onnxruntime c++部署分割网络

#include 
#include 
#include 
#include 
using namespace cv;


void PreProcess(const cv::Mat &image, cv::Mat &image_blob)
{
	int scale = cv::max(image.rows, image.cols);

	/*cv::Mat input = cv::Mat::zeros(scale, scale, image.type());
	cv::Rect roi(0,0, image.cols, image.rows);*/
	cv::Mat input;
	image.copyTo(input);

	cv::Mat inputs;
	cv::resize(input, inputs, cv::Size(256, 256));
	inputs.convertTo(inputs, CV_32F);
	//std::cout << "aaa"< channels, channel_p;
	split(inputs, channels);
	cv::Mat R, G, B;
	B = channels.at(0);
	G = channels.at(1);
	R = channels.at(2);

	B = (B / 255. - 0.406) / 0.225;
	G = (G / 255. - 0.456) / 0.224;
	R = (R / 255. - 0.485) / 0.229;

	channel_p.push_back(R);
	channel_p.push_back(G);
	channel_p.push_back(B);

	cv::Mat outt = cv::Mat(inputs.rows, inputs.cols, CV_32F);
	cv::merge(channel_p, outt);
	image_blob = outt;
}

int main(int argc, char* argv[]) {
  Ort::Env env(ORT_LOGGING_LEVEL_WARNING, "test");
  Ort::SessionOptions session_options;
  session_options.SetIntraOpNumThreads(1);
  
  session_options.SetGraphOptimizationLevel(GraphOptimizationLevel::ORT_ENABLE_EXTENDED);
#ifdef _WIN32
  //erfnet 模型
  const wchar_t* model_path = L"C:\\Users\\Desktop\\switch.onnx";
#else
  const char* model_path = "model.onnx";
#endif

  Ort::Session session(env, model_path, session_options);
  Ort::AllocatorWithDefaultOptions allocator;

  //获取输入name
  const char* inputName = session.GetInputName(0, allocator);
  //std::cout <<"xxxx:"<< inputName << std::endl;
  //获取输出name
  const char* outputName = session.GetOutputName(0, allocator);
  //std::cout << "Output Name: " << outputName << std::endl;
  查看模型定义的输入输出数
  //size_t num_input_nodes = session.GetInputCount();
  //size_t num_out_nodes = session.GetOutputCount();
  //std::cout << num_input_nodes << "  " << num_out_nodes << std::endl;
  std::vector input_node_names{ inputName };
  std::vector output_node_names{ outputName };
  std::vector inputTensors;
  std::vector outputTensors;
  
  //设置输入维度
  std::vector input_node_dims = {1,3,256,256};//(bitch_size,channel,height,width)
  size_t input_tensor_size = 256*256*3*1; 
  std::vector input_tensor_values(input_tensor_size);
  
  //加载图片
  Mat img = imread("C:\\Users\\Desktop\\onnx_test\\roi.jpg");
  //预处理
  Mat dst(256, 256, CV_8UC3);
  Mat dst2,dst3;
  resize(img, dst, Size(256, 256));
  //转换为float型
  dst.convertTo(dst2, CV_32F);
  //减均值除方差，注意B、R通道交换问题
  PreProcess(dst2,dst3);
  //将mat数据转成vector存储形式
  input_tensor_values.assign(dst3.begin(),dst3.end());

  //创建输入tensor
  auto memory_info = Ort::MemoryInfo::CreateCpu(OrtArenaAllocator, OrtMemTypeDefault);
  Ort::Value input_tensor = Ort::Value::CreateTensor(memory_info, input_tensor_values.data(), input_tensor_size, input_node_dims.data(), input_node_dims.size());
  assert(input_tensor.IsTensor());
  inputTensors.push_back(std::move(input_tensor));

  //设置输出维度
  std::vector out_node_dims = { 1,9,256,256 };//9=8+1：类+背景
  size_t out_tensor_size = 256 * 256 * 9*1;
  std::vector out_tensor_values(out_tensor_size);
  //创建输出tensor
  Ort::Value out_tensor = Ort::Value::CreateTensor(memory_info, out_tensor_values.data(), out_tensor_size, out_node_dims.data(), out_node_dims.size());
  assert(out_tensor.IsTensor());
  outputTensors.push_back(out_tensor);

