/// <summary>
/// Predicts the output of the provided data after training has occured.
/// </summary>
/// <param name="outputType">The type of output data.</param>
/// <returns></returns>
public void Predict(EPerceptronOutputType outputType)
{
if (!HasTrained)
{
WorqnetsUtils.PrintMessage("Please Train The Network Before Attempting To Predict",
WorqnetsUtils.MessageType.Error);
return;
}
_bias = TrainingData.CalculatedBias;
switch (outputType)
{
case EPerceptronOutputType.Integer:
PredictedValue =
CalculateWeightedSum(ProblemData.Values, TrainingData.CalculatedWeights).ToString();
break;
case EPerceptronOutputType.String:
PredictedValue = CalculateWeightedSum(ProblemData.Values, TrainingData.CalculatedWeights) > 0
? TrainingData.OutputClass2
: TrainingData.OutputClass1;
break;
default:
throw new ArgumentOutOfRangeException(nameof(outputType), outputType, null);
}
//ProblemData.FloatOutput = PredictedValue > 0 ? 1 : 0;
}
/// <summary>
/// Co-routine to perform the training of the perceptron. This ensures that training is done on a separate thread so as to avoid freezing the program.
/// </summary>
/// <returns>Returns Wait For End Of Frame. Continues at the end of the current frame.</returns>
private IEnumerator DoTrain()
{
InitializeWeights();
EpochsPerformed = 0;
_repeatedAfterConverging = 0;
LastTrainingCouldNotConverge = false;
//Get out of training if train data is null.
if (!TrainingData)
{
yield break;
}
CheckForTrainData();
//Train at least once.
do
{
//Set total error euals zero.
_totalError = 0;
//Get out of training if max number of epochs has been reached (training could not converge).
if (EpochsPerformed == MaxEpochs - 1)
{
WorqnetsUtils.PrintMessage("Could Not Converge", WorqnetsUtils.MessageType.Error);
LastTrainingCouldNotConverge = true;
HasTrained = false;
yield break;
}
for (var i = 0; i < TrainingData.TrainingDataSize; i++)
{
UpdateWeights(i);
if (EnableTrainDebugging)
{
//Debug.Log("========================================");
Debug.Log("W1: " + _weights[0] + " W2: " + _weights[1] + " Bias: " + _bias);
}
}
if (EnableTrainDebugging)
{
Debug.Log("Total Error: " + _totalError);
//Debug.Log("========================================");
}
if (_totalError <= Epsilon)
{
_repeatedAfterConverging++;
}
EpochsPerformed += 1;
} while (EpochsPerformed < MaxEpochs && _repeatedAfterConverging <= RepetitionsAfterConverging);
TrainingData.CalculatedWeights = _weights;
TrainingData.CalculatedBias = _bias;
yield return(new WaitForEndOfFrame());
}
private float CalculateWeightedSum(IList <float> inputs, IList <float> weights)
{
if (inputs.Count != weights.Count)
{
WorqnetsUtils.PrintMessage("Can not train. Inputs and weights do not match",
WorqnetsUtils.MessageType.Warning);
return(float.MinValue);
}
var weightedSum = 0.0f;
for (var i = 0; i < inputs.Count; i += 1)
{
weightedSum += inputs[i] * weights[i];
}
weightedSum += _bias;
return(weightedSum);
}