samer@0: /*
samer@0:  *	Copyright (c) 2000, Samer Abdallah, King's College London.
samer@0:  *	All rights reserved.
samer@0:  *
samer@0:  *	This software is provided AS iS and WITHOUT ANY WARRANTY; 
samer@0:  *	without even the implied warranty of MERCHANTABILITY or 
samer@0:  *	FITNESS FOR A PARTICULAR PURPOSE.
samer@0:  */
samer@0: 
samer@0: package samer.models;
samer@0: import  samer.core.*;
samer@0: import  samer.core.types.*;
samer@0: import  samer.maths.*;
samer@0: // import  samer.tools.*;
samer@0: 
samer@0: /**
samer@0: 	Handles batched delta updates to a matrix.
samer@0: 	The default flush multiplies the accumulated
samer@0: 	delta by the learing rate and adds it to the
samer@0: 	matrix.
samer@0:  */
samer@0: 
samer@0: public class MatrixTrainer implements Model.Trainer, DoubleModel
samer@0: {
samer@0: 	protected VParameter	ratep;		// learning rate
samer@0: 	protected Matrix		A, T;			// A=target matrix, T=matrix of deltas
samer@0: 	protected double[][]	_A, _T;
samer@0: 	protected double		rate=1;
samer@0: 	protected int			n, m;
samer@0: 	protected double []  a, b;
samer@0: 	protected double	count=0;
samer@0: 
samer@0: 	public MatrixTrainer(Vec left, Matrix A, Vec right) {
samer@0: 		// must have left.size()=n, right.size()=m
samer@0: 		this(left.array(),A,right.array());
samer@0: 	}
samer@0: 
samer@0: 	public MatrixTrainer(double [] a, Matrix A, double [] b)
samer@0: 	{
samer@0: 		this.A = A;
samer@0: 		n = A.getRowDimension();
samer@0: 		m = A.getColumnDimension();
samer@0: 		_A = A.getArray();
samer@0: 
samer@0: 		ratep = new VParameter("rate",this);
samer@0: 		T = new Matrix("deltas",n,m);
samer@0: 		_T = T.getArray();
samer@0: 		this.a=a; this.b=b;
samer@0: 	}
samer@0: 
samer@0: 	public void set(double r) { rate=r; }
samer@0: 	public double get() { return rate; }
samer@0: 	public VParameter getRate() { return ratep; }
samer@0: 
samer@0: 	protected void outerProduct(double w, double [] a, double [] b) {
samer@0: 		for (int i=0; i<n; i++) {
samer@0: 			double q=w*a[i], p[] = _T[i]; 
samer@0: 			for (int j=0; j<m; j++) p[j] += q*b[j];
samer@0: 		}
samer@0: 	}
samer@0: 	
samer@0: 	public void dispose() { T.dispose(); ratep.dispose(); }
samer@0: 	public void reset() { T.zero(); T.changed(); count=0; }
samer@0: 	public void oneshot() { accumulate(1); flush(); }
samer@0: 	public void accumulate() { accumulate(1); } 
samer@0: 	public void accumulate(double w) { count+=w; outerProduct(w,a,b); }
samer@0: 	
samer@0: 	public void flush() {
samer@0: 		if (count==0) return;
samer@0: 		double eta=rate/count;
samer@0: 
samer@0: 		T.changed(); // to display accumulated stats
samer@0: 		for (int i=0; i<n; i++) {
samer@0: 			double [] p = _A[i], q = _T[i];
samer@0: 			for (int j=0; j<m; j++) p[j] += eta*q[j];
samer@0: 			Mathx.zero(q);
samer@0: 		}
samer@0: 		A.changed();
samer@0: 		count=0;
samer@0: 	}
samer@0: }