libcruft-util/cruft/util/kmeans.hpp

/*
 * This Source Code Form is subject to the terms of the Mozilla Public
 * License, v. 2.0. If a copy of the MPL was not distributed with this
 * file, You can obtain one at http://mozilla.org/MPL/2.0/.
 *
 * Copyright 2018 Danny Robson <danny@nerdcruft.net>
 */

#pragma once

#include <cruft/util/debug/assert.hpp>
#include <cruft/util/iterator/zip.hpp>
#include <cruft/util/view.hpp>

#include <iterator>


namespace cruft {
    // a simplistic implementation of Lloyd's algorithm
    //
    // returns index of the closest output for each input
    template <typename OutputT, typename InputT, typename FunctionT>
    std::vector<size_t>
    kmeans (cruft::view<InputT> src, cruft::view<OutputT> dst, FunctionT const &&metric)
    {
        CHECK_GE (src.size (), dst.size ());

        using coord_t = typename std::iterator_traits<InputT>::value_type;
        const int iterations = 100;

        std::vector<coord_t> means (src.begin (), src.begin () + dst.size ());
        std::vector<coord_t> accum (dst.size ());
        std::vector<size_t> count (dst.size ());
        std::vector<size_t> closest (src.size ());

        for (auto i = 0; i < iterations; ++i) {
            std::fill (std::begin (accum), std::end (accum), 0);
            std::fill (std::begin (count), std::end (count), 0);

            for (auto const& [j,p]: iterator::izip (src)) {
                size_t bucket = 0;

                for (size_t k = 1; k < dst.size (); ++k) {
                    if (metric (p, means[k]) < metric (p, means[bucket]))
                        bucket = k;
                }

                accum[bucket] += p;
                count[bucket] += 1;
                closest[j] = bucket;
            }

            for (size_t j = 0; j < dst.size (); ++j)
                means[j] = accum[j] / count[j];
        }

        std::copy (std::begin (means), std::end (means), std::begin (dst));

        return closest;
    }


    template <typename OutputT, typename InputT>
    auto
    kmeans (InputT &&src, OutputT &&dst)
    {
        return kmeans (
            std::forward<InputT>  (src),
            std::forward<OutputT> (dst),
            [] (auto a, auto b) {
                return norm (a - b);
        });
    }

}