接上一篇文章，上一篇文章簡單了訓練一個101層的殘差網路並保持為pb格式，下面在c++中讀取圖片進行預測，用Mat轉化為tensor的形式，opencv介面安裝、tensorflow編譯在前面的文章都有說到在這裡不再累贅，進模型之前圖片預處理較簡單，只是做了個去均值操作，灰度什麼的沒做處理，下面看下程式碼:

#include <iostream>
#include <opencv2/opencv.hpp>
#include "tensorflow/cc/saved_model/loader.h"
#include "tensorflow/core/framework/graph.pb.h"
#include "tensorflow/core/protobuf/meta_graph.pb.h"
#include  "tensorflow/cc/saved_model/tag_constants.h"


using namespace std;
using  namespace cv;
using  namespace  tensorflow;

//參考 https://stackoverflow.com/questions/39379747/import-opencv-mat-into-c-tensorflow-without-copying
void  cvmat_to_tensor(cv::Mat  img,Tensor* tensor,int rows,int cols){
    cv::resize(img,img,cv::Size(rows,cols));

    img=img-128.0;

    float *p=tensor->flat<float>().data();

    cv::Mat imagePixels(rows,cols,CV_32FC1,p);
    img.convertTo(imagePixels,CV_32FC1);


}


int main(int argc, char *argv[]) {

    int rows=224;
    int cols=224;
    int chanel=3;

    string modelpath="/Users/zhoumeixu/Documents/python/credit-transform/model/2";
    string picturepath="/Users/zhoumeixu/Desktop/影象測試/美女.jpg";

    string inputname="input";

    string outputname="output";


    cv::Mat img = cv::imread(picturepath);


    tensorflow::SessionOptions sess_options;
    tensorflow::RunOptions run_options;
    tensorflow::SavedModelBundle bundle;
    Status status;

    status =tensorflow::LoadSavedModel(sess_options, run_options, modelpath, {tensorflow::kSavedModelTagServe}, &bundle);

    if(!status.ok()){
        cout<<status.ToString()<<endl;
    }

    tensorflow::MetaGraphDef graph_def = bundle.meta_graph_def;
    std::unique_ptr<tensorflow::Session>& session = bundle.session;

    Tensor x(tensorflow::DT_FLOAT, tensorflow::TensorShape({rows, cols,chanel}));


    cvmat_to_tensor(img,&x,rows,cols);



    std::vector<std::pair<string, tensorflow::Tensor>> inputs;
    inputs.push_back(std::pair<std::string, tensorflow::Tensor>(inputname, x));


    Tensor tensor_out(tensorflow::DT_FLOAT, TensorShape({1,20}));
    std::vector<tensorflow::Tensor> outputs={{ tensor_out }};
    status= session->Run(inputs, {outputname}, {}, &outputs);
    if (!status.ok()) {
        cout<<"failure"<<endl;
        std::cout << status.ToString() << "\n";
        return 1;
    }

    cout<<"ok...."<<endl;

    cout << "Output tensor size:" << outputs.size() << std::endl;
    cout << outputs[0].DebugString()<<endl;


    Tensor t = outputs[0];

    int output_dim = t.shape().dim_size(0);

    cout<<output_dim<<endl;
    auto tmap = t.tensor<float, 1>();

    for(int i=0;i<output_dim;++i){
        cout<<i<<"-->"<<tmap(i)<<" ";
    }
    cout<<endl;


    return 0;
}




 

結果:

ok....
Output tensor size:1
Tensor<type: float shape: [20] values: 0.0127011631 0.0497760847 0.00124033471...>
20
0-->0.0127012 1-->0.0497761 2-->0.00124033 3-->0.0369612 4-->0.0458243 5-->0.0124642 6-->0.00162883 7-->0.00522397 8-->0.0574596 9-->0.130111 10-->0.0435387 11-->0.015093 12-->0.00260962 13-->0.024438 14-->0.00970128 15-->0.0859285 16-->0.168278 17-->0.0681312 18-->0.0497715 19-->0.17912