libtorch+opencv

#include<opencv2/opencv.hpp>
#include <torch/torch.h>
#include <torch/script.h> 

int main(){	
	//定义使用cuda
	auto device = torch::Device(torch::kCPU);
	//读取图片
	auto image = cv::imread("D:/Data/20210730/5348/C3F/094025276_62385543.png");
	//缩放至指定大小
	cv::resize(image, image, cv::Size(224, 56));
	//转成张量
	auto input_tensor = torch::from_blob(image.data, { image.rows, image.cols, 3 }, torch::kByte).permute({ 2, 0, 1 }).unsqueeze(0).to(torch::kFloat32) / 225.0;
	//加载模型
	auto model = torch::jit::load("./resNet50.pt");
	model.to(device);
	model.eval();
	//前向传播
	auto output = model.forward({ input_tensor.to(device) }).toTensor();
	output = torch::softmax(output, 1);
	std::cout << "模型预测结果为第" << torch::argmax(output) << "类，置信度为" << output.max() << std::endl;
	return 0;
}

#include<iostream>
#include <exception> 
#include<opencv2/opencv.hpp>
#include <torch/torch.h>
#include <torch/script.h> 
using namespace std;
using namespace cv;
void getAllImagePath(string imgDir, vector<string>& imgPaths, vector<string>& imgNames) {
	cv::glob(imgDir, imgPaths, true);
	for (int i = 0; i < imgPaths.size(); ++i)
	{
		//1.获取不带路径的文件名,000001.jpg
		string::size_type iPos = imgPaths[i].find_last_of('/') + 1;
		string filename = imgPaths[i].substr(iPos, imgPaths[i].length() - iPos);
		imgNames.emplace_back(filename);
		//cout << filename << endl;
		//2.获取不带后缀的文件名,000001
		string name = filename.substr(0, filename.rfind("."));
		//cout << name << endl;
	}
}
void mat2TensorTransforms(cv::Mat img, torch::Tensor& tensor_image) {
	//1.OpenCV读取的都是BGR格式，通常需要转成RGB的
	cv::cvtColor(img, img, cv::COLOR_BGR2RGB);
	//2.缩放到模型需要的输入尺寸
	cv::Mat image;
	cv::resize(img, image, cv::Size(224, 56));
	//3.将Mat数据转为Tensor张量数据，同时增加了1维
	std::vector<int64_t> sizes = { 1,image.rows, image.cols,3 };
	torch::TensorOptions option(torch::kByte);
    tensor_image = torch::from_blob(image.data, torch::IntList(sizes), option);
	//4.维度换位，来满足模型的输入，就是1X256X256X3 变成 1X3X256X256
	tensor_image = tensor_image.permute({ 0, 3, 1, 2 });
	//5.Normalization
	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);
}
void transforms(torch::Tensor &tensor_image0, torch::Tensor &tensor_image1) {
	// Normalization
	torch::Tensor tensor_image;
	tensor_image = tensor_image0.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);
	tensor_image1 = tensor_image;
}
void getDeviceType(torch::DeviceType& device_type) {
	if (torch::cuda::is_available()) {
		std::cout << "CUDA available! Predicting on GPU." << std::endl;
		device_type = torch::kCUDA;
	}
	else {
		std::cout << "Predicting on CPU." << std::endl;
		device_type = torch::kCPU;
	}
}
int main() {
	ifstream iFile("./path.txt");
	string line;
	if (iFile) {
		while (getline(iFile, line)) {
			cout << line << endl;
		}
	}
	vector<std::string> className = { "B","C" }; //标签输入
	torch::DeviceType device_type;
	getDeviceType(device_type);
	torch::Device device(device_type);
	//加载模型
	auto model = torch::jit::load("./resNet50.pt");
	model.to(device); 
	model.eval();
	//读取图片
	string imgDir = line;
	vector<string> filePaths, fileNames;
	getAllImagePath(imgDir, filePaths, fileNames);
	for (int i = 0; i < filePaths.size(); i++) {
		cv::Mat img = cv::imread(filePaths[i]);
		if (img.empty()) {
			cout << "img is not exist!" << endl;
		}
		cout << fileNames[i] << endl;
		try {
			clock_t time_start = clock();
			torch::Tensor input_tensor;
			mat2TensorTransforms(img, input_tensor);
			//transforms(input_tensor, input_tensor);
			//预测推理
			torch::Tensor result = model.forward({ input_tensor.to(device) }).toTensor();
			//Softmax
			result = torch::softmax(result, 1);
			//预测值		
			int iPreValue = torch::argmax(result).item<int>();
			//置信度
			float fMaxConfidence = result.max().item<float>();
			float fMinConfidence = result.min().item<float>();
			clock_t time_stop = clock();
			cout << "infer time is:" << 1000 * (time_stop - time_start) / (double)CLOCKS_PER_SEC << "ms" << endl;
			cout << "模型预测结果为第" << className[iPreValue] << "类，置信度为" << fMaxConfidence << std::endl;
		}
		catch (const char* msg) {
			cerr << msg << endl;
		}
	}
	return 0;
}