/**
* Create a feed forward network.
*
* @param architecture The architecture string to use.
* @param input The input count.
* @param output The output count.
* @return The feedforward network.
*/
public final MLMethod create(final String architecture, final int input, final int output) {
if (input <= 0) {
throw new EncogError("Must have at least one input for feedforward.");
}
if (output <= 0) {
throw new EncogError("Must have at least one output for feedforward.");
}
final BasicNetwork result = new BasicNetwork();
final List<String> layers = ArchitectureParse.parseLayers(architecture);
ActivationFunction af = new ActivationLinear();
int questionPhase = 0;
for (final String layerStr : layers) {
int defaultCount;
// determine default
if (questionPhase == 0) {
defaultCount = input;
} else {
defaultCount = output;
}
final ArchitectureLayer layer = ArchitectureParse.parseLayer(layerStr, defaultCount);
final boolean bias = layer.isBias();
String part = layer.getName();
if (part != null) {
part = part.trim();
} else {
part = "";
}
ActivationFunction lookup = this.factory.create(part);
if (lookup != null) {
af = lookup;
} else {
if (layer.isUsedDefault()) {
questionPhase++;
if (questionPhase > 2) {
throw new EncogError("Only two ?'s may be used.");
}
}
if (layer.getCount() == 0) {
throw new EncogError(
"Unknown architecture element: " + architecture + ", can't parse: " + part);
}
result.addLayer(new BasicLayer(af, bias, layer.getCount()));
}
}
result.getStructure().finalizeStructure();
result.reset();
return result;
}
/**
* Parse parameters.
*
* @param line The line to parse.
* @return The parsed values.
*/
public static Map<String, String> parseParams(final String line) {
final Map<String, String> result = new HashMap<String, String>();
final SimpleParser parser = new SimpleParser(line);
while (!parser.eol()) {
final String name = ArchitectureParse.parseName(parser).toUpperCase();
parser.eatWhiteSpace();
if (!parser.lookAhead("=", false)) {
throw new EncogError("Missing equals(=) operator.");
} else {
parser.advance();
}
final String value = ArchitectureParse.parseValue(parser);
result.put(name.toUpperCase(), value);
if (!parser.parseThroughComma()) {
break;
}
}
return result;
}
/**
* Create a NEAT population.
*
* @param architecture The architecture string to use.
* @param input The input count.
* @param output The output count.
* @return The population.
*/
public MLMethod create(final String architecture, final int input, final int output) {
if (input <= 0) {
throw new EncogError("Must have at least one input for NEAT.");
}
if (output <= 0) {
throw new EncogError("Must have at least one output for NEAT.");
}
final Map<String, String> args = ArchitectureParse.parseParams(architecture);
final ParamsHolder holder = new ParamsHolder(args);
final int populationSize = holder.getInt(MLMethodFactory.PROPERTY_POPULATION_SIZE, false, 1000);
final int cycles =
holder.getInt(MLMethodFactory.PROPERTY_CYCLES, false, NEATPopulation.DEFAULT_CYCLES);
ActivationFunction af =
this.factory.create(
holder.getString(MLMethodFactory.PROPERTY_AF, false, MLActivationFactory.AF_SSIGMOID));
NEATPopulation pop = new NEATPopulation(input, output, populationSize);
pop.reset();
pop.setActivationCycles(cycles);
pop.setNEATActivationFunction(af);
return pop;
}
/**
* Create a LMA trainer.
*
* @param method The method to use.
* @param training The training data to use.
* @param argsStr The arguments to use.
* @return The newly created trainer.
