tree_model.h

Go to the documentation of this file.

 
#ifndef XGBOOST_TREE_MODEL_H_
#define XGBOOST_TREE_MODEL_H_
 
#include <xgboost/base.h>
#include <xgboost/data.h>
#include <xgboost/feature_map.h>
#include <xgboost/host_device_vector.h>  // for HostDeviceVector
#include <xgboost/linalg.h>              // for VectorView
#include <xgboost/logging.h>
#include <xgboost/model.h>
#include <xgboost/multi_target_tree_model.h>  // for MultiTargetTree
 
#include <algorithm>
#include <cstring>
#include <limits>  // for numeric_limits
#include <memory>  // for unique_ptr
#include <string>
#include <type_traits>  // for is_signed_v
#include <vector>
 
namespace xgboost {
 
namespace tree {
struct ScalarTreeView;
struct MultiTargetTreeView;
}  // namespace tree
 
class Json;
 
struct TreeParam {
  bst_node_t num_nodes{1};
  bst_node_t num_deleted{0};
  bst_feature_t num_feature{0};
  bst_target_t size_leaf_vector{1};
 
  bool operator==(const TreeParam& b) const {
    return num_nodes == b.num_nodes && num_deleted == b.num_deleted &&
           num_feature == b.num_feature && size_leaf_vector == b.size_leaf_vector;
  }
 
  void FromJson(Json const& in);
  void ToJson(Json* p_out) const;
};
 
struct RTreeNodeStat {
  float loss_chg;
  float sum_hess;
  float base_weight;
  int leaf_child_cnt{0};
 
  RTreeNodeStat() = default;
  RTreeNodeStat(float loss_chg, float sum_hess, float weight)
      : loss_chg{loss_chg}, sum_hess{sum_hess}, base_weight{weight} {}
  bool operator==(const RTreeNodeStat& b) const {
    return loss_chg == b.loss_chg && sum_hess == b.sum_hess && base_weight == b.base_weight &&
           leaf_child_cnt == b.leaf_child_cnt;
  }
};
 
class RegTree : public Model {
 public:
  using SplitCondT = float;
  static constexpr bst_node_t kInvalidNodeId{MultiTargetTree::InvalidNodeId()};
  static constexpr uint32_t kDeletedNodeMarker = std::numeric_limits<uint32_t>::max();
  static constexpr bst_node_t kRoot{0};
 
  class Node {
   public:
    XGBOOST_DEVICE Node() {
      // assert compact alignment
      static_assert(sizeof(Node) == 4 * sizeof(int) + sizeof(Info), "Node: 64 bit align");
    }
    Node(int32_t cleft, int32_t cright, int32_t parent, uint32_t split_ind, float split_cond,
         bool default_left)
        : parent_{parent}, cleft_{cleft}, cright_{cright} {
      this->SetParent(parent_);
      this->SetSplit(split_ind, split_cond, default_left);
    }
 
    [[nodiscard]] XGBOOST_DEVICE int LeftChild() const { return this->cleft_; }
    [[nodiscard]] XGBOOST_DEVICE int RightChild() const { return this->cright_; }
    [[nodiscard]] XGBOOST_DEVICE int DefaultChild() const {
      return this->DefaultLeft() ? this->LeftChild() : this->RightChild();
    }
    [[nodiscard]] XGBOOST_DEVICE bst_feature_t SplitIndex() const {
      static_assert(!std::is_signed_v<bst_feature_t>);
      return sindex_ & ((1U << 31) - 1U);
    }
    [[nodiscard]] XGBOOST_DEVICE bool DefaultLeft() const { return (sindex_ >> 31) != 0; }
    [[nodiscard]] XGBOOST_DEVICE bool IsLeaf() const { return cleft_ == kInvalidNodeId; }
    [[nodiscard]] XGBOOST_DEVICE float LeafValue() const { return (this->info_).leaf_value; }
    [[nodiscard]] XGBOOST_DEVICE SplitCondT SplitCond() const { return (this->info_).split_cond; }
    [[nodiscard]] XGBOOST_DEVICE int Parent() const { return parent_ & ((1U << 31) - 1); }
    [[nodiscard]] XGBOOST_DEVICE bool IsLeftChild() const { return (parent_ & (1U << 31)) != 0; }
    [[nodiscard]] XGBOOST_DEVICE bool IsDeleted() const { return sindex_ == kDeletedNodeMarker; }
    [[nodiscard]] XGBOOST_DEVICE bool IsRoot() const { return parent_ == kInvalidNodeId; }
    XGBOOST_DEVICE void SetLeftChild(int nid) { this->cleft_ = nid; }
    XGBOOST_DEVICE void SetRightChild(int nid) { this->cright_ = nid; }
    XGBOOST_DEVICE void SetSplit(unsigned split_index, SplitCondT split_cond,
                                 bool default_left = false) {
      if (default_left) split_index |= (1U << 31);
      this->sindex_ = split_index;
      (this->info_).split_cond = split_cond;
    }
    XGBOOST_DEVICE void SetLeaf(bst_float value, int right = kInvalidNodeId) {
      (this->info_).leaf_value = value;
      this->cleft_ = kInvalidNodeId;
      this->cright_ = right;
    }
    XGBOOST_DEVICE void MarkDelete() { this->sindex_ = kDeletedNodeMarker; }
    XGBOOST_DEVICE void Reuse() { this->sindex_ = 0; }
    // set parent
    XGBOOST_DEVICE void SetParent(int pidx, bool is_left_child = true) {
      if (is_left_child) pidx |= (1U << 31);
      this->parent_ = pidx;
    }
    bool operator==(const Node& b) const {
      return parent_ == b.parent_ && cleft_ == b.cleft_ && cright_ == b.cright_ &&
             sindex_ == b.sindex_ && info_.leaf_value == b.info_.leaf_value;
    }
 
