User:Chase-san/KohonenMap

This is my implementation of a Self-organizing map. It is untested, but it should work just fine.

org.csdgn.nn.SOM

package org.csdgn.nn;

import java.util.Arrays;
import java.util.Random;

import org.csdgn.nn.density.StandardDensity;
import org.csdgn.nn.distance.EulerDistanceSquared;
import org.csdgn.utils.VDA;

/**
 * An Optimized Self-Organizing Map implementation.
 * This makes use of the VDA class.
 * 
 * @author Chase
 * 
 */
public class SOM {
	private static final double cutoff = 1.0e-4;
	private final VDA<double[]> vda;
	private final int inputSize;
	private DensityFunction density = new StandardDensity();
	private DistanceFunction distance = new EulerDistanceSquared();
	private boolean wrap = false;
	private double learningRate = 0.8;
	private int[] BMU;

	/**
	 * @param mapSize
	 *            Size of the neighborhood. Example: {10,10} produces a 2
	 *            dimensional map, each dimension having 10 nodes. Total nodes
	 *            would be 100.
	 * @param input
	 *            The length of the input vector (2D only)
	 * @param output
	 *            The length of the output vector (2D only)
	 */
	public SOM(int[] mapSize, int input, int output) {
		this.vda = new VDA<double[]>(mapSize);
		this.inputSize = input;
		Object[] obj = vda.getBackingArray();
		for(int i = 0; i < obj.length; ++i) {
			obj[i] = new double[input + output];
		}
	}

	/**
	 * Initializes the map to random values
	 */
	public final void initialize() {
		Random r = new Random();
		initialize(r);
	}

	/**
	 * Initializes the map with the given random function. Uses the nextDouble
	 * function.
	 */
	public final void initialize(Random random) {
		Object[] tmp = vda.getBackingArray();
		for(int i = 0; i < tmp.length; ++i) {
			double[] tmp2 = ((double[])tmp[i]);
			for(int j = 0; j < tmp2.length; ++j) {
				tmp2[j] = random.nextDouble();
			}
		}
	}

	/**
	 * Finds the Best Matching Unit for the given input.
	 */
	public final void findBMInput(double[] input) {
		if(input.length != inputSize)
			return;

		double bestDistance = Double.MAX_VALUE;
		Object[] backArray = vda.getBackingArray();
		int[] size = vda.getSize();
		int[] pos = new int[size.length];
		for(Object o : backArray) {
			double[] array = (double[])o;
			double dist = distance.calculate(array, input);
			if(dist < bestDistance) {
				bestDistance = dist;
				BMU = pos.clone();
			}
			next(pos, size);
		}

		// return BMU;
		// return BMU.clone();
	}

	/**
	 * Finds the Worst Matching Unit for the given input.
	 */
	public final void findWMInput(double[] input) {
		if(input.length != inputSize)
			return;

		double worstDistance = Double.MIN_VALUE;
		Object[] backArray = vda.getBackingArray();
		int[] size = vda.getSize();
		int[] pos = new int[size.length];
		for(Object o : backArray) {
			double[] array = (double[])o;
			double dist = distance.calculate(array, input);
			if(dist > worstDistance) {
				worstDistance = dist;
				BMU = pos.clone();
			}
			next(pos, size);
		}
	}

	public final void setMatchingIndex(int... index) {
		if(index.length != vda.getSize().length)
			throw new IllegalArgumentException("Incorrect or Bad Dimensionality.");
		BMU = index.clone();
	}

	/**
	 * This returns the input of the last found BMU or WMU.
	 * 
	 * @return the input vector
	 */
	public final double[] getInput() {
		if(BMU == null)
			throw new UnsupportedOperationException("BMU/WMU must be found first.");
		return Arrays.copyOf(vda.get(BMU), inputSize);
	}

	/**
	 * This returns the output of the last found BMU or WMU.
	 * 
	 * @return the output vector
	 */
	public final double[] getOutput() {
		if(BMU == null)
			throw new UnsupportedOperationException("BMU/WMU must be found first.");
		double[] bmu = vda.get(BMU);
		double[] output = new double[bmu.length - inputSize];
		System.arraycopy(bmu, inputSize, output, 0, bmu.length - inputSize);
		return output;
	}

	/**
	 * Sets the learning rate of this KohonenMap
	 * 
	 * @param rate
	 *            value between 0 and 1
	 */
	public final void setLearningRate(double rate) {
		learningRate = Math.max(Math.min(rate, 1), 0);
	}

	/**
	 * Returns the current rate of learning
	 * 
	 * @return the learning rate
	 */
	public final double getLearningRate() {
		return learningRate;
	}

