c_api.cpp

#include <LightGBM/utils/openmp_wrapper.h>

#include <LightGBM/utils/common.h>
#include <LightGBM/utils/random.h>
#include <LightGBM/utils/threading.h>
#include <LightGBM/c_api.h>
#include <LightGBM/dataset_loader.h>
#include <LightGBM/dataset.h>
#include <LightGBM/boosting.h>
#include <LightGBM/objective_function.h>
#include <LightGBM/metric.h>
#include <LightGBM/config.h>

#include <cstdio>
#include <vector>
#include <string>
#include <cstring>
#include <memory>
#include <stdexcept>
#include <mutex>
#include <functional>

#include "./application/predictor.hpp"
#include "./boosting/gbdt.h"

namespace LightGBM {

class Booster {
public:
  explicit Booster(const char* filename) {
    boosting_.reset(Boosting::CreateBoosting(filename));
  }

  Booster() {
    boosting_.reset(Boosting::CreateBoosting("gbdt", nullptr));
  }

  Booster(const Dataset* train_data,
          const char* parameters) {
    auto param = ConfigBase::Str2Map(parameters);
    config_.Set(param);
    if (config_.num_threads > 0) {
      omp_set_num_threads(config_.num_threads);
    }
    // create boosting
    if (config_.io_config.input_model.size() > 0) {
      Log::Warning("continued train from model is not support for c_api, \
        please use continued train with input score");
    }

    boosting_.reset(Boosting::CreateBoosting(config_.boosting_type, nullptr));

    // initialize the boosting
    boosting_->Init(&config_.boosting_config, nullptr, objective_fun_.get(),
                    Common::ConstPtrInVectorWrapper<Metric>(train_metric_));

    ResetTrainingData(train_data);
  }

  void MergeFrom(const Booster* other) {
    std::lock_guard<std::mutex> lock(mutex_);
    boosting_->MergeFrom(other->boosting_.get());
  }

  ~Booster() {

  }

  void ResetTrainingData(const Dataset* train_data) {
    std::lock_guard<std::mutex> lock(mutex_);
    train_data_ = train_data;
    // create objective function
    objective_fun_.reset(ObjectiveFunction::CreateObjectiveFunction(config_.objective_type,
                                                                    config_.objective_config));
    if (objective_fun_ == nullptr) {
      Log::Warning("Using self-defined objective function");
    }
    // initialize the objective function
    if (objective_fun_ != nullptr) {
      objective_fun_->Init(train_data_->metadata(), train_data_->num_data());
    }

    // create training metric
    train_metric_.clear();
    for (auto metric_type : config_.metric_types) {
      auto metric = std::unique_ptr<Metric>(
        Metric::CreateMetric(metric_type, config_.metric_config));
      if (metric == nullptr) { continue; }
      metric->Init(train_data_->metadata(), train_data_->num_data());
      train_metric_.push_back(std::move(metric));
    }
    train_metric_.shrink_to_fit();
    // reset the boosting
    boosting_->ResetTrainingData(&config_.boosting_config, train_data_,
                                 objective_fun_.get(), Common::ConstPtrInVectorWrapper<Metric>(train_metric_));
  }

  void ResetConfig(const char* parameters) {
    std::lock_guard<std::mutex> lock(mutex_);
    auto param = ConfigBase::Str2Map(parameters);
    if (param.count("num_class")) {
      Log::Fatal("cannot change num class during training");
    }
    if (param.count("boosting_type")) {
      Log::Fatal("cannot change boosting_type during training");
    }
    if (param.count("metric")) {
      Log::Fatal("cannot change metric during training");
    }

    config_.Set(param);
    if (config_.num_threads > 0) {
      omp_set_num_threads(config_.num_threads);
    }

    if (param.count("objective")) {
      // create objective function
      objective_fun_.reset(ObjectiveFunction::CreateObjectiveFunction(config_.objective_type,
                                                                      config_.objective_config));
      if (objective_fun_ == nullptr) {
        Log::Warning("Using self-defined objective function");
      }
      // initialize the objective function
      if (objective_fun_ != nullptr) {
        objective_fun_->Init(train_data_->metadata(), train_data_->num_data());
      }
    }

    boosting_->ResetTrainingData(&config_.boosting_config, train_data_,
                                 objective_fun_.get(), Common::ConstPtrInVectorWrapper<Metric>(train_metric_));

