机器视觉系列5：C++部署pytorch模型onnxruntime

 系列文章目录

第一章：Visual Studio 2019 动态链接库DLL建立

第二章：VS动态链接库DLL调试

第三章：VS2019 OpenCV环境配置 

第四章：C++部署pytorch模型Libtorch

第五章：C++部署pytorch模型onnxruntime

---

目录

 系列文章目录

前言

一、C++部署pytorch？

二、onnxruntime配置

1.下载onnxruntime

2.VS2019配置onnxruntime

2.1配置VC++目录

 2.2配置链接器

 2.3 onnxruntime环境变量配置

三、pytorch模型转换为onnx

四、C++中onnxruntime的使用

五、python中onnxruntime的使用

参考文献

---

前言

环境：visual studio 2019；OpenCV4.5.5；pytorch1.8；onnxruntime1.8.1；

一、C++部署pytorch？

pytorch模型在C++部署，上一章是使用pytorch对应版本的Libtorch部署。其实转onnx部署可能更方便，之前语义分割精度相差太大是因为数据预处理的问题，一般图像在输入网络之前需要标准化和转RGB等相关操作。onnx部署的好处是微软中间格式，兼容各种平台，比较方便。如果部署平台有GPU可能用tensorRT会更好。

二、onnxruntime配置

注意事项：注意pt模型转onnx时的版本，onnxruntime版本对应

1.下载onnxruntime

官网：NuGet Gallery | Home

onnxruntime-win-x64-1.8.1链接如下：

https://objects.githubusercontent.com/github-production-release-asset-2e65be/156939672/7bc42680-deed-11eb-89d9-81963670045b?X-Amz-Algorithm=AWS4-HMAC-SHA256&X-Amz-Credential=AKIAIWNJYAX4CSVEH53A%2F20220809%2Fus-east-1%2Fs3%2Faws4_request&X-Amz-Date=20220809T093837Z&X-Amz-Expires=300&X-Amz-Signature=4c9cc7ec4ba93d83ccb3307533b47e5c40303c43c003b2a791d4d321570dc5b8&X-Amz-SignedHeaders=host&actor_id=0&key_id=0&repo_id=156939672&response-content-disposition=attachment%3B%20filename%3Donnxruntime-win-x64-1.8.1.zip&response-content-type=application%2Foctet-streamhttps://objects.githubusercontent.com/github-production-release-asset-2e65be/156939672/7bc42680-deed-11eb-89d9-81963670045b?X-Amz-Algorithm=AWS4-HMAC-SHA256&X-Amz-Credential=AKIAIWNJYAX4CSVEH53A%2F20220809%2Fus-east-1%2Fs3%2Faws4_request&X-Amz-Date=20220809T093837Z&X-Amz-Expires=300&X-Amz-Signature=4c9cc7ec4ba93d83ccb3307533b47e5c40303c43c003b2a791d4d321570dc5b8&X-Amz-SignedHeaders=host&actor_id=0&key_id=0&repo_id=156939672&response-content-disposition=attachment%3B%20filename%3Donnxruntime-win-x64-1.8.1.zip&response-content-type=application%2Foctet-stream

2.VS2019配置onnxruntime

2.1配置VC++目录

首先配置包含目录和库目录，对应opencv一样的方法。

 2.2配置链接器

依赖项添加所有lib，cmd中进入lib目录，使用dir /b *.lib>1.txt命令可生成目录，复制使用。

 2.3 onnxruntime环境变量配置

可以在系统环境变量中添加，或直接把所有DLL复制进Release或者Debug目录就不用配置环境了。

 

三、pytorch模型转换为onnx

注意事项：模型输入BCHW，推理模式model.eval()，输出版本opset_version=11

import torch

x = torch.randn(1, 3, 512, 512, device="cpu")
model = torch.load('best_model.pth', map_location=torch.device('cpu'))
model.eval()
input_names = ["input"]
output_names = ["output"]

torch.onnx.export(model, x, "GlandUnet.onnx", verbose=True, input_names=input_names, output_names=output_names, opset_version=11)