*/
public MLTrain create(final MLMethod method, final MLDataSet training, final String argsStr) {
if (!(method instanceof SOM)) {
throw new EncogError(
"Neighborhood training cannot be used on a method of type: "
+ method.getClass().getName());
}
final Map<String, String> args = ArchitectureParse.parseParams(argsStr);
final ParamsHolder holder = new ParamsHolder(args);
final double learningRate = holder.getDouble(MLTrainFactory.PROPERTY_LEARNING_RATE, false, 0.7);
final String neighborhoodStr =
holder.getString(MLTrainFactory.PROPERTY_NEIGHBORHOOD, false, "rbf");
final String rbfTypeStr = holder.getString(MLTrainFactory.PROPERTY_RBF_TYPE, false, "gaussian");
RBFEnum t;
if (rbfTypeStr.equalsIgnoreCase("Gaussian")) {
t = RBFEnum.Gaussian;
} else if (rbfTypeStr.equalsIgnoreCase("Multiquadric")) {
t = RBFEnum.Multiquadric;
} else if (rbfTypeStr.equalsIgnoreCase("InverseMultiquadric")) {
t = RBFEnum.InverseMultiquadric;
} else if (rbfTypeStr.equalsIgnoreCase("MexicanHat")) {
t = RBFEnum.MexicanHat;
} else {
t = RBFEnum.Gaussian;
}
NeighborhoodFunction nf = null;
if (neighborhoodStr.equalsIgnoreCase("bubble")) {
nf = new NeighborhoodBubble(1);
} else if (neighborhoodStr.equalsIgnoreCase("rbf")) {
final String str = holder.getString(MLTrainFactory.PROPERTY_DIMENSIONS, true, null);
final int[] size = NumberList.fromListInt(CSVFormat.EG_FORMAT, str);
nf = new NeighborhoodRBF(size, t);
} else if (neighborhoodStr.equalsIgnoreCase("rbf1d")) {
nf = new NeighborhoodRBF1D(t);
}
if (neighborhoodStr.equalsIgnoreCase("single")) {
nf = new NeighborhoodSingle();
}
final BasicTrainSOM result = new BasicTrainSOM((SOM) method, learningRate, training, nf);
if (args.containsKey(MLTrainFactory.PROPERTY_ITERATIONS)) {
final int plannedIterations = holder.getInt(MLTrainFactory.PROPERTY_ITERATIONS, false, 1000);
final double startRate =
holder.getDouble(MLTrainFactory.PROPERTY_START_LEARNING_RATE, false, 0.05);
final double endRate =
holder.getDouble(MLTrainFactory.PROPERTY_END_LEARNING_RATE, false, 0.05);
final double startRadius = holder.getDouble(MLTrainFactory.PROPERTY_START_RADIUS, false, 10);
final double endRadius = holder.getDouble(MLTrainFactory.PROPERTY_END_RADIUS, false, 1);
result.setAutoDecay(plannedIterations, startRate, endRate, startRadius, endRadius);
}
return result;
}
/**
* Create a Manhattan trainer.
*
* @param method The method to use.
* @param training The training data to use.
* @param argsStr The arguments to use.
* @return The newly created trainer.
*/
public final MLTrain create(
final MLMethod method, final MLDataSet training, final String argsStr) {
final Map<String, String> args = ArchitectureParse.parseParams(argsStr);
final ParamsHolder holder = new ParamsHolder(args);
final double learningRate = holder.getDouble(MLTrainFactory.PROPERTY_LEARNING_RATE, false, 0.1);
return new ManhattanPropagation((BasicNetwork) method, training, learningRate);
}
/**
* parse a layer.
*
* @param line The line to parse.
* @param defaultValue The default value.
* @return The parsed ArchitectureLayer.
*/
public static ArchitectureLayer parseLayer(final String line, final int defaultValue) {
final ArchitectureLayer layer = new ArchitectureLayer();
String check = line.trim().toUpperCase();
// first check for bias
if (check.endsWith(":B")) {
check = check.substring(0, check.length() - 2);
layer.setBias(true);
}
// see if simple number
try {
layer.setCount(Integer.parseInt(check));
if (layer.getCount() < 0) {
throw new EncogError("Count cannot be less than zero.");
}
} catch (final NumberFormatException f) {
// don't really care! Just checking to see if its a number.