  }

  void AddValidData(const Dataset* valid_data) {
    std::lock_guard<std::mutex> lock(mutex_);
    valid_metrics_.emplace_back();
    for (auto metric_type : config_.metric_types) {
      auto metric = std::unique_ptr<Metric>(Metric::CreateMetric(metric_type, config_.metric_config));
      if (metric == nullptr) { continue; }
      metric->Init(valid_data->metadata(), valid_data->num_data());
      valid_metrics_.back().push_back(std::move(metric));
    }
    valid_metrics_.back().shrink_to_fit();
    boosting_->AddValidDataset(valid_data,
                               Common::ConstPtrInVectorWrapper<Metric>(valid_metrics_.back()));
  }

  bool TrainOneIter() {
    std::lock_guard<std::mutex> lock(mutex_);
    return boosting_->TrainOneIter(nullptr, nullptr, false);
  }

  bool TrainOneIter(const float* gradients, const float* hessians) {
    std::lock_guard<std::mutex> lock(mutex_);
    return boosting_->TrainOneIter(gradients, hessians, false);
  }

  void RollbackOneIter() {
    std::lock_guard<std::mutex> lock(mutex_);
    boosting_->RollbackOneIter();
  }

  Predictor NewPredictor(int num_iteration, int predict_type) {
    std::lock_guard<std::mutex> lock(mutex_);
    boosting_->SetNumIterationForPred(num_iteration);
    bool is_predict_leaf = false;
    bool is_raw_score = false;
    if (predict_type == C_API_PREDICT_LEAF_INDEX) {
      is_predict_leaf = true;
    } else if (predict_type == C_API_PREDICT_RAW_SCORE) {
      is_raw_score = true;
    } else {
      is_raw_score = false;
    }
    // not threading safe now
    // boosting_->SetNumIterationForPred may be set by other thread during prediction. 
    return Predictor(boosting_.get(), is_raw_score, is_predict_leaf);
  }

  void GetPredictAt(int data_idx, double* out_result, int64_t* out_len) {
    boosting_->GetPredictAt(data_idx, out_result, out_len);
  }

  void SaveModelToFile(int num_iteration, const char* filename) {
    boosting_->SaveModelToFile(num_iteration, filename);
  }

  void LoadModelFromString(const char* model_str) {
    boosting_->LoadModelFromString(model_str);
  }

  std::string SaveModelToString(int num_iteration) {
    return boosting_->SaveModelToString(num_iteration);
  }

  std::string DumpModel(int num_iteration) {
    return boosting_->DumpModel(num_iteration);
  }

  double GetLeafValue(int tree_idx, int leaf_idx) const {
    return dynamic_cast<GBDT*>(boosting_.get())->GetLeafValue(tree_idx, leaf_idx);
  }

  void SetLeafValue(int tree_idx, int leaf_idx, double val) {
    std::lock_guard<std::mutex> lock(mutex_);
    dynamic_cast<GBDT*>(boosting_.get())->SetLeafValue(tree_idx, leaf_idx, val);
  }

  int GetEvalCounts() const {
    int ret = 0;
    for (const auto& metric : train_metric_) {
      ret += static_cast<int>(metric->GetName().size());
    }
    return ret;
  }

  int GetEvalNames(char** out_strs) const {
    int idx = 0;
    for (const auto& metric : train_metric_) {
      for (const auto& name : metric->GetName()) {
        std::strcpy(out_strs[idx], name.c_str());
        ++idx;
      }
    }
    return idx;
  }

  int GetFeatureNames(char** out_strs) const {
    int idx = 0;
    for (const auto& name : boosting_->FeatureNames()) {
      std::strcpy(out_strs[idx], name.c_str());
      ++idx;
    }
    return idx;
  }

  const Boosting* GetBoosting() const { return boosting_.get(); }

private:

  const Dataset* train_data_;
  std::unique_ptr<Boosting> boosting_;
  /*! \brief All configs */
  OverallConfig config_;
  /*! \brief Metric for training data */
  std::vector<std::unique_ptr<Metric>> train_metric_;
  /*! \brief Metrics for validation data */
  std::vector<std::vector<std::unique_ptr<Metric>>> valid_metrics_;
  /*! \brief Training objective function */
  std::unique_ptr<ObjectiveFunction> objective_fun_;
  /*! \brief mutex for threading safe call */
  std::mutex mutex_;
};

}

using namespace LightGBM;

// some help functions used to convert data

std::function<std::vector<double>(int row_idx)>
RowFunctionFromDenseMatric(const void* data, int num_row, int num_col, int data_type, int is_row_major);

std::function<std::vector<std::pair<int, double>>(int row_idx)>
RowPairFunctionFromDenseMatric(const void* data, int num_row, int num_col, int data_type, int is_row_major);

std::function<std::vector<std::pair<int, double>>(int idx)>
RowFunctionFromCSR(const void* indptr, int indptr_type, const int32_t* indices,
                   const void* data, int data_type, int64_t nindptr, int64_t nelem);

