tensorflow/core/kernels/boosted_trees/stats_ops.cc

/* Copyright 2018 The TensorFlow Authors. All Rights Reserved.

Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at

    http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
==============================================================================*/

#include <limits>
#include <vector>

#include "third_party/eigen3/Eigen/Core"
#include "tensorflow/core/framework/op_kernel.h"
#include "tensorflow/core/framework/tensor.h"
#include "tensorflow/core/framework/tensor_shape.h"
#include "tensorflow/core/kernels/boosted_trees/tree_helper.h"
#include "tensorflow/core/platform/logging.h"

namespace tensorflow {

// TODO(tanzheny): Make these const as proto enum.
const char kInequalityDefaultLeft[] = "inequality_default_left";
const char kInequalityDefaultRight[] = "inequality_default_right";
const char kEqualityDefaultLeft[] = "equality_default_left";

using Matrix =
    Eigen::Matrix<float, Eigen::Dynamic, Eigen::Dynamic, Eigen::RowMajor>;
using ConstMatrixMap = Eigen::Map<const Matrix>;
using MatrixMap = Eigen::Map<Matrix>;

using ConstVectorMap = Eigen::Map<const Eigen::VectorXf>;
using VectorMap = Eigen::Map<Eigen::VectorXf>;

// V1 Op. Deprecated. BoostedTreesCalculateBestFeatureSplitOp is V2.
class BoostedTreesCalculateBestGainsPerFeatureOp : public OpKernel {
 public:
  explicit BoostedTreesCalculateBestGainsPerFeatureOp(
      OpKernelConstruction* const context)
      : OpKernel(context) {
    OP_REQUIRES_OK(context, context->GetAttr("max_splits", &max_splits_));
    OP_REQUIRES_OK(context, context->GetAttr("num_features", &num_features_));
  }

  void Compute(OpKernelContext* const context) override {
    // node_id_range
    const Tensor* node_id_range_t;
    OP_REQUIRES_OK(context, context->input("node_id_range", &node_id_range_t));
    const auto node_id_range = node_id_range_t->vec<int32>();
    const int32 node_id_first = node_id_range(0);  // inclusive
    const int32 node_id_last = node_id_range(1);   // exclusive
    // stats_summary_list
    OpInputList stats_summary_list;
    OP_REQUIRES_OK(context, context->input_list("stats_summary_list",
                                                &stats_summary_list));
    const int64 num_buckets = stats_summary_list[0].dim_size(1);
    // Check for single logit: 1 gradient + 1 hessian value.
    DCHECK_EQ(stats_summary_list[0].dim_size(2), 2);
    std::vector<TTypes<float, 3>::ConstTensor> stats_summary;
    stats_summary.reserve(stats_summary_list.size());
    for (const auto& tensor : stats_summary_list) {
      stats_summary.emplace_back(tensor.tensor<float, 3>());
    }
    const Tensor* l1_t;
    OP_REQUIRES_OK(context, context->input("l1", &l1_t));
    const auto l1 = l1_t->scalar<float>()();
    const Tensor* l2_t;
    OP_REQUIRES_OK(context, context->input("l2", &l2_t));
    const auto l2 = l2_t->scalar<float>()();
    const Tensor* tree_complexity_t;
    OP_REQUIRES_OK(context,
                   context->input("tree_complexity", &tree_complexity_t));
    const auto tree_complexity = tree_complexity_t->scalar<float>()();
    const Tensor* min_node_weight_t;
    OP_REQUIRES_OK(context,
                   context->input("min_node_weight", &min_node_weight_t));
    const auto min_node_weight = min_node_weight_t->scalar<float>()();

    // Allocate output lists of tensors:
    OpOutputList output_node_ids_list;
    OP_REQUIRES_OK(
        context, context->output_list("node_ids_list", &output_node_ids_list));
    OpOutputList output_gains_list;
    OP_REQUIRES_OK(context,
                   context->output_list("gains_list", &output_gains_list));
    OpOutputList output_thresholds_list;
    OP_REQUIRES_OK(context, context->output_list("thresholds_list",
                                                 &output_thresholds_list));
    OpOutputList output_left_node_contribs_list;
    OP_REQUIRES_OK(context,
                   context->output_list("left_node_contribs_list",
                                        &output_left_node_contribs_list));
    OpOutputList output_right_node_contribs_list;
    OP_REQUIRES_OK(context,
                   context->output_list("right_node_contribs_list",
                                        &output_right_node_contribs_list));

    // Use identity later to convert float to Eigen::Matrix type for input to
    // CalculateWeightsAndGains. This op only supports single dimension logits.
    Eigen::MatrixXf identity;
    identity.setIdentity(1, 1);
    // Get the best split info per node for each feature.
    for (int feature_idx = 0; feature_idx < num_features_; ++feature_idx) {
      std::vector<float> cum_grad;
      std::vector<float> cum_hess;
      cum_grad.reserve(num_buckets);
      cum_hess.reserve(num_buckets);

      std::vector<int32> output_node_ids;
      std::vector<float> output_gains;
      std::vector<int32> output_thresholds;
      std::vector<float> output_left_node_contribs;
      std::vector<float> output_right_node_contribs;
      for (int node_id = node_id_first; node_id < node_id_last; ++node_id) {
        // Calculate gains.
        cum_grad.clear();
        cum_hess.clear();
        float total_grad = 0.0;
        float total_hess = 0.0;
        for (int bucket = 0; bucket < num_buckets; ++bucket) {
          // TODO(nponomareva): Consider multi-dimensional gradients/hessians.
          total_grad += stats_summary[feature_idx](node_id, bucket, 0);
          total_hess += stats_summary[feature_idx](node_id, bucket, 1);
          cum_grad.push_back(total_grad);
          cum_hess.push_back(total_hess);
        }
        // Check if node has enough of average hessian.
        if (total_hess < min_node_weight) {
          // Do not split the node because not enough avg hessian.
          continue;
        }
        float best_gain = std::numeric_limits<float>::lowest();
        float best_bucket = 0;
        float best_contrib_for_left = 0.0;
        float best_contrib_for_right = 0.0;
        // Parent gain.
        float parent_gain;
        Eigen::VectorXf unused(1);
        CalculateWeightsAndGains(total_grad * identity, total_hess * identity,
                                 l1, l2, &unused, &parent_gain);

        for (int bucket = 0; bucket < num_buckets; ++bucket) {
          const float cum_grad_bucket = cum_grad[bucket];
          const float cum_hess_bucket = cum_hess[bucket];
          // Left child.
          Eigen::VectorXf contrib_for_left(1);
          float gain_for_left;
          CalculateWeightsAndGains(cum_grad_bucket * identity,
                                   cum_hess_bucket * identity, l1, l2,
                                   &contrib_for_left, &gain_for_left);
          // Right child.
          // use contrib_for_right.
          Eigen::VectorXf contrib_for_right(1);
          float gain_for_right;
          CalculateWeightsAndGains((total_grad - cum_grad_bucket) * identity,
                                   (total_hess - cum_hess_bucket) * identity,
                                   l1, l2, &contrib_for_right, &gain_for_right);