  //推理
  session.Run(Ort::RunOptions{nullptr}, input_node_names.data(), inputTensors.data(), 1,output_node_names.data(), outputTensors.data(), 1);
  
  
  //解码推理数据
  cv::Mat out0 = Mat::zeros(256, 256, CV_8UC1);
  cv::Mat out1 = Mat::zeros(256, 256, CV_8UC1);
  cv::Mat out2 = Mat::zeros(256, 256, CV_8UC1);
  cv::Mat out3 = Mat::zeros(256, 256, CV_8UC1);
  cv::Mat out4 = Mat::zeros(256, 256, CV_8UC1);
  cv::Mat out5 = Mat::zeros(256, 256, CV_8UC1);
  cv::Mat out6 = Mat::zeros(256, 256, CV_8UC1);
  cv::Mat out7 = Mat::zeros(256, 256, CV_8UC1);
  cv::Mat out8 = Mat::zeros(256, 256, CV_8UC1);

  for (int i = 0; i < 256; i++)
  {
	  for (int j = 0; j < 256; j++)
	  {
		  float a0 = out_tensor_values[0 * 256 * 256 + i * 256 + j];
		  float a1 = out_tensor_values[1 * 256 * 256 + i * 256 + j];
		  float a2 = out_tensor_values[2 * 256 * 256 + i * 256 + j];
		  float a3 = out_tensor_values[3 * 256 * 256 + i * 256 + j];
		  float a4 = out_tensor_values[4 * 256 * 256 + i * 256 + j];
		  float a5 = out_tensor_values[5 * 256 * 256 + i * 256 + j];
		  float a6 = out_tensor_values[6 * 256 * 256 + i * 256 + j];
		  float a7 = out_tensor_values[7 * 256 * 256 + i * 256 + j];
		  float a8 = out_tensor_values[8 * 256 * 256 + i * 256 + j];
		  if (max(max(max(max(max(max(max(max(a0, a1), a2), a3), a4), a5), a6), a7), a8) == a0) out0.at(i, j) = 255;
		  if (max(max(max(max(max(max(max(max(a0, a1), a2), a3), a4), a5), a6), a7), a8) == a1) out1.at(i, j) = 255;
		  if (max(max(max(max(max(max(max(max(a0, a1), a2), a3), a4), a5), a6), a7), a8) == a2) out2.at(i, j) = 255;
		  if (max(max(max(max(max(max(max(max(a0, a1), a2), a3), a4), a5), a6), a7), a8) == a3) out3.at(i, j) = 255;
		  if (max(max(max(max(max(max(max(max(a0, a1), a2), a3), a4), a5), a6), a7), a8) == a4) out4.at(i, j) = 255;
		  if (max(max(max(max(max(max(max(max(a0, a1), a2), a3), a4), a5), a6), a7), a8) == a5) out5.at(i, j) = 255;
		  if (max(max(max(max(max(max(max(max(a0, a1), a2), a3), a4), a5), a6), a7), a8) == a6) out6.at(i, j) = 255;
		  if (max(max(max(max(max(max(max(max(a0, a1), a2), a3), a4), a5), a6), a7), a8) == a7) out7.at(i, j) = 255;
		  if (max(max(max(max(max(max(max(max(a0, a1), a2), a3), a4), a5), a6), a7), a8) == a8) out8.at(i, j) = 255;
	  }
  }
  //保存每个分割类别的mask
  imwrite("C:\\Users\\Desktop\\roi_0.jpg", out0);
  imwrite("C:\\Users\\Desktop\\roi_1.jpg", out1);
  imwrite("C:\\Users\\Desktop\\roi_2.jpg", out2);
  imwrite("C:\\Users\\Desktop\\roi_3.jpg", out3);
  imwrite("C:\\Users\\Desktop\\roi_4.jpg", out4);
  imwrite("C:\\Users\\Desktop\\roi_5.jpg", out5);
  imwrite("C:\\Users\\Desktop\\roi_6.jpg", out6);
  imwrite("C:\\Users\\Desktop\\roi_7.jpg", out7);
  imwrite("C:\\Users\\Desktop\\roi_8.jpg", out8);
  
  return 0;
}