环境说明

win10
VS2019
OPENCV4.7.0
Litorch1.13
Pytorch 1.12.1

Pytorch 序列化

import torch
from torchvision.models import resnet50
net = resnet50(pretrained=True)
net = net.cuda()
net.eval()

# trace
x = torch.ones(1, 3, 224, 224)
x = x.cuda()
traced_module = torch.jit.trace(net, x)
traced_module.save("resnet50_trace.pt")

Libtorch 下载

[libtorch-win-shared-with-deps-1.13.0+cu117.zip ]
通过网盘分享的文件：libtorch-win-shared-with-deps-1.13.0+cu117.zip
链接: https://pan.baidu.com/s/1NY25r9HnHXhy4kzIn-jlnQ 提取码: v5sq

解压后添加到环境变量里

D:\envPath\libtorch_11.3\libtorch\lib

VS配置

项目右键，选择属性

配置选择Release, 平台选择x64

VC++目录：

D:\envPath\libtorch_11.3\libtorch\include
D:\envPath\libtorch_11.3\libtorch\include\torch\csrc\api\include
C:\opencv\build\include\opencv2
C:\opencv\build\include\

库目录

C:\opencv\build\x64\vc16\lib
D:\envPath\libtorch_11.3\libtorch\lib

链接器-输入-附加依赖项

asmjit.lib
c10.lib
c10_cuda.lib
caffe2_detectron_ops_gpu.lib
caffe2_module_test_dynamic.lib
caffe2_nvrtc.lib
clog.lib
cpuinfo.lib
dnnl.lib
fbgemm.lib
libprotobuf.lib
libprotobuf-lite.lib
libprotoc.lib
mkldnn.lib
torch.lib
torch_cpu.lib
torch_cuda.lib
opencv_world470.lib

链接器-输入-命令行
其他选项添加

/INCLUDE:?ignore_this_library_placeholder@@YAHXZ
 

如图：

主程序

int main(int argc, char** argv)
{
	//Pro_info();
	if (argc < 1) return -1;
	std::string image_path = argv[1];
	//std::string model_path = argv[2];

	// 定义设备类型
	torch::DeviceType* deviceType = new torch::DeviceType();
	if (torch::cuda::is_available())
	{
		*deviceType = torch::kCUDA;
		std::cout << "The cuda is available" << std::endl;
	}
	else
	{
		*deviceType = torch::kCPU;
		std::cout << "The cuda isn't available" << std::endl;
	}

	torch::Device device(*deviceType);
	std::cout << *deviceType << std::endl;

	//加载模型
	using torch::jit::script::Module;
	Module module;
	try
	{
		
		module = torch::jit::load("D:/traced_resnet_model.pt");
		printf("The model load success!\n");


	}
	catch (std::exception& e)
	{
		std::cout << e.what() << std::endl;
		std::cerr << "error loading the model\n";
	}

	
	module.to(device);
	double total_time = 0;
	for (int i = 0; i < 100; i++)
	{
		auto start = std::chrono::system_clock::now();
		cv::Mat img = cv::imread(image_path);

		if (img.empty())
		{
			std::cerr << "The image can't open!\n";
			return -1;
		}

		//图像预处理
		cv::Mat input;
		cv::resize(img, img, cv::Size(input_shape, input_shape));//图片resize成512*512*3
		cv::cvtColor(img, input, cv::COLOR_BGR2RGB);
		//from_blob Mat转Tensor {batchsize,w,h,channles}
		torch::Tensor tensor_image = torch::from_blob(input.data, { 1,input.rows, input.cols,3 }, torch::kByte);
		//shape->(batchsize,channles,w,h)
		tensor_image = tensor_image.permute({ 0,3,1,2 });
		tensor_image = tensor_image.toType(torch::kFloat);
		//image/255.0图像的归一化处理
		tensor_image = tensor_image.div(255);
		tensor_image[0][0] = tensor_image[0][0].sub(0.485).div(0.229);
		tensor_image[0][1] = tensor_image[0][1].sub(0.456).div(0.224);
		tensor_image[0][2] = tensor_image[0][2].sub(0.406).div(0.225);
		auto img_var = torch::autograd::make_variable(tensor_image, false);
		std::vector<torch::jit::IValue> inputs;
		inputs.push_back(img_var.to(device));

		// 网络前向计算
		torch::NoGradGuard no_grad;
		auto output = module.forward({ inputs }).toTensor();
		output = torch::squeeze(torch::argmax(torch::softmax(output, 1), 1), 0);
		//std::cout << output.sizes() << std::endl;
		output = output.to(torch::kU8).to(torch::kCPU);
		//将tensor转为cv::Mat格式，进行展示
		cv::Mat Img(output.sizes()[0], output.sizes()[1], CV_8U, output.data_ptr());
		Img = Img * 255;
		cv::imshow("result", Img);
		cv::waitKey(0);
		return 0;
		auto end = std::chrono::system_clock::now();
		double spend_time = std::chrono::duration_cast<std::chrono::milliseconds>(end - start).count();
		std::cout << spend_time << "ms" << std::endl;
		total_time = total_time + spend_time;

	}

可能遇到的问题

1. pytorch序列化过程要保证不出现警告，否则会导致后续libtorch调用模型出错
2. 环境方面要求libtorch版本要高于pytorch，否则会出现c10 error
3. 不同图像分割方法的预处理和后处理可能不同，需要重新使用C++实现。
4. 运行时如果无法定位程序输入点于动态链接库，可以将libtorch下的dll文件复制到exe对应目录下
5. 显卡没有调用的话，在链接器-命令行添加/INCLUDE:?ignore_this_library_placeholder@@YAHXZ，但不同版本添加的语句不一样，具体参考Terod的总结
6. 'torch/torch.h’头文件提示不存在的话，环境配置中确认“\libtorch\include\torch\csrc\api\include”是否添加
7. 出现其他问题，可以添加以下代码，打印出错误原因

    try
    {
       // your code goes here...
    }
    catch (std::exception &e)
    {
        std::cout << e.what() << std::endl;
    }

参考

1. 经典图像分割网络：Unet 支持libtorch部署推理【附代码】

2. Windows下将PyTorch模型转成Libtorch并使用C++进行读取

3. Windows下PyTorch(LibTorch)配置cuda加速

如有其他问题，欢迎交流！