0

Here https://stackoverflow.com/a/49817506/1277317 there is an example of how to use a convolution network in OpenCV. But this example is in Python. How to do the same in C++? Namely, how to do this in C++: 

net = cv.dnn.readNetFromTensorflow('model.pb')
net.setInput(inp.transpose(0, 3, 1, 2))
cv_out = net.forward()

?

And how to create Mat for the setInput function for an image size: 60x162x1? I use float for the data just like in the python example. Now I have this code and it gives incorrect results:

Net net = readNet("e://xor.pb");

float x0[60][162];

for(int i=0;i<60;i++)
{
    for(int j=0;j<162;j++)
    {
        x0[i][j]=0;
    }
}
x0[5][59]=0.5;
x0[5][60]=1;
x0[5][61]=1;
x0[5][62]=0.5;

Mat aaa = cv::Mat(60,162, CV_32F, x0);

Mat inputBlob = dnn::blobFromImage(aaa, 1.0, Size(60,162));

net.setInput(inputBlob , "conv2d_input");

Mat prob = net.forward("activation_2/Softmax");

for(int i=0;i<prob.cols;i++)
{
    qDebug()<<i<<prob.at<float>(0,i);
}
user1277317
  • 127
  • 11

1 Answers1

0

In openCV almost all functions are designed to work with 3D matrices. So the easiest way for me to work with CV_32F 4D matrices is to work with them directly. The following code works correctly and quickly:

Net net = readNet("e://xor.pb");

const int sizes[] = {1,1,60,162};
Mat tenz = Mat::zeros(4, sizes, CV_32F);


float* dataB=(float*)tenz.data;

int x=1;
int y=2;
dataB[y*tenz.size[2]+x]=0.5f;

x=1;
y=3;
dataB[y*tenz.size[2]+x]=1.0f;


try
{
    net.setInput(tenz , "input_layer_my_input_1");

    Mat prob = net.forward("output_layer_my/MatMul");
}
catch( cv::Exception& e )
{
    const char* err_msg = e.what();
    qDebug()<<"err_msg"<<err_msg;
}
user1277317
  • 127
  • 11