/// <summary>
/// Recognizes the character in the file of the given path.
/// </summary>
/// <param name="p">Path to a file with a character image.</param>
/// <returns>The character in this image.</returns>
public string recognise(string p)
{
// file path 2 array
ArrayInputData aid = new ArrayInputData(DataManipulation.Bmp2Pattern(p), null);
// use the network
Matrix <float> result = mln.use(aid);
return(Mat2String(result));
}
/// <summary>
/// Recognizes the character in the file of the given path.
/// </summary>
/// <param name="p">Path to a file with a character image.</param>
/// <returns>The character in this image.</returns>
public string recognise(string p)
{
// file path 2 array
ArrayInputData aid = new ArrayInputData(DataManipulation.Bmp2Pattern(p), null);
// use the network
Matrix<float> result = mln.use(aid);
return Mat2String(result);
}
/// <summary>
/// Loads images into an input pattern structure.
/// </summary>
/// <param name="p">list of StringInputData, containing the image path as Key and the represented character as Value.</param>
public void loadTrainImages(List<StringInputData> p)
{
trainImages = p;
trainData = new ArrayInputData(new DenseMatrix(patternSize, p.Count), new DenseMatrix(26, p.Count, 0));
Matrix<float> inputPattern;
for (int img = 0; img < p.Count; img++)
{
inputPattern = DataManipulation.Bmp2Pattern(p[img].Key);
trainData.Key.SetColumn(img, inputPattern.Column(0));
// Careful: only upper letters are covered so far!!
trainData.Value.At(p[img].Value.ToUpper()[0] - 'A', img, 1.0f);
}
}
/// <summary>
/// Loads images into an input pattern structure.
/// </summary>
/// <param name="p">list of StringInputData, containing the image path as Key and the represented character as Value.</param>
public void loadTrainImages(List <StringInputData> p)
{
trainImages = p;
trainData = new ArrayInputData(new DenseMatrix(patternSize, p.Count), new DenseMatrix(26, p.Count, 0));
Matrix <float> inputPattern;
for (int img = 0; img < p.Count; img++)
{
inputPattern = DataManipulation.Bmp2Pattern(p[img].Key);
trainData.Key.SetColumn(img, inputPattern.Column(0));
// Careful: only upper letters are covered so far!!
trainData.Value.At(p[img].Value.ToUpper()[0] - 'A', img, 1.0f);
}
}
/// <summary>
/// Clears the training data.
/// </summary>
public void clearTrainImages()
{
trainImages = new List <StringInputData>();
trainData = null;
}
/// <summary>
/// Uses the network with one input pattern to genereate output.
/// </summary>
/// <param name="pat">input pattern. Size (PatternSize X 1)</param>
/// <returns>the output matrix of the network</returns>
internal Matrix<float> use(ArrayInputData pat)
{
// convert
pat.Key = pat.Key.InsertRow(pat.Key.RowCount, new DenseVector(1, 1.0f));
// pass
forwardPass(pat.Key);
// return result
return net_out;
}
/// <summary>
/// Trains the network using common backprop.
/// </summary>
/// <param name="traindata">all training data used in this training session. Matrix must have the size (PatternSize X PatternCount)</param>
/// <param name="epochs">Number of epochs to train the network.</param>
internal void backpropTraining(ArrayInputData traindata, int epochs)
{
// training pattern init
Matrix<float> pat = new DenseMatrix(traindata.Key.RowCount + 1, traindata.Key.ColumnCount, 1.0f);
pat.SetSubMatrix(0, traindata.Key.RowCount, 0, traindata.Key.ColumnCount, traindata.Key);
// target values init
Matrix<float> targets = traindata.Value;
// dw, dv
Matrix<float> dw = new DenseMatrix(HiddenNodes, PatternSize + 1);
Matrix<float> dv = new DenseMatrix(TargetSize, HiddenNodes + 1);
// training
for (int i = 0; i < epochs; i++)
{
// backprop
forwardPass(pat);
backwardPass(pat, targets);
weightUpdate(pat, dw, dv, targets);
}
}
/// <summary>
/// Clears the training data.
/// </summary>
public void clearTrainImages()
{
trainImages = new List<StringInputData>();
trainData = null;
}
