>

using namespace std;
using namespace cv;
using namespace Ort;

// C++表示字符串的方式：char*、string、wchar_t*、wstring、字符串数组
const wchar_t* onnxPath = L"E:/inference-master/models/GoogLeNet/googlenet-pretrained_batch1.onnx";
std::string imagePath = "E:/inference-master/images/catdog";
std::string classNamesPath = "E:/inference-master/imagenet-classes.txt";			// 标签名称列表（类名）

std::vector<std::string> classNameList;		// 标签名，可以从文件读取
int batchSize = 1;

Ort::Env env{ nullptr };
Ort::SessionOptions* sessionOptions;

Ort::Session* session;
size_t inputCount;
size_t outputCount;
std::vector<const char*> inputNames;
std::vector<const char*> outputNames;
std::vector<int64_t> inputShape;
std::vector<int64_t> outputShape;

// 对数组元素求softmax
std::vector<float> softmax(std::vector<float> input)
{
	float total = 0;
	for (auto x : input)
		total += exp(x);
	std::vector<float> result;
	for (auto x : input)
		result.push_back(exp(x) / total);

	return result;
}

int softmax(const cv::Mat& src, cv::Mat& dst)
{
	float max = 0.0;
	float sum = 0.0;

	max = *max_element(src.begin<float>(), src.end<float>());
	cv::exp((src - max), dst);
	sum = cv::sum(dst)[0];
	dst /= sum;

	return 0;
}


// 前（预）处理（通道变换、标准化等）
void PreProcess(cv::Mat srcImage, cv::Mat& dstImage)
{
	// 通道变换，BGR->RGB
	cvtColor(srcImage, dstImage, cv::COLOR_BGR2RGB);
	resize(dstImage, dstImage, Size(224, 224));
	// 图像归一化
	dstImage.convertTo(dstImage, CV_32FC3, 1.0 / 255.0);
	cv::Scalar mean(0.485, 0.456, 0.406);
	cv::Scalar std(0.229, 0.224, 0.225);
	subtract(dstImage, mean, dstImage);
	divide(dstImage, std, dstImage);
}

// 模型初始化
int ModelInit(const wchar_t* onnxPath, bool useCuda, int deviceId)
{
	// 读取标签名称
	std::ifstream fin(classNamesPath.c_str());
	std::string strLine;
	classNameList.clear();
	while (getline(fin, strLine))
		classNameList.push_back(strLine);
	fin.close();

	// 环境设置，控制台输出设置
	env = Ort::Env(ORT_LOGGING_LEVEL_WARNING, "GoogLeNet");

	sessionOptions = new Ort::SessionOptions();
	// 设置线程数
	sessionOptions->SetIntraOpNumThreads(16);
	// 优化等级：启用所有可能的优化
	sessionOptions->SetGraphOptimizationLevel(GraphOptimizationLevel::ORT_ENABLE_ALL);

	if (useCuda) {
		// 开启CUDA加速，需要cuda_provider_factory.h头文件
		OrtSessionOptionsAppendExecutionProvider_CUDA(*sessionOptions, deviceId);
	}	

	// 创建session
	session = new Ort::Session(env, onnxPath, *sessionOptions);

	// 获取输入输出数量
	inputCount = session->GetInputCount();
	outputCount = session->GetOutputCount();
	std::cout << "Number of inputs = " << inputCount << std::endl;
	std::cout << "Number of outputs = " << outputCount << std::endl;

	// 获取输入输出名称
	Ort::AllocatorWithDefaultOptions allocator;
	const char* inputName = session->GetInputName(0, allocator);
	const char* outputName = session->GetOutputName(0, allocator);
	inputNames = { inputName };
	outputNames = { outputName };
	std::cout << "Name of inputs = " << inputName << std::endl;
	std::cout << "Name of outputs = " << outputName << std::endl;

	// 获取输入输出维度信息，返回类型std::vector<int64_t>
	inputShape = session->GetInputTypeInfo(0).GetTensorTypeAndShapeInfo().GetShape();
	outputShape = session->GetOutputTypeInfo(0).GetTensorTypeAndShapeInfo().GetShape();
	std::cout << "Shape of inputs = " << "(" << inputShape[0] << "," << inputShape[1] << "," << inputShape[2] << "," << inputShape[3] << ")" << std::endl;
	std::cout << "Shape of outputs = " << "(" << outputShape[0] << "," << outputShape[1] << ")" << std::endl;

	return 0;
}

// 单图推理
void ModelInference(cv::Mat srcImage, std::string& className, float& confidence)
{
	auto start = chrono::high_resolution_clock::now();
	
	// 输入图像预处理
	cv::Mat image;
	//PreProcess(srcImage, image);  // 这里使用调用函数的方式，处理时间莫名变长很多，很奇怪

	// 通道变换，BGR->RGB
	cvtColor(srcImage, image, cv::COLOR_BGR2RGB);
	resize(image, image, Size(224, 224));
	// 图像归一化
	image.convertTo(image, CV_32FC3, 1.0 / 255.0);
	cv::Scalar mean(0.485, 0.456, 0.406);
	cv::Scalar std(0.229, 0.224, 0.225);
	subtract(image, mean, image);
	divide(image, std, image);

	cv::Mat blob = cv::dnn::blobFromImage(image);

	auto end1 = std::chrono::high_resolution_clock::now();
	auto ms1 = std::chrono::duration_cast<std::chrono::microseconds>(end1 - start);
	std::cout << "PreProcess time: " << (ms1 / 1000.0).count() << "ms" << std::endl;

	// 创建输入tensor
	auto memoryInfo = Ort::MemoryInfo::CreateCpu(OrtAllocatorType::OrtArenaAllocator, OrtMemType::OrtMemTypeDefault);
	std::vector<Ort::Value> inputTensors;
	inputTensors.emplace_back(Ort::Value::CreateTensor<float>(memoryInfo,
		blob.ptr<float>(), blob.total(), inputShape.data(), inputShape.size()));

	// 推理
	auto outputTensors = session->Run(Ort::RunOptions{ nullptr },
		inputNames.data(), inputTensors.data(), inputCount, outputNames.data(), outputCount);

	auto end2 = std::chrono::high_resolution_clock::now();
	auto ms2 = std::chrono::duration_cast<std::chrono::microseconds>(end2 - end1);
	std::cout << "Inference time: " << (ms2 / 1000.0).count() << "ms" << std::endl;

	// 获取输出
	float* preds = outputTensors[0].GetTensorMutableData<float>();	// 也可以使用outputTensors.front();

	int64_t numClasses = outputShape[1];
	cv::Mat output = cv::Mat_<float>(1, numClasses);
	for (int j = 0; j < numClasses; j++) {
		output.at<float>(0, j) = preds[j];
	}

	Point minLoc, maxLoc;
	double minValue = 0, maxValue = 0;
	cv::minMaxLoc(output, &minValue, &maxValue, &minLoc, &maxLoc);
	int labelIndex = maxLoc.x;
	double probability = maxValue;

	className = classNameList[1];
	confidence = probability;

	auto end3 = std::chrono::high_resolution_clock::now();
	auto ms3 = std::chrono::duration_cast<std::chrono::microseconds>(end3 - end2);
	std::cout << "PostProcess time: " << (ms3 / 1000.0).count() << "ms" << std::endl;

	auto ms = chrono::duration_cast<std::chrono::microseconds>(end3 - start);
	std::cout << "onnxruntime单图推理时间：" << (ms / 1000.0).count() << "ms" << std::endl;
}