   private:
    union Info {
      bst_float leaf_value;
      SplitCondT split_cond;
    };
    // pointer to parent, highest bit is used to
    // indicate whether it's a left child or not
    int32_t parent_{kInvalidNodeId};
    // pointer to left, right
    int32_t cleft_{kInvalidNodeId}, cright_{kInvalidNodeId};
    // split feature index, left split or right split depends on the highest bit
    uint32_t sindex_{0};
    // extra info
    Info info_;
  };
 
  void ChangeToLeaf(bst_node_t nidx, float value) {
    auto& h_nodes = nodes_.HostVector();
    CHECK(h_nodes[h_nodes[nidx].LeftChild()].IsLeaf());
    CHECK(h_nodes[h_nodes[nidx].RightChild()].IsLeaf());
    this->DeleteNode(h_nodes[nidx].LeftChild());
    this->DeleteNode(h_nodes[nidx].RightChild());
    h_nodes[nidx].SetLeaf(value);
  }
  void CollapseToLeaf(bst_node_t nidx, float value) {
    auto& h_nodes = nodes_.HostVector();
    if (h_nodes[nidx].IsLeaf()) return;
    if (!h_nodes[h_nodes[nidx].LeftChild()].IsLeaf()) {
      CollapseToLeaf(h_nodes[nidx].LeftChild(), 0.0f);
    }
    if (!h_nodes[h_nodes[nidx].RightChild()].IsLeaf()) {
      CollapseToLeaf(h_nodes[nidx].RightChild(), 0.0f);
    }
    this->ChangeToLeaf(nidx, value);
  }
 
  RegTree() {
    nodes_.HostVector().resize(param_.num_nodes);
    stats_.HostVector().resize(param_.num_nodes);
    split_types_.HostVector().resize(param_.num_nodes, FeatureType::kNumerical);
    split_categories_segments_.HostVector().resize(param_.num_nodes);
    auto& h_nodes = nodes_.HostVector();
    for (int i = 0; i < param_.num_nodes; i++) {
      h_nodes[i].SetLeaf(0.0f);
      h_nodes[i].SetParent(kInvalidNodeId);
    }
  }
  explicit RegTree(bst_target_t n_targets, bst_feature_t n_features) : RegTree{} {
    param_.num_feature = n_features;
    param_.size_leaf_vector = n_targets;
    if (n_targets > 1) {
      this->p_mt_tree_.reset(new MultiTargetTree{&param_});
    }
  }
 
  Node& operator[](bst_node_t nidx) { return nodes_.HostVector()[nidx]; }
 
 public:
  [[nodiscard]] common::Span<Node const> GetNodes(DeviceOrd device) const {
    CHECK(!this->IsMultiTarget());
    return device.IsCPU() ? nodes_.ConstHostSpan()
                          : (nodes_.SetDevice(device), nodes_.ConstDeviceSpan());
  }
 
  [[nodiscard]] common::Span<RTreeNodeStat const> GetStats(DeviceOrd device) const {
    CHECK(!this->IsMultiTarget());
    return device.IsCPU() ? stats_.ConstHostSpan()
                          : (stats_.SetDevice(device), stats_.ConstDeviceSpan());
  }
 
  RTreeNodeStat& Stat(int nid) { return stats_.HostVector()[nid]; }
 
  void LoadModel(Json const& in) override;
  void SaveModel(Json* out) const override;
 
