LinearRegression.cpp

Go to the documentation of this file.

/*
 *  LinearRegression.cpp
 *  hog2
 *
 *  Created by Nathan Sturtevant on 6/1/06.
 *  Copyright 2006 Nathan Sturtevant. All rights reserved.
 *
 */
 
 
#include <math.h>
#include <stdlib.h>
#include <stdio.h>
#include <unistd.h>
#include <assert.h>
#include <string.h>
#include "LinearRegression.h"
#include "SwapEndian.h"
 
static const float VERSION = 1.1;
static const float MINVERSION = 1.0;
 
LinearRegression::LinearRegression(int _inputs, int _outputs, double _rate)
{
        useBinary = false;
        outputActivation = kLinear;
        inputs = _inputs;
        outputs = _outputs;
        rate = _rate;
        weight.resize(0);
        updates.resize(0);
        //weight = 0;
        allocateMemory();
}
 
LinearRegression::LinearRegression(LinearRegression *perp)
{
        useBinary = perp->useBinary;
//      inputs = perp->inputs;
        outputs = perp->outputs;
        rate = perp->rate;
        outputActivation = perp->outputActivation;
        weight.resize(0);
        updates.resize(0);
        //weight = 0;
        //allocateMemory(perp->inputs);
        load(perp);
}
 
LinearRegression::LinearRegression(char *f)
{
        useBinary = false;
        weight.resize(0);
        updates.resize(0);
        //weight = 0;
        inputs = outputs = -1;
        load(f);
}
 
LinearRegression::LinearRegression(FunctionApproximator *fa)
{
        LinearRegression *perp = (LinearRegression *)fa;
        useBinary = perp->useBinary;
        inputs = perp->inputs;
        outputs = perp->outputs;
        outputActivation = perp->outputActivation;      
        rate = perp->rate;
        weight.resize(0);
        updates.resize(0);
        //weight = 0;
        //allocateMemory();
        load(perp);
}
 
LinearRegression::~LinearRegression()
{
        freeMemory();
}
 
void LinearRegression::resizeInputs(int newSize, double newVal)
{
        if (newSize < inputs)
        {
                for (int x = 0; x < outputs; x++)
                {
                        weight[x][newSize] = weight[x][inputs];
                        weight[x].resize(newSize+1);
                        updates[x].resize(newSize);
                }
        }
        else if (newSize == inputs)
        {
                return;
        }
        else {
                for (int x = 0; x < outputs; x++)
                {
                        weight[x].resize(newSize+1);// = new double[inputs+1];
                        weight[x][newSize] = weight[x][inputs];
                        for (int y = inputs; y < newSize; y++)
                                weight[x][y] = newVal;
                        
                        updates[x].resize(newSize);// = new double[inputs+1];
                }
        }
        inputs = newSize;
}
 
void LinearRegression::resizeInputs(int newSize)
{
        if (newSize < inputs)
        {
                for (int x = 0; x < outputs; x++)
                {
                        weight[x][newSize] = weight[x][inputs];
                        weight[x].resize(newSize+1);
                        updates[x].resize(newSize);
                }
        }
        else if (newSize == inputs)
        {
                return;
        }
        else {
                for (int x = 0; x < outputs; x++)
                {
                        weight[x].resize(newSize+1);// = new double[inputs+1];
                        weight[x][newSize] = weight[x][inputs];
                        for (int y = inputs; y < newSize; y++)
                                weight[x][y] = ((double)2*random()/RAND_MAX-1)/(double)newSize;
 
                        updates[x].resize(newSize);// = new double[inputs+1];
                }
        }
        inputs = newSize;
}
 
void LinearRegression::allocateMemory()
{
//      if (weight.size() != 0)
//              freeMemory();
        weight.resize(outputs);// = new double*[outputs];
        updates.resize(outputs);
        for (int x = 0; x < outputs; x++)
        {
                weight[x].resize(inputs+1);// = new double[inputs+1];
                updates[x].resize(inputs);
                for (int y = 0; y <= inputs; y++)
                        weight[x][y] = ((double)2*random()/RAND_MAX-1)/(double)inputs;
        }
        output.resize(outputs);// = new double[outputs];
}
 
void LinearRegression::freeMemory()
{
//      if (weight != 0)
//      {
        weight.resize(0);
        updates.resize(0);
        //              //weight = 0;
//      }
}
 
void LinearRegression::load(const char *fname)
{
        FILE *f;
        f = fopen(fname, "r");
        if (f == 0)
        {
                fprintf(stderr, "REGR Error: could not open file '%s' for loading; trying once more.\n", fname);
                sleep(1);
                f = fopen(fname, "r");
                if (f == 0)
                {
                        fprintf(stderr, "REGR Error: could not open file '%s' for loading.\n", fname);
                        exit(0);
                        return;
                }
        }
        load(f);
        fclose(f);
}
 