// 单图推理
void ModelInference_Batch(std::vector<cv::Mat> srcImages, std::vector<string>& classNames, std::vector<float>& confidences)
{
	auto start = chrono::high_resolution_clock::now();

	// 输入图像预处理
	std::vector<cv::Mat> images;
	for (size_t i = 0; i < srcImages.size(); i++)
	{
		cv::Mat image = srcImages[i].clone();

		// 通道变换，BGR->RGB
		cvtColor(image, image, cv::COLOR_BGR2RGB);
		resize(image, image, Size(224, 224));
		// 图像归一化
		image.convertTo(image, CV_32FC3, 1.0 / 255.0);
		cv::Scalar mean(0.485, 0.456, 0.406);
		cv::Scalar std(0.229, 0.224, 0.225);
		subtract(image, mean, image);
		divide(image, std, image);

		images.push_back(image);
	}

	// 图像转blob格式
	cv::Mat blob = cv::dnn::blobFromImages(images);

	auto end1 = std::chrono::high_resolution_clock::now();
	auto ms1 = std::chrono::duration_cast<std::chrono::microseconds>(end1 - start);
	std::cout << "PreProcess time: " << (ms1 / 1000.0).count() << "ms" << std::endl;

	// 创建输入tensor
	std::vector<Ort::Value> inputTensors;
	auto memoryInfo = Ort::MemoryInfo::CreateCpu(OrtAllocatorType::OrtArenaAllocator, OrtMemType::OrtMemTypeDefault);
	inputTensors.emplace_back(Ort::Value::CreateTensor<float>(memoryInfo,
		blob.ptr<float>(), blob.total(), inputShape.data(), inputShape.size()));
	
	// 推理
	std::vector<Ort::Value> outputTensors = session->Run(Ort::RunOptions{ nullptr },
		inputNames.data(), inputTensors.data(), inputCount, outputNames.data(), outputCount);

	auto end2 = std::chrono::high_resolution_clock::now();
	auto ms2 = std::chrono::duration_cast<std::chrono::microseconds>(end2 - end1)/100;
	std::cout << "inference time: " << (ms2 / 1000.0).count() << "ms" << std::endl;

	// 获取输出
	float* preds = outputTensors[0].GetTensorMutableData<float>();	// 也可以使用outputTensors.front();

	// cout << preds[0] << "," << preds[1] << "," << preds[1000] << "," << preds[1001] << endl;

	int batch = outputShape[0];
	int numClasses = outputShape[1];

	cv::Mat output(batch, numClasses, CV_32FC1, preds);

	int rows = output.size[0];	// batch
	int cols = output.size[1];	// 类别数（每一个类别的得分）
	for (int row = 0; row < rows; row++)
	{
		cv::Mat scores(1, cols, CV_32FC1, output.ptr<float>(row));
		softmax(scores, scores);	// 结果归一化

		Point minLoc, maxLoc;
		double minValue = 0, maxValue = 0;
		cv::minMaxLoc(scores, &minValue, &maxValue, &minLoc, &maxLoc);
		int labelIndex = maxLoc.x;
		double probability = maxValue;

		classNames.push_back(classNameList[labelIndex]);
		confidences.push_back(probability);
	}

	auto end3 = std::chrono::high_resolution_clock::now();
	auto ms3 = std::chrono::duration_cast<std::chrono::microseconds>(end3 - end2);
	std::cout << "PostProcess time: " << (ms3 / 1000.0).count() << "ms" << std::endl;

	auto ms = chrono::duration_cast<std::chrono::microseconds>(end3 - start);
	std::cout << "onnxruntime单图推理时间：" << (ms / 1000.0).count() << "ms" << std::endl;
}

int main(int argc, char** argv)
{
	// 模型初始化
	ModelInit(onnxPath, true, 0);

	// 读取图像
	std::vector<std::string> filenames;
	cv::glob(imagePath, filenames);

	// 单图推理测试
	for (int i = 0; i < filenames.size(); i++)
	{
		// 每张图重复运行100次，计算平均时间
		auto start = chrono::high_resolution_clock::now();
		cv::Mat srcImage = imread(filenames[i]);
		std::string className;
		float confidence;
		for (int n = 0; n < 101; n++) {
			if (n == 1)
				start = chrono::high_resolution_clock::now();
			ModelInference(srcImage, className, confidence);
		}

		// 显示
		cv::putText(srcImage, className + "：" + std::to_string(confidence),
			cv::Point(10, 20), FONT_HERSHEY_SIMPLEX, 0.6, cv::Scalar(0, 0, 255), 1, 1);

		auto end = chrono::high_resolution_clock::now();
		auto ms = chrono::duration_cast<std::chrono::microseconds>(end - start) / 100;
		std::cout << "onnxruntime 平均推理时间：---------------------" << (ms / 1000.0).count() << "ms" << std::endl;
	}

	// 批量推理测试
	std::vector<cv::Mat> srcImages;
	for (int i = 0; i < filenames.size(); i++)
	{
		cv::Mat image = imread(filenames[i]);
		srcImages.push_back(image);
		if ((i + 1) % batchSize == 0 || i == filenames.size() - 1)
		{
			// 重复100次，计算平均时间
			auto start = chrono::high_resolution_clock::now();
			for (int n = 0; n < 101; n++) {
				if (n == 1)
					start = chrono::high_resolution_clock::now();	// 首次推理耗时很久
				std::vector<std::string> classNames;
				std::vector<float> confidences;
				ModelInference_Batch(srcImages, classNames, confidences);
			}
			srcImages.clear();
			auto end = chrono::high_resolution_clock::now();
			auto ms = chrono::duration_cast<std::chrono::microseconds>(end - start) / 100;
			std::cout << "onnxruntime batch" << batchSize << " 平均推理时间：---------------------" << (ms / 1000.0).count() << "ms" << std::endl;
		}
	}

	return 0;
}

注意：ORT支持多图并行推理，但是要求转出onnx的时候batch就要使用固定数值。动态batch（即batch=-1）的onnx文件是不支持推理的。

4.2 选择CPU/GPU

使用GPU推理，只需要添加一行代码：

if (useCuda) {
	// 开启CUDA加速
	OrtSessionOptionsAppendExecutionProvider_CUDA(*sessionOptions, deviceId);
}

4.3 多输入多输出模型推理

推理步骤和单图推理基本一致，需要在输入tensor中依次添加所有的输入。假设模型有两个输入和两个输出：

// 创建session
session2 = new Ort::Session(env1, onnxPath, sessionOptions1);

// 获取模型输入输出信息
inputCount2 = session2->GetInputCount();
outputCount2 = session2->GetOutputCount();

// 输入和输出各有两个
Ort::AllocatorWithDefaultOptions allocator;
const char* inputName1 = session2->GetInputName(0, allocator);
const char* inputName2 = session2->GetInputName(1, allocator);
const char* outputName1 = session2->GetOutputName(0, allocator);
const char* outputName2 = session2->GetOutputName(1, allocator);
intputNames2 = { inputName1, inputName2 };
outputNames2 = { outputName1, outputName2 };

// 获取输入输出维度信息，返回类型std::vector<int64_t>
inputShape2_1 = session2->GetInputTypeInfo(0).GetTensorTypeAndShapeInfo().GetShape();
inputShape2_2 = session2->GetInputTypeInfo(1).GetTensorTypeAndShapeInfo().GetShape();
outputShape2_1 = session2->GetOutputTypeInfo(0).GetTensorTypeAndShapeInfo().GetShape();
outputShape2_2 = session2->GetOutputTypeInfo(1).GetTensorTypeAndShapeInfo().GetShape();

...

