tbb/examples/graph/som/som.cpp

/*
    Copyright 2005-2016 Intel Corporation.  All Rights Reserved.

    This file is part of Threading Building Blocks. Threading Building Blocks is free software;
    you can redistribute it and/or modify it under the terms of the GNU General Public License
    version 2  as  published  by  the  Free Software Foundation.  Threading Building Blocks is
    distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the
    implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
    See  the GNU General Public License for more details.   You should have received a copy of
    the  GNU General Public License along with Threading Building Blocks; if not, write to the
    Free Software Foundation, Inc.,  51 Franklin St,  Fifth Floor,  Boston,  MA 02110-1301 USA

    As a special exception,  you may use this file  as part of a free software library without
    restriction.  Specifically,  if other files instantiate templates  or use macros or inline
    functions from this file, or you compile this file and link it with other files to produce
    an executable,  this file does not by itself cause the resulting executable to be covered
    by the GNU General Public License. This exception does not however invalidate any other
    reasons why the executable file might be covered by the GNU General Public License.
*/

//
// Self-organizing map in TBB flow::graph
//
// we will do a color map (the simple example.)
//
//  serial algorithm
//
//       initialize map with vectors (could be random, gradient, or something else)
//       for some number of iterations
//           update radius r, weight of change L 
//           for each example V
//               find the best matching unit
//               for each part of map within radius of BMU W
//                   update vector:  W(t+1) = W(t) + w(dist)*L*(V - W(t))

#include "som.h"
#include "tbb/task.h"

std::ostream& operator<<( std::ostream &out, const SOM_element &s) {
    out << "(";
    for(int i=0;i<(int)s.w.size();++i) {
        out << s.w[i];
        if(i < (int)s.w.size()-1) {
            out << ",";
        }
    }
    out << ")";
    return out;
}

void remark_SOM_element(const SOM_element &s) {
    printf("(");
    for(int i=0;i<(int)s.w.size();++i) {
        printf("%g",s.w[i]);
        if(i < (int)s.w.size()-1) {
            printf(",");
        }
    }
    printf(")");
}

std::ostream& operator<<( std::ostream &out, const search_result_type &s) {
    out << "<";
    out << get<RADIUS>(s);
    out <<  ", " << get<XV>(s);
    out << ", ";
    out << get<YV>(s);
    out << ">";
    return out;
}

void remark_search_result_type(const search_result_type &s) {
    printf("<%g,%d,%d>", get<RADIUS>(s), get<XV>(s), get<YV>(s));
}

double
randval( double lowlimit, double highlimit) {
    return double(rand()) / double(RAND_MAX) * (highlimit - lowlimit) + lowlimit;
}

void
find_data_ranges(teaching_vector_type &teaching, SOM_element &max_range, SOM_element &min_range ) {
    if(teaching.size() == 0) return;
    max_range = min_range = teaching[0];
    for(int i = 1; i < (int)teaching.size(); ++i) {
        max_range.elementwise_max(teaching[i]);
        min_range.elementwise_min(teaching[i]);
    }
} 

void add_fraction_of_difference( SOM_element &to, SOM_element const &from, double frac) {
    for(int i = 0; i < (int)from.size(); ++i) {
        to[i] += frac*(from[i] - to[i]);
    }
}

double
distance_squared(SOM_element x, SOM_element y) {
    double rval = 0.0; for(int i=0;i<(int)x.size();++i) {
        double diff = x[i] - y[i];
        rval += diff*diff;
    }
    return rval;
}

void SOMap::initialize(InitializeType it, SOM_element &max_range, SOM_element &min_range) {
    for(int x = 0; x < xMax; ++x) {
        for(int y = 0; y < yMax; ++y) {
            for( int i = 0; i < (int)max_range.size(); ++i) {
                if(it == InitializeRandom) {
                    my_map[x][y][i] = (randval(min_range[i], max_range[i]));
                }
                else if(it == InitializeGradient) {
                    my_map[x][y][i] = ((double)(x+y)/(xMax+yMax)*(max_range[i]-min_range[i]) + min_range[i]);
                }
            }
        }
    }
}

