/// <summary>
/// The deep copy constructor
/// </summary>
/// <param name="source">Source instance</param>
public FeedForwardNetworkSettings(FeedForwardNetworkSettings source)
{
OutputLayerActivation = null;
OutputRange = null;
if (source.OutputLayerActivation != null)
{
OutputLayerActivation = ActivationFactory.DeepCloneActivationSettings(source.OutputLayerActivation);
OutputRange = source.OutputRange.DeepClone();
}
RegressionMethod = source.RegressionMethod;
HiddenLayerCollection = new List <HiddenLayerSettings>(source.HiddenLayerCollection.Count);
foreach (HiddenLayerSettings shls in source.HiddenLayerCollection)
{
HiddenLayerCollection.Add(shls.DeepClone());
}
LinRegrTrainerCfg = null;
RPropTrainerCfg = null;
if (source.LinRegrTrainerCfg != null)
{
LinRegrTrainerCfg = source.LinRegrTrainerCfg.DeepClone();
}
if (source.RPropTrainerCfg != null)
{
RPropTrainerCfg = source.RPropTrainerCfg.DeepClone();
}
return;
}
/// <summary>
/// Creates the instance and initialize it from given xml element.
/// This is the preferred way to instantiate reservoir settings.
/// </summary>
/// <param name="elem">
/// Xml data containing feed forward network settings.
/// Content of xml element is always validated against the xml schema.
/// </param>
public FeedForwardNetworkSettings(XElement elem)
{
//Validation
ElemValidator validator = new ElemValidator();
Assembly assemblyRCNet = Assembly.GetExecutingAssembly();
validator.AddXsdFromResources(assemblyRCNet, "RCNet.Neural.Network.FF.FeedForwardNetworkSettings.xsd");
validator.AddXsdFromResources(assemblyRCNet, "RCNet.RCNetTypes.xsd");
XElement feedForwardNetworkSettingsElem = validator.Validate(elem, "rootElem");
//Parsing
OutputLayerActivation = ActivationFactory.LoadSettings(feedForwardNetworkSettingsElem.Descendants().First());
if (!IsAllowedActivation(OutputLayerActivation, out Interval outputRange))
{
throw new ApplicationException($"Activation can't be used in FF network. Activation function has to be stateless and has to support derivative calculation.");
}
OutputRange = outputRange;
RegressionMethod = ParseTrainingMethodType(feedForwardNetworkSettingsElem.Attribute("regressionMethod").Value);
//Hidden layers
HiddenLayerCollection = new List <HiddenLayerSettings>();
XElement hiddenLayersElem = feedForwardNetworkSettingsElem.Descendants("hiddenLayers").FirstOrDefault();
if (hiddenLayersElem != null)
{
foreach (XElement layerElem in hiddenLayersElem.Descendants("layer"))
{
HiddenLayerCollection.Add(new HiddenLayerSettings(layerElem));
}
}
//Trainers
LinRegrTrainerCfg = null;
RPropTrainerCfg = null;
switch (RegressionMethod)
{
case TrainingMethodType.Linear:
XElement linRegrTrainerElem = feedForwardNetworkSettingsElem.Descendants("linRegrTrainer").FirstOrDefault();
if (linRegrTrainerElem != null)
{
LinRegrTrainerCfg = new LinRegrTrainerSettings(linRegrTrainerElem);
}
else
{
LinRegrTrainerCfg = new LinRegrTrainerSettings();
}
break;
case TrainingMethodType.Resilient:
XElement resPropTrainerElem = feedForwardNetworkSettingsElem.Descendants("resPropTrainer").FirstOrDefault();
if (resPropTrainerElem != null)
{
RPropTrainerCfg = new RPropTrainerSettings(resPropTrainerElem);
}
else
{
RPropTrainerCfg = new RPropTrainerSettings();
}
break;
}
return;
}
//Constructors
/// <summary>
/// Creates an uninitialized instance
/// </summary>
public FeedForwardNetworkSettings()
{
OutputLayerActivation = null;
RegressionMethod = TrainingMethodType.Resilient;
HiddenLayerCollection = new List <HiddenLayerSettings>();
LinRegrTrainerCfg = null;
RPropTrainerCfg = new RPropTrainerSettings();
return;
}
/// <summary>
/// Deep copy constructor
/// </summary>
/// <param name="source">Source instance</param>
public RPropTrainerSettings(RPropTrainerSettings source)
{
ZeroTolerance = source.ZeroTolerance;
PositiveEta = source.PositiveEta;
NegativeEta = source.NegativeEta;
IniDelta = source.IniDelta;
MinDelta = source.MinDelta;
MaxDelta = source.MaxDelta;
return;
}
//Constructor
/// <summary>
/// Instantiates the RPropTrainer
/// </summary>
/// <param name="net">
/// The feed forward network to be trained
/// </param>
/// <param name="inputVectorCollection">
/// Collection of the training input vectors
/// </param>
/// <param name="outputVectorCollection">
/// Collection of the desired outputs
/// </param>
/// Trainer parameters
/// <param name="settings">
/// </param>
public RPropTrainer(FeedForwardNetwork net,
List <double[]> inputVectorCollection,
List <double[]> outputVectorCollection,
RPropTrainerSettings settings = null
)
{
if (!net.Finalized)
{
throw new Exception("Can´t create trainer. Network structure was not finalized.");
}
_settings = settings;