// 创建输入tensor
auto memoryInfo = Ort::MemoryInfo::CreateCpu(OrtAllocatorType::OrtArenaAllocator, OrtMemType::OrtMemTypeDefault);
std::vector<Ort::Value> inputTensors;
inputTensors.emplace_back(Ort::Value::CreateTensor<float>(memoryInfo,
	blob1.ptr<float>(), blob1.total(), inputShape2_1.data(), inputShape2_1.size()));
inputTensors.emplace_back(Ort::Value::CreateTensor<float>(memoryInfo,
	blob2.ptr<float>(), blob2.total(), inputShape2_2.data(), inputShape2_2.size()));
	
// 推理
auto outputTensors = session2->Run(Ort::RunOptions{ nullptr },
	intputNames2.data(), inputTensors.data(), inputCount2, outputNames2.data(), outputCount2);

// 获取输出
float* preds1 = outputTensors[0].GetTensorMutableData<float>();
float* preds2 = outputTensors[1].GetTensorMutableData<float>();

5. TensorRT部署GoogLeNet

TRT推理有两种常见的方式：

通过官方安装包里边的提供的trtexec.exe工具，从onnx文件转换得到trt文件，然后执行推理；
由onnx文件转化得到engine文件，再执行推理。

两种方式原理一样，这里我们只介绍第二种方式。推理过程可分为两阶段：使用onnx构建推理engine和加载engine执行推理。

5.1 构建推理引擎（engine文件）

engine的构建是TensorRT推理至关重要的一步，它特定于所构建的确切GPU模型，不能跨平台或TensorRT版本移植。举个简单的例子，如果你在RTX3060上使用TensorRT 8.2.5构建了engine，那么推理部署也必须要在RTX3060上进行，且要具备TensorRT 8.2.5环境。engine构建的大致流程如下：

engine的构建有很多种方式，这里我们介绍常用的三种。我一般会选择直接在Python中构建，这样模型的训练、转onnx、转engine都在Python端完成，方便且省事。

方法一：在Python中构建

import os
import sys
import logging
import argparse
import tensorrt as trt

os.environ['CUDA_VISIBLE_DEVICES'] = '0'
# 延迟加载模式，cuda11.7新功能，设置为LAZY有可能会极大的降低内存和显存的占用
os.environ['CUDA_MODULE_LOADING'] = 'LAZY'

logging.basicConfig(level=logging.INFO)
logging.getLogger("EngineBuilder").setLevel(logging.INFO)
log = logging.getLogger("EngineBuilder")


class EngineBuilder:
    """
    Parses an ONNX graph and builds a TensorRT engine from it.
    """

    def __init__(self, batch_size=1, verbose=False, workspace=8):
        """
        :param verbose: If enabled, a higher verbosity level will be set on the TensorRT logger.
        :param workspace: Max memory workspace to allow, in Gb.
        """
        # 1. 构建builder
        self.trt_logger = trt.Logger(trt.Logger.INFO)
        if verbose:
            self.trt_logger.min_severity = trt.Logger.Severity.VERBOSE

        trt.init_libnvinfer_plugins(self.trt_logger, namespace="")

        self.builder = trt.Builder(self.trt_logger)
        self.config = self.builder.create_builder_config()  # 构造builder.config
        self.config.max_workspace_size = workspace * (2 ** 30)  # workspace分配

        self.batch_size = batch_size
        self.network = None
        self.parser = None

    def create_network(self, onnx_path):
        """
        Parse the ONNX graph and create the corresponding TensorRT network definition.
        :param onnx_path: The path to the ONNX graph to load.
        """
        network_flags = (1 << int(trt.NetworkDefinitionCreationFlag.EXPLICIT_BATCH))

        self.network = self.builder.create_network(network_flags)
        self.parser = trt.OnnxParser(self.network, self.trt_logger)

        onnx_path = os.path.realpath(onnx_path)
        with open(onnx_path, "rb") as f:
            if not self.parser.parse(f.read()):
                log.error("Failed to load ONNX file: {}".format(onnx_path))
                for error in range(self.parser.num_errors):
                    log.error(self.parser.get_error(error))
                sys.exit(1)

        # 获取网络输入输出
        inputs = [self.network.get_input(i) for i in range(self.network.num_inputs)]
        outputs = [self.network.get_output(i) for i in range(self.network.num_outputs)]

        log.info("Network Description")
        for input in inputs:
            self.batch_size = input.shape[0]
            log.info("Input '{}' with shape {} and dtype {}".format(input.name, input.shape, input.dtype))
        for output in outputs:
            log.info("Output '{}' with shape {} and dtype {}".format(output.name, output.shape, output.dtype))
        assert self.batch_size > 0
        self.builder.max_batch_size = self.batch_size

    def create_engine(self, engine_path, precision):
        """
        Build the TensorRT engine and serialize it to disk.
        :param engine_path: The path where to serialize the engine to.
        :param precision: The datatype to use for the engine, either 'fp32', 'fp16' or 'int8'.
        """
        engine_path = os.path.realpath(engine_path)
        engine_dir = os.path.dirname(engine_path)
        os.makedirs(engine_dir, exist_ok=True)
        log.info("Building {} Engine in {}".format(precision, engine_path))

        inputs = [self.network.get_input(i) for i in range(self.network.num_inputs)]

        if precision == "fp16":
            if not self.builder.platform_has_fast_fp16:
                log.warning("FP16 is not supported natively on this platform/device")
            else:
                self.config.set_flag(trt.BuilderFlag.FP16)

        with self.builder.build_engine(self.network, self.config) as engine, open(engine_path, "wb") as f:
            log.info("Serializing engine to file: {:}".format(engine_path))
            f.write(engine.serialize())


def main(args):
    builder = EngineBuilder(args.batch_size, args.verbose, args.workspace)
    builder.create_network(args.onnx)
    builder.create_engine(args.engine, args.precision)


if __name__ == "__main__":
    parser = argparse.ArgumentParser()
    parser.add_argument("-o", "--onnx", default=r'googlenet-pretrained_batch8.onnx', help="The input ONNX model file to load")
    parser.add_argument("-e", "--engine", default=r'googlenet-pretrained_batch8_from_py_3080_FP16.engine', help="The output path for the TRT engine")
    parser.add_argument("-p", "--precision", default="fp16", choices=["fp32", "fp16", "int8"],
                        help="The precision mode to build in, either 'fp32', 'fp16' or 'int8', default: 'fp16'")
    parser.add_argument("-b", "--batch_size", default=8, type=int, help="batch number of input")
    parser.add_argument("-v", "--verbose", action="store_true", help="Enable more verbose log output")
    parser.add_argument("-w", "--workspace", default=8, type=int, help="The max memory workspace size to allow in Gb, "
                                                                       "default: 8")

    args = parser.parse_args()
    main(args)

