/*
* Encog(tm) Core v3.0 - Java Version
* http://www.heatonresearch.com/encog/
* http://code.google.com/p/encog-java/
* Copyright 2008-2011 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.neural.prune;
import java.util.ArrayList;
import java.util.List;
import org.encog.EncogError;
import org.encog.StatusReportable;
import org.encog.ml.data.MLDataSet;
import org.encog.ml.data.buffer.BufferedNeuralDataSet;
import org.encog.ml.train.strategy.StopTrainingStrategy;
import org.encog.neural.networks.BasicNetwork;
import org.encog.neural.networks.training.propagation.Propagation;
import org.encog.neural.networks.training.propagation.resilient.ResilientPropagation;
import org.encog.neural.pattern.NeuralNetworkPattern;
import org.encog.util.concurrency.job.ConcurrentJob;
import org.encog.util.concurrency.job.JobUnitContext;
import org.encog.util.logging.EncogLogging;
/**
* This class is used to help determine the optimal configuration for the hidden
* layers of a neural network. It can accept a pattern, which specifies the type
* of neural network to create, and a list of the maximum and minimum hidden
* layer neurons. It will then attempt to train the neural network at all
* configurations and see which hidden neuron counts work the best.
*
* This method does not simply choose the network with the lowest error rate. A
* specifiable number of best networks are kept, which represent the networks
* with the lowest error rates. From this collection of networks, the best
* network is defined to be the one with the fewest number of connections.
*
* Not all starting random weights are created equal. Because of this, an option
* is provided to allow you to choose how many attempts you want the process to
* make, with different weights. All random weights are created using the
* default Nguyen-Widrow method normally used by Encog.
*
*/
public class PruneIncremental extends ConcurrentJob {
/**
* Format the network as a human readable string that lists the hidden
* layers.
*
* @param network
* The network to format.
* @return A human readable string.
*/
public static String networkToString(final BasicNetwork network) {
final StringBuilder result = new StringBuilder();
int num = 1;
// display only hidden layers
for (int i = 1; i < network.getLayerCount() - 1; i++) {
if (result.length() > 0) {
result.append(",");
}
result.append("H");
result.append(num++);
result.append("=");
result.append(network.getLayerNeuronCount(i));
}
return result.toString();
}
/**
* Are we done?
*/
private boolean done = false;
/**
* The training set to use as different neural networks are evaluated.
*/
private final MLDataSet training;
/**
* The pattern for which type of neural network we would like to create.
*/
private final NeuralNetworkPattern pattern;
/**
* The ranges for the hidden layers.
*/
private final List<HiddenLayerParams> hidden = new ArrayList<HiddenLayerParams>();
/**
* The number if training iterations that should be tried for each network.
*/
private final int iterations;
/**
* An array of the top networks.
*/
private final BasicNetwork[] topNetworks;
/**
* An array of the top errors.
*/
private final double[] topErrors;
/**
* The best network found so far.
*/
private BasicNetwork bestNetwork;
/**
* How many networks have been tried so far?
*/
private int currentTry;
/**
* The object that status should be reported to.
*/
@SuppressWarnings("unused")
private final StatusReportable report;
/**
* Keeps track of how many neurons in each hidden layer as training the
* evaluation progresses.
*/
private int[] hiddenCounts;
/**
* The current highest error.
*/
private double high;
/**
* The current lowest error.
*/
private double low;
/**
* The results in a 2d array.
*/
private double[][] results;
/**
* The size of the first hidden layer.
*/
private int hidden1Size;
/**
* The size of the second hidden layer.
*/
private int hidden2Size;
/**
* The number of tries with random weights.
*/
private final int weightTries;
/**
* Construct an object to determine the optimal number of hidden layers and
* neurons for the specified training data and pattern.
*
* @param training
* The training data to use.
* @param pattern
* The network pattern to use to solve this data.
* @param iterations
* How many iterations to try per network.
* @param weightTries
* The number of random weights to use.
* @param numTopResults
* The number of "top networks" to choose the most simple "best
* network" from.
* @param report
* Object used to report status to.
*/
public PruneIncremental(final MLDataSet training,
final NeuralNetworkPattern pattern, final int iterations,
final int weightTries, final int numTopResults,
final StatusReportable report) {
super(report);
this.training = training;
this.pattern = pattern;
this.iterations = iterations;
this.report = report;
this.weightTries = weightTries;
this.topNetworks = new BasicNetwork[numTopResults];
this.topErrors = new double[numTopResults];
}
/**
* Add a hidden layer's min and max. Call this once per hidden layer.