  [[nodiscard]] bool Equal(RegTree const& b) const;
 
  void ExpandNode(bst_node_t nid, unsigned split_index, bst_float split_value, bool default_left,
                  bst_float base_weight, bst_float left_leaf_weight, bst_float right_leaf_weight,
                  bst_float loss_change, float sum_hess, float left_sum, float right_sum,
                  bst_node_t leaf_right_child = kInvalidNodeId);
  void ExpandNode(bst_node_t nidx, bst_feature_t split_index, float split_cond, bool default_left,
                  linalg::VectorView<float const> base_weight,
                  linalg::VectorView<float const> left_weight,
                  linalg::VectorView<float const> right_weight, float loss_chg, float sum_hess,
                  float left_sum, float right_sum);
  void SetLeaves(std::vector<bst_node_t> leaves, common::Span<float const> weights);
 
  void ExpandCategorical(bst_node_t nid, bst_feature_t split_index,
                         common::Span<const uint32_t> split_cat, bool default_left,
                         bst_float base_weight, bst_float left_leaf_weight,
                         bst_float right_leaf_weight, bst_float loss_change, float sum_hess,
                         float left_sum, float right_sum);
  void ExpandCategorical(bst_node_t nidx, bst_feature_t split_index,
                         common::Span<const uint32_t> split_cat, bool default_left,
                         linalg::VectorView<float const> base_weight,
                         linalg::VectorView<float const> left_weight,
                         linalg::VectorView<float const> right_weight, float loss_chg,
                         float sum_hess, float left_sum, float right_sum);
  [[nodiscard]] bool HasCategoricalSplit() const { return !split_categories_.Empty(); }
  [[nodiscard]] bool IsMultiTarget() const { return static_cast<bool>(p_mt_tree_); }
  [[nodiscard]] bst_target_t NumTargets() const { return param_.size_leaf_vector; }
  [[nodiscard]] auto GetMultiTargetTree() const {
    CHECK(IsMultiTarget());
    return p_mt_tree_.get();
  }
  [[nodiscard]] bst_feature_t NumFeatures() const noexcept { return param_.num_feature; }
  [[nodiscard]] bst_node_t NumNodes() const noexcept { return param_.num_nodes; }
  [[nodiscard]] bst_node_t NumValidNodes() const noexcept {
    return param_.num_nodes - param_.num_deleted;
  }
  [[nodiscard]] bst_node_t NumExtraNodes() const noexcept {
    return param_.num_nodes - 1 - param_.num_deleted;
  }
  /* \brief Count number of leaves in tree. */
  [[nodiscard]] bst_node_t GetNumLeaves() const;
  [[nodiscard]] bst_node_t GetNumSplitNodes() const;
 
  [[nodiscard]] bst_node_t GetDepth(bst_node_t nidx) const;
  void SetRoot(linalg::VectorView<float const> weight, float sum_hess) {
    CHECK(IsMultiTarget());
    return this->p_mt_tree_->SetRoot(weight, sum_hess);
  }
  [[nodiscard]] bst_node_t MaxDepth() const;
 
  struct FVec {
    void Init(size_t size);
    void Fill(SparsePage::Inst const& inst);
 
    void Drop();
    [[nodiscard]] size_t Size() const;
    [[nodiscard]] bst_float GetFvalue(size_t i) const;
    [[nodiscard]] bool IsMissing(size_t i) const;
    [[nodiscard]] bool HasMissing() const;
    void HasMissing(bool has_missing) { this->has_missing_ = has_missing; }
 
    [[nodiscard]] common::Span<float> Data() { return data_; }
 
   private:
    std::vector<float> data_;
    bool has_missing_;
  };
 
  [[nodiscard]] std::string DumpModel(const FeatureMap& fmap, bool with_stats,
                                      std::string format) const;
  [[nodiscard]] common::Span<FeatureType const> GetSplitTypes(DeviceOrd device) const {
    return device.IsCPU() ? split_types_.ConstHostSpan()
                          : (split_types_.SetDevice(device), split_types_.ConstDeviceSpan());
  }
  [[nodiscard]] common::Span<uint32_t const> GetSplitCategories(DeviceOrd device) const {
    return device.IsCPU()
               ? split_categories_.ConstHostSpan()
               : (split_categories_.SetDevice(device), split_categories_.ConstDeviceSpan());
  }
  [[nodiscard]] auto const& GetSplitCategoriesPtr() const {
    return split_categories_segments_.ConstHostVector();
  }
 
  struct CategoricalSplitMatrix {
    struct Segment {
      std::size_t beg{0};
      std::size_t size{0};
    };
    common::Span<FeatureType const> split_type;
    common::Span<uint32_t const> categories;
    common::Span<Segment const> node_ptr;
  };
 
