/*
* Encog(tm) Core v3.4 - Java Version
* http://www.heatonresearch.com/encog/
* https://github.com/encog/encog-java-core
* Copyright 2008-2016 Heaton Research, Inc.
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*
* For more information on Heaton Research copyrights, licenses
* and trademarks visit:
* http://www.heatonresearch.com/copyright
*/
package org.encog.ml.model;
import java.util.ArrayList;
import java.util.HashMap;
import java.util.List;
import java.util.Map;
import org.encog.EncogError;
import org.encog.NullStatusReportable;
import org.encog.StatusReportable;
import org.encog.mathutil.randomize.generate.MersenneTwisterGenerateRandom;
import org.encog.ml.MLClassification;
import org.encog.ml.MLMethod;
import org.encog.ml.MLRegression;
import org.encog.ml.TrainingImplementationType;
import org.encog.ml.data.MLDataSet;
import org.encog.ml.data.cross.DataFold;
import org.encog.ml.data.cross.KFoldCrossvalidation;
import org.encog.ml.data.versatile.MatrixMLDataSet;
import org.encog.ml.data.versatile.VersatileMLDataSet;
import org.encog.ml.data.versatile.columns.ColumnDefinition;
import org.encog.ml.data.versatile.columns.ColumnType;
import org.encog.ml.data.versatile.division.DataDivision;
import org.encog.ml.factory.MLMethodFactory;
import org.encog.ml.factory.MLTrainFactory;
import org.encog.ml.model.config.FeedforwardConfig;
import org.encog.ml.model.config.MethodConfig;
import org.encog.ml.model.config.NEATConfig;
import org.encog.ml.model.config.PNNConfig;
import org.encog.ml.model.config.RBFNetworkConfig;
import org.encog.ml.model.config.SVMConfig;
import org.encog.ml.train.MLTrain;
import org.encog.ml.train.strategy.StopTrainingStrategy;
import org.encog.ml.train.strategy.end.EarlyStoppingStrategy;
import org.encog.util.Format;
import org.encog.util.simple.EncogUtility;
/**
* Encog model is designed to allow you to easily swap between different model
* types and automatically normalize data. It is designed to work with a
* VersatileMLDataSet only.
*/
public class EncogModel {
/**
* The dataset to use.
*/
private final VersatileMLDataSet dataset;
/**
* The input features.
*/
private final List<ColumnDefinition> inputFeatures = new ArrayList<ColumnDefinition>();
/**
* The predicted features.
*/
private final List<ColumnDefinition> predictedFeatures = new ArrayList<ColumnDefinition>();
/**
* The training dataset.
*/
private MatrixMLDataSet trainingDataset;
/**
* The validation dataset.
*/
private MatrixMLDataSet validationDataset;
/**
* The standard configrations for each method type.
*/
private final Map<String, MethodConfig> methodConfigurations = new HashMap<String, MethodConfig>();
/**
* The current method configuration, determined by the selected model.
*/
private MethodConfig config;
/**
* The selected method type.
*/
private String methodType;
/**
* The method arguments for the selected method.
*/
private String methodArgs;
/**
* The selected training type.
*/
private String trainingType;
/**
* The training arguments for the selected training type.
*/
private String trainingArgs;
/**
* The report.
*/
private StatusReportable report = new NullStatusReportable();
/**
* Construct a model for the specified dataset.
* @param theDataset The dataset.
*/
public EncogModel(VersatileMLDataSet theDataset) {
this.dataset = theDataset;
this.methodConfigurations.put(MLMethodFactory.TYPE_FEEDFORWARD,
new FeedforwardConfig());
this.methodConfigurations
.put(MLMethodFactory.TYPE_SVM, new SVMConfig());
this.methodConfigurations.put(MLMethodFactory.TYPE_RBFNETWORK,
new RBFNetworkConfig());
this.methodConfigurations.put(MLMethodFactory.TYPE_NEAT,
new NEATConfig());
this.methodConfigurations
.put(MLMethodFactory.TYPE_PNN, new PNNConfig());
}
/**
* @return the dataset
*/
public VersatileMLDataSet getDataset() {
return dataset;
}
/**
* @return the inputFeatures
*/
public List<ColumnDefinition> getInputFeatures() {
return inputFeatures;
}
/**
* @return the predictedFeatures
*/
public List<ColumnDefinition> getPredictedFeatures() {
return predictedFeatures;
}
/**
* Specify a validation set to hold back.
* @param validationPercent The percent to use for validation.
* @param shuffle True to shuffle.
* @param seed The seed for random generation.