void LinearRegression::load(FILE *f)
{
        int inputs1, outputs1;
        float version;
        int res = fscanf(f, "REGR %f %d %d\n", &version, &inputs1, &outputs1);
        if (res != 3)
        {
                printf("Error: unrecognized LinearRegression file. Expected header 'PERP <version> <inputs> <outputs>'.");
                exit(0);
        }
        if (version > VERSION)
        {
                printf("Error: loaded network is %1.2f, but code can only handle %1.2f to %1.2f.",
                                         version, MINVERSION, VERSION);
                exit(0);
        }
        if (VERSION >= 1.1)
        {
                char text[25];
                res = fscanf(f, "%s\n", text);
                assert(res == 1);
                if (strcmp(text, "binary") == 0)
                        useBinary = true;
                else //if (strcmp(text, "text") == 0)
                        useBinary = false;
        }
 
        if ((inputs1 != inputs) || (outputs1 != outputs))
        {
                freeMemory();
                inputs = inputs1;
                outputs = outputs1;
                allocateMemory();
        }
 
        for (int y = 0; y < outputs; y++)
        {
                for (int x = 0; x <= inputs; x++)
                {
                        if (useBinary)
                        {
                                float shrunk;
                                fread(&shrunk, sizeof(float), 1, f);
                                //fread(&weight[y][x], sizeof(double), 1, f);
                                little2machine(shrunk);
                                weight[y][x] = shrunk;
                        }
                        else
                                fscanf(f, "%le ", &weight[y][x]);
                }
        }
}
 
void LinearRegression::load(const LinearRegression *p)
{
        if (p && ((p->inputs != inputs) || (p->outputs != outputs)))
        {
                freeMemory();
                inputs = p->inputs;
                outputs = p->outputs;
                outputActivation = p->outputActivation;
 
                allocateMemory();
        }
 
        if (p)
        {
                for (int y = 0; y < outputs; y++)
                {
                        for (int x = 0; x <= inputs; x++)
                        {
                                weight[y][x] = p->weight[y][x];
                        }
                }
        }
}
 
bool LinearRegression::validSaveFile(char *fname)
{
        FILE *f;
        f = fopen(fname, "r");
        if (f == 0)
        {
                return false;
        }
        int finput, foutput;
        float version;
        int res = fscanf(f, "REGR %f %d %d\n", &version, &finput, &foutput);
        fclose(f);
        if (res != 3)
        {
                return false;
        }
        if (version > VERSION)
        {
                return false;
        }
        return true;
}
 
void LinearRegression::save(const char *fname)
{
        FILE *f;
        f = fopen(fname, "w+");
        if (f == 0)
        {
                fprintf(stderr, "Error: could not open file for saving.\n");
                return;
        }
        save(f);
        fclose(f);
}
 
void LinearRegression::save(FILE *f)
{
        fprintf(f, "REGR %1.2f %d %d\n", VERSION, inputs, outputs);
        if (useBinary)
                fprintf(f, "binary\n");
        else
                fprintf(f, "text\n");
                
        for (int y = 0; y < outputs; y++)
        {
                for (int x = 0; x <= inputs; x++)
                {
                        if (useBinary)
                        {
                                float val = weight[y][x];
                                little2machine(val);
                                fwrite(&val, sizeof(float), 1, f);
                        }
                        else {
                                fprintf(f, "%le ", weight[y][x]);
                        }
                }
                if (!useBinary)
                        fprintf(f, "\n");
        }
}
 
double LinearRegression::outputerr(std::vector<double> &out, std::vector<double> &expected, int which)
{
        double err = (out[which]-expected[which]);
        //printf("Output error: %1.2f\n", err);
        return err;
}
 
double LinearRegression::train(std::vector<unsigned int> &input, std::vector<double> &target)
{
        double totalErr = 0;
        test(input);
        for (int x = 0; x < outputs; x++)
        {
                double err = outputerr(output,target,x);
                totalErr+=err*err;
                double rateTimesError = rate*err;
                for (unsigned int y = 0; y < input.size(); y++)
                {
                        weight[x][input[y]] -= rateTimesError;
                        updateData &val = updates[x][input[y]];
                        val.n++;
//                      double delta = err-val.mean;
                        double delta = rateTimesError-val.mean;
                        val.mean = val.mean+delta/val.n;
                        val.S += delta*(err-val.mean);
                        val.totErr += rateTimesError*rateTimesError;
                }
                weight[x][inputs] -= rateTimesError; // bias
        }
        return totalErr;
}
 
double LinearRegression::train(std::vector<double> &input, std::vector<double> &target)
{
        double totalErr = 0;
        test(input);
        for (int x = 0; x < outputs; x++)
        {
                double err = outputerr(output,target,x);
                totalErr+=err*err;
                double rateTimesError = rate*err;
                for (int y = 0; y < inputs; y++)
                {
                        weight[x][y] -= rateTimesError*input[y];
 
