machine_learning/kmeans.cpp

/*******************************************************
 * Copyright (c) 2014, ArrayFire
 * All rights reserved.
 *
 * This file is distributed under 3-clause BSD license.
 * The complete license agreement can be obtained at:
 * http://arrayfire.com/licenses/BSD-3-Clause
 ********************************************************/

#include <iostream>
#include <stdio.h>
#include <arrayfire.h>
#include <af/util.h>
#include <cstdlib>

using namespace af;

array distance(array data, array means)
{
    int n = data.dims(0); // Number of features
    int k = means.dims(1); // Number of means

    array data2  = tile(data , 1, k, 1);
    array means2 = tile(means, n, 1, 1);

    // Currently using manhattan distance
    // Can be replaced with other distance measures
    return sum(abs(data2 - means2), 2);
}

// Get cluster id of each location in data
array clusterize(const array data, const array means)
{
    // Get manhattan distance
    array dists = distance(data, means);

    // get the locations of minimum distance
    array idx, val;
    min(val, idx, dists, 1);

    // Return cluster IDs
    return idx;
}

array new_means(array data, array clusters, int k)
{
    int d = data.dims(2);
    array means = constant(0, 1, k, d);
    array clustersd = tile(clusters, 1, 1, d);

    gfor (seq ii, k) {
        means(span, ii, span) = sum(data * (clustersd == ii)) / (sum(clusters == ii) + 1e-5);
    }

    return means;
}

// kmeans(means, clusters, data, k)
// data:  input,  1D or 2D (range > [0-1])
// k:     input,  # desired means (k > 1)
// means: output, vector of means
void kmeans(array &means, array &clusters, const array in, int k, int iter=100)
{
    unsigned n = in.dims(0); // Num features
    unsigned d = in.dims(2); // feature length

    // reshape input
    array data = in * 0;

    // re-center and scale down data to [0, 1]
    array minimum = min(in);
    array maximum = max(in);

    gfor(seq ii, d) {
        data(span, span, ii) = (in(span, span, ii) - minimum(ii).scalar<float>()) / maximum(ii).scalar<float>();
    }

    // Initial guess of means
    means = randu(1, k, d);
    array curr_clusters = constant(0, data.dims(0)) - 1;
    array prev_clusters;

    // Stop updating after specified number of iterations
    for (int i = 0; i < iter; i++) {
        // Store previous cluster ids
        prev_clusters = curr_clusters;

        // Get cluster ids for current means
        curr_clusters = clusterize(data, means);

        // Break early if clusters not changing
        unsigned num_changed = count<unsigned>(prev_clusters != curr_clusters);

        if (num_changed < (n/1000) + 1) break;

        // Update current means for new clusters
        means = new_means(data, curr_clusters, k);
    }

    // Scale up means
    gfor(seq ii, d) {
        means(span, span, ii) = maximum(ii) * means(span, span, ii) + minimum(ii);
    }

    clusters = prev_clusters;

}

// K-Means image recoloring.
// Shifts the hues of an image to the k mean hues.
int kmeans_demo(int k, bool console)
{
    printf("** ArrayFire K-Means Demo (k = %d) **\n\n", k);

    array img = loadImage(ASSETS_DIR"/examples/images/vegetable-woman.jpg", true) / 255; // [0-255]

    int w = img.dims(0), h = img.dims(1), c = img.dims(2);
    array vec = moddims(img, w * h, 1, c);

    array means_full, clusters_full;
    kmeans(means_full, clusters_full, vec, k);

    array means_half, clusters_half;
    kmeans(means_half, clusters_half, vec, k / 2);

    array means_dbl, clusters_dbl;
    kmeans(means_dbl, clusters_dbl, vec, k * 2);

    if (!console) {
#if 0
        array out_full = moddims(means_full(span, clusters_full, span), img.dims());
        array out_half = moddims(means_half(span, clusters_half, span), img.dims());
        array out_dbl  = moddims(means_dbl (span, clusters_dbl , span), img.dims());

        char str_full[32], str_half[32], str_dbl[32];
        sprintf(str_full, "%2d clusters", k);
        sprintf(str_half, "%2d clusters", k/2);
        sprintf(str_dbl , "%2d clusters", k*2);

        fig("color","default");
        fig("sub",2,2,1); image(img); fig("title","input");
        fig("sub",2,2,2); image(out_full); fig("title", str_full);
        fig("sub",2,2,3); image(out_half); fig("title", str_half);
        fig("sub",2,2,4); image(out_dbl ); fig("title", str_dbl );
        printf("Hit enter to finish\n");
        getchar();
#else
        printf("Graphics not implemented yet\n");
#endif
    } else {
        means_full = moddims(means_full, means_full.dims(1), means_full.dims(2));
        means_half = moddims(means_half, means_half.dims(1), means_half.dims(2));
        means_dbl  = moddims(means_dbl , means_dbl.dims(1) , means_dbl.dims(2) );

        af_print(means_full);
        af_print(means_half);
        af_print(means_dbl );
    }

    return 0;
}

int main(int argc, char** argv)
{
    int device = argc > 1 ? atoi(argv[1]) : 0;
    bool console = argc > 2 ? argv[2][0] == '-' : false;
    int k = argc > 3 ? atoi(argv[3]) : 16;

    try {

        af::setDevice(device);
        af::info();
        return kmeans_demo(k, console);

    } catch (af::exception &ae) {
        std::cerr << ae.what() << std::endl;
    }

    return 0;
}