//Methods
/// <inheritdoc/>
protected override void Check()
{
if (!FeedForwardNetwork.IsAllowedOutputAF(OutputActivationCfg))
{
throw new ArgumentException($"Specified output activation function can't be used in FF network's output activation.", "OutputActivationCfg");
}
if (TrainerCfg == null)
{
throw new ArgumentNullException("TrainerCfg", "TrainerCfg can not be null.");
}
Type trainerType = TrainerCfg.GetType();
if (trainerType != typeof(QRDRegrTrainerSettings) &&
trainerType != typeof(RidgeRegrTrainerSettings) &&
trainerType != typeof(ElasticRegrTrainerSettings) &&
trainerType != typeof(RPropTrainerSettings)
)
{
throw new ArgumentException($"Unsupported TrainerCfg {trainerType.Name}.", "TrainerCfg");
}
if ((HiddenLayersCfg.HiddenLayerCfgCollection.Count > 0 || OutputActivationCfg.GetType() != typeof(AFAnalogIdentitySettings)) &&
trainerType != typeof(RPropTrainerSettings)
)
{
throw new ArgumentException($"Improper type of trainer {trainerType.Name}. For FF having other than Identity output activation or containing hidden layers can be used only Resilient back propagation trainer.", "TrainerCfg");
}
return;
}
//Methods
/// <inheritdoc/>
protected override void Check()
{
if (NumOfNeurons < 1)
{
throw new ArgumentException($"Invalid NumOfNeurons {NumOfNeurons.ToString(CultureInfo.InvariantCulture)}. NumOfNeurons must be GT 0.", "NumOfNeurons");
}
if (!FeedForwardNetwork.IsAllowedHiddenAF(ActivationCfg))
{
throw new ArgumentException($"Specified activation function can't be used in the hidden layer of a FF network.", "ActivationCfg");
}
return;
}
/// <inheritdoc/>
public INonRecurrentNetwork DeepClone()
{
FeedForwardNetwork clone = new FeedForwardNetwork(NumOfInputValues, NumOfOutputValues)
{
NumOfNeurons = NumOfNeurons
};
foreach (Layer layer in LayerCollection)
{
clone.LayerCollection.Add(layer.DeepClone());
}
clone._flatWeights = (double[])_flatWeights.Clone();
clone._isAllowedNguyenWidrowRandomization = _isAllowedNguyenWidrowRandomization;
return(clone);
}
//Constructor
/// <summary>
/// Constructs an initialized instance
/// </summary>
/// <param name="net">FF network to be trained</param>
/// <param name="inputVectorCollection">Predictors (input)</param>
/// <param name="outputVectorCollection">Ideal outputs (the same number of rows as number of inputs)</param>
/// <param name="rand">Random object to be used for adding a white-noise to predictors</param>
/// <param name="settings">Startup parameters of the trainer</param>
public QRDRegrTrainer(FeedForwardNetwork net,
List <double[]> inputVectorCollection,
List <double[]> outputVectorCollection,
QRDRegrTrainerSettings settings,
Random rand
)
{
//Check network readyness
if (!net.Finalized)
{
throw new InvalidOperationException($"Can´t create trainer. Network structure was not finalized.");
}
//Check network conditions
if (net.LayerCollection.Count != 1 || !(net.LayerCollection[0].Activation is Identity))
{
throw new InvalidOperationException($"Can´t create trainer. Network structure is not complient (single layer having Identity activation).");
}
//Check samples conditions
if (inputVectorCollection.Count < inputVectorCollection[0].Length + 1)
{
throw new InvalidOperationException($"Can´t create trainer. Insufficient number of training samples {inputVectorCollection.Count}. Minimum is {(inputVectorCollection[0].Length + 1)}.");
}
//Parameters
_settings = settings;
MaxAttempt = _settings.NumOfAttempts;
MaxAttemptEpoch = _settings.NumOfAttemptEpochs;
_net = net;
_rand = rand;
_inputVectorCollection = inputVectorCollection;
_outputVectorCollection = outputVectorCollection;
_outputSingleColMatrixCollection = new List <Matrix>(_net.NumOfOutputValues);
for (int outputIdx = 0; outputIdx < _net.NumOfOutputValues; outputIdx++)
{
Matrix outputSingleColMatrix = new Matrix(_outputVectorCollection.Count, 1);
for (int row = 0; row < _outputVectorCollection.Count; row++)
{
//Output
outputSingleColMatrix.Data[row][0] = _outputVectorCollection[row][outputIdx];
}
_outputSingleColMatrixCollection.Add(outputSingleColMatrix);
}
//Start training attempt
Attempt = 0;
NextAttempt();
return;
}
//Constructor
/// <summary>
/// Instantiates the RPropTrainer
/// </summary>
/// <param name="net">The FF network to be trained.</param>
/// <param name="inputVectorCollection">The input vectors (input).</param>
/// <param name="outputVectorCollection">The output vectors (ideal).</param>
/// <param name="cfg">The configuration of the trainer.</param>
/// <param name="rand">The random object to be used.</param>
public RPropTrainer(FeedForwardNetwork net,
List <double[]> inputVectorCollection,
List <double[]> outputVectorCollection,
RPropTrainerSettings cfg,
Random rand
)
{
if (!net.Finalized)
{
throw new InvalidOperationException($"Can´t create trainer. Network structure was not finalized.");
}
_cfg = cfg;
MaxAttempt = _cfg.NumOfAttempts;
MaxAttemptEpoch = _cfg.NumOfAttemptEpochs;
_net = net;
_rand = rand;
_inputVectorCollection = inputVectorCollection;
_outputVectorCollection = outputVectorCollection;
_weigthsGradsAcc = new double[_net.NumOfWeights];
_weigthsPrevGradsAcc = new double[_net.NumOfWeights];
_weigthsPrevDeltas = new double[_net.NumOfWeights];
_weigthsPrevChanges = new double[_net.NumOfWeights];
//Parallel gradient workers (batch ranges) preparation
int numOfWorkers = Math.Max(1, Math.Min(Environment.ProcessorCount - 1, _inputVectorCollection.Count));
_gradientWorkerDataCollection = new GradientWorkerData[numOfWorkers];
int workerBatchSize = _inputVectorCollection.Count / numOfWorkers;
for (int workerIdx = 0, fromRow = 0; workerIdx < numOfWorkers; workerIdx++, fromRow += workerBatchSize)
{
GradientWorkerData gwd = new GradientWorkerData
(
fromRow: fromRow,
toRow: (workerIdx == numOfWorkers - 1 ? _inputVectorCollection.Count - 1 : (fromRow + workerBatchSize) - 1),
numOfWeights: _net.NumOfWeights
);
_gradientWorkerDataCollection[workerIdx] = gwd;
}
InfoMessage = string.Empty;
//Start training attempt
Attempt = 0;
NextAttempt();
return;
}
//Constructor
/// <summary>
/// Creates an initialized instance.