四、C++中onnxruntime的使用

注意事项：1，一定要注意图像输入模型之前的预处理，是否标准化，是否转RGB

                  2，C++中onnxruntime推理输出为数组的首地址，可以用指针取出生成Mat

/****************************************
@brief     :  分割onnxruntime
@input	   :  图像
@output    :  掩膜
*****************************************/
void SegmentAIONNX(Mat& imgSrc, int width, int height)
{
	模型信息/
	Ort::Env env(ORT_LOGGING_LEVEL_WARNING, "OnnxModel");
	Ort::SessionOptions session_options;
	session_options.SetIntraOpNumThreads(1);
#ifdef _WIN32
	const wchar_t* model_path = L"GlandUnet.onnx";
#else
	const char* model_path = "RedUnet.onnx";
#endif
	Ort::Session session(env, model_path, session_options);
	Ort::AllocatorWithDefaultOptions allocator;
	size_t num_input_nodes = session.GetInputCount();	//batchsize
	size_t num_output_nodes = session.GetOutputCount();
	const char* input_name = session.GetInputName(0, allocator);
	const char* output_name = session.GetOutputName(0, allocator);
	auto input_dims = session.GetInputTypeInfo(0).GetTensorTypeAndShapeInfo().GetShape();	//输入输出维度
	auto output_dims = session.GetOutputTypeInfo(0).GetTensorTypeAndShapeInfo().GetShape();
	std::vector input_names{ input_name };
	std::vector output_names = { output_name };
	输入处理//
	Mat imgBGR = imgSrc;	//输入图片预处理
	Mat imgBGRresize;
	resize(imgBGR, imgBGRresize, Size(input_dims[3], input_dims[2]), InterpolationFlags::INTER_CUBIC);
	Mat imgRGBresize = imgBGRresize;
	//cvtColor(imgBGRresize, imgRGBresize, COLOR_BGR2RGB);	//smp未转RGB
	Mat resize_img;
	imgRGBresize.convertTo(resize_img, CV_32F, 1.0 / 255);  //divided by 255转float
	cv::Mat channels[3]; //分离通道进行HWC->CHW
	cv::split(resize_img, channels);
	std::vector inputTensorValues;
	float mean[] = { 0.485f, 0.456f, 0.406f };	//
	float std_val[] = { 0.229f, 0.224f, 0.225f };
	for (int i = 0; i < resize_img.channels(); i++)	//标准化ImageNet
	{
		channels[i] -= mean[i];  // mean均值
		channels[i] /= std_val[i];   // std方差
	}
	for (int i = 0; i < resize_img.channels(); i++)  //HWC->CHW
	{
		std::vector data = std::vector(channels[i].reshape(1, resize_img.cols * resize_img.rows));
		inputTensorValues.insert(inputTensorValues.end(), data.begin(), data.end());
	}
	Ort::MemoryInfo memoryInfo = Ort::MemoryInfo::CreateCpu(OrtAllocatorType::OrtArenaAllocator, OrtMemType::OrtMemTypeDefault);
	vector inputTensors;
	inputTensors.push_back(Ort::Value::CreateTensor(memoryInfo, inputTensorValues.data(), inputTensorValues.size(), input_dims.data(), input_dims.size()));
	//clock_t startTime, endTime;	//计算推理时间
	//startTime = clock();
	auto outputTensor = session.Run(Ort::RunOptions{ nullptr }, input_names.data(), inputTensors.data(), 1, output_names.data(), 1);   // 开始推理
	//endTime = clock();
	打印模型信息/
	//printf("Using Onnxruntime C++ API\n");
	//printf("Number of inputs = %zu\n", num_input_nodes);
	//printf("Number of output = %zu\n", num_output_nodes);
	//std::cout << "input_name:" << input_name << std::endl;
	//std::cout << "output_name: " << output_name << std::endl;
	//std::cout << "input_dims:" << input_dims[0] << input_dims[1] << input_dims[2] << input_dims[3] << std::endl;
	//std::cout << "output_dims:" << output_dims[0] << output_dims[1] << output_dims[2] << output_dims[3] << std::endl;
	//std::cout << "The run time is:" << (double)(endTime - startTime) / CLOCKS_PER_SEC << "s" << std::endl;
	
	//输出处理//
	float* mask_ptr = outputTensor[0].GetTensorMutableData(); //outtensor首地址
	vector< unsigned char >results(512 * 512);
	for (int i = 0; i < 512 * 512; i++)
	{
		if (mask_ptr[i] >= 0.5)
		{
			results[i] = 0;
		}
		else
		{
			results[i] = 255;
		}
	}
	unsigned char* ptr = &results[0];
	Mat mask = Mat(output_dims[2], output_dims[3], CV_8U, ptr);
	resize(mask, imgSrc, Size(imgBGR.cols, imgBGR.rows));
	//原图展示分割结果//
	//cvtColor(imgSrc, imgSrc, COLOR_GRAY2BGR);
	//Mat imgAdd;
	//addWeighted(imgBGR, 1, imgSrc, 0.3, 0, imgAdd);
}

五、python中onnxruntime的使用

# -*- coding:utf-8 -*-
import cv2
import numpy as np
import onnxruntime as ort
import imgviz
import time

class_names = ['_background_', 'conjunctiva_area']
### 定义一些数据前后处理的工具
def preprocess(input_data):
    # convert the input data into the float32 input
    img_data = input_data.astype('float32')
    # normalize
    mean_vec = np.array([0.485, 0.456, 0.406])
    stddev_vec = np.array([0.229, 0.224, 0.225])
    norm_img_data = np.zeros(img_data.shape).astype('float32')
    for i in range(img_data.shape[0]):
        norm_img_data[i, :, :] = (img_data[i, :, :] / 255 - mean_vec[i]) / stddev_vec[i]
    # add batch channel
    norm_img_data = norm_img_data.reshape(1, 3, 512, 512).astype('float32')
    return norm_img_data

def softmax(x):
    x = x.reshape(-1)
    e_x = np.exp(x - np.max(x))
    return e_x / e_x.sum(axis=0)

def postprocess(result):
    return softmax(np.array(result)).tolist()


session = ort.InferenceSession('GlandUnet.onnx')
img0 = cv2.imread('test.bmp')
h0, w0 = img0.shape[0:2]
img = cv2.cvtColor(img0, cv2.COLOR_BGR2RGB)
img = cv2.resize(img, [512, 512])
image_data = np.array(img).transpose(2, 0, 1) # HWC->CHW
input_data = preprocess(image_data)
time_start = time.time()  # 记录开始时间
raw_result = session.run([], {'input': input_data})
time_end = time.time()  # 记录结束时间
time_sum = time_end - time_start  # 计算的时间差为程序的执行时间，单位为秒/s
print(time_sum)
# label_result = np.argmax(raw_result, dim=1)  # 缺argmax
# out = np.squeeze(raw_result)
# result_img = np.array(out, dtype=np.uint8)
# result_img = cv2.resize(result_img, (w0, h0))

out1 = raw_result[0][0]

cv2.imshow('2', out1[1])
cv2.waitKey(0)

---

参考文献

https://onnxruntime.ai/

https://github.com/leimao/ONNX-Runtime-Inference/blob/main/src/inference.cpp

神经网络语义分割模型C++部署(VS2019+ONNXRuntime+OpenCV)_Shijunfeng00的博客-CSDN博客

opencv 图片HWC格式转CHW格式_wuqingshan2010的博客-CSDN博客