生成fp16模型：参数precision设置为fp16即可。int8模型生成过程比较复杂，且对模型精度影响较大，用的不多，这里暂不介绍。

parser.add_argument("-p", "--precision", default="fp16", choices=["fp32", "fp16", "int8"],
                        help="The precision mode to build in, either 'fp32', 'fp16' or 'int8', default: 'fp16'")

方法二：在C++中构建

#include "NvInfer.h"
#include "NvOnnxParser.h"
#include "cuda_runtime_api.h"
#include "logging.h"
#include <fstream>
#include <map>
#include <chrono>
#include <cmath>
#include <opencv2/opencv.hpp>
#include <fstream>

using namespace nvinfer1;
using namespace nvonnxparser;
using namespace std;
using namespace cv;

std::string onnxPath = "E:/inference-master/models/engine/googlenet-pretrained_batch.onnx";
std::string enginePath = "E:/inference-master/models/engine/googlenet-pretrained_batch_from_cpp.engine";  // 通过C++构建

static const int INPUT_H = 224;
static const int INPUT_W = 224;
static const int OUTPUT_SIZE = 1000;

static const int BATCH_SIZE = 25;

const char* INPUT_BLOB_NAME = "input";
const char* OUTPUT_BLOB_NAME = "output";

static Logger gLogger;

// onnx转engine
void onnx_to_engine(std::string onnx_file_path, std::string engine_file_path, int type) {

    // 创建builder实例，获取cuda内核目录以获取最快的实现，用于创建config、network、engine的其他对象的核心类
    nvinfer1::IBuilder* builder = nvinfer1::createInferBuilder(gLogger);
    const auto explicitBatch = 1U << static_cast<uint32_t>(nvinfer1::NetworkDefinitionCreationFlag::kEXPLICIT_BATCH);

    // 创建网络定义
    nvinfer1::INetworkDefinition* network = builder->createNetworkV2(explicitBatch);

    // 创建onnx解析器来填充网络
    nvonnxparser::IParser* parser = nvonnxparser::createParser(*network, gLogger);

    // 读取onnx模型文件
    parser->parseFromFile(onnx_file_path.c_str(), 2);
    for (int i = 0; i < parser->getNbErrors(); ++i) {
        std::cout << "load error: " << parser->getError(i)->desc() << std::endl;
    }
    printf("tensorRT load mask onnx model successfully!!!...\n");

    // 创建生成器配置对象
    nvinfer1::IBuilderConfig* config = builder->createBuilderConfig();
   
    builder->setMaxBatchSize(BATCH_SIZE);           // 设置最大batch 
    config->setMaxWorkspaceSize(16 * (1 << 20));    // 设置最大工作空间大小

    // 设置模型输出精度，0代表FP32，1代表FP16
    if (type == 1) {
        config->setFlag(nvinfer1::BuilderFlag::kFP16);
    }

    // 创建推理引擎
    nvinfer1::ICudaEngine* engine = builder->buildEngineWithConfig(*network, *config);

    // 将推理引擎保存到本地
    std::cout << "try to save engine file now~~~" << std::endl;
    std::ofstream file_ptr(engine_file_path, std::ios::binary);
    if (!file_ptr) {
        std::cerr << "could not open plan output file" << std::endl;
        return;
    }
    // 将模型转化为文件流数据
    nvinfer1::IHostMemory* model_stream = engine->serialize();
    // 将文件保存到本地
    file_ptr.write(reinterpret_cast<const char*>(model_stream->data()), model_stream->size());
    // 销毁创建的对象
    model_stream->destroy();
    engine->destroy();
    network->destroy();
    parser->destroy();
    std::cout << "convert onnx model to TensorRT engine model successfully!" << std::endl;
}

int main(int argc, char** argv)
{
    // onnx转engine
    onnx_to_engine(onnxPath, enginePath, 0);
    
    return 0；
}

方法三：使用官方安装包bin目录下的trtexec.exe工具构建

trtexec.exe --onnx=googlenet-pretrained_batch.onnx --saveEngine=googlenet-pretrained_batch_from_trt_trt853.engine --shapes=input:25x3x224x224

fp16模型：在后边加--fp16即可

trtexec.exe --onnx=googlenet-pretrained_batch.onnx --saveEngine=googlenet-pretrained_batch_from_trt_trt853.engine --shapes=input:25x3x224x224 --fp16

5.2 读取engine文件并部署模型

推理代码：

#include "NvInfer.h"
#include "NvOnnxParser.h"
#include "cuda_runtime_api.h"
#include "logging.h"
#include <fstream>
#include <map>
#include <chrono>
#include <cmath>
#include <opencv2/opencv.hpp>

#include "cuda.h"
#include "assert.h"
#include "iostream"

using namespace nvinfer1;
using namespace nvonnxparser;
using namespace std;
using namespace cv;

#define CHECK(status) \
    do\
    {\
        auto ret = (status);\
        if (ret != 0)\
        {\
            std::cerr << "Cuda failure: " << ret << std::endl;\
            abort();\
        }\
    } while (0)

std::string enginePath = "E:/inference-master/models/GoogLeNet/googlenet-pretrained_batch1_from_py_3080_FP32.engine";
std::string imagePath = "E:/inference-master/images/catdog";
std::string classNamesPath = "E:/inference-master/imagenet-classes.txt";			// 标签名称列表（类名）

std::vector<std::string> classNameList;		// 标签名列表

static const int INPUT_H = 224;
static const int INPUT_W = 224;
static const int CHANNEL = 3;
static const int OUTPUT_SIZE = 1000;

static const int BATCH_SIZE = 1;

const char* INPUT_BLOB_NAME = "input";
const char* OUTPUT_BLOB_NAME = "output";

static Logger gLogger;

IRuntime* runtime;
ICudaEngine* engine;
IExecutionContext* context;
void* gpu_buffers[2];
cudaStream_t stream;
const int inputIndex = 0;
const int outputIndex = 1;

// 提前申请内存，可节省推理时间
static float mydata[BATCH_SIZE * CHANNEL * INPUT_H * INPUT_W];
static float prob[BATCH_SIZE * OUTPUT_SIZE];

// 逐行求softmax
int softmax(const cv::Mat & src, cv::Mat & dst)
{
    float max = 0.0;
    float sum = 0.0;

    cv::Mat tmpdst = cv::Mat::zeros(src.size(), src.type());

    std::vector<cv::Mat> srcRows;
    // 逐行求softmax
    for (size_t i = 0; i < src.rows; i++)
    {
        cv::Mat tmpRow;
        cv::Mat dataRow = src.row(i).clone();
        max = *std::max_element(dataRow.begin<float>(), dataRow.end<float>());
        cv::exp((dataRow - max), tmpRow);
        sum = cv::sum(tmpRow)[0];
        tmpRow /= sum;

        srcRows.push_back(tmpRow);
        cv::vconcat(srcRows, tmpdst);
    }

    dst = tmpdst.clone();
    return 0;
}

// onnx转engine
void onnx_to_engine(std::string onnx_file_path, std::string engine_file_path, int type) {

