private void listBoxCalculator_SelectedIndexChanged(object sender, EventArgs e)
{
try {
double bias = double.Parse(textBoxComparePrecision.Text);
switch (listBoxCalculator.SelectedIndex)
{
case 0:
_accuracyCalculator = new AccuracyCalculatorDecart(bias);
break;
case 1:
_accuracyCalculator = new AccuracyCalculatorSimpleBiasABSEqual(bias);
break;
case 2:
_accuracyCalculator = new AccuracyCalculatorSimpleMedianBiasEqual(bias);
break;
}
} catch (Exception exception) {
MessageBox.Show(exception.Message);
}
}
/// <summary>
/// returns % of right recognized samples
/// </summary>
/// <param name="_samples">without bias "1"</param>
/// <param name="_answers">without bias "1"</param>
/// <param name="_calculator"></param>
/// <returns></returns>
public double getAccuracy(Matrix _samples, Matrix _answers, IAccuracyCalculator _calculator)
{
if (_samples == null)
{
throw new ArgumentNullException(nameof(_samples));
}
if (_answers == null)
{
throw new ArgumentNullException(nameof(_answers));
}
if (_calculator == null)
{
throw new ArgumentNullException(nameof(_calculator));
}
//adding first "1" in samples and answers for algoritm
Matrix samples1 = _samples.ComposeAddRow(new Matrix(1, _samples.Columns, 1), 0);
Matrix answers1 = _answers.ComposeAddRow(new Matrix(1, _answers.Columns, 1), 0);
Matrix outcome = ForvardPropagation(samples1);
return(_calculator.Calculate(outcome, answers1));
}
/// <summary>
/// train neoronet, returns accuracy on each epoh
/// </summary>
/// <param name="_samples">doesn't contain first 1 for bias</param>
/// <param name="_answers">doesn't contain first 1 for bias</param>
/// <param name="_batchSize"></param>
/// <param name="_maxIteration"></param>
/// <param name="_accuracy">% of recognized samples, 0..1</param>
/// <param name="_testSamples">doesn't contain first 1 for bias</param>
/// <param name="_testAnswers">doesn't contain first 1 for bias</param>
/// <param name="_accuracyCalculator">for testing accuracy</param>
/// <param name="_savingPath">path for saving good nets</param>
/// <returns></returns>
public double[] train(Matrix _samples, Matrix _answers, int _batchSize, int _maxIteration, double _accuracy, Matrix _testSamples, Matrix _testAnswers, IAccuracyCalculator _accuracyCalculator, string _savingPath)
{
if (_samples == null)
{
throw new ArgumentNullException(nameof(_samples));
}
if (_answers == null)
{
throw new ArgumentNullException(nameof(_answers));
}
if (_testSamples == null)
{
throw new ArgumentNullException(nameof(_testSamples));
}
if (_testAnswers == null)
{
throw new ArgumentNullException(nameof(_testAnswers));
}
if (_accuracyCalculator == null)
{
throw new ArgumentNullException(nameof(_accuracyCalculator));
}
if (_samples.Columns != _answers.Columns)
{
throw new Exception("samples doesn't fit to answers");
}
if (_batchSize <= 0 || _batchSize > _samples.Columns)
{
throw new ArgumentOutOfRangeException(nameof(_batchSize));
}
if (_maxIteration <= 0)
{
throw new ArgumentOutOfRangeException(nameof(_maxIteration));
}
if (_accuracy < 0 || _accuracy > 1)
{
throw new ArgumentOutOfRangeException(nameof(_accuracy));
}
//adding first "1" in samples and answers for algoritm
Matrix samples1 = _samples.ComposeAddRow(new Matrix(1, _samples.Columns, 1), 0);
Matrix answers1 = _answers.ComposeAddRow(new Matrix(1, _answers.Columns, 1), 0);
//вибір даних з самплес в батч
List <KeyValuePair <Matrix, Matrix> > batches = new List <KeyValuePair <Matrix, Matrix> >(); //batches (KVP<samples,answers>)
if (samples1.Columns / _batchSize == 0 && samples1.Columns % _batchSize == 0)
{
batches.Add(new KeyValuePair <Matrix, Matrix>(samples1, answers1));
}
else
{
for (int i = 0; i < samples1.Columns / _batchSize; i++)
{
batches.Add(new KeyValuePair <Matrix, Matrix>(samples1.Submatrix(0, samples1.Rows - 1, i * _batchSize, (i + 1) * _batchSize - 1),
answers1.Submatrix(0, answers1.Rows - 1, i * _batchSize, (i + 1) * _batchSize - 1)));
}
if (samples1.Columns % _batchSize != 0)
{
batches.Add(
new KeyValuePair <Matrix, Matrix>(
samples1.Submatrix(0, samples1.Rows - 1, samples1.Columns / _batchSize * _batchSize, samples1.Columns - 1),
answers1.Submatrix(0, answers1.Rows - 1, answers1.Columns / _batchSize * _batchSize, samples1.Columns - 1)));
}
}
//train
List <double> rez = new List <double>();
double curAccur;
for (int i = 0; i < _maxIteration; i++)
{
for (int j = 0; j < batches.Count; j++)
{
BackPropagation(batches[j].Key, batches[j].Value);
}
curAccur = getAccuracy(_testSamples, _testAnswers, _accuracyCalculator);
rez.Add(curAccur);
if (curAccur > _accuracy)
{
break;
}
if (curAccur > _minAccurToSave)
{
this.save(_savingPath + "\\" + curAccur + " " + DateTime.Now.Minute + DateTime.Now.Second + DateTime.Now.Ticks);
_minAccurToSave = curAccur;
}
}
return(rez.ToArray <double>());
}
