annotate src/samer/models/Scaler.java @ 0:bf79fb79ee13

Initial Mercurial check in.

/*
 *	Copyright (c) 2002, 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.core.types.*;
import samer.maths.*;
import samer.maths.opt.*;
import samer.tools.*;

/**
	Automatic gain control for a given input vector.
	Ouput appears in out. Also generates a trace of
	'loudness' of input signal.
  */

public class Scaler extends NullTask implements Model
{
	private Model 		M;
	private int			n;
	private Vec			x;
	private VVector	s;
	private VDouble	multiplier;
	private VDouble	offset;
	private double	logK;

	double []		_x, _s, _g, phi;

	public Scaler( Vec input, Model M) { this(input); setOutputModel(M); M.setInput(s); }
	public Scaler( Vec input) { this(input.size()); setInput(input); }
	public Scaler( int N)
	{
		n = N;

		x = null;
		s = new VVector("output",n);
		multiplier = new VDouble("scale",1.0,VDouble.SIGNAL);
		offset = new VDouble("offset",0.0,VDouble.SIGNAL);
		
		_s = s.array();
		_g = new double[n];
		phi = null;
		reset();
		logK=Math.log(multiplier.value);
	}

	public int getSize() { return n; }
	public VVector output() { return s; }
	public VDouble getScale() { return multiplier; }
	public VDouble getOffset() { return offset; }
	public Model getOutputModel() { return M;	}
	public void setOutputModel(Model m) { M=m;	}
	public void setInput(Vec in) { x=in; _x=x.array(); }
	public void reset() {
		// multiplier.load(Shell.env());
		// offset.load(Shell.env());
	}

	public String toString() { return "Scaler:"+x; } // +"->"+s; }
	public void dispose()
	{
		offset.dispose();
		multiplier.dispose();
		s.dispose();
		super.dispose();
	}

	public void infer() {
		double a=offset.value, k=1/multiplier.value;
		for (int i=0; i<n; i++) _s[i] = k*(_x[i]-a);
		s.changed();
	}

	public void compute() {
		Mathx.mul(_g,M.getGradient(),1/multiplier.value);
	}

	public double	getEnergy() { return M.getEnergy() + n*logK; }
	public double [] getGradient() { return _g; }

	public Functionx functionx() {
		return new Functionx() {
			Functionx fM=M.functionx();
			double [] s=new double[n];

			public void dispose() { fM.dispose(); }
			public void evaluate(Datum P) { P.f=evaluate(P.x,P.g); }
			public double evaluate(double [] x, double [] g) {
				double a=offset.value, k=1/multiplier.value;
				for (int i=0; i<n; i++) s[i] = k*(x[i]-a);
				double E=fM.evaluate(s,g);
				Mathx.mul(g,k);
				return E+n*logK;
			}
		};
	}

	public void starting() { logK=Math.log(multiplier.value); }
	public void stopping() {}
	public void run() { infer(); }

	public Trainer getTrainer() { return new Trainer(); }
	public OffsetTrainer getOffsetTrainer()	{ return new OffsetTrainer(); }
	public ScaleTrainer getScaleTrainer()	{ return new ScaleTrainer(); }

	public class Trainer extends AnonymousTask implements Model.Trainer
	{
		VDouble		rate1=new VDouble("scaleRate",0.001);
		VDouble		rate2=new VDouble("offsetRate",0.000001);
		double		G,H,count;
		double []	_s;
		int				n;


		public Trainer() { _s = s.array(); n=Scaler.this.n; }

		public void reset() { count=0; G=0; H=0; }
		public String toString() { return "Trainer:"+Scaler.this; }

		public VDouble getScaleRate() { return rate1; }
		public VDouble getOffsetRate() { return rate2; }

		public void accumulate() { accumulate(1); }
		public void accumulate(double w) {
//			if (M.getEnergy() > 8000) return;
			double [] phi=M.getGradient();
			double	g=0;
			for (int i=0; i<n; i++) g += phi[i]*_s[i] - 1;
			G += w*g;
			H += w*Mathx.sum(phi);
			count+=w;
		}

		public void flush() {
			if (count==0) return; // nothing to do

			double k=multiplier.value;
			double mu=offset.value;

			mu += (rate2.value/count)*k*H/n;
			G *= rate1.value/(n*count);
			k *= Math.exp(G);
			multiplier.set(k);
			offset.set(mu);
			logK+=G;
			reset();
		}

		public void oneshot() {
			double [] phi=M.getGradient();
			G=0; H = Mathx.sum(phi);
			for (int i=0; i<n; i++) G += phi[i]*_s[i] - 1;

			double k=multiplier.value;
			double mu=offset.value;

			mu += rate2.value*k*H/n;
			G *= rate1.value/n;
			k *= Math.exp(G);
			multiplier.set(k);
			offset.set(mu);
			logK+=G;
		}


		public void dispose() { rate1.dispose(); rate2.dispose(); }
		public void starting() { reset(); logK=Math.log(multiplier.value); }
		public void run() { oneshot(); }
	}

	public class ScaleTrainer extends AnonymousTask implements Model.Trainer
	{
		VDouble	rate1=new VDouble("scaleRate",0.001);
		double		G,count;
		double []	_s;
		int				n;


		public ScaleTrainer() { _s = s.array(); n=Scaler.this.n; }
		public String toString() { return "ScaleTrainer:"+Scaler.this; }

		public void reset() { count=0; G=0; }

		public void accumulate() { accumulate(1); }
		public void accumulate(double w) {
			double [] phi=M.getGradient();
			double	g=0;
			for (int i=0; i<n; i++) g += phi[i]*_s[i] - 1;
			G += w*g;
			count+=w;
		}

		public void flush() {
			if (count==0) return; // nothing to do

			double k=multiplier.value;

			G *= rate1.value/(n*count);
			k *= Math.exp(G);
			multiplier.set(k);
			logK+=G;
			reset();
		}

		public void oneshot() {
			double [] phi=M.getGradient();
			G=0;
			for (int i=0; i<n; i++) G += phi[i]*_s[i] - 1;

			double k=multiplier.value;

			G *= rate1.value/n;
			k *= Math.exp(G);
			multiplier.set(k);
			logK+=G;
		}


		public void dispose() { rate1.dispose(); }
		public void starting() { reset(); logK=Math.log(multiplier.value); }
		public void run() { oneshot(); }
	}

	/** This trains only the offset, not the scale */
	public class OffsetTrainer extends AnonymousTask implements Model.Trainer
	{
		VDouble		rate2=new VDouble("offsetRate",0.000001);
		double		H, count;
		int				n;

		public OffsetTrainer() { n=Scaler.this.n; }
		public String toString() { return "OffsetTrainer:"+Scaler.this; }

		public void reset() { count=0; H=0; }

		public void accumulate() { accumulate(1); }
		public void accumulate(double w) {
			double [] phi=M.getGradient();
			H += w*Mathx.sum(phi);
			count+=w;
		}

		public void flush() {
			if (count==0) return; // nothing to do
			offset.value += (rate2.value/count)*multiplier.value*H/n;
			offset.changed();
			reset();
		}

		public void oneshot() {
			double [] phi=M.getGradient();
			H = Mathx.sum(phi);
			offset.value += rate2.value*multiplier.value*H/n;
			offset.changed();
		}

		public void dispose() { rate2.dispose(); }
		public void starting() { reset(); }
		public void run() { oneshot(); }
	}
	// could have alternative trainers if prior is Gaussian or Laplacian,
	// in which case, parameters can be estimated in closed form
}