searcher.h

Go to the documentation of this file.

// Copyright 2012 Yandex Artem Babenko
#ifndef SEARCHER_H_
#define SEARCHER_H_

#include <algorithm>
#include <map>

#include <boost/archive/binary_iarchive.hpp>
#include <boost/archive/binary_oarchive.hpp>

#include <boost/serialization/serialization.hpp>
#include <boost/serialization/set.hpp>
#include <boost/serialization/vector.hpp>

#include <mkl_cblas.h>

#include "data_util.h"
#include "ordered_lists_merger.h"
#include "perfomance_util.h"

extern int THREADS_COUNT;

extern Dimensions SPACE_DIMENSION;

extern enum PointType point_type;

typedef vector<pair<Distance, ClusterId> > NearestSubspaceCentroids;

template<class Record, class MetaInfo>
class MultiSearcher {
 public:
  MultiSearcher();
  void Init(const string& index_files_prefix,
            const string& coarse_vocabs_filename,
            const string& fine_vocabs_filename,
            const RerankMode& mode,
            const int subspace_centroids_to_consider,
            bool do_rerank);
  void GetNearestNeighbours(const Point& point, int k, 
                            vector<pair<Distance, MetaInfo> >* neighbours) const;
  PerfTester& GetPerfTester();
 private:
  void DeserializeData(const string& index_files_prefix,
                       const string& coarse_vocabs_filename,
                       const string& fine_vocabs_filename);
  void GetNearestSubspacesCentroids(const Point& point,
                                    const int subspace_centroins_count,
                                    vector<NearestSubspaceCentroids>* subspaces_short_lists) const;

  bool TraverseNextMultiIndexCell(const Point& point,
                                  vector<pair<Distance, MetaInfo> >* nearest_subpoints) const;
inline void GetCellEdgesInMultiIndexArray(const vector<int>& cell_coordinates,
                                          int* cell_start, int* cell_finish) const;
  void InitBlasStructures();
  vector<Centroids> coarse_vocabs_;
  vector<Centroids> fine_vocabs_;
  mutable OrderedListsMerger<Distance, ClusterId> merger_;
  bool do_rerank_;
  mutable PerfTester perf_tester_;
  string index_files_prefix_;
  MultiIndex<Record> multiindex_;
  RerankMode rerank_mode_;
  vector<float*> coarse_vocabs_matrices_;
  vector<vector<float> > coarse_centroids_norms_;
  mutable Coord* products_;
  mutable vector<Coord> query_norms_;
  mutable float* residual_;
  int subspace_centroids_to_consider_;
  mutable int found_neghbours_count_;
};

template<class Record, class MetaInfo>
inline void RecordToMetainfoAndDistance(const Coord* point,
                                        const Record& record,
                                        pair<Distance, MetaInfo>* result,
                                        const vector<int>& cell_coordinates,
                                        const vector<Centroids>& fine_vocabs) {
}


template<class Record, class MetaInfo>
MultiSearcher<Record, MetaInfo>::MultiSearcher() {
}

template<class Record, class MetaInfo>
void MultiSearcher<Record, MetaInfo>::DeserializeData(const string& index_files_prefix,
                                                      const string& coarse_vocabs_filename,
                                                      const string& fine_vocabs_filename) {
  cout << "Data deserializing started...\n";
  ifstream cell_edges(string(index_files_prefix + "_cell_edges.bin").c_str(), ios::binary);
  if(!cell_edges.good()) {
    throw std::logic_error("Bad input cell edges stream");
  }
  boost::archive::binary_iarchive arc_cell_edges(cell_edges);
  arc_cell_edges >> multiindex_.cell_edges;
  cout << "Cell edges deserialized...\n";
  ifstream multi_array(string(index_files_prefix + "_multi_array.bin").c_str(), ios::binary);
  if(!multi_array.good()) {
    throw std::logic_error("Bad input cell edges stream");
  }
  boost::archive::binary_iarchive arc_multi_array(multi_array);
  arc_multi_array >> multiindex_.multiindex;
  cout << "Multiindex deserialized...\n";
  ReadVocabularies<float>(coarse_vocabs_filename, SPACE_DIMENSION, &coarse_vocabs_);
  cout << "Coarse vocabs deserialized...\n";
  ReadFineVocabs<float>(fine_vocabs_filename, &fine_vocabs_);
  cout << "Fine vocabs deserialized...\n";
}

