view release on metacpan or  search on metacpan

xgboost/src/gbm/gbtree.cc  view on Meta::CPAN

// gradient boosted trees
class GBTree : public GradientBooster {
 public:
  explicit GBTree(bst_float base_margin) : model_(base_margin) {}

  void InitCache(const std::vector<std::shared_ptr<DMatrix> > &cache) {
    cache_ = cache;
  }

  void Configure(const std::vector<std::pair<std::string, std::string> >& cfg) override {
    this->cfg = cfg;
    model_.Configure(cfg);
    // initialize the updaters only when needed.
    std::string updater_seq = tparam.updater_seq;
    tparam.InitAllowUnknown(cfg);
    if (updater_seq != tparam.updater_seq) updaters.clear();
    for (const auto& up : updaters) {
      up->Init(cfg);
    }
    // for the 'update' process_type, move trees into trees_to_update
    if (tparam.process_type == kUpdate) {
      model_.InitTreesToUpdate();
    }

    // configure predictor
    predictor = std::unique_ptr<Predictor>(Predictor::Create(tparam.predictor));
    predictor->Init(cfg, cache_);
  }

  void Load(dmlc::Stream* fi) override {
    model_.Load(fi);

    this->cfg.clear();
    this->cfg.push_back(std::make_pair(std::string("num_feature"),
                                       common::ToString(model_.param.num_feature)));
  }

  void Save(dmlc::Stream* fo) const override {
    model_.Save(fo);
  }

  bool AllowLazyCheckPoint() const override {
    return model_.param.num_output_group == 1 ||
        tparam.updater_seq.find("distcol") != std::string::npos;
  }

  void DoBoost(DMatrix* p_fmat,
               std::vector<bst_gpair>* in_gpair,
               ObjFunction* obj) override {
    const std::vector<bst_gpair>& gpair = *in_gpair;
    std::vector<std::vector<std::unique_ptr<RegTree> > > new_trees;
    const int ngroup = model_.param.num_output_group;
    if (ngroup == 1) {
      std::vector<std::unique_ptr<RegTree> > ret;
      BoostNewTrees(gpair, p_fmat, 0, &ret);
      new_trees.push_back(std::move(ret));
    } else {
      CHECK_EQ(gpair.size() % ngroup, 0U)
          << "must have exactly ngroup*nrow gpairs";
      std::vector<bst_gpair> tmp(gpair.size() / ngroup);
      for (int gid = 0; gid < ngroup; ++gid) {
        bst_omp_uint nsize = static_cast<bst_omp_uint>(tmp.size());
        #pragma omp parallel for schedule(static)
        for (bst_omp_uint i = 0; i < nsize; ++i) {
          tmp[i] = gpair[i * ngroup + gid];
        }
        std::vector<std::unique_ptr<RegTree> > ret;
        BoostNewTrees(tmp, p_fmat, gid, &ret);
        new_trees.push_back(std::move(ret));
      }
    }
    double tstart = dmlc::GetTime();
    for (int gid = 0; gid < ngroup; ++gid) {
      this->CommitModel(std::move(new_trees[gid]), gid);
    }
    if (tparam.debug_verbose > 0) {
      LOG(INFO) << "CommitModel(): " << dmlc::GetTime() - tstart << " sec";
    }
  }

  void PredictBatch(DMatrix* p_fmat,
               std::vector<bst_float>* out_preds,
               unsigned ntree_limit) override {
    predictor->PredictBatch(p_fmat, out_preds, model_, 0, ntree_limit);
  }

  void PredictInstance(const SparseBatch::Inst& inst,
               std::vector<bst_float>* out_preds,
               unsigned ntree_limit,
               unsigned root_index) override {
    predictor->PredictInstance(inst, out_preds, model_,
                               ntree_limit, root_index);
  }

  void PredictLeaf(DMatrix* p_fmat,
                   std::vector<bst_float>* out_preds,
                   unsigned ntree_limit) override {
    predictor->PredictLeaf(p_fmat, out_preds, model_, ntree_limit);
  }

  void PredictContribution(DMatrix* p_fmat,
                           std::vector<bst_float>* out_contribs,
                           unsigned ntree_limit) override {
    predictor->PredictContribution(p_fmat, out_contribs, model_, ntree_limit);
  }

  std::vector<std::string> DumpModel(const FeatureMap& fmap,
                                     bool with_stats,
                                     std::string format) const override {
    return model_.DumpModel(fmap, with_stats, format);
  }