* Specify a zero min if it is possible to remove this hidden layer.
*
* @param min
* The minimum number of neurons for this layer.
* @param max
* The maximum number of neurons for this layer.
*/
public final void addHiddenLayer(final int min, final int max) {
final HiddenLayerParams param = new HiddenLayerParams(min, max);
this.hidden.add(param);
}
/**
* Generate a network according to the current hidden layer counts.
*
* @return The network based on current hidden layer counts.
*/
private BasicNetwork generateNetwork() {
this.pattern.clear();
for (final int element : this.hiddenCounts) {
if (element > 0) {
this.pattern.addHiddenLayer(element);
}
}
return (BasicNetwork) this.pattern.generate();
}
/**
* @return The network being processed.
*/
public final BasicNetwork getBestNetwork() {
return this.bestNetwork;
}
/**
* @return The hidden layer max and min.
*/
public final List<HiddenLayerParams> getHidden() {
return this.hidden;
}
/**
* @return The size of the first hidden layer.
*/
public final int getHidden1Size() {
return this.hidden1Size;
}
/**
* @return The size of the second hidden layer.
*/
public final int getHidden2Size() {
return this.hidden2Size;
}
/**
* @return The higest error so far.
*/
public final double getHigh() {
return this.high;
}
/**
* @return The number of training iterations to try for each network.
*/
public final int getIterations() {
return this.iterations;
}
/**
* @return The lowest error so far.
*/
public final double getLow() {
return this.low;
}
/**
* @return The network pattern to use.
*/
public final NeuralNetworkPattern getPattern() {
return this.pattern;
}
/**
* @return The error results.
*/
public final double[][] getResults() {
return this.results;
}
/**
* @return the topErrors
*/
public final double[] getTopErrors() {
return this.topErrors;
}
/**
* @return the topNetworks
*/
public final BasicNetwork[] getTopNetworks() {
return this.topNetworks;
}
/**
* @return The training set to use.
*/
public final MLDataSet getTraining() {
return this.training;
}
/**
* Increase the hidden layer counts according to the hidden layer
* parameters. Increase the first hidden layer count by one, if it is maxed
* out, then set it to zero and increase the next hidden layer.
*
* @return False if no more increases can be done, true otherwise.
*/
private boolean increaseHiddenCounts() {
int i = 0;
do {
final HiddenLayerParams param = this.hidden.get(i);
this.hiddenCounts[i]++;
// is this hidden layer still within the range?
if (this.hiddenCounts[i] <= param.getMax()) {
return true;
}
// increase the next layer if we've maxed out this one
this.hiddenCounts[i] = param.getMin();
i++;
} while (i < this.hiddenCounts.length);
// can't increase anymore, we're done!
return false;
}
/**
* Init for prune.
*/
public final void init() {
// handle display for one layer
if (this.hidden.size() == 1) {
this.hidden1Size = (this.hidden.get(0).getMax() - this.hidden
.get(0).getMin()) + 1;
this.hidden2Size = 0;
this.results = new double[this.hidden1Size][1];
} else if (this.hidden.size() == 2) {
// handle display for two layers
this.hidden1Size = (this.hidden.get(0).getMax() - this.hidden
.get(0).getMin()) + 1;
this.hidden2Size = (this.hidden.get(1).getMax() - this.hidden
.get(1).getMin()) + 1;
this.results = new double[this.hidden1Size][this.hidden2Size];
} else {
// we don't handle displays for more than two layers
this.hidden1Size = 0;
this.hidden2Size = 0;
this.results = null;
}
// reset min and max
this.high = Double.NEGATIVE_INFINITY;
this.low = Double.POSITIVE_INFINITY;
}
/**
* Get the next workload. This is the number of hidden neurons. This is the
* total amount of work to be processed.
*
* @return The amount of work to be processed by this.
*/
@Override
public final int loadWorkload() {
int result = 1;
for (final HiddenLayerParams param : this.hidden) {
result *= (param.getMax() - param.getMin()) + 1;
}
init();
return result;
}
/**
* Perform an individual job unit, which is a single network to train and
* evaluate.
*
* @param context
* Contains information about the job unit.