*/
public void holdBackValidation(double validationPercent, boolean shuffle,
int seed) {
List<DataDivision> dataDivisionList = new ArrayList<DataDivision>();
dataDivisionList.add(new DataDivision(1.0 - validationPercent));// Training
dataDivisionList.add(new DataDivision(validationPercent));// Validation
this.dataset.divide(dataDivisionList, shuffle,
new MersenneTwisterGenerateRandom(seed));
this.trainingDataset = dataDivisionList.get(0).getDataset();
this.validationDataset = dataDivisionList.get(1).getDataset();
}
/**
* Fit the model using cross validation.
* @param k The number of folds total.
* @param foldNum The current fold.
* @param fold The current fold.
*/
private void fitFold(int k, int foldNum, DataFold fold) {
MLMethod method = this.createMethod();
MLTrain train = this.createTrainer(method, fold.getTraining());
if (train.getImplementationType() == TrainingImplementationType.Iterative) {
EarlyStoppingStrategy earlyStop = new EarlyStoppingStrategy(
fold.getValidation());
train.addStrategy(earlyStop);
StringBuilder line = new StringBuilder();
while (!train.isTrainingDone()) {
train.iteration();
line.setLength(0);
line.append("Fold #");
line.append(foldNum);
line.append("/");
line.append(k);
line.append(": Iteration #");
line.append(train.getIteration());
line.append(", Training Error: ");
line.append(Format.formatDouble(train.getError(), 8));
line.append(", Validation Error: ");
line.append(Format.formatDouble(earlyStop.getValidationError(),
8));
report.report(k, foldNum, line.toString());
}
fold.setScore(earlyStop.getValidationError());
fold.setMethod(method);
} else if (train.getImplementationType() == TrainingImplementationType.OnePass) {
train.iteration();
double validationError = calculateError(method,
fold.getValidation());
this.report.report(k, k,
"Trained, Training Error: " + train.getError()
+ ", Validatoin Error: " + validationError);
fold.setScore(validationError);
fold.setMethod(method);
} else {
throw new EncogError("Unsupported training type for EncogModel: "
+ train.getImplementationType());
}
}
/**
* Calculate the error for the given method and dataset.
* @param method The method to use.
* @param data The data to use.
* @return The error.
*/
public double calculateError(MLMethod method, MLDataSet data) {
if (this.dataset.getNormHelper().getOutputColumns().size() == 1) {
ColumnDefinition cd = this.dataset.getNormHelper()
.getOutputColumns().get(0);
if (cd.getDataType() == ColumnType.nominal) {
return EncogUtility.calculateClassificationError(
(MLClassification) method, data);
}
}
return EncogUtility.calculateRegressionError((MLRegression) method,
data);
}
/**
* Create a trainer.
* @param method The method to train.
* @param dataset The dataset.
* @return The trainer.
*/
private MLTrain createTrainer(MLMethod method, MLDataSet dataset) {
if (this.trainingType == null) {
throw new EncogError(
"Please call selectTraining first to choose how to train.");
}
MLTrainFactory trainFactory = new MLTrainFactory();
MLTrain train = trainFactory.create(method, dataset, this.trainingType,
this.trainingArgs);
return train;
}
/**
* Crossvalidate and fit.
* @param k The number of folds.
* @param shuffle True if we should shuffle.
* @return The trained method.
*/
public MLMethod crossvalidate(int k, boolean shuffle) {
KFoldCrossvalidation cross = new KFoldCrossvalidation(
this.trainingDataset, k);
cross.process(shuffle);
int foldNumber = 0;
for (DataFold fold : cross.getFolds()) {
foldNumber++;
report.report(k, foldNumber, "Fold #" + foldNumber);
fitFold(k, foldNumber, fold);
}
double sum = 0;
double bestScore = Double.POSITIVE_INFINITY;
MLMethod bestMethod = null;
for (DataFold fold : cross.getFolds()) {
sum += fold.getScore();
if (fold.getScore() < bestScore) {
bestScore = fold.getScore();
bestMethod = fold.getMethod();
}
}
sum = sum / cross.getFolds().size();
report.report(k, k, "Cross-validated score:" + sum);
return bestMethod;
}
/**
* @return the trainingDataset
*/
public MatrixMLDataSet getTrainingDataset() {
return trainingDataset;
}
/**
* @param trainingDataset
* the trainingDataset to set
*/
public void setTrainingDataset(MatrixMLDataSet trainingDataset) {
this.trainingDataset = trainingDataset;
}
/**
* @return the validationDataset
*/
public MatrixMLDataSet getValidationDataset() {
return validationDataset;
}
/**
* @param validationDataset
* the validationDataset to set
*/
public void setValidationDataset(MatrixMLDataSet validationDataset) {
this.validationDataset = validationDataset;
}
/**
* Select the method to use.
* @param dataset The dataset.
* @param methodType The type of method.
* @param methodArgs The method arguments.
* @param trainingType The training type.
* @param trainingArgs The training arguments.
