view src/samer/models/GaussianStats.java @ 0:bf79fb79ee13

Initial Mercurial check in.

/*
 *	Copyright (c) 2000, Samer Abdallah, King's College London.
 *	All rights reserved.
 *
 *	This software is provided AS iS and WITHOUT ANY WARRANTY; 
 *	without even the implied warranty of MERCHANTABILITY or 
 *	FITNESS FOR A PARTICULAR PURPOSE.
 */

package samer.models;
import  samer.core.*;
import  samer.maths.*;
import  samer.tools.*;
import  java.util.*;

/**
	Collect statistics about a vector: accumulates
	sums and sums of products, then computes mean
	and covariance.

	This is NOT a model! Computing energies and probabilities
	directly from the first and second moments is horrible, since
	you need the inverse of the covariance matrix. You could,
	however, use this to train a Gaussian model.
 */

public class GaussianStats extends NullTask
{
	VVector	vector;
	VVector	sumx;
	Matrix		sumxx;
	int			n, count;
	double	[] _sumx, x;
	double	[][] _sumxx;
	double	[] sig;

	public GaussianStats(VVector vec) throws Exception
	{
		n=vec.size();
		sumx = new VVector("t1",n);	// first moment: sum
		sumxx = new Matrix("t2",n,n);	// second moment: sum products

		_sumx=sumx.array();
		_sumxx=sumxx.getArray();
		x=vec.array();
		vector=vec;
		sig= new double[n];
		
	}

	public void dispose() {
		sumx.dispose();
		sumxx.dispose();
	}

	public void getMean(VVector mu) { getMean(mu.array()); mu.changed(); }
	public void getMean(double [] mu) { Mathx.mul(mu, _sumx, 1.0/count);	}
	public void getCovariance(Matrix C) {
		// compute mean products
		C.assign(sumxx);
		C.timesEquals(1.0/count);

		// subtract square mean from mean square
		double [][] _C=C.getArray();

		for (int i=0; i<n; i++) {
			double tmp=_sumx[i]/count; // diagonal
			for (int j=0; j<i; j++) { // off diagonal
				_C[i][j] -= tmp*(_sumx[j]/count);
				_C[j][i] = _C[i][j];
			}
			_C[i][i] -= tmp*tmp;
		}
		C.changed();
	}

	public void getCorrelation(Matrix R) {
		// compute mean products
		R.assign(sumxx);
		R.timesEquals(1.0/count);

		double [][] _R=R.getArray();

		// first get std devs and put 1s down diagonal
		for (int i=0; i<n; i++) {
			double tmp=_sumx[i]/count; // diagonal
			sig[i] = Math.sqrt(_R[i][i] - tmp*tmp);
			_R[i][i]=1;
		}

		for (int i=0; i<n; i++) {
			double tmp=_sumx[i]/count;
			for (int j=0; j<i; j++) { // off diagonal
				_R[i][j] -= tmp*(_sumx[j]/count);
				_R[i][j] /= sig[i]*sig[j];
				_R[j][i] = _R[i][j];
			}
		}
		R.changed();
	}

	public void reset() {
		Mathx.zero(_sumx);
		sumxx.zero(); count=0;
	}

	public void starting() {}
	public void stopping() {}
	public void run()
	{
		Mathx.add(_sumx,x);

		for (int i=0; i<n; i++) {
			double  [] a=_sumxx[i];
			double  k=x[i];
			for (int j=0; j<=i; j++) {
				a[j] += k*x[j];
			}
		}
		sumx.changed();
		sumxx.changed();
		count++;
	}
}