template<class Record, class MetaInfo>
void MultiSearcher<Record, MetaInfo>::Init(const string& index_files_prefix,
                                           const string& coarse_vocabs_filename,
                                           const string& fine_vocabs_filename,
                                           const RerankMode& mode,
                                           const int subspace_centroids_to_consider,
                                           const bool do_rerank) {
  do_rerank_ = do_rerank;
  index_files_prefix_ = index_files_prefix;
  subspace_centroids_to_consider_ = subspace_centroids_to_consider;
  DeserializeData(index_files_prefix, coarse_vocabs_filename, fine_vocabs_filename);
  rerank_mode_ = mode;
  merger_.GetYieldedItems().table.resize(std::pow((float)subspace_centroids_to_consider,
                                                         (int)coarse_vocabs_.size()));
  for(int i = 0; i < coarse_vocabs_.size(); ++i) {
    merger_.GetYieldedItems().dimensions.push_back(subspace_centroids_to_consider);
  }
  InitBlasStructures();
}

template<class Record, class MetaInfo>
void MultiSearcher<Record, MetaInfo>::InitBlasStructures(){
  coarse_vocabs_matrices_.resize(coarse_vocabs_.size());
  coarse_centroids_norms_.resize(coarse_vocabs_.size(), vector<float>(coarse_vocabs_[0].size()));
  for(int coarse_id = 0; coarse_id < coarse_vocabs_matrices_.size(); ++coarse_id) {
    coarse_vocabs_matrices_[coarse_id] = new float[coarse_vocabs_[0].size() * coarse_vocabs_[0][0].size()];
    for(int i = 0; i < coarse_vocabs_[0].size(); ++i) {
      Coord norm = 0;
      for(int j = 0; j < coarse_vocabs_[0][0].size(); ++j) {
        coarse_vocabs_matrices_[coarse_id][coarse_vocabs_[0][0].size() * i + j] = coarse_vocabs_[coarse_id][i][j];
        norm += coarse_vocabs_[coarse_id][i][j] * coarse_vocabs_[coarse_id][i][j];
      }
      coarse_centroids_norms_[coarse_id][i] = norm;
    }
  }
  products_ = new Coord[coarse_vocabs_[0].size()];
  query_norms_.resize(coarse_vocabs_[0].size());
  residual_ = new Coord[coarse_vocabs_[0][0].size() * coarse_vocabs_.size()];
}

template<class Record, class MetaInfo>
PerfTester& MultiSearcher<Record, MetaInfo>::GetPerfTester() {
  return perf_tester_;
}

template<class Record, class MetaInfo>
void MultiSearcher<Record, MetaInfo>::GetNearestSubspacesCentroids(const Point& point,
                                                                   const int subspace_centroins_count,
                                                                   vector<NearestSubspaceCentroids>*
                                                                   subspaces_short_lists) const {
  subspaces_short_lists->resize(coarse_vocabs_.size());
  Dimensions subspace_dimension = point.size() / coarse_vocabs_.size();
  for(int subspace_index = 0; subspace_index < coarse_vocabs_.size(); ++subspace_index) {
    Dimensions start_dim = subspace_index * subspace_dimension;
    Dimensions final_dim = std::min((Dimensions)point.size(), start_dim + subspace_dimension);
    Coord query_norm = cblas_sdot(final_dim - start_dim, &(point[start_dim]), 1, &(point[start_dim]), 1);
    std::fill(query_norms_.begin(), query_norms_.end(), query_norm);
    cblas_saxpy(coarse_vocabs_[0].size(), 1, &(coarse_centroids_norms_[subspace_index][0]), 1, &(query_norms_[0]), 1);
    cblas_sgemv(CblasRowMajor, CblasNoTrans, coarse_vocabs_[0].size(), subspace_dimension, -2.0,
                coarse_vocabs_matrices_[subspace_index], subspace_dimension, &(point[start_dim]), 1, 1, &(query_norms_[0]), 1);
    subspaces_short_lists->at(subspace_index).resize(query_norms_.size());
    for(int i = 0; i < query_norms_.size(); ++i) {
      subspaces_short_lists->at(subspace_index)[i] = std::make_pair(query_norms_[i], i);
    }
    std::nth_element(subspaces_short_lists->at(subspace_index).begin(),
                     subspaces_short_lists->at(subspace_index).begin() + subspace_centroins_count,
                     subspaces_short_lists->at(subspace_index).end());
    subspaces_short_lists->at(subspace_index).resize(subspace_centroins_count);
    std::sort(subspaces_short_lists->at(subspace_index).begin(),
              subspaces_short_lists->at(subspace_index).end());
  }
}