/// </summary>
/// <param name="net">The FF network to be trained.</param>
/// <param name="inputVectorCollection">The input vectors (input).</param>
/// <param name="outputVectorCollection">The output vectors (ideal).</param>
/// <param name="cfg">The configuration of the trainer.</param>
public ElasticRegrTrainer(FeedForwardNetwork net,
List <double[]> inputVectorCollection,
List <double[]> outputVectorCollection,
ElasticRegrTrainerSettings cfg
)
{
//Check network readyness
if (!net.Finalized)
{
throw new InvalidOperationException($"Can´t create trainer. Network structure was not finalized.");
}
//Check network conditions
if (net.LayerCollection.Count != 1 || !(net.LayerCollection[0].Activation is AFAnalogIdentity))
{
throw new InvalidOperationException($"Can´t create trainer. Network structure is not complient (single layer having Identity activation).");
}
//Check samples conditions
if (inputVectorCollection.Count == 0)
{
throw new InvalidOperationException($"Can´t create trainer. Missing training samples.");
}
//Collections
_inputVectorCollection = new List <double[]>(inputVectorCollection);
_outputVectorCollection = new List <double[]>(outputVectorCollection);
var rangePartitioner = Partitioner.Create(0, _inputVectorCollection.Count);
_parallelRanges = new List <Tuple <int, int> >(rangePartitioner.GetDynamicPartitions());
//Parameters
_cfg = cfg;
MaxAttempt = _cfg.NumOfAttempts;
MaxAttemptEpoch = _cfg.NumOfAttemptEpochs;
Attempt = 1;
AttemptEpoch = 0;
_net = net;
_gamma = _cfg.Lambda * _cfg.Alpha;
return;
}
//Constructor
/// <summary>
/// Constructs an initialized instance
/// </summary>
/// <param name="net">FF network to be trained</param>
/// <param name="inputVectorCollection">Predictors (input)</param>
/// <param name="outputVectorCollection">Ideal outputs (the same number of rows as number of inputs)</param>
/// <param name="settings">Optional startup parameters of the trainer</param>
public RidgeRegrTrainer(FeedForwardNetwork net,
List <double[]> inputVectorCollection,
List <double[]> outputVectorCollection,
RidgeRegrTrainerSettings settings
)
{
//Check network readyness
if (!net.Finalized)
{
throw new InvalidOperationException($"Can´t create trainer. Network structure was not finalized.");
}
//Check network conditions
if (net.LayerCollection.Count != 1 || !(net.LayerCollection[0].Activation is Identity))
{
throw new InvalidOperationException($"Can´t create trainer. Network structure is not complient (single layer having Identity activation).");
}
//Check samples conditions
if (inputVectorCollection.Count == 0)
{
throw new InvalidOperationException($"Can´t create trainer. Missing training samples.");
}
//Collections
_inputVectorCollection = new List <double[]>(inputVectorCollection);
_outputVectorCollection = new List <double[]>(outputVectorCollection);
//Parameters
_settings = settings;
MaxAttempt = _settings.NumOfAttempts;
MaxAttemptEpoch = _settings.NumOfAttemptEpochs;
Attempt = 1;
AttemptEpoch = 0;
_net = net;
_outputSingleColVectorCollection = new List <Vector>(_net.NumOfOutputValues);
for (int outputIdx = 0; outputIdx < _net.NumOfOutputValues; outputIdx++)
{
Vector outputSingleColVector = new Vector(outputVectorCollection.Count);
for (int row = 0; row < outputVectorCollection.Count; row++)
{
//Output
outputSingleColVector.Data[row] = outputVectorCollection[row][outputIdx];
}
_outputSingleColVectorCollection.Add(outputSingleColVector);
}
//Lambda seeker
_lambdaSeeker = new ParamSeeker(_settings.LambdaSeekerCfg);
_currLambda = 0;
//Matrix setup
Matrix X = new Matrix(inputVectorCollection.Count, _net.NumOfInputValues + 1);
for (int row = 0; row < inputVectorCollection.Count; row++)
{
//Add constant bias
X.Data[row][0] = 1d;
//Add predictors
inputVectorCollection[row].CopyTo(X.Data[row], 1);
}
_XT = X.Transpose();
_XTdotX = _XT * X;
_XTdotY = new Vector[_net.NumOfOutputValues];
for (int outputIdx = 0; outputIdx < _net.NumOfOutputValues; outputIdx++)
{
_XTdotY[outputIdx] = _XT * _outputSingleColVectorCollection[outputIdx];
}
return;
}