          if (GainIsLarger(gain_for_left + gain_for_right, best_gain)) {
            best_gain = gain_for_left + gain_for_right;
            best_bucket = bucket;
            best_contrib_for_left = contrib_for_left[0];
            best_contrib_for_right = contrib_for_right[0];
          }
        }  // for bucket
        output_node_ids.push_back(node_id);
        // Remove the parent gain for the parent node.
        output_gains.push_back(best_gain - parent_gain);
        output_thresholds.push_back(best_bucket);
        output_left_node_contribs.push_back(best_contrib_for_left);
        output_right_node_contribs.push_back(best_contrib_for_right);
      }  // for node_id
      const int num_nodes = output_node_ids.size();
      // output_node_ids
      Tensor* output_node_ids_t;
      OP_REQUIRES_OK(context,
                     output_node_ids_list.allocate(feature_idx, {num_nodes},
                                                   &output_node_ids_t));
      auto output_node_ids_vec = output_node_ids_t->vec<int32>();
      // output_gains
      Tensor* output_gains_t;
      OP_REQUIRES_OK(context, output_gains_list.allocate(
                                  feature_idx, {num_nodes}, &output_gains_t));
      auto output_gains_vec = output_gains_t->vec<float>();
      // output_thresholds
      Tensor* output_thresholds_t;
      OP_REQUIRES_OK(context,
                     output_thresholds_list.allocate(feature_idx, {num_nodes},
                                                     &output_thresholds_t));
      auto output_thresholds_vec = output_thresholds_t->vec<int32>();
      // output_left_node_contribs
      Tensor* output_left_node_contribs_t;
      OP_REQUIRES_OK(context, output_left_node_contribs_list.allocate(
                                  feature_idx, {num_nodes, 1},
                                  &output_left_node_contribs_t));
      auto output_left_node_contribs_matrix =
          output_left_node_contribs_t->matrix<float>();
      // output_right_node_contribs
      Tensor* output_right_node_contribs_t;
      OP_REQUIRES_OK(context, output_right_node_contribs_list.allocate(
                                  feature_idx, {num_nodes, 1},
                                  &output_right_node_contribs_t));
      auto output_right_node_contribs_matrix =
          output_right_node_contribs_t->matrix<float>();
      // Sets output tensors from vectors.
      for (int i = 0; i < num_nodes; ++i) {
        output_node_ids_vec(i) = output_node_ids[i];
        // Adjust the gains to penalize by tree complexity.
        output_gains_vec(i) = output_gains[i] - tree_complexity;
        output_thresholds_vec(i) = output_thresholds[i];
        output_left_node_contribs_matrix(i, 0) = output_left_node_contribs[i];
        // This op only supports 1-dimensional logits.
        output_right_node_contribs_matrix(i, 0) = output_right_node_contribs[i];
      }
    }  // for f
  }

 private:
  int max_splits_;
  int num_features_;
};

// V1 op that only supports single dimensional logit.
REGISTER_KERNEL_BUILDER(
    Name("BoostedTreesCalculateBestGainsPerFeature").Device(DEVICE_CPU),
    BoostedTreesCalculateBestGainsPerFeatureOp);

// V2 Op.
class BoostedTreesCalculateBestFeatureSplitOp : public OpKernel {
 public:
  explicit BoostedTreesCalculateBestFeatureSplitOp(
      OpKernelConstruction* const context)
      : OpKernel(context) {
    OP_REQUIRES_OK(context, context->GetAttr("logits_dimension", &logits_dim_));
    OP_REQUIRES_OK(context, context->GetAttr("split_type", &split_type_));
  }

  void Compute(OpKernelContext* const context) override {
    // node_id_range
    const Tensor* node_id_range_t;
    OP_REQUIRES_OK(context, context->input("node_id_range", &node_id_range_t));
    const auto node_id_range = node_id_range_t->vec<int32>();
    const int32 node_id_first = node_id_range(0);  // inclusive
    const int32 node_id_last = node_id_range(1);   // exclusive

    const Tensor* stats_summary_t;
    OP_REQUIRES_OK(context, context->input("stats_summary", &stats_summary_t));
    TTypes<float, 4>::ConstTensor stats_summary =
        stats_summary_t->tensor<float, 4>();
    const int64 feature_dims = stats_summary_t->dim_size(1);
    const int64 num_buckets = stats_summary_t->dim_size(2);
    const int64 logits_dim = logits_dim_;
    const int64 hessian_dim = stats_summary_t->dim_size(3) - logits_dim;
    DCHECK_GT(hessian_dim, 0);
    DCHECK_LE(hessian_dim, logits_dim * logits_dim);

    const Tensor* l1_t;
    OP_REQUIRES_OK(context, context->input("l1", &l1_t));
    const auto l1 = l1_t->scalar<float>()();
    DCHECK_GE(l1, 0);
    if (logits_dim_ > 1) {
      // Multi-class L1 regularization not supported yet.
      DCHECK_EQ(l1, 0);
    }

    const Tensor* l2_t;
    OP_REQUIRES_OK(context, context->input("l2", &l2_t));
    const auto l2 = l2_t->scalar<float>()();
    DCHECK_GE(l2, 0);

    const Tensor* tree_complexity_t;
    OP_REQUIRES_OK(context,
                   context->input("tree_complexity", &tree_complexity_t));
    const auto tree_complexity = tree_complexity_t->scalar<float>()();

    const Tensor* min_node_weight_t;
    OP_REQUIRES_OK(context,
                   context->input("min_node_weight", &min_node_weight_t));
    const auto min_node_weight = min_node_weight_t->scalar<float>()();

    std::vector<int32> output_node_ids;
    std::vector<float> output_gains;
    std::vector<int32> output_feature_dimensions;
    std::vector<int32> output_thresholds;
    std::vector<Eigen::VectorXf> output_left_node_contribs;
    std::vector<Eigen::VectorXf> output_right_node_contribs;
    std::vector<string> output_split_types;

    // Iterate each node and find the best gain per node.
    for (int node_id = node_id_first; node_id < node_id_last; ++node_id) {
      float best_gain = std::numeric_limits<float>::lowest();
      int32 best_bucket = 0;
      int32 best_f_dim = 0;
      string best_split_type;
      Eigen::VectorXf best_contrib_for_left(logits_dim);
      Eigen::VectorXf best_contrib_for_right(logits_dim);
      float parent_gain;