// Row iterator of on column for CSC matrix
class CSC_RowIterator {
public:
  CSC_RowIterator(const void* col_ptr, int col_ptr_type, const int32_t* indices,
                  const void* data, int data_type, int64_t ncol_ptr, int64_t nelem, int col_idx);
  ~CSC_RowIterator() {}
  // return value at idx, only can access by ascent order
  double Get(int idx);
  // return next non-zero pair, if index < 0, means no more data
  std::pair<int, double> NextNonZero();
private:
  int nonzero_idx_ = 0;
  int cur_idx_ = -1;
  double cur_val_ = 0.0f;
  bool is_end_ = false;
  std::function<std::pair<int, double>(int idx)> iter_fun_;
};

// start of c_api functions

LIGHTGBM_C_EXPORT const char* LGBM_GetLastError() {
  return LastErrorMsg();
}

LIGHTGBM_C_EXPORT int LGBM_DatasetCreateFromFile(const char* filename,
                                                 const char* parameters,
                                                 const DatasetHandle reference,
                                                 DatasetHandle* out) {
  API_BEGIN();
  auto param = ConfigBase::Str2Map(parameters);
  IOConfig io_config;
  io_config.Set(param);
  DatasetLoader loader(io_config, nullptr, 1, filename);
  if (reference == nullptr) {
    *out = loader.LoadFromFile(filename);
  } else {
    *out = loader.LoadFromFileAlignWithOtherDataset(filename,
                                                    reinterpret_cast<const Dataset*>(reference));
  }
  API_END();
}

LIGHTGBM_C_EXPORT int LGBM_DatasetCreateFromSampledMat(const void* data,
                                                       int data_type,
                                                       int32_t num_sample_row,
                                                       int32_t ncol,
                                                       int32_t num_total_row,
                                                       const char* parameters,
                                                       DatasetHandle* out) {
  if (num_sample_row == num_total_row) {
    return LGBM_DatasetCreateFromMat(data, data_type, num_total_row, ncol, 1, parameters, nullptr, out);
  } else {
    API_BEGIN();
    auto param = ConfigBase::Str2Map(parameters);
    IOConfig io_config;
    io_config.Set(param);
    auto get_row_fun = RowFunctionFromDenseMatric(data, num_sample_row, ncol, data_type, 1);
    std::vector<std::vector<double>> sample_values(ncol);
    std::vector<std::vector<int>> sample_idx(ncol);
    for (int i = 0; i < num_sample_row; ++i) {
      auto row = get_row_fun(i);
      for (size_t idx = 0; idx < row.size(); ++idx) {
        if (std::fabs(row[idx]) > kEpsilon) {
          sample_values[idx].emplace_back(row[idx]);
          sample_idx[idx].emplace_back(i);
        }
      }
    }
    DatasetLoader loader(io_config, nullptr, 1, nullptr);
    *out = loader.CostructFromSampleData(sample_values, sample_idx,
                                         num_sample_row,
                                         static_cast<data_size_t>(num_total_row));
    API_END();
  }
}

LIGHTGBM_C_EXPORT int LGBM_DatasetCreateFromSampledCSR(const void* indptr,
                                                       int indptr_type,
                                                       const int32_t* indices,
                                                       const void* data,
                                                       int data_type,
                                                       int64_t nindptr,
                                                       int64_t n_sample_elem,
                                                       int64_t num_col,
                                                       int64_t num_total_row,
                                                       const char* parameters,
                                                       DatasetHandle* out) {
  if (nindptr - 1 == num_total_row) {
    return LGBM_DatasetCreateFromCSR(indptr, indptr_type, indices, data,
                                     data_type, nindptr, n_sample_elem, num_col, parameters, nullptr, out);
  } else {
    API_BEGIN();
    auto param = ConfigBase::Str2Map(parameters);
    IOConfig io_config;
    io_config.Set(param);
    auto get_row_fun = RowFunctionFromCSR(indptr, indptr_type, indices, data, data_type, nindptr, n_sample_elem);
    int32_t num_sample_row = static_cast<int32_t>(nindptr - 1);
    std::vector<std::vector<double>> sample_values(num_col);
    std::vector<std::vector<int>> sample_idx(num_col);
    for (int i = 0; i < num_sample_row; ++i) {
      auto row = get_row_fun(i);
      for (std::pair<int, double>& inner_data : row) {
        if (static_cast<size_t>(inner_data.first) >= sample_values.size()) {
          sample_values.resize(inner_data.first + 1);
          sample_idx.resize(inner_data.first + 1);
        }
        if (std::fabs(inner_data.second) > kEpsilon) {
          sample_values[inner_data.first].emplace_back(inner_data.second);
          sample_idx[inner_data.first].emplace_back(i);
        }
      }
    }
    CHECK(num_col >= static_cast<int>(sample_values.size()));
    DatasetLoader loader(io_config, nullptr, 1, nullptr);
    *out = loader.CostructFromSampleData(sample_values, sample_idx,
                                         num_sample_row,
                                         static_cast<data_size_t>(num_total_row));
    API_END();
  }
}