    // 创建builder实例，获取cuda内核目录以获取最快的实现，用于创建config、network、engine的其他对象的核心类
    nvinfer1::IBuilder* builder = nvinfer1::createInferBuilder(gLogger);
    const auto explicitBatch = 1U << static_cast<uint32_t>(nvinfer1::NetworkDefinitionCreationFlag::kEXPLICIT_BATCH);

    // 创建网络定义
    nvinfer1::INetworkDefinition* network = builder->createNetworkV2(explicitBatch);

    // 创建onnx解析器来填充网络
    nvonnxparser::IParser* parser = nvonnxparser::createParser(*network, gLogger);

    // 读取onnx模型文件
    parser->parseFromFile(onnx_file_path.c_str(), 2);
    for (int i = 0; i < parser->getNbErrors(); ++i) {
        std::cout << "load error: " << parser->getError(i)->desc() << std::endl;
    }
    printf("tensorRT load mask onnx model successfully!!!...\n");

    // 创建生成器配置对象
    nvinfer1::IBuilderConfig* config = builder->createBuilderConfig();
   
    builder->setMaxBatchSize(BATCH_SIZE);           // 设置最大batch 
    config->setMaxWorkspaceSize(16 * (1 << 20));    // 设置最大工作空间大小

    // 设置模型输出精度
    if (type == 1) {
        config->setFlag(nvinfer1::BuilderFlag::kFP16);
    }
    if (type == 2) {
        config->setFlag(nvinfer1::BuilderFlag::kINT8);
    }
    // 创建推理引擎
    nvinfer1::ICudaEngine* engine = builder->buildEngineWithConfig(*network, *config);

    // 将推理引擎保存到本地
    std::cout << "try to save engine file now~~~" << std::endl;
    std::ofstream file_ptr(engine_file_path, std::ios::binary);
    if (!file_ptr) {
        std::cerr << "could not open plan output file" << std::endl;
        return;
    }
    // 将模型转化为文件流数据
    nvinfer1::IHostMemory* model_stream = engine->serialize();
    // 将文件保存到本地
    file_ptr.write(reinterpret_cast<const char*>(model_stream->data()), model_stream->size());
    // 销毁创建的对象
    model_stream->destroy();
    engine->destroy();
    network->destroy();
    parser->destroy();
    std::cout << "convert onnx model to TensorRT engine model successfully!" << std::endl;
}

// 模型推理：包括创建GPU显存缓冲区、配置模型输入及模型推理
void doInference(IExecutionContext& context, const void* input, float* output, int batchSize)
{
    //auto start = chrono::high_resolution_clock::now();

    // DMA input batch data to device, infer on the batch asynchronously, and DMA output back to host
    CHECK(cudaMemcpyAsync(gpu_buffers[inputIndex], input, batchSize * 3 * INPUT_H * INPUT_W * sizeof(float), cudaMemcpyHostToDevice, stream));

    // context.enqueue(batchSize, buffers, stream, nullptr);
    context.enqueueV2(gpu_buffers, stream, nullptr);

    //auto end1 = std::chrono::high_resolution_clock::now();
    //auto ms1 = std::chrono::duration_cast<std::chrono::microseconds>(end1 - start);
    //std::cout << "推理: " << (ms1 / 1000.0).count() << "ms" << std::endl;

    CHECK(cudaMemcpyAsync(output, gpu_buffers[outputIndex], batchSize * OUTPUT_SIZE * sizeof(float), cudaMemcpyDeviceToHost, stream));
    //size_t dest_pitch = 0;
    //CHECK(cudaMemcpy2D(output, dest_pitch, buffers[outputIndex], batchSize * sizeof(float), batchSize, OUTPUT_SIZE, cudaMemcpyDeviceToHost));
    cudaStreamSynchronize(stream);

    
    //auto end2 = std::chrono::high_resolution_clock::now();
    //auto ms2 = std::chrono::duration_cast<std::chrono::microseconds>(end2 - start)/100.0;
    //std::cout << "cuda-host: " << (ms2 / 1000.0).count() << "ms" << std::endl;

}

// 结束推理，释放资源
void GpuMemoryRelease()
{
    // Release stream and buffers
    cudaStreamDestroy(stream);
    CHECK(cudaFree(gpu_buffers[0]));
    CHECK(cudaFree(gpu_buffers[1]));

    // Destroy the engine
    context->destroy();
    engine->destroy();
    runtime->destroy();
}

// GoogLeNet模型初始化
void ModelInit(std::string enginePath, int deviceId)
{
    // 设置GPU
    cudaSetDevice(deviceId);

    // 从本地读取engine模型文件
    char* trtModelStream{ nullptr };
    size_t size{ 0 };
    std::ifstream file(enginePath, std::ios::binary);
    if (file.good()) {
        file.seekg(0, file.end);    // 将读指针从文件末尾开始移动0个字节
        size = file.tellg();    // 返回读指针的位置，此时读指针的位置就是文件的字节数
        file.seekg(0, file.beg);    // 将读指针从文件开头开始移动0个字节
        trtModelStream = new char[size];
        assert(trtModelStream);
        file.read(trtModelStream, size);
        file.close();
    }

    // 创建推理运行环境实例
    runtime = createInferRuntime(gLogger);
    assert(runtime != nullptr);
    // 反序列化模型
    engine = runtime->deserializeCudaEngine(trtModelStream, size, nullptr);
    assert(engine != nullptr);
    // 创建推理上下文
    context = engine->createExecutionContext();
    assert(context != nullptr);
    delete[] trtModelStream;

    // Create stream
    CHECK(cudaStreamCreate(&stream));

    // Pointers to input and output device buffers to pass to engine.
    // Engine requires exactly IEngine::getNbBindings() number of buffers.
    assert(engine.getNbBindings() == 2);

    // In order to bind the buffers, we need to know the names of the input and output tensors.
    // Note that indices are guaranteed to be less than IEngine::getNbBindings()
    const int inputIndex = engine->getBindingIndex(INPUT_BLOB_NAME);
    const int outputIndex = engine->getBindingIndex(OUTPUT_BLOB_NAME);

    // Create GPU buffers on device
    CHECK(cudaMalloc(&gpu_buffers[inputIndex], BATCH_SIZE * 3 * INPUT_H * INPUT_W * sizeof(float)));
    CHECK(cudaMalloc(&gpu_buffers[outputIndex], BATCH_SIZE * OUTPUT_SIZE * sizeof(float)));

    // 读取标签名称
    ifstream fin(classNamesPath.c_str());
    string strLine;
    classNameList.clear();
    while (getline(fin, strLine))
        classNameList.push_back(strLine);
    fin.close();
}

// 单图推理
bool ModelInference(cv::Mat srcImage, std::string& className, float& confidence)
{
    auto start = chrono::high_resolution_clock::now();

    cv::Mat image = srcImage.clone();