      ConstMatrixMap stats_mat(&stats_summary(node_id, 0, 0, 0), num_buckets,
                               logits_dim + hessian_dim);
      const Eigen::VectorXf total_grad =
          stats_mat.leftCols(logits_dim).colwise().sum();
      const Eigen::VectorXf total_hess =
          stats_mat.rightCols(hessian_dim).colwise().sum();
      if (total_hess.norm() < min_node_weight) {
        continue;
      }
      Eigen::VectorXf parent_weight(logits_dim);
      CalculateWeightsAndGains(total_grad, total_hess, l1, l2, &parent_weight,
                               &parent_gain);

      if (split_type_ == "inequality") {
        best_split_type = kInequalityDefaultLeft;
        CalculateBestInequalitySplit(
            stats_summary, node_id, feature_dims, logits_dim, hessian_dim,
            num_buckets, min_node_weight, l1, l2, &best_gain, &best_bucket,
            &best_f_dim, &best_contrib_for_left, &best_contrib_for_right);
      } else {
        best_split_type = kEqualityDefaultLeft;
        CalculateBestEqualitySplit(
            stats_summary, total_grad, total_hess, node_id, feature_dims,
            logits_dim, hessian_dim, num_buckets, l1, l2, &best_gain,
            &best_bucket, &best_f_dim, &best_contrib_for_left,
            &best_contrib_for_right);
      }

      if (best_gain == std::numeric_limits<float>::lowest()) {
        // Do not add the node if not split if found.
        continue;
      }
      output_node_ids.push_back(node_id);
      // Remove the parent gain for the parent node.
      output_gains.push_back(best_gain - parent_gain);
      output_feature_dimensions.push_back(best_f_dim);
      // default direction is fixed for dense splits.
      // TODO(tanzheny) account for default values.
      output_split_types.push_back(best_split_type);
      output_thresholds.push_back(best_bucket);
      output_left_node_contribs.push_back(best_contrib_for_left);
      output_right_node_contribs.push_back(best_contrib_for_right);
    }  // for node id
    const int num_nodes = output_node_ids.size();
    // output_node_ids
    Tensor* output_node_ids_t = nullptr;
    OP_REQUIRES_OK(context, context->allocate_output("node_ids", {num_nodes},
                                                     &output_node_ids_t));
    auto output_node_ids_vec = output_node_ids_t->vec<int32>();

    // output_gains
    Tensor* output_gains_t;
    OP_REQUIRES_OK(context, context->allocate_output("gains", {num_nodes},
                                                     &output_gains_t));
    auto output_gains_vec = output_gains_t->vec<float>();

    // output_feature_dimensions
    Tensor* output_feature_dimension_t;
    OP_REQUIRES_OK(context,
                   context->allocate_output("feature_dimensions", {num_nodes},
                                            &output_feature_dimension_t));
    auto output_feature_dimensions_vec =
        output_feature_dimension_t->vec<int32>();

    // output_thresholds
    Tensor* output_thresholds_t;
    OP_REQUIRES_OK(context, context->allocate_output("thresholds", {num_nodes},
                                                     &output_thresholds_t));
    auto output_thresholds_vec = output_thresholds_t->vec<int32>();

    // output_left_node_contribs
    Tensor* output_left_node_contribs_t;
    OP_REQUIRES_OK(context, context->allocate_output(
                                "left_node_contribs", {num_nodes, logits_dim},
                                &output_left_node_contribs_t));
    auto output_left_node_contribs_matrix =
        output_left_node_contribs_t->matrix<float>();

    // output_right_node_contribs
    Tensor* output_right_node_contribs_t;
    OP_REQUIRES_OK(context, context->allocate_output(
                                "right_node_contribs", {num_nodes, logits_dim},
                                &output_right_node_contribs_t));
    auto output_right_node_contribs_matrix =
        output_right_node_contribs_t->matrix<float>();

    // split type
    Tensor* output_split_types_t;
    OP_REQUIRES_OK(
        context, context->allocate_output("split_with_default_directions",
                                          {num_nodes}, &output_split_types_t));
    auto output_split_types_vec = output_split_types_t->vec<tstring>();

    // Sets output tensors from vectors.
    for (int i = 0; i < num_nodes; ++i) {
      output_node_ids_vec(i) = output_node_ids[i];
      // Adjust the gains to penalize by tree complexity.
      output_gains_vec(i) = output_gains[i] - tree_complexity;
      output_feature_dimensions_vec(i) = output_feature_dimensions[i];
      output_thresholds_vec(i) = output_thresholds[i];
      for (int j = 0; j < logits_dim; ++j) {
        output_left_node_contribs_matrix(i, j) =
            output_left_node_contribs[i][j];
        output_right_node_contribs_matrix(i, j) =
            output_right_node_contribs[i][j];
      }
      output_split_types_vec(i) = output_split_types[i];
    }
  }

 private:
  // TODO(crawles): Simplify inequality path just like equality b/138329196
  // Currently this is not simplify-able due to numerical instability in math
  // i.e. gain = -g.transpose() * hessian_and_reg.colPivHouseholderQr().solve(g)
  // It caused gain to be Inf when g is approaching 0 but not exactly 0 while
  // there is no regularization.
  // Calculate the best inequality split per node.
  void CalculateBestInequalitySplit(TTypes<float, 4>::ConstTensor stats_summary,
                                    const int node_id, const int feature_dims,
                                    const int logits_dim, const int hessian_dim,
                                    const int num_buckets,
                                    const float min_node_weight, const float l1,
                                    const float l2, float* best_gain,
                                    int* best_bucket, int* best_f_dim,
                                    Eigen::VectorXf* best_contrib_for_left,
                                    Eigen::VectorXf* best_contrib_for_right) {
    std::vector<Eigen::VectorXf> cum_grad;
    std::vector<Eigen::VectorXf> cum_hess;
    cum_grad.reserve(num_buckets);
    cum_hess.reserve(num_buckets);

    for (int f_dim = 0; f_dim < feature_dims; ++f_dim) {
      cum_grad.clear();
      cum_hess.clear();
      Eigen::VectorXf total_grad = Eigen::VectorXf::Zero(logits_dim);
      Eigen::VectorXf total_hess = Eigen::VectorXf::Zero(hessian_dim);
      for (int bucket = 0; bucket < num_buckets; ++bucket) {
        for (int i = 0; i < logits_dim; ++i) {
          total_grad[i] += stats_summary(node_id, f_dim, bucket, i);
        }
        for (int i = 0; i < hessian_dim; ++i) {
          // Full hessian.
          total_hess[i] +=
              stats_summary(node_id, f_dim, bucket, logits_dim + i);
        }
        cum_grad.push_back(total_grad);
        cum_hess.push_back(total_hess);
      }

      for (int bucket = 0; bucket < num_buckets; ++bucket) {
        MaybeUpdateBestSplit(cum_grad[bucket], total_grad, cum_hess[bucket],
                             total_hess, logits_dim, bucket, f_dim, l1, l2,
                             best_gain, best_bucket, best_f_dim,
                             best_contrib_for_left, best_contrib_for_right);
      }  // for bucket
    }
  }

