Revision history [back]

I have been using the tutorial on this site titled "how to start implementing a neural network in C"( cant post link yet due to rep) and others: to create a neural network to process images of a track I have made for a car to be able to drive round it autonomously.

The process of reading in the training images (converted as required) and their respective classes works perfectly and stores to the data file Values.xml with the below code:

void readScanStore(){
cv::Mat trainingData;//mat collectionn of images to train with 
char const *path = "/home/pi/selfdrivingcarV1/train_data/";//needs folders of 0 1 2 3  holdig 50 pics each for representing outputs
int const numFilesDirs[]={110,95,63}; //number of photos for each direction (fwd l r)
char const strDirs[]={'0','1','2'}; //optional outputs " 0=go,1 right 2 left
int const numDirs = 3;//number of directions


cv::Mat trainingLabels (0,0,(CV_32S));
//same as svm??
for (int i=0;i!=numDirs; i++){//outer for loop to go through all 4 output options
    int numFiles = numFilesDirs[i];//assign inner loop based on size of samples from current outerloop val
        for (int j=0;j!=numFiles;j++){//loop through all files within current output value
            std::cout << "direction" << strDirs[i] << "file: " << j <<".jpg" << "\n";// print current output val and files associated with that direction
            std::stringstream ss;
            ss << path << strDirs[i] << "/" << j << ".jpg";//print current working image
            cv::Mat img = cv::imread(ss.str(),0);


        if (!img.data)
        cout << "error no file found " << ss << endl;


            Size size(10,10);
            Mat ImgCon;
            resize(img,ImgCon,size);
            ImgCon.reshape(1,1);
        //assume img is continous
            //reshape image to 1xtotal res 

            trainingData.push_back(ImgCon); //push back image
            trainingLabels.push_back(i);//assign instruction corresponding to image in label set
        }
    }  

trainingData.convertTo(trainingData, CV_32F);//,1/255.0);//convert all images to cv32f (numeric values)

cv::FileStorage fs("VALUES.xml",FileStorage::WRITE);//store numeric values in xml file as training data values for each image.
fs << "TrainingData" <<trainingData;//assign each image 
fs << "classes" << trainingLabels;//assign associated classes

printf("complete");
    }

which is then proceeded by the training procedure which loads the values of the images and uses them as training data to train the NN:

void trainNetwork(){
    int nclasses = 3;

    cv::FileStorage fs;
fs.open("VALUES.xml", cv::FileStorage::READ);
cv::Mat trainData;
cv::Mat classes;
fs["TrainingData"] >> trainData;
fs["classes"] >> classes;
    Mat confusion(nclasses,nclasses,CV_32S, Scalar(0)); // will hold our test results

 train_test(nclasses, trainData, classes, confusion);
}

which calls the following:

    void train_test(int nclasses, const Mat &train_data, const Mat &train_labels, Mat &confusion) {

        int nfeatures = train_data.cols;
        Ptr<ml::ANN_MLP> ann = ml::ANN_MLP::create();
        Mat_<int> layers(4,1);
        layers(0) = nfeatures;     // input
        layers(1) = nclasses * 8;  // hidden
        layers(2) = nclasses * 4;  // hidden
        layers(3) = nclasses;      // output, 1 pin per class.
        ann->setLayerSizes(layers);
        ann->setActivationFunction(ml::ANN_MLP::SIGMOID_SYM,0,0);
        ann->setTermCriteria(TermCriteria(TermCriteria::MAX_ITER+TermCriteria::EPS, 300, 0.0001));
        ann->setTrainMethod(ml::ANN_MLP::BACKPROP, 0.0001);
    printf("sending data to train_test"); // setup the ann:

        // ann requires "one-hot" encoding of class labels:
        Mat train_classes = Mat::zeros(train_data.rows, nclasses, CV_32FC1);


        printf("%i\n",train_classes.rows);
        //////////////////////////////////////////////////////////////////////////
        ///for(int i=0; i<train_classes.rows; i++)
        //{
        //train_classes.at<float>(i, train_labels.at<int>(i)) = 1.f;
        //}
//////////////////////////////////////////////////////////////////    

    cerr << train_data.size() << " " << train_classes.size() << endl;

        ann->train(train_data, ml::ROW_SAMPLE, train_classes);


        cv::FileStorage fs("NNPARAMS.xml",FileStorage::WRITE);//store numeric values in xml file as training data values for each image.
        ann->write(fs);
    }

It all works with obviously incorrect values due to the fact that the above code which is all commented out(3 lines) in train_test causes the program to halt when it is uncommented. I am not sure what the error could be however I suspect it may be the XML file with the image files (Hence why I have posted the process that creates this also).

Any suggestions on where to start?