    // 预处理（尺寸变换、通道变换、归一化）
    cv::cvtColor(image, image, cv::COLOR_BGR2RGB);
    cv::resize(image, image, cv::Size(224, 224));
    image.convertTo(image, CV_32FC3, 1.0 / 255.0);
    cv::Scalar mean(0.485, 0.456, 0.406);
    cv::Scalar std(0.229, 0.224, 0.225);
    cv::subtract(image, mean, image);
    cv::divide(image, std, image);

    // cv::Mat blob = cv::dnn::blobFromImage(image);

    // 下边代码比上边blobFromImages速度更快
    for (int r = 0; r < INPUT_H; r++)
    {
        float* rowData = image.ptr<float>(r);
        for (int c = 0; c < INPUT_W; c++)
        {
            mydata[0 * INPUT_H * INPUT_W + r * INPUT_W + c] = rowData[CHANNEL * c];
            mydata[1 * INPUT_H * INPUT_W + r * INPUT_W + c] = rowData[CHANNEL * c + 1];
            mydata[2 * INPUT_H * INPUT_W + r * INPUT_W + c] = rowData[CHANNEL * c + 2];
        }
    }

    // 模型推理
    // doInference(*context, blob.data, prob, BATCH_SIZE);
    doInference(*context, mydata, prob, BATCH_SIZE);

    // 推理结果后处理
    cv::Mat preds = cv::Mat(BATCH_SIZE, OUTPUT_SIZE, CV_32FC1, (float*)prob);
 
    softmax(preds, preds);

    Point minLoc, maxLoc;
    double minValue = 0, maxValue = 0;
    cv::minMaxLoc(preds, &minValue, &maxValue, &minLoc, &maxLoc);
    int labelIndex = maxLoc.x;
    double probability = maxValue;

    className = classNameList[labelIndex];
    confidence = probability;

    std::cout << "class：" << className << endl << "confidence：" << confidence << endl;
    auto end = chrono::high_resolution_clock::now();
    auto ms = chrono::duration_cast<std::chrono::microseconds>(end - start);
    std::cout << "Inference time by TensorRT：" << (ms / 1000.0).count() << "ms" << std::endl;
   
    return 0;
}

// GoogLeNet模型推理
bool ModelInference_Batch(std::vector<cv::Mat> srcImages, std::vector<std::string>& classNames, std::vector<float>& confidences)
{
    auto start = std::chrono::high_resolution_clock::now();

    if (srcImages.size() != BATCH_SIZE) return false;

    // 预处理（尺寸变换、通道变换、归一化）
    std::vector<cv::Mat> images;
    for (size_t i = 0; i < srcImages.size(); i++)
    {
        cv::Mat image = srcImages[i].clone();
        cv::cvtColor(image, image, cv::COLOR_BGR2RGB);
        cv::resize(image, image, cv::Size(224, 224));
        image.convertTo(image, CV_32FC3, 1.0 / 255.0);
        cv::Scalar mean(0.485, 0.456, 0.406);
        cv::Scalar std(0.229, 0.224, 0.225);
        cv::subtract(image, mean, image);
        cv::divide(image, std, image);

        images.push_back(image);
    }

    // 图像转blob格式
    // cv::Mat blob = cv::dnn::blobFromImages(images);

    // 下边代码比上边blobFromImages速度更快
    for (int b = 0; b < BATCH_SIZE; b++)
    { 
        cv::Mat image = images[b];
        for (int r = 0; r < INPUT_H; r++)
        {
            float* rowData = image.ptr<float>(r);
            for (int c = 0; c < INPUT_W; c++)
            {
                mydata[b * CHANNEL * INPUT_H * INPUT_W + 0 * INPUT_H * INPUT_W + r * INPUT_W + c] = rowData[CHANNEL * c];
                mydata[b * CHANNEL * INPUT_H * INPUT_W + 1 * INPUT_H * INPUT_W + r * INPUT_W + c] = rowData[CHANNEL * c + 1];
                mydata[b * CHANNEL * INPUT_H * INPUT_W + 2 * INPUT_H * INPUT_W + r * INPUT_W + c] = rowData[CHANNEL * c + 2];
            }   
        }
    }
     
    auto end1 = std::chrono::high_resolution_clock::now();
    auto ms1 = std::chrono::duration_cast<std::chrono::microseconds>(end1 - start);
    std::cout << "PreProcess time: " << (ms1 / 1000.0).count() << "ms" << std::endl;

    // 执行推理
    doInference(*context, mydata, prob, BATCH_SIZE);

    auto end2 = std::chrono::high_resolution_clock::now();
    auto ms2 = std::chrono::duration_cast<std::chrono::microseconds>(end2 - end1);
    std::cout << "Inference time: " << (ms2 / 1000.0).count() << "ms" << std::endl;

    // 推理结果后处理
    cv::Mat result = cv::Mat(BATCH_SIZE, OUTPUT_SIZE, CV_32FC1, (float*)prob);
    softmax(result, result);

    for (int r = 0; r < BATCH_SIZE; r++)
    {
        cv::Mat scores = result.row(r).clone();

        cv::Point minLoc, maxLoc;
        double minValue = 0, maxValue = 0;
        cv::minMaxLoc(scores, &minValue, &maxValue, &minLoc, &maxLoc);
        int labelIndex = maxLoc.x;
        double probability = maxValue;

        classNames.push_back(classNameList[labelIndex]);
        confidences.push_back(probability);
    }

    auto end3 = std::chrono::high_resolution_clock::now();
    auto ms3 = std::chrono::duration_cast<std::chrono::microseconds>(end3 - end2);
    std::cout << "PostProcess time: " << (ms3 / 1000.0).count() << "ms" << std::endl;

    auto ms = std::chrono::duration_cast<std::chrono::microseconds>(end3 - start);
    std::cout << "TensorRT batch" << BATCH_SIZE << " 推理时间：" << (ms / 1000.0).count() << "ms" << std::endl;

    return true;
}

int main(int argc, char** argv)
{
    // onnx转engine
    // onnx_to_engine(onnxPath, enginePath, 0);

    // 模型初始化
    ModelInit(enginePath, 0);

    // 读取图像
    vector<string> filenames;
    cv::glob(imagePath, filenames);

    // 单图推理测试
    for (int n = 0; n < filenames.size(); n++)
    {
        // 重复100次，计算平均时间
        auto start = chrono::high_resolution_clock::now();
        cv::Mat src = imread(filenames[n]);
        std::string className;
        float confidence;
        for (int i = 0; i < 101; i++) {
            if (i == 1)
                start = chrono::high_resolution_clock::now();
            ModelInference(src, className, confidence);
        }
        auto end = chrono::high_resolution_clock::now();
        auto ms = chrono::duration_cast<std::chrono::microseconds>(end - start) / 100;
        std::cout << "TensorRT 平均推理时间：---------------------" << (ms / 1000.0).count() << "ms" << std::endl;
    }