  // Calculate the best equality split per node.
  void CalculateBestEqualitySplit(TTypes<float, 4>::ConstTensor stats_summary,
                                  const Eigen::VectorXf& total_grad,
                                  const Eigen::VectorXf& total_hess,
                                  const int node_id, const int feature_dims,
                                  const int logits_dim, const int hessian_dim,
                                  const int num_buckets, const float l1,
                                  const float l2, float* best_gain,
                                  int* best_bucket, int* best_f_dim,
                                  Eigen::VectorXf* best_contrib_for_left,
                                  Eigen::VectorXf* best_contrib_for_right) {
    for (int f_dim = 0; f_dim < feature_dims; ++f_dim) {
      for (int bucket = 0; bucket < num_buckets; ++bucket) {
        ConstVectorMap stats_vec(&stats_summary(node_id, f_dim, bucket, 0),
                                 logits_dim + hessian_dim);
        Eigen::VectorXf curr_grad = stats_vec.head(logits_dim);
        Eigen::VectorXf curr_hess = stats_vec.tail(hessian_dim);
        MaybeUpdateBestSplit(curr_grad, total_grad, curr_hess, total_hess,
                             logits_dim, bucket, f_dim, l1, l2, best_gain,
                             best_bucket, best_f_dim, best_contrib_for_left,
                             best_contrib_for_right);
      }
    }
  }

  void MaybeUpdateBestSplit(const Eigen::VectorXf& grad_for_left,
                            const Eigen::VectorXf& total_grad,
                            const Eigen::VectorXf& hess_for_left,
                            const Eigen::VectorXf& total_hess,
                            const int logits_dim, const int bucket,
                            const int f_dim, const float l1, const float l2,
                            float* best_gain, int* best_bucket, int* best_f_dim,
                            Eigen::VectorXf* best_contrib_for_left,
                            Eigen::VectorXf* best_contrib_for_right) {
    // Left child.
    Eigen::VectorXf contrib_for_left(logits_dim);
    float gain_for_left;
    CalculateWeightsAndGains(grad_for_left, hess_for_left, l1, l2,
                             &contrib_for_left, &gain_for_left);
    // Right child.
    const auto grad_for_right = total_grad - grad_for_left;
    const auto hess_for_right = total_hess - hess_for_left;
    Eigen::VectorXf contrib_for_right(logits_dim);
    float gain_for_right;
    CalculateWeightsAndGains(grad_for_right, hess_for_right, l1, l2,
                             &contrib_for_right, &gain_for_right);
    if (GainIsLarger(gain_for_left + gain_for_right, *best_gain)) {
      *best_gain = gain_for_left + gain_for_right;
      *best_bucket = bucket;
      *best_f_dim = f_dim;
      *best_contrib_for_left = contrib_for_left;
      *best_contrib_for_right = contrib_for_right;
    }
  }

  int logits_dim_;
  string split_type_;
};

// v2 op that supports multi-class.
REGISTER_KERNEL_BUILDER(
    Name("BoostedTreesCalculateBestFeatureSplit").Device(DEVICE_CPU),
    BoostedTreesCalculateBestFeatureSplitOp);

// Map from bucket id to vector of statistics.
typedef std::map<int32, std::vector<float>> BucketMap;
typedef BucketMap::iterator BucketMapIterator;
// Map from feature dimension to BucketMap.
typedef std::map<int32, BucketMap> FeatureMap;
typedef FeatureMap::iterator FeatureMapIterator;

class BoostedTreesSparseCalculateBestFeatureSplitOp : public OpKernel {
 public:
  explicit BoostedTreesSparseCalculateBestFeatureSplitOp(
      OpKernelConstruction* const context)
      : OpKernel(context) {
    // TODO(crawles): Using logits_dim_ for multi-class split.
    OP_REQUIRES_OK(context, context->GetAttr("logits_dimension", &logits_dim_));
    // TODO(tanzheny): Using this for equality split.
    OP_REQUIRES_OK(context, context->GetAttr("split_type", &split_type_));
  }

  void Compute(OpKernelContext* const context) override {
    // node_id_range
    const Tensor* node_id_range_t;
    OP_REQUIRES_OK(context, context->input("node_id_range", &node_id_range_t));
    const auto node_id_range = node_id_range_t->vec<int32>();
    const int32 node_id_first = node_id_range(0);  // inclusive
    const int32 node_id_last = node_id_range(1);   // exclusive

    const Tensor* stats_summary_indices_t;
    OP_REQUIRES_OK(context, context->input("stats_summary_indices",
                                           &stats_summary_indices_t));
    const auto stats_summary_indices = stats_summary_indices_t->matrix<int32>();
    const int32 num_sparse_entries = stats_summary_indices_t->dim_size(0);

    const Tensor* stats_summary_values_t;
    OP_REQUIRES_OK(context, context->input("stats_summary_values",
                                           &stats_summary_values_t));
    const auto stats_summary_values = stats_summary_values_t->vec<float>();

    const Tensor* stats_summary_shape_t;
    OP_REQUIRES_OK(
        context, context->input("stats_summary_shape", &stats_summary_shape_t));
    const auto stats_summary_shape = stats_summary_shape_t->vec<int32>();
    const int32 num_buckets = stats_summary_shape(2) - 1;
    const int32 stats_dims = stats_summary_shape(3);

    const Tensor* l1_t;
    OP_REQUIRES_OK(context, context->input("l1", &l1_t));
    const auto l1 = l1_t->scalar<float>()();

    const Tensor* l2_t;
    OP_REQUIRES_OK(context, context->input("l2", &l2_t));
    const auto l2 = l2_t->scalar<float>()();

    const Tensor* tree_complexity_t;
    OP_REQUIRES_OK(context,
                   context->input("tree_complexity", &tree_complexity_t));
    const auto tree_complexity = tree_complexity_t->scalar<float>()();

    const Tensor* min_node_weight_t;
    OP_REQUIRES_OK(context,
                   context->input("min_node_weight", &min_node_weight_t));
    const auto min_node_weight = min_node_weight_t->scalar<float>()();

    std::vector<int32> output_node_ids;
    std::vector<float> output_gains;
    std::vector<int32> output_feature_dimensions;
    std::vector<int32> output_thresholds;
    std::vector<float> output_left_node_contribs;
    std::vector<float> output_right_node_contribs;
    std::vector<string> output_split_types;

