Annoy.hpp

#ifndef KNNCOLLE_ANNOY_ANNOY_HPP
#define KNNCOLLE_ANNOY_ANNOY_HPP
 
#include <vector>
#include <algorithm>
#include <memory>
#include <cstddef>
#include <cstring>
#include <stdexcept>
#include <limits>
#include <filesystem>
 
#include "knncolle/knncolle.hpp"
#include "annoy/annoylib.h"
#include "annoy/kissrandom.h"
 
#include "utils.hpp"
 
namespace knncolle_annoy {
 
inline static constexpr const char* annoy_prebuilt_save_name = "knncolle_annoy::Annoy";
 
struct AnnoyOptions {
    int num_trees = 50;
 
    double search_mult = -1;
};
 
template<typename Index_, typename Data_, typename Distance_, typename AnnoyDistance_, typename AnnoyIndex_, typename AnnoyData_, class AnnoyRng_, class AnnoyThreadPolicy_>
class AnnoyPrebuilt;
 
template<
    typename Index_,
    typename Data_,
    typename Distance_, 
    class AnnoyDistance_,
    typename AnnoyIndex_ = Index_,
    typename AnnoyData_ = float,
    class AnnoyRng_ = Annoy::Kiss64Random,
    class AnnoyThreadPolicy_ = Annoy::AnnoyIndexSingleThreadedBuildPolicy
>
class AnnoySearcher final : public knncolle::Searcher<Index_, Data_, Distance_> {
private:
    const AnnoyPrebuilt<Index_, Data_, Distance_, AnnoyDistance_, AnnoyIndex_, AnnoyData_, AnnoyRng_, AnnoyThreadPolicy_>& my_parent;
 
    static constexpr bool same_internal_data = std::is_same<Data_, AnnoyData_>::value;
    typename std::conditional<!same_internal_data, std::vector<AnnoyData_>, bool>::type my_buffer;
 
    static constexpr bool same_internal_index = std::is_same<Index_, AnnoyIndex_>::value;
    std::vector<AnnoyIndex_> my_indices;
 
    static constexpr bool same_internal_distance = std::is_same<Distance_, AnnoyData_>::value;
    typename std::conditional<!same_internal_distance, std::vector<AnnoyData_>, bool>::type my_distances;
 
    typedef int SearchKType;
 
    Index_ my_capped_k = 0;
 
    SearchKType get_search_k(Index_ k) const {
        if (my_parent.my_search_mult < 1) {
            return -1; // instructs Annoy to use k * num_trees. 
        } else if (k <= my_capped_k) {
            return my_parent.my_search_mult * static_cast<double>(k) + 0.5; // rounded.
        } else {
            return std::numeric_limits<SearchKType>::max(); // cap to avoid overflow.
        }
    }
 
public:
    AnnoySearcher(const AnnoyPrebuilt<Index_, Data_, Distance_, AnnoyDistance_, AnnoyIndex_, AnnoyData_, AnnoyRng_, AnnoyThreadPolicy_>& parent) : my_parent(parent) {
        if constexpr(!same_internal_data) {
            sanisizer::resize(my_buffer, my_parent.my_dim);
        }
 
        if (my_parent.my_search_mult >= 1) {
            my_capped_k = static_cast<double>(std::numeric_limits<SearchKType>::max()) / my_parent.my_search_mult;
        }
    }
 
private:
    std::pair<std::vector<AnnoyIndex_>*, std::vector<AnnoyData_>*> obtain_pointers(std::vector<Index_>* output_indices, std::vector<Distance_>* output_distances, Index_ k) {
        std::vector<AnnoyIndex_>* icopy_ptr = &my_indices;
        if (output_indices) {
            if constexpr(same_internal_index) {
                icopy_ptr = output_indices;
            }
        }
        icopy_ptr->clear();
        icopy_ptr->reserve(k);
 