	/**
	 * Sets the map to wrap its updates (slightly more costly)
	 */
	public final void setWraps(boolean n) {
		wrap = n;
	}

	/**
	 * Returns if the current map wraps
	 * 
	 * @return
	 */
	public final boolean isWrapping() {
		return wrap;
	}

	/**
	 * Sets the density function this map uses for updating nearby nodes. If
	 * unset it uses the StandardDensity class.
	 * 
	 * @param func
	 *            the Density Function
	 */
	public final void setDensityFunction(DensityFunction func) {
		this.density = func;
	}

	/**
	 * Sets the distance function used to find the best or worst matching unit.
	 * If not set, the map uses the EulerDistanceSquared class.<br>
	 * The neighborhood distance is Manhattan Distance.
	 * 
	 * @param func
	 */
	public final void setDistanceFunction(DistanceFunction func) {
		this.distance = func;
	}

	/**
	 * Updates the map with the given data. Uses the last found BMU or WMU.
	 * 
	 * @param input
	 *            input vector
	 * @param output
	 *            expected output vector
	 */
	public final void train(double input[], double output[]) {
		if(BMU == null)
			throw new UnsupportedOperationException("BMU/WMU must be found first.");

		Object[] backArray = vda.getBackingArray();
		int[] size = vda.getSize();
		int[] pos = new int[size.length];

		for(Object o : backArray) {
			double[] array = (double[])o;
			double distance = neighborhood(pos, BMU);

			if(wrap) {
				int[] tpos = pos.clone();
				int[] npos = BMU.clone();
				for(int i = 0; i < tpos.length; ++i) {
					tpos[i] += size[i] / 2;
					npos[i] += size[i] / 2;
					if(tpos[i] > size[i])
						tpos[i] -= size[i];
					if(npos[i] > size[i])
						npos[i] -= size[i];
				}
				double ndist = neighborhood(tpos, npos);
				if(ndist < distance)
					distance = ndist;
			}

			double neighborhood = density.calculate(distance);

			/* Changes below this point benefits are negligible */
			if(neighborhood < cutoff) {
				next(pos, size);
				continue;
			}

			for(int i = 0; i < array.length; ++i) {
				if(i < inputSize) {
					array[i] = weight(array[i], input[i], neighborhood);
				} else {
					array[i] = weight(array[i], output[i - inputSize], neighborhood);
				}
			}

			next(pos, size);
		}
	}

	private static final double neighborhood(int[] p, int[] q) {
		if(p == null || q == null)
			return 0;
		int len = Math.min(p.length, q.length);
		int output = 0;
		for(int i = 0; i < len; ++i)
			output += Math.abs(p[i] - q[i]);
		return output;
	}

	private final double weight(double c, double t, double n) {
		return c + n * (t - c) * learningRate;
	}

	private static final void next(int[] pos, int[] size) {
		++pos[pos.length - 1];
		for(int i = pos.length - 1; i > 0; --i) {
			if(pos[i] >= size[i]) {
				++pos[i - 1];
				pos[i] = 0;
			}
		}
	}
}

org.csdgn.nn.KohonenMap

package org.csdgn.nn;

import java.util.Random;
import org.csdgn.nn.density.StandardDensity;
import org.csdgn.nn.distance.EulerDistanceSquared;

/**
 * A Self-Organizing Map implementation.
 * 
 * Requires: <br>
 * org.csdgn.nn.DensityFunction<br>
 * org.csdgn.nn.DistanceFunction<br>
 * org.csdgn.nn.density.StandardDensity<br>
 * org.csdgn.nn.distance.EulerDistanceSquared
 * 
 * TODO: Optimize for speed.
 * 
 * @author Chase
 *
 */
public class KohonenMap {
	private static final double cutoff = 1e-4;
	
	/**
	 * Holds the neighborhood layout;
	 */
	public final Node[] map;
	public final int[] mapSize;
	public double learningRate = 0.8;
	private DensityFunction density;
	private DistanceFunction distance;
	private boolean wrap = false;
	private int BMU;
	
	/**
	 * @param mapSize Size of the neighborhood. Example: {10,10} produces a 2 dimensional 
	 * map, each dimension having 10 nodes. Total nodes would be 100.
	 * @param input The length of the input vector (2D only)
	 * @param output The length of the output vector (2D only)
	 */
	public KohonenMap(int[] mapSize, int input, int output) {
		/* Setup the map */
		int size = 1;
		for(int m : mapSize) size *= m;
		this.map = new Node[size];
		
		this.mapSize = mapSize.clone();
		
		this.density = new StandardDensity();
		this.distance = new EulerDistanceSquared();
		
		int[] pos = new int[mapSize.length];
		for(int i=0; i<map.length; ++i) {
			this.map[i] = new Node(mapSize.length,input,output);
			/* Setup the location of each node, for speed reasons. */
			System.arraycopy(pos, 0, this.map[i].position, 0, pos.length);
			