LIGHTGBM_C_EXPORT int LGBM_DatasetCreateByReference(const DatasetHandle reference,
                                                    int64_t num_total_row,
                                                    DatasetHandle* out) {
  API_BEGIN();
  std::unique_ptr<Dataset> ret;
  ret.reset(new Dataset(static_cast<data_size_t>(num_total_row)));
  ret->CreateValid(reinterpret_cast<const Dataset*>(reference));
  *out = ret.release();
  API_END();
}

LIGHTGBM_C_EXPORT int LGBM_DatasetPushRows(DatasetHandle dataset,
                                           const void* data,
                                           int data_type,
                                           int32_t nrow,
                                           int32_t ncol,
                                           int32_t start_row) {
  API_BEGIN();
  auto p_dataset = reinterpret_cast<Dataset*>(dataset);
  auto get_row_fun = RowFunctionFromDenseMatric(data, nrow, ncol, data_type, 1);
#pragma omp parallel for schedule(static)
  for (int i = 0; i < nrow; ++i) {
    const int tid = omp_get_thread_num();
    auto one_row = get_row_fun(i);
    p_dataset->PushOneRow(tid, start_row + i, one_row);
  }
  if (start_row + nrow == p_dataset->num_data()) {
    p_dataset->FinishLoad();
  }
  API_END();
}

LIGHTGBM_C_EXPORT int LGBM_DatasetPushRowsByCSR(DatasetHandle dataset,
                                                const void* indptr,
                                                int indptr_type,
                                                const int32_t* indices,
                                                const void* data,
                                                int data_type,
                                                int64_t nindptr,
                                                int64_t nelem,
                                                int64_t,
                                                int64_t start_row) {
  API_BEGIN();
  auto p_dataset = reinterpret_cast<Dataset*>(dataset);
  auto get_row_fun = RowFunctionFromCSR(indptr, indptr_type, indices, data, data_type, nindptr, nelem);
  int32_t nrow = static_cast<int32_t>(nindptr - 1);
#pragma omp parallel for schedule(static)
  for (int i = 0; i < nrow; ++i) {
    const int tid = omp_get_thread_num();
    auto one_row = get_row_fun(i);
    p_dataset->PushOneRow(tid,
                          static_cast<data_size_t>(start_row + i), one_row);
  }
  if (start_row + nrow == static_cast<int64_t>(p_dataset->num_data())) {
    p_dataset->FinishLoad();
  }
  API_END();
}

LIGHTGBM_C_EXPORT int LGBM_DatasetCreateFromMat(const void* data,
                                                int data_type,
                                                int32_t nrow,
                                                int32_t ncol,
                                                int is_row_major,
                                                const char* parameters,
                                                const DatasetHandle reference,
                                                DatasetHandle* out) {
  API_BEGIN();
  auto param = ConfigBase::Str2Map(parameters);
  IOConfig io_config;
  io_config.Set(param);
  std::unique_ptr<Dataset> ret;
  auto get_row_fun = RowFunctionFromDenseMatric(data, nrow, ncol, data_type, is_row_major);
  if (reference == nullptr) {
    // sample data first
    Random rand(io_config.data_random_seed);
    const int sample_cnt = static_cast<int>(nrow < io_config.bin_construct_sample_cnt ? nrow : io_config.bin_construct_sample_cnt);
    auto sample_indices = rand.Sample(nrow, sample_cnt);
    std::vector<std::vector<double>> sample_values(ncol);
    std::vector<std::vector<int>> sample_idx(ncol);
    for (size_t i = 0; i < sample_indices.size(); ++i) {
      auto idx = sample_indices[i];
      auto row = get_row_fun(static_cast<int>(idx));
      for (size_t j = 0; j < row.size(); ++j) {
        if (std::fabs(row[j]) > kEpsilon) {
          sample_values[j].emplace_back(row[j]);
          sample_idx[j].emplace_back(static_cast<int>(i));
        }
      }
    }
    DatasetLoader loader(io_config, nullptr, 1, nullptr);
    ret.reset(loader.CostructFromSampleData(sample_values, sample_idx, sample_cnt, nrow));
  } else {
    ret.reset(new Dataset(nrow));
    ret->CreateValid(
      reinterpret_cast<const Dataset*>(reference));
  }

#pragma omp parallel for schedule(static)
  for (int i = 0; i < nrow; ++i) {
    const int tid = omp_get_thread_num();
    auto one_row = get_row_fun(i);
    ret->PushOneRow(tid, i, one_row);
  }
  ret->FinishLoad();
  *out = ret.release();
  API_END();
}