                        updateData &val = updates[x][y];
                        val.n++;
//                      double delta = err-val.mean;
                        double delta = rateTimesError-val.mean;
                        val.mean = val.mean+delta/val.n;
                        val.S += delta*(err-val.mean);
                        val.totErr += rateTimesError*rateTimesError;
                }
                weight[x][inputs] -= rateTimesError*(1); // bias
        }
        return totalErr;
}
 
double *LinearRegression::test(const std::vector<unsigned int> &input)
{
        for (int y = 0; y < outputs; y++)
        {
                output[y] = weight[y][inputs]; // bias
                for (unsigned int x = 0; x < input.size(); x++)
                {
                        output[y] += weight[y][input[x]];
                }
        }
        return &output[0];
}
 
double *LinearRegression::test(const std::vector<double> &input)
{
        for (int y = 0; y < outputs; y++)
        {
                output[y] = weight[y][inputs];
                for (int x = 0; x < inputs; x++)
                {
                        output[y] += weight[y][x]*input[x];
                }
        }
        return &output[0];
}
 
void LinearRegression::getWeightUpdateVariance(std::vector<double> &var, unsigned int which)
{
        assert(which < weight.size());
        var.resize(weight[which].size()-1);
        for (unsigned int x = 0; x < weight[which].size()-1; x++) // ignore bias!
        {
                updateData &val = updates[which][x];
                if (val.n > 1)
                        var[x] = val.S/(val.n-1);
                else
                        var[x] = 0;
        }
}
 
double LinearRegression::getWeightUpdateVariance(unsigned int weightNum, unsigned int whichOutput)
{
        assert(whichOutput < weight.size());
        assert(weightNum < weight[whichOutput].size());
 
        updateData &val = updates[whichOutput][weightNum];
        if (val.n > 1)
                return val.S/(val.n-1);
        return 0;
}
 
void LinearRegression::getWeightUpdateAverage(std::vector<double> &var, unsigned int which)
{
        assert(which < weight.size());
        var.resize(weight[which].size()-1);
        for (unsigned int x = 0; x < weight[which].size()-1; x++) // ignore bias!
        {
                updateData &val = updates[which][x];
                if (val.n > 1)
                        var[x] = val.totErr/val.n;
                else
                        var[x] = 0;
        }
}
 
double LinearRegression::getWeightUpdateAverage(unsigned int weightNum, unsigned int whichOutput)
{
        assert(whichOutput < weight.size());
        assert(weightNum < weight[whichOutput].size());
        
        updateData &val = updates[whichOutput][weightNum];
        if (val.n > 1)
                return val.totErr/val.n;
        return 0;
}
 
void LinearRegression::getWeightUpdateSum(std::vector<double> &var, unsigned int which)
{
        assert(which < weight.size());
        var.resize(weight[which].size()-1);
        for (unsigned int x = 0; x < weight[which].size()-1; x++) // ignore bias!
        {
                var[x] = updates[which][x].totErr;
        }
}
 
double LinearRegression::getWeightUpdateSum(unsigned int weightNum, unsigned int whichOutput)
{
        assert(whichOutput < weight.size());
        assert(weightNum < weight[whichOutput].size());
        
        return updates[whichOutput][weightNum].totErr;
}
 
void LinearRegression::resetWeightVariance(unsigned int weightNum, unsigned int whichOutput)
{
        updates[whichOutput][weightNum].reset();
}
 
void LinearRegression::resetWeightVariance()
{
        for (unsigned int y = 0; y < updates.size(); y++)
        {
                for (unsigned int x = 0; x < updates[y].size(); x++)
                {
                        updates[y][x].reset();
                }
        }
}
 
int LinearRegression::getWeightFrequency(unsigned int weightNum, unsigned int whichOutput)
{
        return updates[whichOutput][weightNum].n;
}
 
void LinearRegression::setInputWeight(double value, unsigned int weightNum, unsigned int whichOutput)
{
        updates[whichOutput][weightNum].reset();
        weight[whichOutput][weightNum] = value;
}
 
void LinearRegression::Print()
{
        for (int y = 0; y < outputs; y++)
                for (int x = 0; x <= inputs; x++)
                {
                        printf("%1.3f  ", weight[y][x]);
                }
        printf("\n");
}