if (_settings == null)
{
//Default parameters
_settings = new RPropTrainerSettings();
}
_net = net;
_inputVectorCollection = inputVectorCollection;
_outputVectorCollection = outputVectorCollection;
_weigthsGradsAcc = new double[_net.NumOfWeights];
_weigthsGradsAcc.Populate(0);
_weigthsPrevGradsAcc = new double[_net.NumOfWeights];
_weigthsPrevGradsAcc.Populate(0);
_weigthsPrevDeltas = new double[_net.NumOfWeights];
_weigthsPrevDeltas.Populate(_settings.IniDelta);
_weigthsPrevChanges = new double[_net.NumOfWeights];
_weigthsPrevChanges.Populate(0);
_prevMSE = 0;
_lastMSE = 0;
_epoch = 0;
//Parallel gradient workers / batch ranges preparation
_workerRangeCollection = new List <WorkerRange>();
int numOfWorkers = Math.Min(Environment.ProcessorCount, _inputVectorCollection.Count);
numOfWorkers = Math.Max(1, numOfWorkers);
int workerBatchSize = _inputVectorCollection.Count / numOfWorkers;
for (int workerIdx = 0, fromRow = 0; workerIdx < numOfWorkers; workerIdx++, fromRow += workerBatchSize)
{
WorkerRange workerRange = new WorkerRange();
workerRange.FromRow = fromRow;
if (workerIdx == numOfWorkers - 1)
{
workerRange.ToRow = _inputVectorCollection.Count - 1;
}
else
{
workerRange.ToRow = (fromRow + workerBatchSize) - 1;
}
_workerRangeCollection.Add(workerRange);
}
return;
}
/// <summary>
/// Deep copy constructor
/// </summary>
/// <param name="source">Source instance</param>
public RPropTrainerSettings(RPropTrainerSettings source)
{
NumOfAttempts = source.NumOfAttempts;
NumOfAttemptEpochs = source.NumOfAttemptEpochs;
ZeroTolerance = source.ZeroTolerance;
PositiveEta = source.PositiveEta;
NegativeEta = source.NegativeEta;
IniDelta = source.IniDelta;
MinDelta = source.MinDelta;
MaxDelta = source.MaxDelta;
return;
}
//Constructor
/// <summary>
/// Instantiates the RPropTrainer
/// </summary>
/// <param name="net">
/// The feed forward network to be trained
/// </param>
/// <param name="inputVectorCollection">
/// Collection of the training input vectors
/// </param>
/// <param name="outputVectorCollection">
/// Collection of the desired outputs
/// </param>
/// Trainer parameters
/// <param name="settings">
/// </param>
public RPropTrainer(FeedForwardNetwork net,
List <double[]> inputVectorCollection,
List <double[]> outputVectorCollection,
RPropTrainerSettings settings = null
)
{
if (!net.Finalized)
{
throw new Exception("Can´t create trainer. Network structure was not finalized.");
}
_settings = settings;
if (_settings == null)
{
//Default parameters
_settings = new RPropTrainerSettings();
}
_net = net;
_inputVectorCollection = inputVectorCollection;
_outputVectorCollection = outputVectorCollection;
_weigthsGradsAcc = new double[_net.NumOfWeights];
_weigthsGradsAcc.Populate(0);
_weigthsPrevGradsAcc = new double[_net.NumOfWeights];
_weigthsPrevGradsAcc.Populate(0);
_weigthsPrevDeltas = new double[_net.NumOfWeights];
_weigthsPrevDeltas.Populate(_settings.IniDelta);
_weigthsPrevChanges = new double[_net.NumOfWeights];
_weigthsPrevChanges.Populate(0);
_prevMSE = 0;
MSE = 0;
Epoch = 0;
//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;
}
return;
}
//Constructor
/// <summary>
/// Instantiates the RPropTrainer
/// </summary>
/// <param name="net"> The feed forward network to be trained </param>
/// <param name="inputVectorCollection"> Collection of the training input vectors </param>
/// <param name="outputVectorCollection"> Collection of the desired outputs </param>
/// <param name="settings"> Trainer parameters </param>
/// <param name="rand">Random object to be used</param>
public RPropTrainer(FeedForwardNetwork net,
List <double[]> inputVectorCollection,
List <double[]> outputVectorCollection,
RPropTrainerSettings settings,
Random rand
)
{
if (!net.Finalized)
{
throw new Exception("Can´t create trainer. Network structure was not finalized.");
}
_settings = settings;
MaxAttempt = _settings.NumOfAttempts;
MaxAttemptEpoch = _settings.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;
}
//Methods
/// <summary>
/// See the base.
/// </summary>
public override bool Equals(object obj)
{
if (obj == null)
{
return(false);
}
RPropTrainerSettings cmpSettings = obj as RPropTrainerSettings;
if (ZeroTolerance != cmpSettings.ZeroTolerance ||
PositiveEta != cmpSettings.PositiveEta ||
NegativeEta != cmpSettings.NegativeEta ||
IniDelta != cmpSettings.IniDelta ||
MinDelta != cmpSettings.MinDelta ||
MaxDelta != cmpSettings.MaxDelta
)
{
return(false);
}
return(true);
}
/// <summary>
/// Creates the deep copy instance of this instance
/// </summary>
public RPropTrainerSettings DeepClone()
{
RPropTrainerSettings clone = new RPropTrainerSettings(this);
return(clone);
}
