/// <summary>
/// Creates an initialized instance
/// </summary>
/// <param name="networkName">Name of the network</param>
/// <param name="binBorder">If specified, it indicates that the whole network output is binary and specifies numeric border where GE network output is decided as a 1 and LT output as a 0.</param>
/// <param name="foldNum">Current fold number</param>
/// <param name="numOfFolds">Total number of the folds</param>
/// <param name="foldNetworkNum">Current fold network number</param>
/// <param name="numOfFoldNetworks">Total number of the fold networks</param>
/// <param name="regrAttemptNumber">Current regression attempt number</param>
/// <param name="regrMaxAttempts">Maximum number of regression attempts</param>
/// <param name="epoch">Current epoch number within the current regression attempt</param>
/// <param name="maxEpochs">Maximum number of epochs</param>
/// <param name="currNetwork">Current network and related important error statistics.</param>
/// <param name="bestNetwork">The best network for now and related important error statistics.</param>
/// <param name="lastImprovementEpoch">Specifies when was lastly found an improvement (bestNetwork=currNetwork).</param>
public BuildingState(string networkName,
double binBorder,
int foldNum,
int numOfFolds,
int foldNetworkNum,
int numOfFoldNetworks,
int regrAttemptNumber,
int regrMaxAttempts,
int epoch,
int maxEpochs,
TrainedNetwork currNetwork,
TrainedNetwork bestNetwork,
int lastImprovementEpoch
)
{
NetworkName = networkName;
BinBorder = binBorder;
FoldNum = foldNum;
NumOfFolds = numOfFolds;
FoldNetworkNum = foldNetworkNum;
NumOfFoldNetworks = numOfFoldNetworks;
RegrAttemptNumber = regrAttemptNumber;
RegrMaxAttempts = regrMaxAttempts;
Epoch = epoch;
MaxEpochs = maxEpochs;
CurrNetwork = currNetwork;
BestNetwork = bestNetwork;
LastImprovementEpoch = lastImprovementEpoch;
return;
}
//Static methods
/// <summary>
/// This is a default implementation of an evaluation whether the "candidate" network
/// achieved a better result than the best network so far
/// </summary>
/// <param name="binaryOutput">Indicates the whole network output is binary</param>
/// <param name="candidate">Network to be evaluated</param>
/// <param name="currentBest">The best network so far</param>
public static bool IsBetter(bool binaryOutput, TrainedNetwork candidate, TrainedNetwork currentBest)
{
if (binaryOutput)
{
if (candidate.CombinedBinaryError > currentBest.CombinedBinaryError)
{
return(false);
}
else if (candidate.CombinedBinaryError < currentBest.CombinedBinaryError)
{
return(true);
}
//CombinedBinaryError is the same
else if (candidate.TestingBinErrorStat.BinValErrStat[0].Sum > currentBest.TestingBinErrorStat.BinValErrStat[0].Sum)
{
return(false);
}
else if (candidate.TestingBinErrorStat.BinValErrStat[0].Sum < currentBest.TestingBinErrorStat.BinValErrStat[0].Sum)
{
return(true);
}
//CombinedBinaryError is the same
//TestingBinErrorStat.BinValErrStat[0].Sum is the same
else if (candidate.TrainingBinErrorStat.BinValErrStat[0].Sum > currentBest.TrainingBinErrorStat.BinValErrStat[0].Sum)
{
return(false);
}
else if (candidate.TrainingBinErrorStat.BinValErrStat[0].Sum < currentBest.TrainingBinErrorStat.BinValErrStat[0].Sum)
{
return(true);
}
//CombinedBinaryError is the same
//TestingBinErrorStat.BinValErrStat[0].Sum is the same
//TrainingBinErrorStat.BinValErrStat[0].Sum is the same
else if (candidate.CombinedPrecisionError < currentBest.CombinedPrecisionError)
{
return(true);
}
else
{
return(false);
}
}
else
{
return(candidate.CombinedPrecisionError < currentBest.CombinedPrecisionError);
}
}
/// <summary>
/// The deep copy constructor.