LIGHTGBM_C_EXPORT int LGBM_DatasetCreateFromCSR(const void* indptr,
                                                int indptr_type,
                                                const int32_t* indices,
                                                const void* data,
                                                int data_type,
                                                int64_t nindptr,
                                                int64_t nelem,
                                                int64_t num_col,
                                                const char* parameters,
                                                const DatasetHandle reference,
                                                DatasetHandle* out) {
  API_BEGIN();
  auto param = ConfigBase::Str2Map(parameters);
  IOConfig io_config;
  io_config.Set(param);
  std::unique_ptr<Dataset> ret;
  auto get_row_fun = RowFunctionFromCSR(indptr, indptr_type, indices, data, data_type, nindptr, nelem);
  int32_t nrow = static_cast<int32_t>(nindptr - 1);
  if (reference == nullptr) {
    // sample data first
    Random rand(io_config.data_random_seed);
    const int sample_cnt = static_cast<int>(nrow < io_config.bin_construct_sample_cnt ? nrow : io_config.bin_construct_sample_cnt);
    auto sample_indices = rand.Sample(nrow, sample_cnt);
    std::vector<std::vector<double>> sample_values;
    std::vector<std::vector<int>> sample_idx;
    for (size_t i = 0; i < sample_indices.size(); ++i) {
      auto idx = sample_indices[i];
      auto row = get_row_fun(static_cast<int>(idx));
      for (std::pair<int, double>& inner_data : row) {
        if (static_cast<size_t>(inner_data.first) >= sample_values.size()) {
          sample_values.resize(inner_data.first + 1);
          sample_idx.resize(inner_data.first + 1);
        }
        if (std::fabs(inner_data.second) > kEpsilon) {
          sample_values[inner_data.first].emplace_back(inner_data.second);
          sample_idx[inner_data.first].emplace_back(static_cast<int>(i));
        }
      }
    }
    CHECK(num_col >= static_cast<int>(sample_values.size()));
    DatasetLoader loader(io_config, nullptr, 1, nullptr);
    ret.reset(loader.CostructFromSampleData(sample_values, sample_idx, sample_cnt, nrow));
  } else {
    ret.reset(new Dataset(nrow));
    ret->CreateValid(
      reinterpret_cast<const Dataset*>(reference));
  }

#pragma omp parallel for schedule(static)
  for (int i = 0; i < nindptr - 1; ++i) {
    const int tid = omp_get_thread_num();
    auto one_row = get_row_fun(i);
    ret->PushOneRow(tid, i, one_row);
  }
  ret->FinishLoad();
  *out = ret.release();
  API_END();
}

LIGHTGBM_C_EXPORT int LGBM_DatasetCreateFromCSC(const void* col_ptr,
                                                int col_ptr_type,
                                                const int32_t* indices,
                                                const void* data,
                                                int data_type,
                                                int64_t ncol_ptr,
                                                int64_t nelem,
                                                int64_t num_row,
                                                const char* parameters,
                                                const DatasetHandle reference,
                                                DatasetHandle* out) {
  API_BEGIN();
  auto param = ConfigBase::Str2Map(parameters);
  IOConfig io_config;
  io_config.Set(param);
  std::unique_ptr<Dataset> ret;
  int32_t nrow = static_cast<int32_t>(num_row);
  if (reference == nullptr) {
    // sample data first
    Random rand(io_config.data_random_seed);
    const int sample_cnt = static_cast<int>(nrow < io_config.bin_construct_sample_cnt ? nrow : io_config.bin_construct_sample_cnt);
    auto sample_indices = rand.Sample(nrow, sample_cnt);
    std::vector<std::vector<double>> sample_values(ncol_ptr - 1);
    std::vector<std::vector<int>> sample_idx(ncol_ptr - 1);
  #pragma omp parallel for schedule(static)
    for (int i = 0; i < static_cast<int>(sample_values.size()); ++i) {
      CSC_RowIterator col_it(col_ptr, col_ptr_type, indices, data, data_type, ncol_ptr, nelem, i);
      for (int j = 0; j < sample_cnt; j++) {
        auto val = col_it.Get(sample_indices[j]);
        if (std::fabs(val) > kEpsilon) {
          sample_values[i].emplace_back(val);
          sample_idx[i].emplace_back(j);
        }
      }
    }
    DatasetLoader loader(io_config, nullptr, 1, nullptr);
    ret.reset(loader.CostructFromSampleData(sample_values, sample_idx, sample_cnt, nrow));
  } else {
    ret.reset(new Dataset(nrow));
    ret->CreateValid(
      reinterpret_cast<const Dataset*>(reference));
  }

