public void Process(String methodName, String methodArchitecture, String trainerName, String trainerArgs,
int outputNeurons)
{
// first, create the machine learning method
var methodFactory = new MLMethodFactory();
IMLMethod method = methodFactory.Create(methodName, methodArchitecture, 2, outputNeurons);
// second, create the data set
IMLDataSet dataSet = new BasicMLDataSet(XORInput, XORIdeal);
// third, create the trainer
var trainFactory = new MLTrainFactory();
IMLTrain train = trainFactory.Create(method, dataSet, trainerName, trainerArgs);
// reset if improve is less than 1% over 5 cycles
if (method is IMLResettable && !(train is ManhattanPropagation))
{
train.AddStrategy(new RequiredImprovementStrategy(500));
}
// fourth, train and evaluate.
EncogUtility.TrainToError(train, 0.01);
method = train.Method;
EncogUtility.Evaluate((IMLRegression)method, dataSet);
// finally, write out what we did
Console.WriteLine(@"Machine Learning Type: " + methodName);
Console.WriteLine(@"Machine Learning Architecture: " + methodArchitecture);
Console.WriteLine(@"Training Method: " + trainerName);
Console.WriteLine(@"Training Args: " + trainerArgs);
}
/// <summary>
/// Perform the training option.
/// </summary>
public void Train()
{
// first, create the machine learning method
var methodFactory = new MLMethodFactory();
IMLMethod method = methodFactory.Create(Config.MethodType, Config.MethodArchitecture, Config.InputWindow, 1);
// second, create the data set
string filename = FileUtil.CombinePath(new FileInfo(_path), Config.FilenameTrain).ToString();
IMLDataSet dataSet = EncogUtility.LoadEGB2Memory(new FileInfo(filename));
// third, create the trainer
var trainFactory = new MLTrainFactory();
IMLTrain train = trainFactory.Create(method, dataSet, Config.TrainType, Config.TrainParams);
// reset if improve is less than 1% over 5 cycles
if (method is IMLResettable && !(train is ManhattanPropagation))
{
train.AddStrategy(new RequiredImprovementStrategy(500));
}
// fourth, train and evaluate.
EncogUtility.TrainToError(train, Config.TargetError);
method = train.Method;
EncogDirectoryPersistence.SaveObject(FileUtil.CombinePath(new FileInfo(_path), Config.MethodName), method);
// finally, write out what we did
Console.WriteLine(@"Machine Learning Type: " + Config.MethodType);
Console.WriteLine(@"Machine Learning Architecture: " + Config.MethodArchitecture);
Console.WriteLine(@"Training Method: " + Config.TrainType);
Console.WriteLine(@"Training Args: " + Config.TrainParams);
}
/// <inheritdoc />
public INormalizationStrategy SuggestNormalizationStrategy(VersatileMLDataSet dataset, String architecture)
{
double inputLow = -1;
double inputHigh = 1;
double outputLow = -1;
double outputHigh = 1;
// Create a basic neural network, just to examine activation functions.
var methodFactory = new MLMethodFactory();
var network = (BasicNetwork)methodFactory.Create(MethodName, architecture, 1, 1);
if (network.LayerCount < 1)
{
throw new EncogError("Neural network does not have an output layer.");
}
IActivationFunction outputFunction = network.GetActivation(network.LayerCount - 1);
double[] d = { -1000, -100, -50 };
outputFunction.ActivationFunction(d, 0, d.Length);
if (d[0] > 0 && d[1] > 0 && d[2] > 0)
{
inputLow = 0;
}
INormalizationStrategy result = new BasicNormalizationStrategy(
inputLow,
inputHigh,
outputLow,
outputHigh);
return(result);
}
/// <summary>
/// Create the selected method.
/// </summary>
/// <returns>The created method.</returns>
public IMLMethod CreateMethod()
{
if (_methodType == null)
{
throw new EncogError(
"Please call selectMethod first to choose what type of method you wish to use.");
}
var methodFactory = new MLMethodFactory();
IMLMethod method = methodFactory.Create(_methodType, _methodArgs, _dataset
.NormHelper.CalculateNormalizedInputCount(), _config
.DetermineOutputCount(_dataset));
return(method);
}
/// <inheritdoc />
public override sealed bool ExecuteCommand(String args)
{
// get filenames
String trainingID = Prop.GetPropertyString(
ScriptProperties.MlConfigTrainingFile);
String resourceID = Prop.GetPropertyString(
ScriptProperties.MlConfigMachineLearningFile);
FileInfo trainingFile = Script.ResolveFilename(trainingID);
FileInfo resourceFile = Script.ResolveFilename(resourceID);
String type = Prop.GetPropertyString(
ScriptProperties.MlConfigType);
String arch = Prop.GetPropertyString(
ScriptProperties.MlConfigArchitecture);
EncogLogging.Log(EncogLogging.LevelDebug, "Beginning create");
EncogLogging.Log(EncogLogging.LevelDebug, "training file:"
+ trainingID);
EncogLogging.Log(EncogLogging.LevelDebug, "resource file:"
+ resourceID);
EncogLogging.Log(EncogLogging.LevelDebug, "type:" + type);
EncogLogging.Log(EncogLogging.LevelDebug, "arch:" + arch);
var egb = new EncogEGBFile(trainingFile.ToString());
egb.Open();
int input = egb.InputCount;
int ideal = egb.IdealCount;
egb.Close();
var factory = new MLMethodFactory();
IMLMethod obj = factory.Create(type, arch, input, ideal);
if (obj is BayesianNetwork)
{
string query = Prop.GetPropertyString(ScriptProperties.MLConfigQuery);
((BayesianNetwork)obj).DefineClassificationStructure(query);
}
EncogDirectoryPersistence.SaveObject(resourceFile, obj);
return(false);
}
private void CreateCommand()
{
if (_cmd.Args.Count != 3)
{
Console.WriteLine(@"Must specify method filename, type and args.");
return;
}
String methodFile = _cmd.Args[0];
String type = _cmd.Args[1];
String args = _cmd.Args[2];
_sw.Start();
var factory = new MLMethodFactory();
IMLMethod method = factory.Create(type, args, 0, 0);
EncogDirectoryPersistence.SaveObject(new FileInfo(methodFile), method);
}