/// </summary>
/// <param name="source">Source instance</param>
public TrainedNetwork(TrainedNetwork source)
{
NetworkName = source.NetworkName;
BinBorder = source.BinBorder;
Network = null;
if (source.Network != null)
{
Network = source.Network.DeepClone();
}
TrainerInfoMessage = source.TrainerInfoMessage;
TrainingErrorStat = source.TrainingErrorStat?.DeepClone();
TrainingBinErrorStat = source.TrainingBinErrorStat?.DeepClone();
TestingErrorStat = source.TestingErrorStat?.DeepClone();
TestingBinErrorStat = source.TestingBinErrorStat?.DeepClone();
OutputWeightsStat = source.OutputWeightsStat?.DeepClone();
CombinedPrecisionError = source.CombinedPrecisionError;
CombinedBinaryError = source.CombinedBinaryError;
ExpectedPrecisionAccuracy = source.ExpectedPrecisionAccuracy;
ExpectedBinaryAccuracy = source.ExpectedBinaryAccuracy;
return;
}
/// <summary>
/// Builds trained network
/// </summary>
/// <returns>Trained network</returns>
public TrainedNetwork Build()
{
TrainedNetwork bestNetwork = null;
int lastImprovementEpoch = 0;
double lastImprovementCombinedPrecisionError = 0d;
double lastImprovementCombinedBinaryError = 0d;
//Create network and trainer
NonRecurrentNetUtils.CreateNetworkAndTrainer(_networkSettings,
_trainingBundle.InputVectorCollection,
_trainingBundle.OutputVectorCollection,
_rand,
out INonRecurrentNetwork net,
out INonRecurrentNetworkTrainer trainer
);
//Iterate training cycles
while (trainer.Iteration())
{
//Compute current error statistics after training iteration
//Training data part
TrainedNetwork currNetwork = new TrainedNetwork
{
NetworkName = _networkName,
BinBorder = _binBorder,
Network = net,
TrainerInfoMessage = trainer.InfoMessage,
TrainingErrorStat = net.ComputeBatchErrorStat(_trainingBundle.InputVectorCollection, _trainingBundle.OutputVectorCollection, out List <double[]> trainingComputedOutputsCollection)
};
if (BinaryOutput)
{
currNetwork.TrainingBinErrorStat = new BinErrStat(_binBorder, trainingComputedOutputsCollection, _trainingBundle.OutputVectorCollection);
currNetwork.CombinedBinaryError = currNetwork.TrainingBinErrorStat.TotalErrStat.Sum;
}
currNetwork.CombinedPrecisionError = currNetwork.TrainingErrorStat.ArithAvg;
//Testing data part
currNetwork.TestingErrorStat = net.ComputeBatchErrorStat(_testingBundle.InputVectorCollection, _testingBundle.OutputVectorCollection, out List <double[]> testingComputedOutputsCollection);
currNetwork.CombinedPrecisionError = Math.Max(currNetwork.CombinedPrecisionError, currNetwork.TestingErrorStat.ArithAvg);
if (BinaryOutput)
{
currNetwork.TestingBinErrorStat = new BinErrStat(_binBorder, testingComputedOutputsCollection, _testingBundle.OutputVectorCollection);
currNetwork.CombinedBinaryError = Math.Max(currNetwork.CombinedBinaryError, currNetwork.TestingBinErrorStat.TotalErrStat.Sum);
}
//Expected precision accuracy
currNetwork.ExpectedPrecisionAccuracy = Math.Min((1d - (currNetwork.TrainingErrorStat.ArithAvg / currNetwork.Network.OutputRange.Span)), (1d - (currNetwork.TestingErrorStat.ArithAvg / currNetwork.Network.OutputRange.Span)));
//Expected binary accuracy
if (BinaryOutput)
{
currNetwork.ExpectedBinaryAccuracy = Math.Min((1d - currNetwork.TrainingBinErrorStat.TotalErrStat.ArithAvg), (1d - currNetwork.TestingBinErrorStat.TotalErrStat.ArithAvg));
}
else
{
currNetwork.ExpectedBinaryAccuracy = double.NaN;
}
//Restart lastImprovementEpoch when new trainer's attempt started
if (trainer.AttemptEpoch == 1)
{
lastImprovementEpoch = trainer.AttemptEpoch;
lastImprovementCombinedPrecisionError = currNetwork.CombinedPrecisionError;
lastImprovementCombinedBinaryError = currNetwork.CombinedBinaryError;
}
//First initialization of the best network
bestNetwork = bestNetwork ?? currNetwork.DeepClone();
//RegrState instance
BuildingState regrState = new BuildingState(_networkName, _binBorder, _foldNum, _numOfFolds, _foldNetworkNum, _numOfFoldNetworks, trainer.Attempt, trainer.MaxAttempt, trainer.AttemptEpoch, trainer.MaxAttemptEpoch, currNetwork, bestNetwork, lastImprovementEpoch);
//Call controller
BuildingInstr instructions = _controller(regrState);
//Better?
if (instructions.CurrentIsBetter)
{
//Adopt current regression unit as a best one
bestNetwork = currNetwork.DeepClone();
regrState.BestNetwork = bestNetwork;
lastImprovementEpoch = trainer.AttemptEpoch;
lastImprovementCombinedPrecisionError = currNetwork.CombinedPrecisionError;
lastImprovementCombinedBinaryError = currNetwork.CombinedBinaryError;
}
if (currNetwork.CombinedBinaryError < lastImprovementCombinedBinaryError || currNetwork.CombinedPrecisionError < lastImprovementCombinedPrecisionError)
{
lastImprovementEpoch = trainer.AttemptEpoch;
lastImprovementCombinedPrecisionError = currNetwork.CombinedPrecisionError;
lastImprovementCombinedBinaryError = currNetwork.CombinedBinaryError;
}
//Raise notification event
RegressionEpochDone(regrState, instructions.CurrentIsBetter);
//Process instructions
if (instructions.StopProcess)
{
break;
}
else if (instructions.StopCurrentAttempt)
{
if (!trainer.NextAttempt())
{
break;
}
}
}//while (iteration)
//Create statistics of the best network weights
bestNetwork.OutputWeightsStat = bestNetwork.Network.ComputeWeightsStat();
return(bestNetwork);
}