    FeatureMap f_map;

    int32 previous_node_id = -1;
    for (int idx = 0; idx < num_sparse_entries; ++idx) {
      int32 node_id = stats_summary_indices(idx, 0);
      if (node_id != previous_node_id) {
        process_node(f_map, &output_node_ids, &output_gains,
                     &output_feature_dimensions, &output_thresholds,
                     &output_left_node_contribs, &output_right_node_contribs,
                     &output_split_types, previous_node_id, min_node_weight, l1,
                     l2, num_buckets);
        f_map.clear();
      }
      previous_node_id = node_id;
      DCHECK_LE(node_id_first, node_id);
      DCHECK_LT(node_id, node_id_last);
      const int32 feature_dim = stats_summary_indices(idx, 1);
      const int32 bucket_id = stats_summary_indices(idx, 2);
      const int32 stat_dim = stats_summary_indices(idx, 3);
      std::pair<FeatureMapIterator, bool> const& f_insert_result = f_map.insert(
          FeatureMapIterator::value_type(feature_dim, BucketMap()));
      auto& b_map = f_insert_result.first->second;
      std::pair<BucketMapIterator, bool> const& b_insert_result =
          b_map.insert(BucketMapIterator::value_type(
              bucket_id, std::vector<float>(stats_dims)));
      auto& stats = b_insert_result.first->second;
      stats[stat_dim] = stats_summary_values(idx);
    }  // for node_id
    // process the last node id
    process_node(f_map, &output_node_ids, &output_gains,
                 &output_feature_dimensions, &output_thresholds,
                 &output_left_node_contribs, &output_right_node_contribs,
                 &output_split_types, previous_node_id, min_node_weight, l1, l2,
                 num_buckets);

    const int num_nodes = output_node_ids.size();
    // output_node_ids
    Tensor* output_node_ids_t = nullptr;
    OP_REQUIRES_OK(context, context->allocate_output("node_ids", {num_nodes},
                                                     &output_node_ids_t));
    auto output_node_ids_vec = output_node_ids_t->vec<int32>();

    // output_gains
    Tensor* output_gains_t;
    OP_REQUIRES_OK(context, context->allocate_output("gains", {num_nodes},
                                                     &output_gains_t));
    auto output_gains_vec = output_gains_t->vec<float>();

    // output_feature_dimensions
    Tensor* output_feature_dimension_t;
    OP_REQUIRES_OK(context,
                   context->allocate_output("feature_dimensions", {num_nodes},
                                            &output_feature_dimension_t));
    auto output_feature_dimensions_vec =
        output_feature_dimension_t->vec<int32>();

    // output_thresholds
    Tensor* output_thresholds_t;
    OP_REQUIRES_OK(context, context->allocate_output("thresholds", {num_nodes},
                                                     &output_thresholds_t));
    auto output_thresholds_vec = output_thresholds_t->vec<int32>();

    // output_left_node_contribs
    Tensor* output_left_node_contribs_t;
    OP_REQUIRES_OK(
        context, context->allocate_output("left_node_contribs", {num_nodes, 1},
                                          &output_left_node_contribs_t));
    auto output_left_node_contribs_matrix =
        output_left_node_contribs_t->matrix<float>();

    // output_right_node_contribs
    Tensor* output_right_node_contribs_t;
    OP_REQUIRES_OK(
        context, context->allocate_output("right_node_contribs", {num_nodes, 1},
                                          &output_right_node_contribs_t));
    auto output_right_node_contribs_matrix =
        output_right_node_contribs_t->matrix<float>();

    // split type
    Tensor* output_split_types_t;
    OP_REQUIRES_OK(
        context, context->allocate_output("split_with_default_directions",
                                          {num_nodes}, &output_split_types_t));
    auto output_split_types_vec = output_split_types_t->vec<tstring>();

    // Sets output tensors from vectors.
    for (int i = 0; i < num_nodes; ++i) {
      output_node_ids_vec(i) = output_node_ids[i];
      // Adjust the gains to penalize by tree complexity.
      output_gains_vec(i) = output_gains[i] - tree_complexity;
      output_feature_dimensions_vec(i) = output_feature_dimensions[i];
      output_thresholds_vec(i) = output_thresholds[i];
      // TODO(crawles): change this for multi-class.
      output_left_node_contribs_matrix(i, 0) = output_left_node_contribs[i];
      output_right_node_contribs_matrix(i, 0) = output_right_node_contribs[i];
      output_split_types_vec(i) = output_split_types[i];
    }
  }

 protected:
  void process_node(const FeatureMap& f_map,
                    std::vector<int32>* output_node_ids,
                    std::vector<float>* output_gains,
                    std::vector<int32>* output_feature_dimensions,
                    std::vector<int32>* output_thresholds,
                    std::vector<float>* output_left_node_contribs,
                    std::vector<float>* output_right_node_contribs,
                    std::vector<string>* output_split_types,
                    const int32 node_id, const float min_node_weight,
                    const float l1, const float l2, const int32 num_buckets) {
    float parent_gain;
    Eigen::VectorXf unused(logits_dim_);
    Eigen::MatrixXf identity;
    identity.setIdentity(1, 1);

    // start processing for previous node id.
    float best_gain = std::numeric_limits<float>::lowest();
    float best_bucket = 0;
    float best_f_dim = 0;
    string best_split_type = kInequalityDefaultLeft;
    float best_contrib_for_left = 0.0;
    float best_contrib_for_right = 0.0;
    // the sum of gradients including default bucket.
    float total_grad = 0;
    // the sum of hessians including default bucket.
    float total_hess = 0;

    for (auto f_iter = f_map.begin(); f_iter != f_map.end(); ++f_iter) {
      const int32 feature_dim = f_iter->first;
      const auto buckets_to_stats_map = f_iter->second;