    // 批量（动态batch）推理测试
    std::vector<cv::Mat> srcImages;
    int okNum = 0, ngNum = 0;
    for (int n = 0; n < filenames.size(); n++)
    {
        cv::Mat image = imread(filenames[n]);
        srcImages.push_back(image);
        if ((n + 1) % BATCH_SIZE == 0 || n == filenames.size() - 1)
        {
            // 重复100次，计算平均时间
            auto start = chrono::high_resolution_clock::now();
            for (int i = 0; i < 101; i++) {
                if (i == 1)
                    start = chrono::high_resolution_clock::now();
                std::vector<std::string> classNames;
                std::vector<float> confidences;
                ModelInference_Batch(srcImages, classNames, confidences);

                for (int j = 0; j < classNames.size(); j++)
                {
                    if (classNames[j] == "0")
                        okNum++;
                    else
                        ngNum++;
                }

            }
            srcImages.clear();
            auto end = chrono::high_resolution_clock::now();
            auto ms = chrono::duration_cast<std::chrono::microseconds>(end - start) / 100;
            std::cout << "TensorRT " << BATCH_SIZE << " 平均推理时间：---------------------" << (ms / 1000.0).count() << "ms" << std::endl;
        }        
    }

    GpuMemoryRelease();

    std::cout << "all_num = " << filenames.size() << endl << "okNum = " << okNum << endl << "ngNum = " << ngNum << endl;

    return 0;
}

5.3 fp32、fp16模型对比测试

fp16模型推理结果几乎和fp32一致，但是却较大的节约了显存和内存占用，同时推理速度也有明显的提升。

6. OpenVINO部署GoogLeNet

6.1 推理过程及代码

代码：

/* 推理过程
* 1. Create OpenVINO-Runtime Core
* 2. Compile Model
* 3. Create Inference Request
* 4. Set Inputs
* 5. Start Inference
* 6. Process inference Results
*/

#include <opencv2/opencv.hpp>
#include <openvino/openvino.hpp>
#include <inference_engine.hpp>
#include <chrono>
#include <fstream>

using namespace std;
using namespace InferenceEngine;
using namespace cv;

std::string onnxPath = "E:/inference-master/models/GoogLeNet/googlenet-pretrained_batch1.onnx";
std::string imagePath = "E:/inference-master/images/catdog";
std::string classNamesPath = "E:/inference-master/imagenet-classes.txt";			// 标签名称列表（类名）

ov::InferRequest inferRequest;
std::vector<std::string> classNameList;		// 标签名，可以从文件读取
int batchSize = 1;

// softmax，输入输出为数组
std::vector<float> softmax(std::vector<float> input)
{
	float total = 0;
	for (auto x : input)
		total += exp(x);
	std::vector<float> result;
	for (auto x : input)
		result.push_back(exp(x) / total);

	return result;
}

// softmax,输入输出为Mat
int softmax(const cv::Mat& src, cv::Mat& dst)
{
	float max = 0.0;
	float sum = 0.0;

	max = *max_element(src.begin<float>(), src.end<float>());
	cv::exp((src - max), dst);
	sum = cv::sum(dst)[0];
	dst /= sum;

	return 0;
}

// 模型初始化
void ModelInit(string onnxPath)
{
	// Step 1: 创建一个Core对象
	ov::Core core;

	// 打印当前设备
	std::vector<std::string> availableDevices = core.get_available_devices();
	for (int i = 0; i < availableDevices.size(); i++)
		printf("supported device name: %s\n", availableDevices[i].c_str());

	// Step 2: 读取模型
	std::shared_ptr<ov::Model> model = core.read_model(onnxPath);

	// Step 3: 加载模型到CPU
	ov::CompiledModel compiled_model = core.compile_model(model, "CPU");

	// 设置推理实例并发数为5个
	//core.set_property("CPU", ov::streams::num(10));
	// 设置推理实例数为自动分配
	//core.set_property("CPU", ov::streams::num(ov::streams::AUTO));
	// 推理实例数按计算资源平均分配
	//core.set_property("CPU", ov::streams::num(ov::streams::NUMA));

	// 设置推理实例的线程并发数为10
	// core.set_property("CPU", ov::inference_num_threads(20));

	// Step 4: 创建推理请求
	inferRequest = compiled_model.create_infer_request();

	// 读取标签名称
	ifstream fin(classNamesPath.c_str());
	string strLine;
	classNameList.clear();
	while (getline(fin, strLine))
		classNameList.push_back(strLine);
	fin.close();
}

// 单图推理
void ModelInference(cv::Mat srcImage, std::string& className, float& confidence )
{
	auto start = chrono::high_resolution_clock::now();

	// Step 5: 将输入数据填充到输入tensor
	// 通过索引获取输入tensor
	ov::Tensor input_tensor = inferRequest.get_input_tensor(0);
	// 通过名称获取输入tensor
	// ov::Tensor input_tensor = infer_request.get_tensor("input");

	// 预处理
	cv::Mat image = srcImage.clone();
	cv::cvtColor(image, image, cv::COLOR_BGR2RGB);
	resize(image, image, Size(224, 224));
	image.convertTo(image, CV_32FC3, 1.0 / 255.0);
	Scalar mean(0.485, 0.456, 0.406);
	Scalar std(0.229, 0.224, 0.225);
	subtract(image, mean, image);
	divide(image, std, image);

	// HWC -> NCHW
	ov::Shape tensor_shape = input_tensor.get_shape();
	const size_t channels = tensor_shape[1];
	const size_t height = tensor_shape[2];
	const size_t width = tensor_shape[3];
	float* image_data = input_tensor.data<float>();
	for (size_t r = 0; r < height; r++) {
		for (size_t c = 0; c < width * channels; c++) {
			int w = (r * width * channels + c) / channels;
			int mod = (r * width * channels + c) % channels;  // 0,1,2
			image_data[mod * width * height + w] = image.at<float>(r, c);
		}
	}

	// --------------- Step 6: Start inference ---------------
	inferRequest.infer();

	// --------------- Step 7: Process the inference results ---------------
	// model has only one output
	auto output_tensor = inferRequest.get_output_tensor();
	float* detection = (float*)output_tensor.data();
	ov::Shape out_shape = output_tensor.get_shape();
	int batch = output_tensor.get_shape()[0];
	int num_classes = output_tensor.get_shape()[1];

	cv::Mat result(batch, num_classes, CV_32F, detection);

	softmax(result, result);

	Point minLoc, maxLoc;
	double minValue = 0, maxValue = 0;
	cv::minMaxLoc(result, &minValue, &maxValue, &minLoc, &maxLoc);
	int labelIndex = maxLoc.x;
	double probability = maxValue;

	auto end = chrono::high_resolution_clock::now();
	auto ms = chrono::duration_cast<std::chrono::milliseconds>(end - start);
	std::cout << "openvino单张推理时间：" << ms.count() << "ms" << std::endl;
}

// 多图并行推理（动态batch）
void ModelInference_Batch(std::vector<cv::Mat> srcImages, std::vector<string>& classNames, std::vector<float>& confidences)
{
	auto start = chrono::high_resolution_clock::now();

	// Step 5: 将输入数据填充到输入tensor
	// 通过索引获取输入tensor
	ov::Tensor input_tensor = inferRequest.get_input_tensor(0);
	// 通过名称获取输入tensor
	// ov::Tensor input_tensor = infer_request.get_tensor("input");

