opencv C++ neural network predict() function throws "Bad argument" error - c++

Question

I have managed to train a neural network to recognize numbers in an image and have saved the network parameters to an .xml file.
However, when testing the network against a new image the code fails at the predict() stage with the error:
OpenCV Error: Bad argument (Both input and output must be floating-point matrices of the same type and have the same number of rows) in CvANN_MLP::predict, file ........\opencv\modules\ml\src\ann_mlp.cpp, line 279.
ann_mlp.cpp line 279 is:
if( !CV_IS_MAT(_inputs) || !CV_IS_MAT(_outputs) ||
!CV_ARE_TYPES_EQ(_inputs,_outputs) ||
(CV_MAT_TYPE(_inputs->type) != CV_32FC1 &&
CV_MAT_TYPE(_inputs->type) != CV_64FC1) ||
_inputs->rows != _outputs->rows )
CV_Error( CV_StsBadArg, "Both input and output must be floating-point matrices "
"of the same type and have the same number of rows" );
I have checked input rows by running this code:
cv::Size s = newVec.size();
int rows = s.height;
int cols = s.width;
cout << "newVec dimensions: " << rows << " x " << cols << endl;
...and it comes out with the expected 1 x 900 vector / matrix.
I have set the input and output matrices to be CV_32FC1 as per the error dialog like this:
Input matrix
cv::Mat newVec(1, 900, CV_32FC1);
newVec = crop_img.reshape(0, 1); //reshape / unroll image to vector
CvMat n = newVec;
newVec = cv::Mat(&n);
Output matrix
cv::Mat classOut = cvCreateMatHeader(1, CLASSES, CV_32FC1);
And I try to run the prediction function like this:
CvANN_MLP* nnetwork = new CvANN_MLP;
nnetwork->load("nnetwork.xml", "nnetwork");
int maxIndex = 0;
cv::Mat classOut = cvCreateMatHeader(1, CLASSES, CV_32FC1);
//prediction
nnetwork->predict(newVec, classOut);
float value;
float maxValue = classOut.at<float>(0, 0);
for (int index = 1; index<CLASSES; index++)
{
value = classOut.at<float>(0, index);
if (value>maxValue)
{
maxValue = value;
maxIndex = index;
}
}
Any ideas? Much appreciated...

Answer 1

I suspect the problem is your input, not your output.
First it's important to understand that OpenCV deserves a lot of the blame for this, not you. Their C++ API is quite mediocre, and it caused major confusion to you.
See, normally in C++ when you define a 1x900 matrix of floats, it stays a matrix of floats. C++ has strong type safety.
OpenCV does not. If you assign a matrix of bytes to a matrix of floats, the latter will change its type (!).
Your code initializes newVec to such a matrix of floats, then assigns a second matrix, and then yet another matrix. I suspect that crop_img is still an image, i.e. 8 bits. Reshaping it will make it 1x900, but not floating point. That's the job of .convertTo.