      // The very last bucket contains stats for missing values.
      // TODO(crawles): use vector for multi-class.
      const float default_grad =
          (buckets_to_stats_map.find(num_buckets) == buckets_to_stats_map.end()
               ? 0
               : buckets_to_stats_map.at(num_buckets)[0]);
      const float default_hess =
          (buckets_to_stats_map.find(num_buckets) == buckets_to_stats_map.end()
               ? 0
               : buckets_to_stats_map.at(num_buckets)[1]);

      if (f_iter == f_map.begin()) {
        // first get the sum of grads, including default bucket.
        for (auto b_iter = buckets_to_stats_map.begin();
             b_iter != buckets_to_stats_map.end(); ++b_iter) {
          total_grad += b_iter->second[0];
          total_hess += b_iter->second[1];
        }
        if (total_hess < min_node_weight) {
          // Do not split the node because not enough avg hessian.
          break;
        }
        CalculateWeightsAndGains(total_grad * identity, total_hess * identity,
                                 l1, l2, &unused, &parent_gain);
      }

      float total_left_grad = 0;
      float total_left_hess = 0;
      for (auto b_iter = buckets_to_stats_map.begin();
           b_iter != buckets_to_stats_map.end(); ++b_iter) {
        const int32 bucket_id = b_iter->first;
        // total_left_stats should exclude stats from default bucket.
        if (bucket_id == num_buckets) {
          break;
        }
        // TODO(crawles): vector for multi-class.
        total_left_grad += b_iter->second[0];
        total_left_hess += b_iter->second[1];
        // From left to right, default right.
        // Left child.
        Eigen::VectorXf contrib_for_left(1);
        float gain_for_left;
        CalculateWeightsAndGains(total_left_grad * identity,
                                 total_left_hess * identity, l1, l2,
                                 &contrib_for_left, &gain_for_left);
        // Right child.
        Eigen::VectorXf contrib_for_right(1);
        float gain_for_right;
        CalculateWeightsAndGains((total_grad - total_left_grad) * identity,
                                 (total_hess - total_left_hess) * identity, l1,
                                 l2, &contrib_for_right, &gain_for_right);
        if (GainIsLarger(gain_for_left + gain_for_right, best_gain)) {
          best_gain = gain_for_left + gain_for_right;
          best_bucket = bucket_id;
          best_f_dim = feature_dim;
          best_split_type = kInequalityDefaultRight;
          best_contrib_for_left = contrib_for_left[0];
          best_contrib_for_right = contrib_for_right[0];
        }

        // From right to left, default left.
        CalculateWeightsAndGains((total_left_grad + default_grad) * identity,
                                 (total_left_hess + default_hess) * identity,
                                 l1, l2, &contrib_for_left, &gain_for_left);
        CalculateWeightsAndGains(
            (total_grad - default_grad - total_left_grad) * identity,
            (total_hess - default_hess - total_left_hess) * identity, l1, l2,
            &contrib_for_right, &gain_for_right);
        if (GainIsLarger(gain_for_left + gain_for_right, best_gain)) {
          best_gain = gain_for_left + gain_for_right;
          best_bucket = bucket_id;
          best_f_dim = feature_dim;
          best_split_type = kInequalityDefaultLeft;
          best_contrib_for_left = contrib_for_left[0];
          best_contrib_for_right = contrib_for_right[0];
        }
      }  // for bucket_id
    }    // for feature_dim
    if (best_gain != std::numeric_limits<float>::lowest()) {
      output_node_ids->push_back(node_id);
      // Remove the parent gain.
      output_gains->push_back(best_gain - parent_gain);
      output_feature_dimensions->push_back(best_f_dim);
      output_split_types->push_back(best_split_type);
      output_thresholds->push_back(best_bucket);
      output_left_node_contribs->push_back(best_contrib_for_left);
      output_right_node_contribs->push_back(best_contrib_for_right);
    }
  }

 private:
  int logits_dim_;
  string split_type_;
};

REGISTER_KERNEL_BUILDER(
    Name("BoostedTreesSparseCalculateBestFeatureSplit").Device(DEVICE_CPU),
    BoostedTreesSparseCalculateBestFeatureSplitOp);

class BoostedTreesMakeStatsSummaryOp : public OpKernel {
 public:
  explicit BoostedTreesMakeStatsSummaryOp(OpKernelConstruction* const context)
      : OpKernel(context) {
    OP_REQUIRES_OK(context, context->GetAttr("max_splits", &max_splits_));
    OP_REQUIRES_OK(context, context->GetAttr("num_buckets", &num_buckets_));
    OP_REQUIRES_OK(context, context->GetAttr("num_features", &num_features_));
  }

  void Compute(OpKernelContext* const context) override {
    // node_ids
    const Tensor* node_ids_t;
    OP_REQUIRES_OK(context, context->input("node_ids", &node_ids_t));
    const auto node_ids = node_ids_t->vec<int32>();
    // gradients
    const Tensor* gradients_t;
    OP_REQUIRES_OK(context, context->input("gradients", &gradients_t));
    const auto gradients = gradients_t->matrix<float>();
    // hessians
    const Tensor* hessians_t;
    OP_REQUIRES_OK(context, context->input("hessians", &hessians_t));
    const auto hessians = hessians_t->matrix<float>();
    // bucketized_features
    OpInputList bucketized_features_list;
    OP_REQUIRES_OK(context, context->input_list("bucketized_features_list",
                                                &bucketized_features_list));
    // Infer batch size.
    const int64 batch_size = node_ids_t->dim_size(0);

    // Allocate temporary stats tensor (Rank 4).
    Tensor temp_stats_double_t;
    OP_REQUIRES_OK(context, context->allocate_temp(
                                DT_DOUBLE,
                                {num_features_, max_splits_, num_buckets_, 2},
                                &temp_stats_double_t));
    auto temp_stats_double = temp_stats_double_t.tensor<double, 4>();
    temp_stats_double.setZero();

    // Partition by node, and then bucketize.
    for (int feature_idx = 0; feature_idx < num_features_; ++feature_idx) {
      const auto& features = bucketized_features_list[feature_idx].vec<int32>();
      for (int i = 0; i < batch_size; ++i) {
        const int32 node = node_ids(i);
        const int32 bucket = features(i);
        temp_stats_double(feature_idx, node, bucket, 0) += gradients(i, 0);
        temp_stats_double(feature_idx, node, bucket, 1) += hessians(i, 0);
      }
    }

    // Copy temp tensor over to output tensor.
    Tensor* output_stats_summary_t = nullptr;
    OP_REQUIRES_OK(context, context->allocate_output(
                                "stats_summary", temp_stats_double_t.shape(),
                                &output_stats_summary_t));
    output_stats_summary_t->tensor<float, 4>() =
        temp_stats_double.template cast<float>();
  }

 private:
  int max_splits_;
  int num_buckets_;
  int num_features_;
};

REGISTER_KERNEL_BUILDER(Name("BoostedTreesMakeStatsSummary").Device(DEVICE_CPU),
                        BoostedTreesMakeStatsSummaryOp);

class BoostedTreesAggregateStatsOp : public OpKernel {
 public:
  explicit BoostedTreesAggregateStatsOp(OpKernelConstruction* const context)
      : OpKernel(context) {
    OP_REQUIRES_OK(context, context->GetAttr("max_splits", &max_splits_));
    OP_REQUIRES_OK(context, context->GetAttr("num_buckets", &num_buckets_));
  }

  void Compute(OpKernelContext* const context) override {
    // node_ids.
    const Tensor* node_ids_t;
    OP_REQUIRES_OK(context, context->input("node_ids", &node_ids_t));
    const auto node_ids = node_ids_t->vec<int32>();

