Method divideFeatureBlocks

src/learner/bcd.h:78–122

  }

  void divideFeatureBlocks() {
    // partition feature blocks
    for (int i = 0; i < g_train_info_.slot_size(); ++i) {
      auto info = g_train_info_.slot(i);
      CHECK(info.has_id());
      if (info.id() == 0) continue;  // it's the label
      CHECK(info.has_nnz_ele());
      CHECK(info.has_nnz_ex());
      double nnz_per_row = (double)info.nnz_ele() / (double)info.nnz_ex();
      int n = 1;  // number of blocks for a feature group
      if (nnz_per_row > 1 + 1e-6) {
        // only parititon feature group whose features are correlated
        n = std::max((int)std::ceil(nnz_per_row * bcd_conf_.feature_block_ratio()), 1);
      }
      for (int i = 0; i < n; ++i) {
        auto block = Range<Key>(info.min_key(), info.max_key()).evenDivide(n, i);
        if (block.empty()) continue;
        fea_blk_.push_back(std::make_pair(info.id(), block));
      }
    }
    LI << "Features are partitioned into " << fea_blk_.size() << " blocks";

    // a simple block order
    for (int i = 0; i < fea_blk_.size(); ++i) blk_order_.push_back(i);

    // blocks for important feature groups
    std::vector<string> hit_blk;
    for (int i = 0; i < bcd_conf_.prior_fea_group_size(); ++i) {
      int grp_id = bcd_conf_.prior_fea_group(i);
      std::vector<int> tmp;
      for (int k = 0; k < fea_blk_.size(); ++k) {
        if (fea_blk_[k].first == grp_id) tmp.push_back(k);
      }
      if (tmp.empty()) continue;
      hit_blk.push_back(std::to_string(grp_id));
      for (int j = 0; j < bcd_conf_.num_iter_for_prior_fea_group(); ++j) {
        if (bcd_conf_.random_feature_block_order()) {
          std::random_shuffle(tmp.begin(), tmp.end());
        }
        prior_blk_order_.insert(prior_blk_order_.end(), tmp.begin(), tmp.end());
      }
    }
    if (!hit_blk.empty()) LI << "Prior feature groups: " + join(hit_blk, ", ");
    total_timer_.restart();
  }

  void saveModel(const DataConfig& data) {
    Task task;