}
// see if it is a default
if ("?".equals(check)) {
if (defaultValue < 0) {
throw new EncogError("Default (?) in an invalid location.");
} else {
layer.setCount(defaultValue);
layer.setUsedDefault(true);
return layer;
}
}
// single item, no function
final int startIndex = check.indexOf('(');
final int endIndex = check.lastIndexOf(')');
if (startIndex == -1) {
layer.setName(check);
return layer;
}
// function
if (endIndex == -1) {
throw new EncogError("Illegal parentheses.");
}
layer.setName(check.substring(0, startIndex).trim());
final String paramStr = check.substring(startIndex + 1, endIndex);
final Map<String, String> params = ArchitectureParse.parseParams(paramStr);
layer.getParams().putAll(params);
return layer;
}
/**
* Create a RBF network.
*
* @param architecture THe architecture string to use.
* @param input The input count.
* @param output The output count.
* @return The RBF network.
*/
public MLMethod create(final String architecture, final int input, final int output) {
final List<String> layers = ArchitectureParse.parseLayers(architecture);
if (layers.size() != MAX_LAYERS) {
throw new EncogError("RBF Networks must have exactly three elements, " + "separated by ->.");
}
final ArchitectureLayer inputLayer = ArchitectureParse.parseLayer(layers.get(0), input);
final ArchitectureLayer rbfLayer = ArchitectureParse.parseLayer(layers.get(1), -1);
final ArchitectureLayer outputLayer = ArchitectureParse.parseLayer(layers.get(2), output);
final int inputCount = inputLayer.getCount();
final int outputCount = outputLayer.getCount();
RBFEnum t;
if (rbfLayer.getName().equalsIgnoreCase("Gaussian")) {
t = RBFEnum.Gaussian;
} else if (rbfLayer.getName().equalsIgnoreCase("Multiquadric")) {
t = RBFEnum.Multiquadric;
} else if (rbfLayer.getName().equalsIgnoreCase("InverseMultiquadric")) {
t = RBFEnum.InverseMultiquadric;
} else if (rbfLayer.getName().equalsIgnoreCase("MexicanHat")) {
t = RBFEnum.MexicanHat;
} else {
throw new NeuralNetworkError("Unknown RBF: " + rbfLayer.getName());
}
final ParamsHolder holder = new ParamsHolder(rbfLayer.getParams());
final int rbfCount = holder.getInt("C", true, 0);
final RBFNetwork result = new RBFNetwork(inputCount, rbfCount, outputCount, t);
return result;
}
/**
* Create a SVM trainer.
*
* @param method The method to use.
* @param training The training data to use.
* @param argsStr The arguments to use.
* @return The newly created trainer.
*/
public MLTrain create(final MLMethod method, final MLDataSet training, final String argsStr) {
if (!(method instanceof SVM)) {
throw new EncogError(
"SVM Train training cannot be used on a method of type: " + method.getClass().getName());
}
final double defaultGamma = 1.0 / ((SVM) method).getInputCount();
final double defaultC = 1.0;
final Map<String, String> args = ArchitectureParse.parseParams(argsStr);
final ParamsHolder holder = new ParamsHolder(args);
final double gamma = holder.getDouble(MLTrainFactory.PROPERTY_GAMMA, false, defaultGamma);
final double c = holder.getDouble(MLTrainFactory.PROPERTY_C, false, defaultC);
final SVMTrain result = new SVMTrain((SVM) method, training);
result.setGamma(gamma);
result.setC(c);
return result;
}
/**
* Create a SVM trainer.
*
* @param method The method to use.
* @param training The training data to use.
* @param argsStr The arguments to use.
* @return The newly created trainer.
