Problems with the ML model using OpenCV

AI Discussions
1

I am using the OpenCV library to generate an ML model that recognises the handwritten digits from the MNIST dataset.

During the process of training the model, the results in the output layer almost always end up as Nan or Inf. I tried changing the parameters of the model like number of iterations or learning rate but with little or no success. I also tried to use a smaller set of data for training, but that doesn’t seem to work as well. Does anybody have an idea what I’m doing wrong?

Here is the code for the model setup

cv::Ptr<cv::ml::ANN_MLP> mlp = cv::ml::ANN_MLP::create();

 mlp->setActivationFunction(cv::ml::ANN_MLP::SIGMOID_SYM, 1, 1);

 int inputLayerSize = imagesData[0].total();
 int hiddenLayerSize = 100;
 int outputLayerSize = 10;
 cv::Mat layers = (cv::Mat_<int>(3, 1)<<inputLayerSize, hiddenLayerSize, outputLayerSize);
 mlp->setLayerSizes(layers);

 int numSamples = imagesData.size();

 cv::Mat trainingData(numSamples, inputLayerSize, CV_32F);
 cv::Mat labelData(numSamples, outputLayerSize, CV_32F);

 for (int i = 0; i < numSamples; i++) {

     cv::Mat image = imagesData[i].reshape(1, 1);
     image.convertTo(trainingData.row(i), CV_32F);

     cv::Mat label = cv::Mat::zeros(1, outputLayerSize, CV_32F);
     label.at<float>(0, labelsData[i]) = 1.0;
     label.copyTo(labelData.row(i));

 }

 cv::TermCriteria termCrit(cv::TermCriteria::MAX_ITER + cv::TermCriteria::EPS, 100, 0.001);
 mlp->setTermCriteria(termCrit);

 mlp->setTrainMethod(cv::ml::ANN_MLP::BACKPROP, 0.001, 0.1);

 mlp->train(trainingData, cv::ml::ROW_SAMPLE, labelData);
2

I’ve never used OpenCV for ML training, so not a lot I can say specifically.

In general, if you get Inf or NAN results, it’s typically due to using non-normalized data, or using too high a learning rate.

Or your model is broken. I can’t say much about that without understanding those tools specifically.

3

Thx for reply. Ill try to work on the data and see if the normalization is wrong