goss.hpp

#ifndef LIGHTGBM_BOOSTING_GOSS_H_
#define LIGHTGBM_BOOSTING_GOSS_H_

#include <LightGBM/utils/array_args.h>
#include <LightGBM/utils/log.h>
#include <LightGBM/utils/openmp_wrapper.h>
#include <LightGBM/boosting.h>

#include "score_updater.hpp"
#include "gbdt.h"

#include <cstdio>
#include <vector>
#include <string>
#include <fstream>
#include <chrono>

namespace LightGBM {

class GOSS: public GBDT {
public:
  /*!
  * \brief Constructor
  */
  GOSS() : GBDT() { 

  }

  ~GOSS() {

  }

  void Init(const BoostingConfig* config, const Dataset* train_data, const ObjectiveFunction* object_function,
    const std::vector<const Metric*>& training_metrics) override {
    GBDT::Init(config, train_data, object_function, training_metrics);
    CHECK(gbdt_config_->top_rate + gbdt_config_->other_rate <= 1.0f);
    CHECK(gbdt_config_->top_rate > 0.0f && gbdt_config_->other_rate > 0.0f);
    if (gbdt_config_->bagging_freq > 0 && gbdt_config_->bagging_fraction != 1.0f) {
      Log::Fatal("cannot use bagging in GOSS");
    }
    Log::Info("using GOSS");
  }

  void ResetTrainingData(const BoostingConfig* config, const Dataset* train_data, const ObjectiveFunction* object_function,
    const std::vector<const Metric*>& training_metrics) override {
    if (config->bagging_freq > 0 && config->bagging_fraction != 1.0f) {
      Log::Fatal("cannot use bagging in GOSS");
    }
    GBDT::ResetTrainingData(config, train_data, object_function, training_metrics);
    if (train_data_ == nullptr) { return; }
    bag_data_indices_.resize(num_data_);
    tmp_indices_.resize(num_data_);
    tmp_indice_right_.resize(num_data_);
    offsets_buf_.resize(num_threads_);
    left_cnts_buf_.resize(num_threads_);
    right_cnts_buf_.resize(num_threads_);
    left_write_pos_buf_.resize(num_threads_);
    right_write_pos_buf_.resize(num_threads_);

    is_use_subset_ = false;
    if (config->top_rate + config->other_rate <= 0.5) {
      auto bag_data_cnt = static_cast<data_size_t>((config->top_rate + config->other_rate) * num_data_);
      tmp_subset_.reset(new Dataset(bag_data_cnt));
      tmp_subset_->CopyFeatureMapperFrom(train_data_);
      is_use_subset_ = true;
    }
    // flag to not bagging first
    bag_data_cnt_ = num_data_;
  }

  data_size_t BaggingHelper(Random& cur_rand, data_size_t start, data_size_t cnt, data_size_t* buffer, data_size_t* buffer_right) {
    std::vector<score_t> tmp_gradients(cnt, 0.0f);
    for (data_size_t i = 0; i < cnt; ++i) {
      for (int curr_class = 0; curr_class < num_class_; ++curr_class) {
        int idx = curr_class * num_data_ + start + i;
        tmp_gradients[i] += std::fabs(gradients_[idx] * hessians_[idx]);
      }
    }
    data_size_t top_k = static_cast<data_size_t>(cnt * gbdt_config_->top_rate);
    data_size_t other_k = static_cast<data_size_t>(cnt * gbdt_config_->other_rate);
    top_k = std::max(1, top_k);
    ArrayArgs<score_t>::ArgMaxAtK(&tmp_gradients, 0, static_cast<int>(tmp_gradients.size()), top_k);
    score_t threshold = tmp_gradients[top_k - 1];