*/
public MLTrain create(final MLMethod method, final MLDataSet training, final String argsStr) {
if (!(method instanceof SVM)) {
throw new EncogError(
"SVM Train training cannot be used on a method of type: " + method.getClass().getName());
}
final Map<String, String> args = ArchitectureParse.parseParams(argsStr);
new ParamsHolder(args);
final ParamsHolder holder = new ParamsHolder(args);
final double gammaStart =
holder.getDouble(
SVMSearchFactory.PROPERTY_GAMMA1, false, SVMSearchTrain.DEFAULT_GAMMA_BEGIN);
final double cStart =
holder.getDouble(SVMSearchFactory.PROPERTY_C1, false, SVMSearchTrain.DEFAULT_CONST_BEGIN);
final double gammaStop =
holder.getDouble(SVMSearchFactory.PROPERTY_GAMMA2, false, SVMSearchTrain.DEFAULT_GAMMA_END);
final double cStop =
holder.getDouble(SVMSearchFactory.PROPERTY_C2, false, SVMSearchTrain.DEFAULT_CONST_END);
final double gammaStep =
holder.getDouble(
SVMSearchFactory.PROPERTY_GAMMA_STEP, false, SVMSearchTrain.DEFAULT_GAMMA_STEP);
final double cStep =
holder.getDouble(
SVMSearchFactory.PROPERTY_C_STEP, false, SVMSearchTrain.DEFAULT_CONST_STEP);
final SVMSearchTrain result = new SVMSearchTrain((SVM) method, training);
result.setGammaBegin(gammaStart);
result.setGammaEnd(gammaStop);
result.setGammaStep(gammaStep);
result.setConstBegin(cStart);
result.setConstEnd(cStop);
result.setConstStep(cStep);
return result;
}
/**
* Create the SVM.
*
* @param architecture The architecture string.
* @param input The input count.
* @param output The output count.
* @return The newly created SVM.
*/
public final MLMethod create(final String architecture, final int input, final int output) {
final List<String> layers = ArchitectureParse.parseLayers(architecture);
if (layers.size() != MAX_LAYERS) {
throw new EncogError("SVM's must have exactly three elements, separated by ->.");
}
final ArchitectureLayer inputLayer = ArchitectureParse.parseLayer(layers.get(0), input);
final ArchitectureLayer paramsLayer = ArchitectureParse.parseLayer(layers.get(1), input);
final ArchitectureLayer outputLayer = ArchitectureParse.parseLayer(layers.get(2), output);
final String name = paramsLayer.getName();
final String kernelStr = paramsLayer.getParams().get("KERNEL");
final String svmTypeStr = paramsLayer.getParams().get("TYPE");
SVMType svmType = SVMType.NewSupportVectorClassification;
KernelType kernelType = KernelType.RadialBasisFunction;
boolean useNew = true;
if (svmTypeStr == null) {
useNew = true;
} else if (svmTypeStr.equalsIgnoreCase("NEW")) {
useNew = true;
} else if (svmTypeStr.equalsIgnoreCase("OLD")) {
useNew = false;
} else {
throw new EncogError("Unsupported type: " + svmTypeStr + ", must be NEW or OLD.");
}
if (name.equalsIgnoreCase("C")) {
if (useNew) {
svmType = SVMType.NewSupportVectorClassification;
} else {
svmType = SVMType.SupportVectorClassification;
}
} else if (name.equalsIgnoreCase("R")) {
if (useNew) {
svmType = SVMType.NewSupportVectorRegression;
} else {
svmType = SVMType.EpsilonSupportVectorRegression;
}
} else {
throw new EncogError(
"Unsupported mode: " + name + ", must be C for classify or R for regression.");
}
if (kernelStr == null) {
kernelType = KernelType.RadialBasisFunction;
} else if ("linear".equalsIgnoreCase(kernelStr)) {
kernelType = KernelType.Linear;
} else if ("poly".equalsIgnoreCase(kernelStr)) {
kernelType = KernelType.Poly;
} else if ("precomputed".equalsIgnoreCase(kernelStr)) {
kernelType = KernelType.Precomputed;
} else if ("rbf".equalsIgnoreCase(kernelStr)) {
kernelType = KernelType.RadialBasisFunction;
} else if ("sigmoid".equalsIgnoreCase(kernelStr)) {
kernelType = KernelType.Sigmoid;
} else {
throw new EncogError(
"Unsupported kernel: " + kernelStr + ", must be linear,poly,precomputed,rbf or sigmoid.");
}
final int inputCount = inputLayer.getCount();
final int outputCount = outputLayer.getCount();
if (outputCount != 1) {
throw new EncogError("SVM can only have an output size of 1.");
}
final SVM result = new SVM(inputCount, svmType, kernelType);
return result;
}