    // gradients.
    const Tensor* gradients_t;
    OP_REQUIRES_OK(context, context->input("gradients", &gradients_t));
    const auto gradients = gradients_t->matrix<float>();

    // hessians.
    const Tensor* hessians_t;
    OP_REQUIRES_OK(context, context->input("hessians", &hessians_t));
    const auto hessians = hessians_t->matrix<float>();

    // feature.
    const Tensor* feature_t;
    OP_REQUIRES_OK(context, context->input("feature", &feature_t));
    const auto feature = feature_t->matrix<int32>();

    // Infer batch size, feature dimension and stats dimension.
    const int64 batch_size = node_ids_t->dim_size(0);
    const int64 logits_dims = gradients_t->dim_size(1);
    const int64 hessians_dims = hessians_t->dim_size(1);
    const int64 stats_dims = logits_dims + hessians_dims;
    const int64 feature_dims = feature_t->dim_size(1);

    // Allocate temporary stats tensor (Rank 4), upcasting to double.
    Tensor temp_stats_double_t;
    OP_REQUIRES_OK(context, context->allocate_temp(DT_DOUBLE,
                                                   {max_splits_, feature_dims,
                                                    num_buckets_, stats_dims},
                                                   &temp_stats_double_t));
    auto temp_stats_double = temp_stats_double_t.tensor<double, 4>();
    temp_stats_double.setZero();

    for (int i = 0; i < batch_size; ++i) {
      const int32 node = node_ids(i);
      for (int feature_dim = 0; feature_dim < feature_dims; ++feature_dim) {
        const int32 bucket = feature(i, feature_dim);
        for (int stat_dim = 0; stat_dim < logits_dims; ++stat_dim) {
          temp_stats_double(node, feature_dim, bucket, stat_dim) +=
              gradients(i, stat_dim);
        }
        for (int stat_dim = logits_dims; stat_dim < stats_dims; ++stat_dim) {
          temp_stats_double(node, feature_dim, bucket, stat_dim) +=
              hessians(i, stat_dim - logits_dims);
        }
      }
    }

    // Copy temp tensor over to output tensor, downcasting to float.
    Tensor* output_stats_summary_t = nullptr;
    OP_REQUIRES_OK(context, context->allocate_output(
                                "stats_summary", temp_stats_double_t.shape(),
                                &output_stats_summary_t));
    output_stats_summary_t->tensor<float, 4>() =
        temp_stats_double.template cast<float>();
  }

 private:
  int max_splits_;
  int num_buckets_;
};

REGISTER_KERNEL_BUILDER(Name("BoostedTreesAggregateStats").Device(DEVICE_CPU),
                        BoostedTreesAggregateStatsOp);

// Key based on node id, feature dimension and bucket id.
struct StatsPartitionKey {
  StatsPartitionKey(const int32 node_id, const int32 feature_dim,
                    const int32 bucket_id)
      : node_id(node_id), feature_dim(feature_dim), bucket_id(bucket_id) {}

  bool operator==(const StatsPartitionKey& other) const {
    return (node_id == other.node_id) && (feature_dim == other.feature_dim) &&
           (bucket_id == other.bucket_id);
  }

  // Compare for StatsPartitionKey.
  struct Less {
    bool operator()(const StatsPartitionKey& a,
                    const StatsPartitionKey& b) const {
      if (a.node_id < b.node_id) {
        return true;
      }
      if ((a.node_id == b.node_id) && (a.feature_dim < b.feature_dim)) {
        return true;
      }
      if ((a.node_id == b.node_id) && (a.feature_dim == b.feature_dim) &&
          (a.bucket_id < b.bucket_id)) {
        return true;
      }
      return false;
    }
  };

  // Tree node id.
  int32 node_id;
  // Dimension within feature column.
  int32 feature_dim;
  // bucketized feature value .
  int32 bucket_id;
};

typedef std::map<StatsPartitionKey, std::vector<float>, StatsPartitionKey::Less>
    StatsPartitionMap;
typedef StatsPartitionMap::iterator StatsPartitionIterator;

// Key based on instance and feature dimension.
struct InstanceFeatureDimKey {
  InstanceFeatureDimKey() : instance(-1), feature_dim(-1) {}

  InstanceFeatureDimKey(const int32 instance, const int32 feature_dim)
      : instance(instance), feature_dim(feature_dim) {}

  bool operator==(const InstanceFeatureDimKey& other) const {
    return (instance == other.instance) && (feature_dim == other.feature_dim);
  }

  // Compare for InstanceFeatureDimKey.
  struct Less {
    bool operator()(const InstanceFeatureDimKey& a,
                    const InstanceFeatureDimKey& b) const {
      if (a.instance < b.instance) {
        return true;
      }
      if ((a.instance == b.instance) && (a.feature_dim < b.feature_dim)) {
        return true;
      }
      return false;
    }
  };

  // Instance id within a batch.
  int32 instance;
  // Dimension within feature column.
  int32 feature_dim;
};

// Add statistics to StatsPartitionMap for (instance, feature dim, bucket id).
static void AddInstanceStatsToMap(const int32 instance, const int32 feature_dim,
                                  const int32 bucket_id,
                                  const int32 logits_dims,
                                  const int32 stats_dims,
                                  StatsPartitionMap* stats_map,
                                  const TTypes<float>::ConstMatrix& gradients,
                                  const TTypes<float>::ConstMatrix& hessians,
                                  const TTypes<int32>::ConstVec& node_ids) {
  const int32 node_id = node_ids(instance);
  const auto key = StatsPartitionKey(node_id, feature_dim, bucket_id);
  std::pair<StatsPartitionIterator, bool> const& insert_result =
      stats_map->insert(StatsPartitionIterator::value_type(
          key, std::vector<float>(stats_dims, 0.0f)));
  auto& stats = insert_result.first->second;
  for (int stat_dim = 0; stat_dim < logits_dims; ++stat_dim) {
    stats[stat_dim] += gradients(instance, stat_dim);
  }
  for (int stat_dim = logits_dims; stat_dim < stats_dims; ++stat_dim) {
    stats[stat_dim] += hessians(instance, stat_dim - logits_dims);
  }
}