    score_t multiply = static_cast<score_t>(cnt - top_k) / other_k;
    data_size_t cur_left_cnt = 0;
    data_size_t cur_right_cnt = 0;
    data_size_t big_weight_cnt = 0;
    for (data_size_t i = 0; i < cnt; ++i) {
      score_t grad = 0.0f;
      for (int curr_class = 0; curr_class < num_class_; ++curr_class) {
        int idx = curr_class * num_data_ + start + i;
        grad += std::fabs(gradients_[idx] * hessians_[idx]);
      }
      if (grad >= threshold) {
        buffer[cur_left_cnt++] = start + i;
        ++big_weight_cnt;
      } else {
        data_size_t sampled = cur_left_cnt - big_weight_cnt;
        data_size_t rest_need = other_k - sampled;
        data_size_t rest_all = (cnt - i) - (top_k - big_weight_cnt);
        double prob = (rest_need) / static_cast<double>(rest_all);
        if (cur_rand.NextFloat() < prob) {
          buffer[cur_left_cnt++] = start + i;
          gradients_[start + i] *= multiply;
          hessians_[start + i] *= multiply;
        } else {
          buffer_right[cur_right_cnt++] = start + i;
        }
      }
    }
    return cur_left_cnt;
  }

  void Bagging(int iter) override {
    bag_data_cnt_ = num_data_;
    // not subsample for first iterations
    if (iter < static_cast<int>(1.0f / gbdt_config_->learning_rate)) { return; }

    const data_size_t min_inner_size = 100;
    data_size_t inner_size = (num_data_ + num_threads_ - 1) / num_threads_;
    if (inner_size < min_inner_size) { inner_size = min_inner_size; }

#pragma omp parallel for schedule(static, 1)
    for (int i = 0; i < num_threads_; ++i) {
      left_cnts_buf_[i] = 0;
      right_cnts_buf_[i] = 0;
      data_size_t cur_start = i * inner_size;
      if (cur_start > num_data_) { continue; }
      data_size_t cur_cnt = inner_size;
      if (cur_start + cur_cnt > num_data_) { cur_cnt = num_data_ - cur_start; }
      Random cur_rand(gbdt_config_->bagging_seed + iter * num_threads_ + i);
      data_size_t cur_left_count = BaggingHelper(cur_rand, cur_start, cur_cnt,
        tmp_indices_.data() + cur_start, tmp_indice_right_.data() + cur_start);
      offsets_buf_[i] = cur_start;
      left_cnts_buf_[i] = cur_left_count;
      right_cnts_buf_[i] = cur_cnt - cur_left_count;
    }
    data_size_t left_cnt = 0;
    left_write_pos_buf_[0] = 0;
    right_write_pos_buf_[0] = 0;
    for (int i = 1; i < num_threads_; ++i) {
      left_write_pos_buf_[i] = left_write_pos_buf_[i - 1] + left_cnts_buf_[i - 1];
      right_write_pos_buf_[i] = right_write_pos_buf_[i - 1] + right_cnts_buf_[i - 1];
    }
    left_cnt = left_write_pos_buf_[num_threads_ - 1] + left_cnts_buf_[num_threads_ - 1];

#pragma omp parallel for schedule(static, 1)
    for (int i = 0; i < num_threads_; ++i) {
      if (left_cnts_buf_[i] > 0) {
        std::memcpy(bag_data_indices_.data() + left_write_pos_buf_[i],
          tmp_indices_.data() + offsets_buf_[i], left_cnts_buf_[i] * sizeof(data_size_t));
      }
      if (right_cnts_buf_[i] > 0) {
        std::memcpy(bag_data_indices_.data() + left_cnt + right_write_pos_buf_[i],
          tmp_indice_right_.data() + offsets_buf_[i], right_cnts_buf_[i] * sizeof(data_size_t));
      }
    }
    bag_data_cnt_ = left_cnt;
    // set bagging data to tree learner
    if (!is_use_subset_) {
      tree_learner_->SetBaggingData(bag_data_indices_.data(), bag_data_cnt_);
    } else {
      // get subset
      tmp_subset_->ReSize(bag_data_cnt_);
      tmp_subset_->CopySubset(train_data_, bag_data_indices_.data(), bag_data_cnt_, false);
      tree_learner_->ResetTrainingData(tmp_subset_.get());
    }
  }

  /*!
  * \brief Get Type name of this boosting object
  */
  const char* SubModelName() const override { return "tree"; }

private:
  std::vector<data_size_t> tmp_indice_right_;
};

}  // namespace LightGBM
#endif   // LIGHTGBM_BOOSTING_GOSS_H_