        std::vector<AnnoyData_>* dcopy_ptr = NULL;
        if (output_distances) {
            if constexpr(same_internal_distance) {
                dcopy_ptr = output_distances;
            } else {
                dcopy_ptr = &my_distances;
            }
            dcopy_ptr->clear();
            dcopy_ptr->reserve(k);
        }
 
        return std::make_pair(icopy_ptr, dcopy_ptr);
    }
 
    template<typename Type_>
    static void remove_self(std::vector<Type_>& vec, std::size_t at) {
        if (at < vec.size()) {
            vec.erase(vec.begin() + at);
        } else {
            vec.pop_back();
        }
    }
 
    template<typename Source_, typename Dest_>
    static void copy_skip_self(const std::vector<Source_>& source, std::vector<Dest_>& dest, std::size_t at) {
        auto sIt = source.begin();
        auto end = source.size();
        dest.clear();
        dest.reserve(end - 1);
 
        if (at < end) {
            dest.insert(dest.end(), sIt, sIt + at);
            dest.insert(dest.end(), sIt + at + 1, source.end());
        } else {
            // Just in case we're full of ties at duplicate points, such that 'c'
            // is not in the set.  Note that, if self_found=false, we must have at
            // least 'k+2' points for 'c' to not be detected as its own neighbor.
            // Thus there is no need to worry whether 'end - 1 != k'; we
            // are guaranteed to return 'k' elements in this case.
            dest.insert(dest.end(), sIt, sIt + end - 1);
        }
    }
 
public:
    void search(Index_ i, Index_ k, std::vector<Index_>* output_indices, std::vector<Distance_>* output_distances) {
        // +1, as it forgets to discard 'self'. This should not overflow as k < num_obs.
        const auto kp1 = k + 1;
 
        auto ptrs = obtain_pointers(output_indices, output_distances, kp1);
        auto icopy_ptr = ptrs.first;
        auto dcopy_ptr = ptrs.second;
 
        my_parent.my_index.get_nns_by_item(
            i,
            sanisizer::cast<std::size_t>(sanisizer::attest_gez(kp1)),
            get_search_k(kp1),
            icopy_ptr,
            dcopy_ptr
        );
 
        std::size_t at;
        {
            const auto& cur_i = *icopy_ptr;
            at = cur_i.size();
            const AnnoyIndex_ icopy = i; // cast to AnnoyIndex_ to avoid signed/unsigned comparisons.
            for (std::size_t x = 0, end = cur_i.size(); x < end; ++x) {
                if (cur_i[x] == icopy) {
                    at = x;
                    break;
                }
            }
        }
 
        if (output_indices) {
            if constexpr(same_internal_index) {
                remove_self(*output_indices, at);
            } else {
                copy_skip_self(my_indices, *output_indices, at);
            }
        }
 
        if (output_distances) {
            if constexpr(same_internal_distance) {
                remove_self(*output_distances, at);
            } else {
                copy_skip_self(my_distances, *output_distances, at);
            }
        }
    }
 
private:
    void search_raw(const AnnoyData_* query, Index_ k, std::vector<Index_>* output_indices, std::vector<Distance_>* output_distances) {
        auto ptrs = obtain_pointers(output_indices, output_distances, k);
        auto icopy_ptr = ptrs.first;
        auto dcopy_ptr = ptrs.second;
 
        my_parent.my_index.get_nns_by_vector(
            query,
            sanisizer::cast<std::size_t>(sanisizer::attest_gez(k)),
            get_search_k(k),
            icopy_ptr,
            dcopy_ptr
        );
 
        if (output_indices) {
            if constexpr(!same_internal_index) {
                output_indices->clear();
                output_indices->insert(output_indices->end(), my_indices.begin(), my_indices.end());
            }
        }
 
        if (output_distances) {
            if constexpr(!same_internal_distance) {
                output_distances->clear();
                output_distances->insert(output_distances->end(), my_distances.begin(), my_distances.end());
            }
        }
    }
 