#pragma omp parallel for schedule(static)
  for (int i = 0; i < ncol_ptr - 1; ++i) {
    const int tid = omp_get_thread_num();
    int feature_idx = ret->InnerFeatureIndex(i);
    if (feature_idx < 0) { continue; }
    int group = ret->Feature2Group(feature_idx);
    int sub_feature = ret->Feture2SubFeature(feature_idx);
    CSC_RowIterator col_it(col_ptr, col_ptr_type, indices, data, data_type, ncol_ptr, nelem, i);
    int row_idx = 0;
    while (row_idx < nrow) {
      auto pair = col_it.NextNonZero();
      row_idx = pair.first;
      // no more data
      if (row_idx < 0) { break; }
      ret->PushOneData(tid, row_idx, group, sub_feature, pair.second);
    }
  }
  ret->FinishLoad();
  *out = ret.release();
  API_END();
}

LIGHTGBM_C_EXPORT int LGBM_DatasetGetSubset(
  const DatasetHandle handle,
  const int32_t* used_row_indices,
  int32_t num_used_row_indices,
  const char* parameters,
  DatasetHandle* out) {
  API_BEGIN();
  auto param = ConfigBase::Str2Map(parameters);
  IOConfig io_config;
  io_config.Set(param);
  auto full_dataset = reinterpret_cast<const Dataset*>(handle);
  auto ret = std::unique_ptr<Dataset>(new Dataset(num_used_row_indices));
  ret->CopyFeatureMapperFrom(full_dataset);
  ret->CopySubset(full_dataset, used_row_indices, num_used_row_indices, true);
  *out = ret.release();
  API_END();
}

LIGHTGBM_C_EXPORT int LGBM_DatasetSetFeatureNames(
  DatasetHandle handle,
  const char** feature_names,
  int num_feature_names) {
  API_BEGIN();
  auto dataset = reinterpret_cast<Dataset*>(handle);
  std::vector<std::string> feature_names_str;
  for (int i = 0; i < num_feature_names; ++i) {
    feature_names_str.emplace_back(feature_names[i]);
  }
  dataset->set_feature_names(feature_names_str);
  API_END();
}

LIGHTGBM_C_EXPORT int LGBM_DatasetGetFeatureNames(
  DatasetHandle handle,
  char** feature_names,
  int* num_feature_names) {
  API_BEGIN();
  auto dataset = reinterpret_cast<Dataset*>(handle);
  auto inside_feature_name = dataset->feature_names();
  *num_feature_names = static_cast<int>(inside_feature_name.size());
  for (int i = 0; i < *num_feature_names; ++i) {
    std::strcpy(feature_names[i], inside_feature_name[i].c_str());
  }
  API_END();
}

LIGHTGBM_C_EXPORT int LGBM_DatasetFree(DatasetHandle handle) {
  API_BEGIN();
  delete reinterpret_cast<Dataset*>(handle);
  API_END();
}

LIGHTGBM_C_EXPORT int LGBM_DatasetSaveBinary(DatasetHandle handle,
                                             const char* filename) {
  API_BEGIN();
  auto dataset = reinterpret_cast<Dataset*>(handle);
  dataset->SaveBinaryFile(filename);
  API_END();
}

LIGHTGBM_C_EXPORT int LGBM_DatasetSetField(DatasetHandle handle,
                                           const char* field_name,
                                           const void* field_data,
                                           int num_element,
                                           int type) {
  API_BEGIN();
  auto dataset = reinterpret_cast<Dataset*>(handle);
  bool is_success = false;
  if (type == C_API_DTYPE_FLOAT32) {
    is_success = dataset->SetFloatField(field_name, reinterpret_cast<const float*>(field_data), static_cast<int32_t>(num_element));
  } else if (type == C_API_DTYPE_INT32) {
    is_success = dataset->SetIntField(field_name, reinterpret_cast<const int*>(field_data), static_cast<int32_t>(num_element));
  } else if (type == C_API_DTYPE_FLOAT64) {
    is_success = dataset->SetDoubleField(field_name, reinterpret_cast<const double*>(field_data), static_cast<int32_t>(num_element));
  }
  if (!is_success) { throw std::runtime_error("Input data type erorr or field not found"); }
  API_END();
}

LIGHTGBM_C_EXPORT int LGBM_DatasetGetField(DatasetHandle handle,
                                           const char* field_name,
                                           int* out_len,
                                           const void** out_ptr,
                                           int* out_type) {
  API_BEGIN();
  auto dataset = reinterpret_cast<Dataset*>(handle);
  bool is_success = false;
  if (dataset->GetFloatField(field_name, out_len, reinterpret_cast<const float**>(out_ptr))) {
    *out_type = C_API_DTYPE_FLOAT32;
    is_success = true;
  } else if (dataset->GetIntField(field_name, out_len, reinterpret_cast<const int**>(out_ptr))) {
    *out_type = C_API_DTYPE_INT32;
    is_success = true;
  } else if (dataset->GetDoubleField(field_name, out_len, reinterpret_cast<const double**>(out_ptr))) {
    *out_type = C_API_DTYPE_FLOAT64;
    is_success = true;
  }
  if (!is_success) { throw std::runtime_error("Field not found"); }
  if (*out_ptr == nullptr) { *out_len = 0; }
  API_END();
}