 protected:
  // initialize updater before using them
  inline void InitUpdater() {
    if (updaters.size() != 0) return;
    std::string tval = tparam.updater_seq;
    std::vector<std::string> ups = common::Split(tval, ',');
    for (const std::string& pstr : ups) {
      std::unique_ptr<TreeUpdater> up(TreeUpdater::Create(pstr.c_str()));
      up->Init(this->cfg);
      updaters.push_back(std::move(up));
    }
  }
  // do group specific group
  inline void
  BoostNewTrees(const std::vector<bst_gpair> &gpair,
                DMatrix *p_fmat,
                int bst_group,
                std::vector<std::unique_ptr<RegTree> >* ret) {
    this->InitUpdater();
    std::vector<RegTree*> new_trees;
    ret->clear();
    // create the trees

xgboost/src/gbm/gbtree.cc  view on Meta::CPAN

  // ----training fields----
  // configurations for tree
  std::vector<std::pair<std::string, std::string> > cfg;
  // the updaters that can be applied to each of tree
  std::vector<std::unique_ptr<TreeUpdater>> updaters;
  // Cached matrices
  std::vector<std::shared_ptr<DMatrix>> cache_;
  std::unique_ptr<Predictor> predictor;
};

// dart
class Dart : public GBTree {
 public:
  explicit Dart(bst_float base_margin) : GBTree(base_margin) {}

  void Configure(const std::vector<std::pair<std::string, std::string> >& cfg) override {
    GBTree::Configure(cfg);
    if (model_.trees.size() == 0) {
      dparam.InitAllowUnknown(cfg);
    }
  }

  void Load(dmlc::Stream* fi) override {
    GBTree::Load(fi);
    weight_drop.resize(model_.param.num_trees);
    if (model_.param.num_trees != 0) {
      fi->Read(&weight_drop);
    }
  }

  void Save(dmlc::Stream* fo) const override {
    GBTree::Save(fo);
    if (weight_drop.size() != 0) {
      fo->Write(weight_drop);
    }
  }

  // predict the leaf scores with dropout if ntree_limit = 0
  void PredictBatch(DMatrix* p_fmat,
               std::vector<bst_float>* out_preds,
               unsigned ntree_limit) override {
    DropTrees(ntree_limit);
    PredLoopInternal<Dart>(p_fmat, out_preds, 0, ntree_limit, true);
  }

  void PredictInstance(const SparseBatch::Inst& inst,
               std::vector<bst_float>* out_preds,
               unsigned ntree_limit,
               unsigned root_index) override {
    DropTrees(1);
    if (thread_temp.size() == 0) {
      thread_temp.resize(1, RegTree::FVec());
      thread_temp[0].Init(model_.param.num_feature);
    }
    out_preds->resize(model_.param.num_output_group);
    ntree_limit *= model_.param.num_output_group;
    if (ntree_limit == 0 || ntree_limit > model_.trees.size()) {
      ntree_limit = static_cast<unsigned>(model_.trees.size());
    }
    // loop over output groups
    for (int gid = 0; gid < model_.param.num_output_group; ++gid) {
      (*out_preds)[gid]
          = PredValue(inst, gid, root_index,
                      &thread_temp[0], 0, ntree_limit) + model_.base_margin;
    }
  }

 protected:
  friend class GBTree;
  // internal prediction loop
  // add predictions to out_preds
  template<typename Derived>
  inline void PredLoopInternal(
      DMatrix* p_fmat,
      std::vector<bst_float>* out_preds,
      unsigned tree_begin,
      unsigned ntree_limit,
      bool init_out_preds) {
    int num_group = model_.param.num_output_group;
    ntree_limit *= num_group;
    if (ntree_limit == 0 || ntree_limit > model_.trees.size()) {
      ntree_limit = static_cast<unsigned>(model_.trees.size());
    }

    if (init_out_preds) {
      size_t n = num_group * p_fmat->info().num_row;
      const std::vector<bst_float>& base_margin = p_fmat->info().base_margin;
      out_preds->resize(n);
      if (base_margin.size() != 0) {
        CHECK_EQ(out_preds->size(), n);
        std::copy(base_margin.begin(), base_margin.end(), out_preds->begin());
      } else {
        std::fill(out_preds->begin(), out_preds->end(), model_.base_margin);
      }
    }

    if (num_group == 1) {
      PredLoopSpecalize<Derived>(p_fmat, out_preds, 1,
                                 tree_begin, ntree_limit);
    } else {
      PredLoopSpecalize<Derived>(p_fmat, out_preds, num_group,
                                 tree_begin, ntree_limit);
    }
  }