  [[nodiscard]] CategoricalSplitMatrix GetCategoriesMatrix(DeviceOrd device) const {
    CategoricalSplitMatrix view;
    view.split_type = this->GetSplitTypes(device);
    view.categories = this->GetSplitCategories(device);
    if (device.IsCPU()) {
      view.node_ptr = split_categories_segments_.ConstHostSpan();
    } else {
      split_categories_segments_.SetDevice(device);
      view.node_ptr = split_categories_segments_.ConstDeviceSpan();
    }
    return view;
  }
 
  [[nodiscard]] bst_node_t LeftChild(bst_node_t nidx) const {
    if (IsMultiTarget()) {
      return this->p_mt_tree_->LeftChild(nidx);
    }
    return nodes_.ConstHostVector()[nidx].LeftChild();
  }
  [[nodiscard]] bst_node_t RightChild(bst_node_t nidx) const {
    if (IsMultiTarget()) {
      return this->p_mt_tree_->RightChild(nidx);
    }
    return nodes_.ConstHostVector()[nidx].RightChild();
  }
  [[nodiscard]] bst_node_t Size() const {
    if (IsMultiTarget()) {
      return this->p_mt_tree_->Size();
    }
    return this->nodes_.Size();
  }
 
  [[nodiscard]] RegTree* Copy() const;
  tree::ScalarTreeView HostScView() const;
  tree::MultiTargetTreeView HostMtView() const;
 
 private:
  template <bool typed>
  void LoadCategoricalSplit(Json const& in);
  void SaveCategoricalSplit(Json* p_out) const;
  TreeParam param_;
  // vector of nodes
  HostDeviceVector<Node> nodes_;
  // free node space, used during training process
  std::vector<int> deleted_nodes_;
  // stats of nodes
  HostDeviceVector<RTreeNodeStat> stats_;
  HostDeviceVector<FeatureType> split_types_;
 
  // Categories for each internal node.
  HostDeviceVector<uint32_t> split_categories_;
  // Ptr to split categories of each node.
  HostDeviceVector<CategoricalSplitMatrix::Segment> split_categories_segments_;
  // ptr to multi-target tree with vector leaf.
  std::unique_ptr<MultiTargetTree> p_mt_tree_;
  // allocate a new node,
  // !!!!!! NOTE: may cause BUG here, nodes.resize
  bst_node_t AllocNode() {
    if (param_.num_deleted != 0) {
      int nid = deleted_nodes_.back();
      deleted_nodes_.pop_back();
      nodes_.HostVector()[nid].Reuse();
      --param_.num_deleted;
      return nid;
    }
    int nd = param_.num_nodes++;
    CHECK_LT(param_.num_nodes, std::numeric_limits<int>::max())
        << "number of nodes in the tree exceed 2^31";
    nodes_.HostVector().resize(param_.num_nodes);
    stats_.HostVector().resize(param_.num_nodes);
    split_types_.HostVector().resize(param_.num_nodes, FeatureType::kNumerical);
    split_categories_segments_.HostVector().resize(param_.num_nodes);
    return nd;
  }
  // delete a tree node, keep the parent field to allow trace back
  void DeleteNode(int nid) {
    CHECK_GE(nid, 1);
    auto pid = (*this)[nid].Parent();
    if (nid == (*this)[pid].LeftChild()) {
      (*this)[pid].SetLeftChild(kInvalidNodeId);
    } else {
      (*this)[pid].SetRightChild(kInvalidNodeId);
    }
 
    deleted_nodes_.push_back(nid);
    nodes_.HostVector()[nid].MarkDelete();
    ++param_.num_deleted;
  }
};
 
inline void RegTree::FVec::Init(size_t size) {
  data_.resize(size);
  std::fill(data_.begin(), data_.end(), std::numeric_limits<float>::quiet_NaN());
  has_missing_ = true;
}
 
inline void RegTree::FVec::Fill(SparsePage::Inst const& inst) {
  auto p_data = inst.data();
  auto p_out = data_.data();
 
  for (std::size_t i = 0, n = inst.size(); i < n; ++i) {
    auto const& entry = p_data[i];
    p_out[entry.index] = entry.fvalue;
  }
  has_missing_ = data_.size() != inst.size();
}
 
inline void RegTree::FVec::Drop() { this->Init(this->Size()); }
 
inline size_t RegTree::FVec::Size() const { return data_.size(); }
 
inline float RegTree::FVec::GetFvalue(size_t i) const { return data_[i]; }
 
inline bool RegTree::FVec::IsMissing(size_t i) const { return std::isnan(data_[i]); }
 
inline bool RegTree::FVec::HasMissing() const { return has_missing_; }
 
// Multi-target tree not yet implemented error
inline StringView MTNotImplemented() {
  return " support for multi-target tree is not yet implemented.";
}
}  // namespace xgboost
#endif  // XGBOOST_TREE_MODEL_H_