	// 预处理（尺寸变换、通道变换、归一化）
	std::vector<cv::Mat> images;
	for (size_t i = 0; i < srcImages.size(); i++)
	{
		cv::Mat image = srcImages[i].clone();
		cv::cvtColor(image, image, cv::COLOR_BGR2RGB);
		cv::resize(image, image, cv::Size(224, 224));
		image.convertTo(image, CV_32FC3, 1.0 / 255.0);
		cv::Scalar mean(0.485, 0.456, 0.406);
		cv::Scalar std(0.229, 0.224, 0.225);
		cv::subtract(image, mean, image);
		cv::divide(image, std, image);

		images.push_back(image);
	}

	ov::Shape tensor_shape = input_tensor.get_shape();
	const size_t batch = tensor_shape[0];
	const size_t channels = tensor_shape[1];
	const size_t height = tensor_shape[2];
	const size_t width = tensor_shape[3];
	float* image_data = input_tensor.data<float>();

	// 图像转blob格式（速度比下边像素操作方式更快）
	cv::Mat blob = cv::dnn::blobFromImages(images);
	memcpy(image_data, blob.data, batch * 3 * height * width * sizeof(float));

	// NHWC -> NCHW
	//for (size_t b = 0; b < batch; b++){
	//	for (size_t r = 0; r < height; r++) {
	//		for (size_t c = 0; c < width * channels; c++) {
	//			int w = (r * width * channels + c) / channels;
	//			int mod = (r * width * channels + c) % channels;  // 0,1,2
	//			image_data[b * 3 * width * height + mod * width * height + w] = images[b].at<float>(r, c);
	//		}
	//	}
	//}

	auto end1 = std::chrono::high_resolution_clock::now();
	auto ms1 = std::chrono::duration_cast<std::chrono::microseconds>(end1 - start);
	std::cout << "PreProcess time: " << (ms1 / 1000.0).count() << "ms" << std::endl;
	
	// --------------- Step 6: Start inference ---------------
	inferRequest.infer();

	auto end2 = std::chrono::high_resolution_clock::now();
	auto ms2 = std::chrono::duration_cast<std::chrono::microseconds>(end2 - end1)/100;
	std::cout << "Inference time: " << (ms2 / 1000.0).count() << "ms" << std::endl;

	// --------------- Step 7: Process the inference results ---------------
	// model has only one output
	auto output_tensor = inferRequest.get_output_tensor();
	float* detection = (float*)output_tensor.data();
	ov::Shape out_shape = output_tensor.get_shape();
	int num_classes = output_tensor.get_shape()[1];

	cv::Mat output(batch, num_classes, CV_32F, detection);

	int rows = output.size[0];	// batch
	int cols = output.size[1];	// 类别数（每一个类别的得分）
	for (int row = 0; row < rows; row++)
	{
		cv::Mat scores(1, cols, CV_32FC1, output.ptr<float>(row));
		softmax(scores, scores);	// 结果归一化

		Point minLoc, maxLoc;
		double minValue = 0, maxValue = 0;
		cv::minMaxLoc(scores, &minValue, &maxValue, &minLoc, &maxLoc);
		int labelIndex = maxLoc.x;
		double probability = maxValue;

		classNames.push_back(classNameList[labelIndex]);
		confidences.push_back(probability);
	}

	auto end3 = std::chrono::high_resolution_clock::now();
	auto ms3 = std::chrono::duration_cast<std::chrono::microseconds>(end3 - end2);
	std::cout << "PostProcess time: " << (ms3 / 1000.0).count() << "ms" << std::endl;

	auto ms = chrono::duration_cast<std::chrono::milliseconds>(end3 - start);
	std::cout << "openvino单张推理时间：" << ms.count() << "ms" << std::endl;
}

int main(int argc, char** argv)
{
	// 模型初始化
	ModelInit(onnxPath);

	// 读取图像
	vector<string> filenames;
	glob(imagePath, filenames);

	// 单图推理测试
	for (int n = 0; n < filenames.size(); n++)
	{
		// 重复100次，计算平均时间
		auto start = chrono::high_resolution_clock::now();
		for (int i = 0; i < 101; i++) {
			if (i == 1)
				start = chrono::high_resolution_clock::now();
			cv::Mat src = imread(filenames[n]);
			std::string className;
			float confidence;
			ModelInference(src, className, confidence);
		}
		auto end = chrono::high_resolution_clock::now();
		auto ms = chrono::duration_cast<std::chrono::microseconds>(end - start) / 100.0;
		std::cout << "opencv_dnn 单图平均推理时间：---------------------" << (ms / 1000.0).count() << "ms" << std::endl;
	}

	std::vector<cv::Mat> srcImages;
	for (int i = 0; i < filenames.size(); i++)
	{
		cv::Mat image = imread(filenames[i]);
		srcImages.push_back(image);
		if ((i + 1) % batchSize == 0 || i == filenames.size() - 1)
		{
			// 重复100次，计算平均时间
			auto start = chrono::high_resolution_clock::now();
			for (int i = 0; i < 101; i++) {
				if (i == 1)
					start = chrono::high_resolution_clock::now();
				std::vector<std::string> classNames;
				std::vector<float> confidences;
				ModelInference_Batch(srcImages, classNames, confidences);
			}
			srcImages.clear();
			auto end = chrono::high_resolution_clock::now();
			auto ms = chrono::duration_cast<std::chrono::microseconds>(end - start) / 100;
			std::cout << "openvino batch" << batchSize << " 平均推理时间：---------------------" << (ms / 1000.0).count() << "ms" << std::endl;
		}
	}
	return 0;
}

注意：OV支持多图并行推理，但是要求转出onnx的时候batch就要使用固定数值。动态batch（即batch=-1）的onnx文件会报错。

6.2 遇到的问题

理论：OpenVINO是基于CPU推理最佳的方式。

实测：在测试OpenVINO的过程中，我们发现OpenVINO推理对于CPU的利用率远没有OpenCV DNN和ONNXRuntime高，这也是随着batch数量增加，OV在CPU上的推理速度反而不如DNN和ORT的主要原因。尝试过网上的多种优化方式，比如设置线程数并发数等等，未取得任何改善。如下图，在OpenVINO推理过程中，始终只有一半的CPU处于活跃状态；而OnnxRuntime或者OpenCV DNN推理时，所有的CPU均处于活跃状态。

7. 四种推理方式对比测试

以GoogLeNet模型为例，测得几种推理方式在推理部分的耗时如下：

基于CPU推理：

基于GPU推理：

不论采用何种推理方式，同一网络的前处理和后处理过程基本都是一致的。所以，为了更直观的对比几种推理方式的速度，我们抛去前后处理，只统计图中实际推理部分，即3、4、5这三个过程的执行时间。

同样是GoogLeNet网络，步骤3-5的执行时间对比如下：

注：OpenVINO-CPU测试中始终只使用了一半数量的内核，各种优化设置都没有改善。

最终结论：

GPU加速首选TensorRT；
CPU加速，单图推理首选OpenVINO，多图并行推理可选择ONNXRuntime;
如果需要兼具CPU和GPU推理功能，可选择ONNXRuntime。

参考资料

1. openvino2022版安装配置与C++SDK开发详解

2. https://github.com/NVIDIA/TensorRT

3. https://github.com/wang-xinyu/tensorrtx

4. 【TensorRT】TensorRT 部署Yolov5模型（C++）