  template<typename Derived>
  inline void PredLoopSpecalize(
      DMatrix* p_fmat,
      std::vector<bst_float>* out_preds,
      int num_group,
      unsigned tree_begin,
      unsigned tree_end) {
    const MetaInfo& info = p_fmat->info();
    const int nthread = omp_get_max_threads();
    CHECK_EQ(num_group, model_.param.num_output_group);
    InitThreadTemp(nthread);
    std::vector<bst_float>& preds = *out_preds;
    CHECK_EQ(model_.param.size_leaf_vector, 0)
        << "size_leaf_vector is enforced to 0 so far";
    CHECK_EQ(preds.size(), p_fmat->info().num_row * num_group);
    // start collecting the prediction
    dmlc::DataIter<RowBatch>* iter = p_fmat->RowIterator();
    Derived* self = static_cast<Derived*>(this);
    iter->BeforeFirst();
    while (iter->Next()) {
      const RowBatch &batch = iter->Value();
      // parallel over local batch
      const int K = 8;
      const bst_omp_uint nsize = static_cast<bst_omp_uint>(batch.size);
      const bst_omp_uint rest = nsize % K;
      #pragma omp parallel for schedule(static)
      for (bst_omp_uint i = 0; i < nsize - rest; i += K) {
        const int tid = omp_get_thread_num();
        RegTree::FVec& feats = thread_temp[tid];
        int64_t ridx[K];
        RowBatch::Inst inst[K];
        for (int k = 0; k < K; ++k) {
          ridx[k] = static_cast<int64_t>(batch.base_rowid + i + k);
        }
        for (int k = 0; k < K; ++k) {
          inst[k] = batch[i + k];
        }
        for (int k = 0; k < K; ++k) {
          for (int gid = 0; gid < num_group; ++gid) {
            const size_t offset = ridx[k] * num_group + gid;
            preds[offset] +=
                self->PredValue(inst[k], gid, info.GetRoot(ridx[k]),
                                &feats, tree_begin, tree_end);
          }
        }
      }
      for (bst_omp_uint i = nsize - rest; i < nsize; ++i) {
        RegTree::FVec& feats = thread_temp[0];
        const int64_t ridx = static_cast<int64_t>(batch.base_rowid + i);
        const RowBatch::Inst inst = batch[i];
        for (int gid = 0; gid < num_group; ++gid) {
          const size_t offset = ridx * num_group + gid;
          preds[offset] +=
              self->PredValue(inst, gid, info.GetRoot(ridx),
                              &feats, tree_begin, tree_end);
        }
      }
    }
  }
  // commit new trees all at once
  void CommitModel(std::vector<std::unique_ptr<RegTree> >&& new_trees,
                   int bst_group) override {
    for (size_t i = 0; i < new_trees.size(); ++i) {
      model_.trees.push_back(std::move(new_trees[i]));
      model_.tree_info.push_back(bst_group);
    }
    model_.param.num_trees += static_cast<int>(new_trees.size());
    size_t num_drop = NormalizeTrees(new_trees.size());
    if (dparam.silent != 1) {
      LOG(INFO) << "drop " << num_drop << " trees, "
                << "weight = " << weight_drop.back();
    }
  }
  // predict the leaf scores without dropped trees
  inline bst_float PredValue(const RowBatch::Inst &inst,
                             int bst_group,
                             unsigned root_index,
                             RegTree::FVec *p_feats,
                             unsigned tree_begin,
                             unsigned tree_end) {
    bst_float psum = 0.0f;
    p_feats->Fill(inst);
    for (size_t i = tree_begin; i < tree_end; ++i) {
      if (model_.tree_info[i] == bst_group) {
        bool drop = (std::binary_search(idx_drop.begin(), idx_drop.end(), i));
        if (!drop) {
          int tid = model_.trees[i]->GetLeafIndex(*p_feats, root_index);
          psum += weight_drop[i] * (*model_.trees[i])[tid].leaf_value();
        }
      }
    }
    p_feats->Drop(inst);
    return psum;
  }

  // select dropped trees
  inline void DropTrees(unsigned ntree_limit_drop) {
    std::uniform_real_distribution<> runif(0.0, 1.0);
    auto& rnd = common::GlobalRandom();
    // reset
    idx_drop.clear();
    // sample dropped trees
    bool skip = false;
    if (dparam.skip_drop > 0.0) skip = (runif(rnd) < dparam.skip_drop);
    if (ntree_limit_drop == 0 && !skip) {
      if (dparam.sample_type == 1) {
        bst_float sum_weight = 0.0;
        for (size_t i = 0; i < weight_drop.size(); ++i) {
          sum_weight += weight_drop[i];
        }
        for (size_t i = 0; i < weight_drop.size(); ++i) {
          if (runif(rnd) < dparam.rate_drop * weight_drop.size() * weight_drop[i] / sum_weight) {
            idx_drop.push_back(i);
          }