			/* Update the position marker */
			++pos[0];
			for(int j=0;j<pos.length-1;++j) {
				if(pos[j] >= mapSize[j]) {
					++pos[j+1];
					pos[j] = 0;
				}
			}
		}
	}
	
	/**
	 * Initializes the map to random values
	 */
	public final void initialize() {
		Random r = new Random();
		initialize(r);
	}
	
	/**
	 * Initializes the map with the given random function.
	 * Uses the nextDouble function.
	 */
	public final void initialize(Random random) {
		for(Node n : map) {
			for(int i = 0; i < n.input.length; ++i)
				n.input[i] = random.nextDouble();
			for(int i = 0; i < n.output.length; ++i)
				n.output[i] = random.nextDouble(); 
		}
	}
	
	/**
	 * Finds the Best Matching Unit for the given input.
	 * @return the BMUs identifier
	 */
	public final int findBMInput(double[] input) {
		BMU = 0;
		
		double distance = Double.MAX_VALUE;
		for(int i=0; i<map.length; ++i) {
			double dist = this.distance.calculate(map[i].input, input);
			
			if(dist < distance) {
				distance = dist;
				BMU = i;
			}
		}
		return BMU;
	}
	
	/**
	 * Finds the Best Matching Unit for the given output.
	 * @return the BMUs identifier
	 */
	public final int findBMOutput(double[] output) {
		BMU = 0;
		double distance = Double.MAX_VALUE;
		for(int i=0; i<map.length; ++i) {
			double dist = this.distance.calculate(map[i].output, output);
			if(dist < distance) {
				distance = dist;
				BMU = i;
			}
		}
		return BMU;
	}
	
	/**
	 * Finds the Worst Matching Unit for the given input
	 * @return the WMUs identifier
	 */
	public final int findWMInput(double[] input) {
		BMU = 0;
		double distance = Double.MIN_VALUE;
		for(int i=0; i<map.length; ++i) {
			double dist = this.distance.calculate(map[i].input, input);
			if(dist > distance) {
				distance = dist;
				BMU = i;
			}
		}
		return BMU;
	}
	
	/**
	 * Finds the Worst Matching Unit for the given output
	 * @return the WMUs identifier
	 */
	public final int findWMOutput(double[] output) {
		BMU = 0;
		double distance = Double.MIN_VALUE;
		for(int i=0; i<map.length; ++i) {
			double dist = this.distance.calculate(map[i].output, output);
			if(dist > distance) {
				distance = dist;
				BMU = i;
			}
		}
		return BMU;
	}
	
	/**
	 * Sets the Matched index to the set value.
	 * @param index
	 */
	public final void setMatchIndex(int index) {
		BMU = Math.max(0,Math.min(index, map.length-1));
	}
	
	/**
	 * This returns the input of the last found BMU or WMU.
	 * @return the input vector
	 */
	public final double[] getInput() {
		return this.map[BMU].input;
	}
	
	/**
	 * This returns the output of the last found BMU or WMU.
	 * @return the output vector
	 */
	public final double[] getOutput() {
		return this.map[BMU].output;
	}
	
	/**
	 * This returns the input of the given ID.
	 * @return the input vector
	 */
	public final double[] getInput(int id) {
		if(id > 0 && id < map.length)
			return this.map[id].input;
		return new double[0];
	}
	
	/**
	 * This returns the output of the given ID.
	 * @return the output vector
	 */
	public final double[] getOutput(int id) {
		if(id > 0 && id < map.length)
			return this.map[id].output;
		return new double[0];
	}
	
	/**
	 * Sets the learning rate of this KohonenMap
	 * @param rate value between 0 and 1
	 */
	public final void setLearningRate(double rate) {
		learningRate = Math.max(Math.min(rate, 1),0);
	}
	
	/**
	 * Returns the current rate of learning
	 * @return the learning rate 
	 */
	public final double getLearningRate() {
		return learningRate;
	}
	
	/**
	 * Sets the map to wrap its updates (slightly more costly)
	 */
	public final void setWraps(boolean n) {
		wrap = n;
	}
	
	/**
	 * Returns if the current map wraps
	 * @return
	 */
	public final boolean isWrapping() {
		return wrap;
	}
	
	/**
	 * Sets the density function this map uses for updating nearby nodes.
	 * If unset it uses the StandardDensity class.
	 * @param func the Density Function
	 */
	public final void setDensityFunction(DensityFunction func) {
		this.density = func;
	}
	