*/
@Override
public final void performJobUnit(final JobUnitContext context) {
final BasicNetwork network = (BasicNetwork) context.getJobUnit();
BufferedNeuralDataSet buffer = null;
MLDataSet useTraining = this.training;
if (this.training instanceof BufferedNeuralDataSet) {
buffer = (BufferedNeuralDataSet) this.training;
useTraining = buffer.openAdditional();
}
// train the neural network
double error = Double.POSITIVE_INFINITY;
for (int z = 0; z < this.weightTries; z++) {
network.reset();
final Propagation train = new ResilientPropagation(network,
useTraining);
final StopTrainingStrategy strat = new StopTrainingStrategy(0.001,
5);
train.addStrategy(strat);
train.setNumThreads(1); // force single thread mode
for (int i = 0; (i < this.iterations) && !getShouldStop()
&& !strat.shouldStop(); i++) {
train.iteration();
}
error = Math.min(error, train.getError());
}
if (buffer != null) {
buffer.close();
}
if (!getShouldStop()) {
// update min and max
this.high = Math.max(this.high, error);
this.low = Math.min(this.low, error);
if (this.hidden1Size > 0) {
int networkHidden1Count;
int networkHidden2Count;
if (network.getLayerCount() > 3) {
networkHidden2Count = network.getLayerNeuronCount(1);
networkHidden1Count = network.getLayerNeuronCount(2);
} else {
networkHidden2Count = 0;
networkHidden1Count = network.getLayerNeuronCount(1);
}
int row, col;
if (this.hidden2Size == 0) {
row = networkHidden1Count - this.hidden.get(0).getMin();
col = 0;
} else {
row = networkHidden1Count - this.hidden.get(0).getMin();
col = networkHidden2Count - this.hidden.get(1).getMin();
}
if ((row < 0) || (col < 0)) {
System.out.println("STOP");
}
this.results[row][col] = error;
}
// report status
this.currentTry++;
updateBest(network, error);
reportStatus(
context,
"Current: "
+ PruneIncremental.networkToString(network)
+ "; Best: "
+ PruneIncremental
.networkToString(this.bestNetwork));
}
}
/**
* Begin the prune process.
*/
@Override
public final void process() {
if (this.hidden.size() == 0) {
throw new EncogError(
"To calculate the optimal hidden size, at least "
+ "one hidden layer must be defined.");
}
this.hiddenCounts = new int[this.hidden.size()];
// set the best network
this.bestNetwork = null;
// set to minimums
int i = 0;
for (final HiddenLayerParams parm : this.hidden) {
this.hiddenCounts[i++] = parm.getMin();
}
// make sure hidden layer 1 has at least one neuron
if (this.hiddenCounts[0] == 0) {
throw new EncogError(
"To calculate the optimal hidden size, at least "
+ "one neuron must be the minimum for the first hidden layer.");
}
super.process();
}
/**
* Request the next task. This is the next network to attempt to train.
*
* @return The next network to train.
*/
@Override
public final Object requestNextTask() {
if (this.done || getShouldStop()) {
return null;
}
final BasicNetwork network = generateNetwork();
if (!increaseHiddenCounts()) {
this.done = true;
}
return network;
}
/**
* Update the best network.
*
* @param network
* The network to consider.
* @param error
* The error for this network.
*/
private synchronized void updateBest(final BasicNetwork network,
final double error) {
this.high = Math.max(this.high, error);
this.low = Math.min(this.low, error);
int selectedIndex = -1;
// find a place for this in the top networks, if it is a top network
for (int i = 0; i < this.topNetworks.length; i++) {
if (this.topNetworks[i] == null) {
selectedIndex = i;
break;
} else if (this.topErrors[i] > error) {
// this network might be worth replacing, see if the one
// already selected is a better option.
if ((selectedIndex == -1)
|| (this.topErrors[selectedIndex] < this.topErrors[i])) {
selectedIndex = i;
}
}
}
// replace the selected index
if (selectedIndex != -1) {
this.topErrors[selectedIndex] = error;
this.topNetworks[selectedIndex] = network;
}
// now select the best network, which is the most simple of the
// top networks.
BasicNetwork choice = null;
for (final BasicNetwork n : this.topNetworks) {
if (n == null) {
continue;
}
if (choice == null) {
choice = n;
} else {
if (n.getStructure().calculateSize() < choice.getStructure()
.calculateSize()) {
choice = n;
}
}
}
if (choice != this.bestNetwork) {
EncogLogging.log(EncogLogging.LEVEL_DEBUG,
"Prune found new best network: error=" + error
+ ", network=" + choice);
}
}
}