*/
public void selectMethod(VersatileMLDataSet dataset, String methodType,
String methodArgs, String trainingType, String trainingArgs) {
if (!this.methodConfigurations.containsKey(methodType)) {
throw new EncogError("Don't know how to autoconfig method: "
+ methodType);
}
this.methodType = methodType;
this.methodArgs = methodArgs;
this.config = this.methodConfigurations.get(methodType);
dataset.getNormHelper().setStrategy(
this.methodConfigurations.get(methodType)
.suggestNormalizationStrategy(dataset, methodArgs));
}
/**
* Create the selected method.
* @return The created method.
*/
public MLMethod createMethod() {
if (this.methodType == null) {
throw new EncogError(
"Please call selectMethod first to choose what type of method you wish to use.");
}
MLMethodFactory methodFactory = new MLMethodFactory();
MLMethod method = methodFactory.create(methodType, methodArgs, dataset
.getNormHelper().calculateNormalizedInputCount(), this.config
.determineOutputCount(dataset));
return method;
}
/**
* Select the method to create.
* @param dataset The dataset.
* @param methodType The method type.
*/
public void selectMethod(VersatileMLDataSet dataset, String methodType) {
if (!this.methodConfigurations.containsKey(methodType)) {
throw new EncogError("Don't know how to autoconfig method: "
+ methodType);
}
this.config = this.methodConfigurations.get(methodType);
this.methodType = methodType;
this.methodArgs = this.config.suggestModelArchitecture(dataset);
dataset.getNormHelper().setStrategy(
this.config.suggestNormalizationStrategy(dataset, methodArgs));
}
/**
* Select the training type.
* @param dataset The dataset.
*/
public void selectTrainingType(VersatileMLDataSet dataset) {
if (this.methodType == null) {
throw new EncogError(
"Please select your training method, before your training type.");
}
MethodConfig config = this.methodConfigurations.get(methodType);
selectTraining(dataset, config.suggestTrainingType(),
config.suggestTrainingArgs(trainingType));
}
/**
* Select the training to use.
* @param dataset The dataset.
* @param trainingType The type of training.
* @param trainingArgs The training arguments.
*/
public void selectTraining(VersatileMLDataSet dataset, String trainingType,
String trainingArgs) {
if (this.methodType == null) {
throw new EncogError(
"Please select your training method, before your training type.");
}
this.trainingType = trainingType;
this.trainingArgs = trainingArgs;
}
/**
* @return the methodConfigurations
*/
public Map<String, MethodConfig> getMethodConfigurations() {
return methodConfigurations;
}
/**
* @return the report
*/
public StatusReportable getReport() {
return report;
}
/**
* @param report
* the report to set
*/
public void setReport(StatusReportable report) {
this.report = report;
}
public MLRegression train(boolean shuffle) {
MLMethod method = this.createMethod();
MLTrain train = this.createTrainer(method, this.trainingDataset);
if (train.getImplementationType() == TrainingImplementationType.Iterative) {
StopTrainingStrategy earlyStop = new StopTrainingStrategy(0.01,10);
train.addStrategy(earlyStop);
StringBuilder line = new StringBuilder();
while (!train.isTrainingDone()) {
train.iteration();
line.setLength(0);
line.append("Iteration #");
line.append(train.getIteration());
line.append(", Training Error: ");
line.append(Format.formatDouble(train.getError(), 8));
report.report(0, 0, line.toString());
}
} else if (train.getImplementationType() == TrainingImplementationType.OnePass) {
train.iteration();
this.report.report(0, 0,
"Trained, Training Error: " + train.getError());
} else {
throw new EncogError("Unsupported training type for EncogModel: "
+ train.getImplementationType());
}
return (MLRegression)method;
}
public MLRegression train(int iterations, boolean shuffle) {
MLMethod method = this.createMethod();
MLTrain train = this.createTrainer(method, this.trainingDataset);
if (train.getImplementationType() == TrainingImplementationType.Iterative) {
StopTrainingStrategy earlyStop = new StopTrainingStrategy(0.01,10);
train.addStrategy(earlyStop);
StringBuilder line = new StringBuilder();
for(int i=1;i<=iterations;i++) {
train.iteration();
line.setLength(0);
line.append("Iteration #");
line.append(train.getIteration());
line.append("/");
line.append(iterations);
line.append(", Training Error: ");
line.append(Format.formatDouble(train.getError(), 8));
report.report(0, 0, line.toString());
}
} else if (train.getImplementationType() == TrainingImplementationType.OnePass) {
train.iteration();
this.report.report(0, 0,
"Trained, Training Error: " + train.getError());
} else {
throw new EncogError("Unsupported training type for EncogModel: "
+ train.getImplementationType());
}
return (MLRegression)method;
}
}