tree_model.h

前往此文件文档。

 
#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/linalg.h> // 用于 VectorView
#include <xgboost/logging.h>
#include <xgboost/model.h>
#include <xgboost/multi_target_tree_model.h> // 用于 MultiTargetTree
 
#include <algorithm>
#include <cstring>
#include <limits>
#include <memory> // 用于 make_unique
#include <stack>
#include <string>
#include <vector>
 
namespace xgboost {
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 {
 bst_float loss_chg;
 bst_float sum_hess;
 bst_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;
  }
};
 
template <typename T>
class CopyUniquePtr {
 std::unique_ptr<T> ptr_{nullptr};
 
 public
 CopyUniquePtr() = default;
 CopyUniquePtr(CopyUniquePtr const& that) {
 ptr_.reset(nullptr);
 if (that.ptr_) {
 ptr_ = std::make_unique<T>(*that);
    }
  }
 T* get() const noexcept { return ptr_.get(); } // NOLINT
 
 T& operator*() { return *ptr_; }
 T* operator->() noexcept { return this->get(); }
 
 T const& operator*() const { return *ptr_; }
 T const* operator->() const noexcept { return this->get(); }
 
 explicit operator bool() const { return static_cast<bool>(ptr_); }
 bool operator!() const { return !ptr_; }
 void reset(T* ptr) { ptr_.reset(ptr); } // NOLINT
};
 
class RegTree : public Model {
 public
 using SplitCondT = bst_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() {
 // 断言紧凑对齐
 static_assert(sizeof(Node) == 4 * sizeof(int) + sizeof(Info), "Node: 64 位对齐");
    }
 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;
    }
 // 设置父节点
 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;
    };
 // 指向父节点的指针，最高位用于
 // 指示它是否为左子节点
 int32_t parent_{kInvalidNodeId};
 // 指向左子节点、右子节点的指针
 int32_t cleft_{kInvalidNodeId}, cright_{kInvalidNodeId};
 // 分裂特征索引，左分裂或右分裂取决于最高位
 uint32_t sindex_{0};
 // 额外信息
 Info info_;
  };
 
 void ChangeToLeaf(int rid, bst_float value) {
 CHECK(nodes_[nodes_[rid].LeftChild() ].IsLeaf());
 CHECK(nodes_[nodes_[rid].RightChild()].IsLeaf());
 this->DeleteNode(nodes_[rid].LeftChild());
 this->DeleteNode(nodes_[rid].RightChild());
 nodes_[rid].SetLeaf(value);
  }
 void CollapseToLeaf(int rid, bst_float value) {
 if (nodes_[rid].IsLeaf()) return;
 if (!nodes_[nodes_[rid].LeftChild() ].IsLeaf()) {
 CollapseToLeaf(nodes_[rid].LeftChild(), 0.0f);
    }
 if (!nodes_[nodes_[rid].RightChild() ].IsLeaf()) {
 CollapseToLeaf(nodes_[rid].RightChild(), 0.0f);
    }
 this->ChangeToLeaf(rid, value);
  }
 
 RegTree() {
 nodes_.resize(param_.num_nodes);
 stats_.resize(param_.num_nodes);
 split_types_.resize(param_.num_nodes, FeatureType::kNumerical);
 split_categories_segments_.resize(param_.num_nodes);
 for (int i = 0; i < param_.num_nodes; i++) {
 nodes_[i].SetLeaf(0.0f);
 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[](int nid) {
 return nodes_[nid];
  }
 const Node& operator[](int nid) const {
 return nodes_[nid];
  }
 
 [[nodiscard]] const std::vector<Node>& GetNodes() const { return nodes_; }
 
 [[nodiscard]] const std::vector<RTreeNodeStat>& GetStats() const { return stats_; }
 
 RTreeNodeStat& Stat(int nid) {
 return stats_[nid];
  }
 [[nodiscard]] const RTreeNodeStat& Stat(int nid) const {
 return stats_[nid];
  }
 
 void LoadModel(Json const& in) override;
 void SaveModel(Json* out) const override;
 
 bool operator==(const RegTree& b) const {
 return nodes_ == b.nodes_ && stats_ == b.stats_ &&
 deleted_nodes_ == b.deleted_nodes_ && param_ == b.param_;
  }
 /* \brief 遍历此树中的所有节点。
   *
 * \param Function 接受节点索引，并在迭代应停止时返回 false，否则返回 true。
 * 停止，否则返回 true。
   */
 template <typename Func> void WalkTree(Func func) const {
 std::stack<bst_node_t> nodes;
 nodes.push(kRoot);
 auto &self = *this;
 while (!nodes.empty()) {
 auto nidx = nodes.top();
 nodes.pop();
 if (!func(nidx)) {
 return;
      }
 auto left = self.LeftChild(nidx);
 auto right = self.RightChild(nidx);
 if (left != RegTree::kInvalidNodeId) {
 nodes.push(left);
      }
 if (right != RegTree::kInvalidNodeId) {
 nodes.push(right);
      }
    }
  }
 [[nodiscard]] bool Equal(const RegTree& 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);
 
 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);
 [[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 计算树中的叶子数量。 */
 [[nodiscard]] bst_node_t GetNumLeaves() const;
 [[nodiscard]] bst_node_t GetNumSplitNodes() const;
 