public:
    void search(const Data_* query, Index_ k, std::vector<Index_>* output_indices, std::vector<Distance_>* output_distances) {
        if constexpr(same_internal_data) {
            search_raw(query, k, output_indices, output_distances);
        } else {
            std::copy_n(query, my_parent.my_dim, my_buffer.begin());
            search_raw(my_buffer.data(), k, output_indices, output_distances);
        }
    }
};
 
template<
    typename Index_,
    typename Data_,
    typename Distance_, 
    class AnnoyDistance_,
    typename AnnoyIndex_ = Index_,
    typename AnnoyData_ = float,
    class AnnoyRng_ = Annoy::Kiss64Random,
    class AnnoyThreadPolicy_ = Annoy::AnnoyIndexSingleThreadedBuildPolicy
>
class AnnoyPrebuilt final : public knncolle::Prebuilt<Index_, Data_, Distance_> {
public:
    template<class Matrix_>
    AnnoyPrebuilt(const Matrix_& data, const AnnoyOptions& options) :
        my_dim(data.num_dimensions()),
        my_obs(data.num_observations()),
        my_search_mult(options.search_mult),
        my_index(my_dim)
    {
        // check that we can, in fact, represent the number of observations in the specified AnnoyIndex_ type.
        sanisizer::cast<AnnoyIndex_>(sanisizer::attest_gez(my_obs));
 
        auto work = data.new_known_extractor();
        if constexpr(std::is_same<Data_, AnnoyData_>::value) {
            for (Index_ i = 0; i < my_obs; ++i) {
                auto ptr = work->next();
                my_index.add_item(i, ptr);
            }
        } else {
            auto incoming = sanisizer::create<std::vector<AnnoyData_> >(my_dim);
            for (Index_ i = 0; i < my_obs; ++i) {
                auto ptr = work->next();
                std::copy_n(ptr, my_dim, incoming.begin());
                my_index.add_item(i, incoming.data());
            }
        }
 
        my_index.build(options.num_trees);
        return;
    }
 
private:
    std::size_t my_dim;
    Index_ my_obs;
    double my_search_mult;
 
    // Unfortunately, AnnoyIndex::save is not const (and also unloads and
    // reloads the index, for reasons I don't understand). So we manually
    // handle the saving by looking into the protected members.
    struct SuperHackyThing final : public Annoy::AnnoyIndex<AnnoyIndex_, AnnoyData_, AnnoyDistance_, AnnoyRng_, AnnoyThreadPolicy_> {
        template<typename ... Args_>
        SuperHackyThing(Args_&& ... args) : Annoy::AnnoyIndex<AnnoyIndex_, AnnoyData_, AnnoyDistance_, AnnoyRng_, AnnoyThreadPolicy_>(std::forward<Args_>(args)...) {}
 
        auto get_nodes() const {
            return this->_nodes;
        }
 
        auto get_s() const {
            return this->_s;
        }
 
        auto get_n_nodes() const {
            return this->_n_nodes;
        }
    };
    SuperHackyThing my_index;
 
    friend class AnnoySearcher<Index_, Data_, Distance_, AnnoyDistance_, AnnoyIndex_, AnnoyData_, AnnoyRng_, AnnoyThreadPolicy_>;
 
public:
    std::size_t num_dimensions() const {
        return my_dim;
    }
 
    Index_ num_observations() const {
        return my_obs;
    }
 
    std::unique_ptr<knncolle::Searcher<Index_, Data_, Distance_> > initialize() const {
        return initialize_known();
    }
 
    auto initialize_known() const {
        return std::make_unique<AnnoySearcher<Index_, Data_, Distance_, AnnoyDistance_, AnnoyIndex_, AnnoyData_, AnnoyRng_, AnnoyThreadPolicy_> >(*this);
    }
 