LIGHTGBM_C_EXPORT int LGBM_DatasetGetNumData(DatasetHandle handle,
                                             int* out) {
  API_BEGIN();
  auto dataset = reinterpret_cast<Dataset*>(handle);
  *out = dataset->num_data();
  API_END();
}

LIGHTGBM_C_EXPORT int LGBM_DatasetGetNumFeature(DatasetHandle handle,
                                                int* out) {
  API_BEGIN();
  auto dataset = reinterpret_cast<Dataset*>(handle);
  *out = dataset->num_total_features();
  API_END();
}

// ---- start of booster

LIGHTGBM_C_EXPORT int LGBM_BoosterCreate(const DatasetHandle train_data,
                                         const char* parameters,
                                         BoosterHandle* out) {
  API_BEGIN();
  const Dataset* p_train_data = reinterpret_cast<const Dataset*>(train_data);
  auto ret = std::unique_ptr<Booster>(new Booster(p_train_data, parameters));
  *out = ret.release();
  API_END();
}

LIGHTGBM_C_EXPORT int LGBM_BoosterCreateFromModelfile(
  const char* filename,
  int* out_num_iterations,
  BoosterHandle* out) {
  API_BEGIN();
  auto ret = std::unique_ptr<Booster>(new Booster(filename));
  *out_num_iterations = ret->GetBoosting()->GetCurrentIteration();
  *out = ret.release();
  API_END();
}

LIGHTGBM_C_EXPORT int LGBM_BoosterLoadModelFromString(
  const char* model_str,
  int* out_num_iterations,
  BoosterHandle* out) {
  API_BEGIN();
  auto ret = std::unique_ptr<Booster>(new Booster());
  ret->LoadModelFromString(model_str);
  *out_num_iterations = ret->GetBoosting()->GetCurrentIteration();
  *out = ret.release();
  API_END();
}

LIGHTGBM_C_EXPORT int LGBM_BoosterFree(BoosterHandle handle) {
  API_BEGIN();
  delete reinterpret_cast<Booster*>(handle);
  API_END();
}

LIGHTGBM_C_EXPORT int LGBM_BoosterMerge(BoosterHandle handle,
                                        BoosterHandle other_handle) {
  API_BEGIN();
  Booster* ref_booster = reinterpret_cast<Booster*>(handle);
  Booster* ref_other_booster = reinterpret_cast<Booster*>(other_handle);
  ref_booster->MergeFrom(ref_other_booster);
  API_END();
}

LIGHTGBM_C_EXPORT int LGBM_BoosterAddValidData(BoosterHandle handle,
                                               const DatasetHandle valid_data) {
  API_BEGIN();
  Booster* ref_booster = reinterpret_cast<Booster*>(handle);
  const Dataset* p_dataset = reinterpret_cast<const Dataset*>(valid_data);
  ref_booster->AddValidData(p_dataset);
  API_END();
}

LIGHTGBM_C_EXPORT int LGBM_BoosterResetTrainingData(BoosterHandle handle,
                                                    const DatasetHandle train_data) {
  API_BEGIN();
  Booster* ref_booster = reinterpret_cast<Booster*>(handle);
  const Dataset* p_dataset = reinterpret_cast<const Dataset*>(train_data);
  ref_booster->ResetTrainingData(p_dataset);
  API_END();
}

LIGHTGBM_C_EXPORT int LGBM_BoosterResetParameter(BoosterHandle handle, const char* parameters) {
  API_BEGIN();
  Booster* ref_booster = reinterpret_cast<Booster*>(handle);
  ref_booster->ResetConfig(parameters);
  API_END();
}

LIGHTGBM_C_EXPORT int LGBM_BoosterGetNumClasses(BoosterHandle handle, int* out_len) {
  API_BEGIN();
  Booster* ref_booster = reinterpret_cast<Booster*>(handle);
  *out_len = ref_booster->GetBoosting()->NumberOfClasses();
  API_END();
}

LIGHTGBM_C_EXPORT int LGBM_BoosterUpdateOneIter(BoosterHandle handle, int* is_finished) {
  API_BEGIN();
  Booster* ref_booster = reinterpret_cast<Booster*>(handle);
  if (ref_booster->TrainOneIter()) {
    *is_finished = 1;
  } else {
    *is_finished = 0;
  }
  API_END();
}