 [[nodiscard]] std::int32_t GetDepth(bst_node_t nid) const {
 if (IsMultiTarget()) {
 return this->p_mt_tree_->Depth(nid);
    }
 int depth = 0;
 while (!nodes_[nid].IsRoot()) {
 ++depth;
 nid = nodes_[nid].Parent();
    }
 return depth;
  }
 void SetLeaf(bst_node_t nidx, linalg::VectorView<float const> weight) {
 CHECK(IsMultiTarget());
 return this->p_mt_tree_->SetLeaf(nidx, weight);
  }
 
 [[nodiscard]] int MaxDepth(int nid) const {
 if (nodes_[nid].IsLeaf()) return 0;
 return std::max(MaxDepth(nodes_[nid].LeftChild()) + 1, MaxDepth(nodes_[nid].RightChild()) + 1);
  }
 
 int MaxDepth() { return MaxDepth(0); }
 
 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]] FeatureType NodeSplitType(bst_node_t nidx) const { return split_types_.at(nidx); }
 [[nodiscard]] std::vector<FeatureType> const& GetSplitTypes() const {
 return split_types_;
  }
 [[nodiscard]] common::Span<uint32_t const> GetSplitCategories() const {
 return split_categories_;
  }
 [[nodiscard]] common::Span<uint32_t const> NodeCats(bst_node_t nidx) const {
 auto node_ptr = GetCategoriesMatrix().node_ptr;
 auto categories = GetCategoriesMatrix().categories;
 auto segment = node_ptr[nidx];
 auto node_cats = categories.subspan(segment.beg, segment.size);
 return node_cats;
  }
 [[nodiscard]] auto const& GetSplitCategoriesPtr() const { return split_categories_segments_; }
 
 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() const {
 CategoricalSplitMatrix view;
 view.split_type = common::Span<FeatureType const>(this->GetSplitTypes());
 view.categories = this->GetSplitCategories();
 view.node_ptr = common::Span<CategoricalSplitMatrix::Segment const>(split_categories_segments_);
 return view;
  }
 
 [[nodiscard]] bst_feature_t SplitIndex(bst_node_t nidx) const {
 if (IsMultiTarget()) {
 return this->p_mt_tree_->SplitIndex(nidx);
    }
 return (*this)[nidx].SplitIndex();
  }
 [[nodiscard]] float SplitCond(bst_node_t nidx) const {
 if (IsMultiTarget()) {
 return this->p_mt_tree_->SplitCond(nidx);
    }
 return (*this)[nidx].SplitCond();
  }
 [[nodiscard]] bool DefaultLeft(bst_node_t nidx) const {
 if (IsMultiTarget()) {
 return this->p_mt_tree_->DefaultLeft(nidx);
    }
 return (*this)[nidx].DefaultLeft();
  }
 [[nodiscard]] bst_node_t DefaultChild(bst_node_t nidx) const {
 return this->DefaultLeft(nidx) ? this->LeftChild(nidx) : this->RightChild(nidx);
  }
 [[nodiscard]] bool IsRoot(bst_node_t nidx) const {
 if (IsMultiTarget()) {
 return nidx == kRoot;
    }
 return (*this)[nidx].IsRoot();
  }
 [[nodiscard]] bool IsLeaf(bst_node_t nidx) const {
 if (IsMultiTarget()) {
 return this->p_mt_tree_->IsLeaf(nidx);
    }
 return (*this)[nidx].IsLeaf();
  }
 [[nodiscard]] bst_node_t Parent(bst_node_t nidx) const {
 if (IsMultiTarget()) {
 return this->p_mt_tree_->Parent(nidx);
    }
 return (*this)[nidx].Parent();
  }
 [[nodiscard]] bst_node_t LeftChild(bst_node_t nidx) const {
 if (IsMultiTarget()) {
 return this->p_mt_tree_->LeftChild(nidx);
    }
 return (*this)[nidx].LeftChild();
  }
 [[nodiscard]] bst_node_t RightChild(bst_node_t nidx) const {
 if (IsMultiTarget()) {
 return this->p_mt_tree_->RightChild(nidx);
    }
 return (*this)[nidx].RightChild();
  }
 [[nodiscard]] bool IsLeftChild(bst_node_t nidx) const {
 if (IsMultiTarget()) {
 CHECK_NE(nidx, kRoot);
 auto p = this->p_mt_tree_->Parent(nidx);
 return nidx == this->p_mt_tree_->LeftChild(p);
    }
 return (*this)[nidx].IsLeftChild();
  }
 [[nodiscard]] bst_node_t Size() const {
 if (IsMultiTarget()) {
 return this->p_mt_tree_->Size();
    }
 return this->nodes_.size();
  }
 
 private
 template <bool typed>
 void LoadCategoricalSplit(Json const& in);
 void SaveCategoricalSplit(Json* p_out) const;
 TreeParam param_;
 // vector of nodes
 std::vector<Node> nodes_;
 // free node space, used during training process
 std::vector<int> deleted_nodes_;
 // stats of nodes
 std::vector<RTreeNodeStat> stats_;
 std::vector<FeatureType> split_types_;
 
 // Categories for each internal node.
 std::vector<uint32_t> split_categories_;
 // Ptr to split categories of each node.
 std::vector<CategoricalSplitMatrix::Segment> split_categories_segments_;
 // ptr to multi-target tree with vector leaf.
 CopyUniquePtr<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_[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_.resize(param_.num_nodes);
 stats_.resize(param_.num_nodes);
 split_types_.resize(param_.num_nodes, FeatureType::kNumerical);
 split_categories_segments_.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_[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_