tensorflow2.5.0-GPU测试工程

/* * linkuang.cpp * * Created on: 2020年4月20日 * Author: root */ #include "linkuang.h" using namespace std; using namespace cv; /*initialize CNN model in front of the system begin * input arg:imgsize-----------size of standard input image in tensorflow model * wastethresh_prob--the second time judge the waste. * If the CNN inference prob for waste < wastethresh_prob, * we consider it as ore,not waste; * */ /*CNN的初始化部分：在系统初始化时调用一次即可。 * 输入：imgsize-----------CNN模型要求的图像大小，不能太大(<=100)。建议：32、48、64 * lesswaste_prob----对废石的二次判断阈值(0.0,0.5~1)。如果CNN结果判为废石，且概率<=lesswaste_prob, * 那么我们认为这不是真正的废石，其实是矿石。 * lesswaste_prob==0.0,即表示不做阈值判断，完全相信CNN废石的结果; * lesswaste_prob越大，则抛废率越小; * morewaste_prob----对矿石的二次判断阈值(0.0,0.5~1)。如果CNN结果为矿石，且概率<=morewaste_prob， * 那么我们认为这不是真正的矿石，其实是废石。 * morewaste_prob=0.0，即表示不做阈值判断，完全相信CNN矿石的结果; * morewaste_prob越大，则抛废率越大; * */ linkuang::linkuang(int imgsize,float lesswaste_prob,float morewaste_prob) { // TODO Auto-generated constructor stub standard_rows=imgsize;//48; standard_cols=imgsize;//64 lesswastethresh=lesswaste_prob; morewastethresh=morewaste_prob; getimgsindex=0; showtestindex=0; std::string graphpath="/home/jumper/xrt/reference/cnnmodel/my_unet_shitou_youhua5.pb"; // std::string modelpath="/home/jumper/xrt/reference/model/model_jinhuo0.7/model"; ///////CNN initiation-- tensorflow::Status status = NewSession(tensorflow::SessionOptions(), &session); if (!status.ok()) { throw std::runtime_error("ERROR: linkuang CNN NewSession() init failed..."); } tensorflow::GraphDef graphdef; tensorflow::Status status_load = ReadBinaryProto(tensorflow::Env::Default(), graphpath, &graphdef); if (!status_load.ok()) { std::cout << status_load.ToString() <<std::endl; throw std::runtime_error("ERROR: Loading model failed..."); } tensorflow::Status status_create = session->Create(graphdef); if (!status_create.ok()) { std::cerr <<status_create.ToString() << std::endl; throw std::runtime_error("ERROR: Creating graph in session failed..."); } } /*单个石头CNN预测 *输入:highlittleimg------单独石头的高能小图,CV_16UC1,背景是原图 * lowlittleimg-------单独石头的低能小图，CV_16UC1,必须是已经与高能对齐的小图,背景是0 *输出:label---表示此石头的类别 * 0:废石; * 1:矿石; *return :-1:模型预测失败 * -2:预测成功，但比较概率得到类别失败 * 0 ：成功 * */ int linkuang::linkuangSingleInference(cv::Mat &cnncolorimg,cv::Mat &outputimg) { Mat standardinput(standard_rows, standard_cols, CV_8UC1); resize(cnncolorimg,standardinput,Size(standard_cols,standard_rows),0,0,1); //CNN start...和训练做相同的预处理 tensorflow::Tensor resized_tensor(tensorflow::DT_FLOAT, tensorflow::TensorShape({1,standard_rows,standard_cols,1})); // float *imgdata = resized_tensor.flat<float>().data(); // Mat cnninputImg(standard_rows, standard_cols, CV_32FC1, imgdata); // standardinput.convertTo(cnninputImg,CV_32FC1); // cnninputImg=cnninputImg/255; auto outputMap =resized_tensor.tensor<float,4>();//获取tensor指针,注意这里outputMap是Eigen::tensor类型 for(int r=0;r<standard_rows;r++)//遍历行数 { for(int c=0;c<standard_cols;c++)//遍历列数 { outputMap(0,r,c,0)=float(standardinput.ptr<uchar>(r)[c])/255; } } //CNN input std::vector<std::pair<std::string, tensorflow::Tensor> > inputs; std::string Input1Name = "input_1_1";//"input_1_1:0"; inputs.push_back(std::make_pair(Input1Name, resized_tensor)); //CNN predict std::vector<tensorflow::Tensor> outputs; std::string output="output_1";//"output_1:0"; tensorflow::Status status_run = session->Run({{Input1Name,resized_tensor}}, {output}, {}, &outputs); if (!status_run.ok()) { std::cout <<"ERROR: RUN failed in real inference()..."<< status_run.ToString() << "\n"; return -1; } int flag=getOutputImg(outputs[0],outputimg); if(flag!=0) { std::cout <<"ERROR: RUN failed in getCnnRealLabel()..."<<std::endl; return -2; } showtestindex++; return 0; } int linkuang::getOutputImg(tensorflow::Tensor &output,cv::Mat &outputimg) { memcpy(outputimg.data, output.tensor_data().data(), output.tensor_data().size()); // for(int r=0;r!=standard_rows;r++) // { // for(int c=0;c!=standard_cols;c++) // { // cout<<outputimg.ptr<float>(r)[c]<<"\t"; // } // cout<<endl; // } // cout<<endl<<endl; return 0; } /* int linkuang::getOutputImg(tensorflow::Tensor &output,cv::Mat &outputimg) { int ndim2 = output.shape().dims(); auto tmap = output.tensor<float, 3>(); //<float,ndim2> cout<<tmap(2000,2000,2000)<<endl;//why not limited in [standard_rows,standard_cols,1] as python ?? for(int r=0;r!=standard_rows;r++) { for(int c=0;c!=standard_cols;c++) { float value=tmap(r,c,0); cout<<value<<"\t"; outputimg.ptr<uchar>(r)[c]=255*value; } cout<<endl; } cout<<endl<<endl; return 0; } */ //得到真正的标签 int linkuang::getCnnRealLabel(tensorflow::Tensor &probabilities,int &output_class_id,float &output_prob) { int ndim2 = probabilities.shape().dims(); // Get the dimension of the tensor auto tmap = probabilities.tensor<float, 2>(); // Tensor Shape: [batch_size, target_class_num] int output_dim = probabilities.shape().dim_size(1); // Get the target_class_num from 1st dimension output_class_id=0; float primerprob=tmap(0, 0); if(tmap(0, 1)>primerprob) { primerprob=tmap(0, 1); output_class_id=1; } output_prob=primerprob; return 0; } linkuang::~linkuang() { // TODO Auto-generated destructor stub tensorflow::Status freestatus=session->Close(); if (!freestatus.ok()) { throw std::runtime_error("ERROR: close session..."); } }