	/**
	 * Sets the distance function used to find the best or worst matching unit.
	 * If unset, this map uses the EulerDistanceSquared class.<br>
	 * The neighborhood distance is Manhattan Distance.
	 * @param func
	 */
	public final void setDistanceFunction(DistanceFunction func) {
		this.distance = func;
	}
	
	/**
	 * Updates the map with the given data. Uses the last found
	 * BMU or WMU.
	 * @param input input vector
	 * @param output expected output vector
	 */
	public final void train(double input[], double output[]) {
		Node bmu = map[BMU];
		for(int i=0; i<map.length; ++i) {
			map[i].update(bmu.position, input, output);
		}
	}
	
	/**
	 * This uses Manhattan Distance.
	 */
	private static final double neighborhood(int[] p, int[] q) {
		if(p == null || q == null) return 0;
		int len = Math.min(p.length, q.length);
		int output = 0;
		for(int i=0; i<len; ++i)
			output += Math.abs(p[i] - q[i]);
		return output; 
	}
	
	private final class Node {
		/** Location in the neighborhood */
		private final int[] position;
		/** Input vector */
		private final double[] input;
		/** Output vector */
		private final double[] output;
		
		public Node(int mapSize, int inputSize, int outputSize) {
			position = new int[mapSize];
			input = new double[inputSize];
			output = new double[outputSize];
		}
		
		private final double weight(double c, double t, double n) {
			return c + n * (t - c) * learningRate;
		}
		
		private final void update(int[] pos, double[] in, double[] out) {
			double distance = neighborhood(pos,position);
			if(wrap) {
				int[] tpos = pos.clone();
				int[] npos = position.clone();
				for(int i=0; i<tpos.length; ++i) {
					tpos[i] += mapSize[i]/2;
					npos[i] += mapSize[i]/2;
					if(tpos[i] > mapSize[i]) tpos[i] -= mapSize[i];
					if(npos[i] > mapSize[i]) npos[i] -= mapSize[i];
				}
				double ndist = neighborhood(tpos,npos);
				if(ndist < distance) distance = ndist;
			}
			
			double neighborhood = density.calculate(distance);
			
			/* Changes below this point benefits are negligible */
			if(neighborhood < cutoff) return;
			
			for(int i=0; i<input.length; ++i)
				input[i] = weight(input[i], in[i], neighborhood);
			
			for(int i=0; i<output.length; ++i)
				output[i] = weight(output[i], out[i], neighborhood);
		}
		
	}
}

org.csdgn.nn.DensityFunction

package org.csdgn.nn;

public interface DensityFunction {
	/**
	 * Calculates the density at the given point, where x is a certain distance from the center of the distribution.
	 */
	public double calculate(double x);
}

org.csdgn.nn.DistanceFunction

package org.csdgn.nn;

/**
 * A function for determining the distance between two double arrays.
 * @author Chase
 */
public interface DistanceFunction {
	public double calculate(double[] p, double[] q);
}

org.csdgn.nn.density.StandardDensity

package org.csdgn.nn.density;

import org.csdgn.nn.DensityFunction;

/**
 * <math>density(x) = 2^{-x^2}</math>
 */
public final class StandardDensity implements DensityFunction {
	/**
	 * <math>density(x) = 2^{-x^2}</math>
	 */
	@Override
	public double calculate(double x) {
		return Math.pow(2, -(x*x));
	}
}

org.csdgn.nn.density.NormalDistribution

package org.csdgn.nn.density;

import org.csdgn.nn.DensityFunction;

public final class NormalDistribution implements DensityFunction {
	private final double multi;
	private final double variance;
	private final double mean;
	
	public NormalDistribution() {
		this(1,0);
	}
	public NormalDistribution(double variance, double mean) {
		this.multi = 1.0 / Math.sqrt(2*Math.PI*variance);
		this.variance = variance;
		this.mean = mean;
	}
	@Override
	public double calculate(double x) {
		double e = ((x - mean)*(x - mean)) / (2*variance);
		return multi*Math.exp(-e);
	}

}

org.csdgn.nn.distance.EulerDistanceSquared

package org.csdgn.nn.distance;

import org.csdgn.nn.DistanceFunction;

public class EulerDistanceSquared implements DistanceFunction {
	/**
	 * <math>distSqr(p,q) = \sum_{i=0}^n (p_i - q_i)^2</math> where <math>n</math> is the
	 * size of the smaller of <math>p</math> or <math>q</math>
	 */
	@Override
	public final double calculate(double[] p, double[] q) {
		if(p == null || q == null) return 0;
		int len = Math.min(p.length, q.length);
		double k,output = 0;
		for(int i=0; i<len; ++i)
			output += (k=(p[i] - q[i]))*k;
		return output;
	}

}