template<class Record, class MetaInfo>
void MultiSearcher<Record, MetaInfo>::GetCellEdgesInMultiIndexArray(const vector<int>& cell_coordinates,
                                                                    int* cell_start, int* cell_finish) const {
  int global_index = multiindex_.cell_edges.GetCellGlobalIndex(cell_coordinates);
  *cell_start = multiindex_.cell_edges.table[global_index];
  if(global_index + 1 == multiindex_.cell_edges.table.size()) {
    *cell_finish = multiindex_.multiindex.size();
  } else {
    *cell_finish = multiindex_.cell_edges.table[global_index + 1];
  }
}

template<class Record, class MetaInfo>
bool MultiSearcher<Record, MetaInfo>::TraverseNextMultiIndexCell(const Point& point,
                                                                 vector<pair<Distance, MetaInfo> >*
                                                                             nearest_subpoints) const {
  MergedItemIndices cell_inner_indices;
  clock_t before = clock();
  if(!merger_.GetNextMergedItemIndices(&cell_inner_indices)) {
    return false;
  }
  clock_t after = clock();
  perf_tester_.cell_coordinates_time += after - before;
  vector<int> cell_coordinates(cell_inner_indices.size());
  for(int list_index = 0; list_index < merger_.lists_ptr->size(); ++list_index) {
    cell_coordinates[list_index] = merger_.lists_ptr->at(list_index)[cell_inner_indices[list_index]].second;
  }
  int cell_start, cell_finish;
  before = clock();
  GetCellEdgesInMultiIndexArray(cell_coordinates, &cell_start, &cell_finish);
  after = clock();
  perf_tester_.cell_edges_time += after - before;
  if(cell_start >= cell_finish) {
    return true;
  }
  vector<Record>::const_iterator it = multiindex_.multiindex.begin() + cell_start;
  GetResidual(point, cell_coordinates, coarse_vocabs_, residual_);
  cell_finish = std::min((int)cell_finish, cell_start + (int)nearest_subpoints->size() - found_neghbours_count_);
  for(int array_index = cell_start; array_index < cell_finish; ++array_index) {
    if(rerank_mode_ == USE_RESIDUALS) {
      RecordToMetainfoAndDistance<Record, MetaInfo>(residual_, *it,
                                                    &(nearest_subpoints->at(found_neghbours_count_)),
                                                    cell_coordinates, fine_vocabs_);
    } else if(rerank_mode_ == USE_INIT_POINTS) {
      RecordToMetainfoAndDistance<Record, MetaInfo>(&(point[0]), *it,
                                                    &(nearest_subpoints->at(found_neghbours_count_)),
                                                    cell_coordinates, fine_vocabs_);
    }
    perf_tester_.NextNeighbour();
    ++found_neghbours_count_;
    ++it;
  }
  return true;
}