// subsquare [low,high)
double
SOMap::BMU_range( const SOM_element &s, int &xval, int &yval, subsquare_type &r) {
    double min_distance_squared = DBL_MAX;
    task &my_task = task::self();
    int min_x = -1;
    int min_y = -1;
    for(int x = r.rows().begin(); x != r.rows().end(); ++x) {
        for( int y = r.cols().begin(); y != r.cols().end(); ++y) {
            double dist = distance_squared(s,my_map[x][y]);
            if(dist < min_distance_squared) {
                min_distance_squared = dist;
                min_x = x;
                min_y = y;
            }
            if(cancel_test && my_task.is_cancelled()) {
                xval = r.rows().begin();
                yval = r.cols().begin();
                return DBL_MAX;
            }
        }
    }
    xval = min_x;
    yval = min_y;
    return sqrt(min_distance_squared);
}

void
SOMap::epoch_update_range( SOM_element const &s, int epoch, int min_x, int min_y, double radius, double learning_rate, blocked_range<int> &r) {
    int min_xiter = (int)((double)min_x - radius);
    if(min_xiter < 0) min_xiter = 0;
    int max_xiter = (int)((double)min_x + radius);
    if(max_xiter > (int)my_map.size()-1) max_xiter = (int)my_map.size()-1;
    for(int xx = r.begin(); xx <= r.end(); ++xx) {
        double xrsq = (xx-min_x)*(xx-min_x);
        double ysq = radius*radius - xrsq;  // max extent of y influence
        double yd;
        if(ysq > 0) {
            yd = sqrt(ysq);
            int lb = (int)(min_y - yd);
            int ub = (int)(min_y + yd);
            for(int yy = lb; yy < ub; ++yy) {
                if(yy >= 0 && yy < (int)my_map[xx].size()) {
                    // [xx, yy] is in the range of the update.
                    double my_rsq = xrsq + (yy-min_y)*(yy-min_y);  // distance from BMU squared
                    double theta = exp(-(radius*radius) /(2.0* my_rsq)); 
                    add_fraction_of_difference(my_map[xx][yy], s, theta * learning_rate);
                }
            }
        }
    }
}

void SOMap::teach(teaching_vector_type &in) {
    for(int i = 0; i < nPasses; ++i ) {
        int j = (int)(randval(0, (double)in.size()));  // this won't be reproducible.
        if(j == in.size()) --j;
        
        int min_x = -1;
        int min_y = -1;
        subsquare_type br2(0, (int)my_map.size(), 1, 0, (int)my_map[0].size(), 1);
        (void) BMU_range(in[j],min_x, min_y, br2);  // just need min_x, min_y
        // radius of interest
        double radius = max_radius * exp(-(double)i*radius_decay_rate);
        // update circle is min_xiter to max_xiter inclusive.
        double learning_rate = max_learning_rate * exp( -(double)i * learning_decay_rate);
        epoch_update(in[j], i, min_x, min_y, radius, learning_rate);
    }
}

void SOMap::debug_output() {
    printf("SOMap:\n");
    for(int i = 0; i < (int)(this->my_map.size()); ++i) {
        for(int j = 0; j < (int)(this->my_map[i].size()); ++j) {
            printf( "map[%d, %d] == ", i, j );
            remark_SOM_element( this->my_map[i][j] );
            printf("\n");
        }
    }
}

#define RED 0
#define GREEN 1
#define BLUE 2

void readInputData() {
    my_teaching.push_back(SOM_element());
    my_teaching.push_back(SOM_element());
    my_teaching.push_back(SOM_element());
    my_teaching.push_back(SOM_element());
    my_teaching.push_back(SOM_element());
    my_teaching[0][RED] = 1.0; my_teaching[0][GREEN] = 0.0; my_teaching[0][BLUE] = 0.0;
    my_teaching[1][RED] = 0.0; my_teaching[1][GREEN] = 1.0; my_teaching[1][BLUE] = 0.0;
    my_teaching[2][RED] = 0.0; my_teaching[2][GREEN] = 0.0; my_teaching[2][BLUE] = 1.0;
    my_teaching[3][RED] = 0.3; my_teaching[3][GREEN] = 0.3; my_teaching[3][BLUE] = 0.0;
    my_teaching[4][RED] = 0.5; my_teaching[4][GREEN] = 0.5; my_teaching[4][BLUE] = 0.9;
}