// Add statistics to StatsPartitionMap for bucket_id ranging from
// (start_instance, start_feature_dim) to (end_instance, end_feature_dim),
// inclusive on start and end instances, exclusive on end feature dim.
static void AddRangeStats(const int start_instance, const int end_instance,
                          const int start_feature_dim,
                          const int end_feature_dim,
                          StatsPartitionMap* stats_map,
                          const TTypes<float>::ConstMatrix& gradients,
                          const TTypes<float>::ConstMatrix& hessians,
                          const TTypes<int32>::ConstVec& node_ids,
                          const int32 feature_dims, const int32 bucket_id,
                          const int32 logits_dims, const int32 stats_dims) {
  DCHECK_LE(start_instance, end_instance);
  if (start_instance == end_instance) {
    DCHECK_LT(start_feature_dim, end_feature_dim);
  }
  for (int32 instance = start_instance; instance <= end_instance; ++instance) {
    const int32 start_f_dim =
        (instance == start_instance) ? start_feature_dim + 1 : 0;
    const int32 end_f_dim =
        (instance == end_instance) ? end_feature_dim : feature_dims;
    for (int32 f_dim = start_f_dim; f_dim < end_f_dim; ++f_dim) {
      AddInstanceStatsToMap(instance, f_dim, bucket_id, logits_dims, stats_dims,
                            stats_map, gradients, hessians, node_ids);
    }
  }
}

class BoostedTreesSparseAggregateStatsOp : public OpKernel {
 public:
  explicit BoostedTreesSparseAggregateStatsOp(
      OpKernelConstruction* const context)
      : OpKernel(context) {
    OP_REQUIRES_OK(context, context->GetAttr("max_splits", &max_splits_));
    OP_REQUIRES_OK(context, context->GetAttr("num_buckets", &num_buckets_));
  }

  void Compute(OpKernelContext* const context) override {
    // node_ids.
    const Tensor* node_ids_t;
    OP_REQUIRES_OK(context, context->input("node_ids", &node_ids_t));
    const auto node_ids = node_ids_t->vec<int32>();

    // gradients.
    const Tensor* gradients_t;
    OP_REQUIRES_OK(context, context->input("gradients", &gradients_t));
    const auto gradients = gradients_t->matrix<float>();

    // hessians.
    const Tensor* hessians_t;
    OP_REQUIRES_OK(context, context->input("hessians", &hessians_t));
    const auto hessians = hessians_t->matrix<float>();

    // feature indices.
    const Tensor* feature_indices_t;
    OP_REQUIRES_OK(context,
                   context->input("feature_indices", &feature_indices_t));
    const auto feature_indices = feature_indices_t->matrix<int32>();

    // feature values.
    const Tensor* feature_values_t;
    OP_REQUIRES_OK(context,
                   context->input("feature_values", &feature_values_t));
    const auto feature_values = feature_values_t->vec<int32>();

    // feature shape.
    const Tensor* feature_shape_t;
    OP_REQUIRES_OK(context, context->input("feature_shape", &feature_shape_t));
    OP_REQUIRES(context, TensorShapeUtils::IsVector(feature_shape_t->shape()),
                errors::InvalidArgument(
                    "Input shapes should be a vector but received shapes ",
                    feature_shape_t->shape().DebugString()));
    const auto feature_shape = feature_shape_t->vec<int32>();

    const int64 batch_size = gradients_t->dim_size(0);
    const int64 logits_dims = gradients_t->dim_size(1);
    const int64 hessians_dims = hessians_t->dim_size(1);
    const int64 stats_dims = logits_dims + hessians_dims;
    const int64 num_sparse_entries = feature_indices_t->dim_size(0);
    const int32 feature_dims = feature_shape(1);
    DCHECK_LE(num_sparse_entries, batch_size * feature_dims);

    // Aggregate statistics info to map.
    StatsPartitionMap stats_map;

    int prev_instance = 0;
    int prev_f_dim = -1;

    for (int i = 0; i < num_sparse_entries; ++i) {
      // the instance number within a batch
      const int32 instance = feature_indices(i, 0);
      DCHECK_LE(instance, batch_size);
      DCHECK_GE(instance, prev_instance);
      // the node id within a tree.
      const int32 node_id = node_ids(instance);
      DCHECK_LE(node_id, max_splits_);
      // the feature dimension.
      const int32 f_dim = feature_indices(i, 1);
      DCHECK_LE(f_dim, feature_dims);
      // the bucket id of the value.
      const int32 bucket_id = feature_values(i);
      DCHECK_LE(bucket_id, num_buckets_);

      // Add statistics for the missing entries into default bucket.
      // The last bucket is default bucket.
      const int missing_entry_bucket = num_buckets_;
      AddRangeStats(prev_instance, instance, prev_f_dim, f_dim, &stats_map,
                    gradients, hessians, node_ids, feature_dims,
                    missing_entry_bucket, logits_dims, stats_dims);
      prev_instance = instance;
      prev_f_dim = f_dim;
      // Add statistics for the non-missing entry into
      // (cur_instance, cur_f_dim, bucket_id).
      AddInstanceStatsToMap(instance, f_dim, bucket_id, logits_dims, stats_dims,
                            &stats_map, gradients, hessians, node_ids);
    }
    AddRangeStats(prev_instance, batch_size - 1, prev_f_dim, feature_dims,
                  &stats_map, gradients, hessians, node_ids, feature_dims,
                  num_buckets_, logits_dims, stats_dims);

    // Serialize statistics info map to tensor output.
    const int64 num_slots = stats_map.size() * stats_dims;
    Tensor* summary_indices_t = nullptr;
    OP_REQUIRES_OK(context,
                   context->allocate_output("stats_summary_indices",
                                            TensorShape({num_slots, 4}),
                                            &summary_indices_t));
    auto summary_indices = summary_indices_t->matrix<int32>();
    Tensor* summary_values_t = nullptr;
    OP_REQUIRES_OK(context, context->allocate_output("stats_summary_values",
                                                     TensorShape({num_slots}),
                                                     &summary_values_t));
    auto summary_values = summary_values_t->vec<float>();
    int entry_index = 0;
    for (auto& iter : stats_map) {
      for (int stat_dim = 0; stat_dim < stats_dims; ++stat_dim) {
        summary_indices(entry_index, 0) = iter.first.node_id;
        summary_indices(entry_index, 1) = iter.first.feature_dim;
        summary_indices(entry_index, 2) = iter.first.bucket_id;
        summary_indices(entry_index, 3) = stat_dim;
        summary_values(entry_index) = iter.second[stat_dim];
        ++entry_index;
      }
    }

    Tensor* summary_shape_t = nullptr;
    OP_REQUIRES_OK(
        context, context->allocate_output("stats_summary_shape",
                                          TensorShape({4}), &summary_shape_t));
    auto summary_shape = summary_shape_t->vec<int32>();
    summary_shape(0) = max_splits_;
    summary_shape(1) = feature_dims;
    summary_shape(2) = num_buckets_ + 1;
    summary_shape(3) = stats_dims;
  }

 private:
  int max_splits_;
  int num_buckets_;
};

REGISTER_KERNEL_BUILDER(
    Name("BoostedTreesSparseAggregateStats").Device(DEVICE_CPU),
    BoostedTreesSparseAggregateStatsOp);

}  // namespace tensorflow