template<class Record, class MetaInfo>
void MultiSearcher<Record, MetaInfo>::GetNearestNeighbours(const Point& point, int k, 
                                                           vector<pair<Distance, MetaInfo> >* neighbours) const {
  perf_tester_.handled_queries_count += 1;
  neighbours->resize(k);
  perf_tester_.ResetQuerywiseStatistic();
  clock_t start = clock();
  perf_tester_.search_start = start;
  clock_t before = clock();
  vector<NearestSubspaceCentroids> subspaces_short_lists;
  GetNearestSubspacesCentroids(point, subspace_centroids_to_consider_, &subspaces_short_lists);
  clock_t after = clock();
  perf_tester_.nearest_subcentroids_time += after - before;
  clock_t before_merger = clock();
  merger_.setLists(subspaces_short_lists);
  clock_t after_merger = clock();
  perf_tester_.merger_init_time += after_merger - before_merger;
  clock_t before_traversal = clock();
  found_neghbours_count_ = 0;
  bool traverse_next_cell = true;
  int cells_visited = 0;
  while(found_neghbours_count_ < k && traverse_next_cell) {
    perf_tester_.cells_traversed += 1;
    traverse_next_cell = TraverseNextMultiIndexCell(point, neighbours);
    cells_visited += 1;
  }
  clock_t after_traversal = clock();
  perf_tester_.full_traversal_time += after_traversal - before_traversal;
  if(do_rerank_) {
    if(neighbours->size() > 10000) {
      std::nth_element(neighbours->begin(), neighbours->begin() + 10000, neighbours->end());
      neighbours->resize(10000);
    }
    std::sort(neighbours->begin(), neighbours->end());
  }
  clock_t finish = clock();
  perf_tester_.full_search_time += finish - start;
}

template<>
inline void RecordToMetainfoAndDistance<RerankADC8, PointId>(const Coord* point, const RerankADC8& record,
                                                             pair<Distance, PointId>* result,
                                                             const vector<int>& cell_coordinates,
                                                             const vector<Centroids>& fine_vocabs) {
  result->second = record.pid;
  int coarse_clusters_count = cell_coordinates.size();
  int fine_clusters_count = fine_vocabs.size();
  int coarse_to_fine_ratio = fine_clusters_count / coarse_clusters_count;
  int subvectors_dim = SPACE_DIMENSION / fine_clusters_count;
  char* rerank_info_ptr = (char*)&record + sizeof(record.pid);
  for(int centroid_index = 0; centroid_index < fine_clusters_count; ++centroid_index) {
    int start_dim = centroid_index * subvectors_dim;
    int final_dim = start_dim + subvectors_dim;
    FineClusterId pid_nearest_centroid = *((FineClusterId*)rerank_info_ptr);
    rerank_info_ptr += sizeof(FineClusterId);
    int current_coarse_index = centroid_index / coarse_to_fine_ratio;
    Distance subvector_distance = 0;
    for(int i = start_dim; i < final_dim; ++i) {
      Coord diff = fine_vocabs[centroid_index][pid_nearest_centroid][i - start_dim] - point[i];
        subvector_distance += diff * diff;
    }
    result->first += subvector_distance;
  }
}

template<>
inline void RecordToMetainfoAndDistance<RerankADC16, PointId>(const Coord* point, const RerankADC16& record,
                                                              pair<Distance, PointId>* result,
                                                              const vector<int>& cell_coordinates,
                                                              const vector<Centroids>& fine_vocabs) {
  result->second = record.pid;
  int coarse_clusters_count = cell_coordinates.size();
  int fine_clusters_count = fine_vocabs.size();
  int coarse_to_fine_ratio = fine_clusters_count / coarse_clusters_count;
  int subvectors_dim = SPACE_DIMENSION / fine_clusters_count;
  char* rerank_info_ptr = (char*)&record + sizeof(record.pid);
  for(int centroid_index = 0; centroid_index < fine_clusters_count; ++centroid_index) {
    int start_dim = centroid_index * subvectors_dim;
    int final_dim = start_dim + subvectors_dim;
    FineClusterId pid_nearest_centroid = *((FineClusterId*)rerank_info_ptr);
    rerank_info_ptr += sizeof(FineClusterId);
    int current_coarse_index = centroid_index / coarse_to_fine_ratio;
    Distance subvector_distance = 0;
    for(int i = start_dim; i < final_dim; ++i) {
      Coord diff = fine_vocabs[centroid_index][pid_nearest_centroid][i - start_dim] - point[i];
      subvector_distance += diff * diff;
    }
    result->first += subvector_distance;
  }
}

template class MultiSearcher<RerankADC8, PointId>;
template class MultiSearcher<RerankADC16, PointId>;
template class MultiSearcher<PointId, PointId>;

#endif