public:
    void save(const std::filesystem::path& dir) const {
        knncolle::quick_save(dir / "ALGORITHM", annoy_prebuilt_save_name, std::strlen(annoy_prebuilt_save_name));
        knncolle::quick_save(dir / "NUM_OBS", &my_obs, 1);
        knncolle::quick_save(dir / "NUM_DIM", &my_dim, 1);
        knncolle::quick_save(dir / "SEARCH_MULT", &my_search_mult, 1);
 
        knncolle::NumericType types[2];
        types[0] = knncolle::get_numeric_type<AnnoyIndex_>();
        types[1] = knncolle::get_numeric_type<AnnoyData_>();
        knncolle::quick_save(dir / "TYPES", types, 2);
 
        const auto dname = get_distance_name<AnnoyDistance_>();
        knncolle::quick_save(dir / "DISTANCE", dname, std::strlen(dname));
 
        // Handling custom Annoy types. 
        auto& icust = custom_save_for_annoy_index<AnnoyIndex_>(); 
        if (icust) {
            icust(dir); 
        }
 
        auto& dcust = custom_save_for_annoy_data<AnnoyData_>();
        if (dcust) {
            dcust(dir); 
        }
 
        auto& dscust = custom_save_for_annoy_distance<AnnoyDistance_>();
        if (dscust) {
            dscust(dir);
        }
 
        // Not bothering to save anything else; the RNG and thread policy
        // should not be relevant once the index is built.
 
        // For reasons unknown to us, AnnoyIndex::save() is not const, so we have to do it manually.
        // Hopefully this will be fixed in the future.
        const auto idxpath = dir / "INDEX";
        knncolle::quick_save(idxpath, reinterpret_cast<char*>(my_index.get_nodes()), sanisizer::product<std::streamsize>(my_index.get_s(), my_index.get_n_nodes()));
    }
 
    AnnoyPrebuilt(const std::filesystem::path& dir, std::size_t ndim) : my_dim(ndim), my_index(ndim) {
        knncolle::quick_load(dir / "NUM_OBS", &my_obs, 1);
        knncolle::quick_load(dir / "SEARCH_MULT", &my_search_mult, 1);
 
        const auto idxpath = (dir / "INDEX").string();
        char* errbuf = NULL;
        if (!my_index.load(idxpath.c_str(), true, &errbuf)) {
            std::runtime_error ex(errbuf);
            delete errbuf;
            throw ex;
        }
    }
};
template<
    typename Index_,
    typename Data_,
    typename Distance_, 
    class AnnoyDistance_,
    typename AnnoyIndex_ = Index_,
    typename AnnoyData_ = float,
    class AnnoyRng_ = Annoy::Kiss64Random,
    class AnnoyThreadPolicy_ = Annoy::AnnoyIndexSingleThreadedBuildPolicy,
    class Matrix_ = knncolle::Matrix<Index_, Data_>
>
class AnnoyBuilder : public knncolle::Builder<Index_, Data_, Distance_, Matrix_> {
private:
    AnnoyOptions my_options;
 
public:
    AnnoyBuilder(AnnoyOptions options) : my_options(std::move(options)) {}
 
    AnnoyBuilder() = default;
 
    AnnoyOptions& get_options() {
        return my_options;
    }
 
public:
    knncolle::Prebuilt<Index_, Data_, Distance_>* build_raw(const Matrix_& data) const {
        return build_known_raw(data);
    }
public:
    auto build_known_raw(const Matrix_& data) const {
        return new AnnoyPrebuilt<Index_, Data_, Distance_, AnnoyDistance_, AnnoyIndex_, AnnoyData_, AnnoyRng_, AnnoyThreadPolicy_>(data, my_options);
    }
 
    auto build_known_unique(const Matrix_& data) const {
        return std::unique_ptr<I<decltype(*build_known_raw(data))> >(build_known_raw(data));
    }
 
    auto build_known_shared(const Matrix_& data) const {
        return std::shared_ptr<I<decltype(*build_known_raw(data))> >(build_known_raw(data));
    }
};
 
}
 
#endif