LIGHTGBM_C_EXPORT int LGBM_BoosterUpdateOneIterCustom(BoosterHandle handle,
                                                      const float* grad,
                                                      const float* hess,
                                                      int* is_finished) {
  API_BEGIN();
  Booster* ref_booster = reinterpret_cast<Booster*>(handle);
  if (ref_booster->TrainOneIter(grad, hess)) {
    *is_finished = 1;
  } else {
    *is_finished = 0;
  }
  API_END();
}

LIGHTGBM_C_EXPORT int LGBM_BoosterRollbackOneIter(BoosterHandle handle) {
  API_BEGIN();
  Booster* ref_booster = reinterpret_cast<Booster*>(handle);
  ref_booster->RollbackOneIter();
  API_END();
}

LIGHTGBM_C_EXPORT int LGBM_BoosterGetCurrentIteration(BoosterHandle handle, int* out_iteration) {
  API_BEGIN();
  Booster* ref_booster = reinterpret_cast<Booster*>(handle);
  *out_iteration = ref_booster->GetBoosting()->GetCurrentIteration();
  API_END();
}

LIGHTGBM_C_EXPORT int LGBM_BoosterGetEvalCounts(BoosterHandle handle, int* out_len) {
  API_BEGIN();
  Booster* ref_booster = reinterpret_cast<Booster*>(handle);
  *out_len = ref_booster->GetEvalCounts();
  API_END();
}

LIGHTGBM_C_EXPORT int LGBM_BoosterGetEvalNames(BoosterHandle handle, int* out_len, char** out_strs) {
  API_BEGIN();
  Booster* ref_booster = reinterpret_cast<Booster*>(handle);
  *out_len = ref_booster->GetEvalNames(out_strs);
  API_END();
}

LIGHTGBM_C_EXPORT int LGBM_BoosterGetFeatureNames(BoosterHandle handle, int* out_len, char** out_strs) {
  API_BEGIN();
  Booster* ref_booster = reinterpret_cast<Booster*>(handle);
  *out_len = ref_booster->GetFeatureNames(out_strs);
  API_END();
}

LIGHTGBM_C_EXPORT int LGBM_BoosterGetNumFeature(BoosterHandle handle, int* out_len) {
  API_BEGIN();
  Booster* ref_booster = reinterpret_cast<Booster*>(handle);
  *out_len = ref_booster->GetBoosting()->MaxFeatureIdx() + 1;
  API_END();
}

LIGHTGBM_C_EXPORT int LGBM_BoosterGetEval(BoosterHandle handle,
                                          int data_idx,
                                          int* out_len,
                                          double* out_results) {
  API_BEGIN();
  Booster* ref_booster = reinterpret_cast<Booster*>(handle);
  auto boosting = ref_booster->GetBoosting();
  auto result_buf = boosting->GetEvalAt(data_idx);
  *out_len = static_cast<int>(result_buf.size());
  for (size_t i = 0; i < result_buf.size(); ++i) {
    (out_results)[i] = static_cast<double>(result_buf[i]);
  }
  API_END();
}

LIGHTGBM_C_EXPORT int LGBM_BoosterGetNumPredict(BoosterHandle handle,
                                                int data_idx,
                                                int64_t* out_len) {
  API_BEGIN();
  auto boosting = reinterpret_cast<Booster*>(handle)->GetBoosting();
  *out_len = boosting->GetNumPredictAt(data_idx);
  API_END();
}

LIGHTGBM_C_EXPORT int LGBM_BoosterGetPredict(BoosterHandle handle,
                                             int data_idx,
                                             int64_t* out_len,
                                             double* out_result) {
  API_BEGIN();
  Booster* ref_booster = reinterpret_cast<Booster*>(handle);
  ref_booster->GetPredictAt(data_idx, out_result, out_len);
  API_END();
}

LIGHTGBM_C_EXPORT int LGBM_BoosterPredictForFile(BoosterHandle handle,
                                                 const char* data_filename,
                                                 int data_has_header,
                                                 int predict_type,
                                                 int num_iteration,
                                                 const char* result_filename) {
  API_BEGIN();
  Booster* ref_booster = reinterpret_cast<Booster*>(handle);
  auto predictor = ref_booster->NewPredictor(static_cast<int>(num_iteration), predict_type);
  bool bool_data_has_header = data_has_header > 0 ? true : false;
  predictor.Predict(data_filename, result_filename, bool_data_has_header);
  API_END();
}

int64_t GetNumPredOneRow(const Booster* ref_booster, int predict_type, int64_t num_iteration) {
  int64_t num_preb_in_one_row = ref_booster->GetBoosting()->NumberOfClasses();
  if (predict_type == C_API_PREDICT_LEAF_INDEX) {
    int64_t max_iteration = ref_booster->GetBoosting()->GetCurrentIteration();
    if (num_iteration > 0) {
      num_preb_in_one_row *= static_cast<int>(std::min(max_iteration, num_iteration));
    } else {
      num_preb_in_one_row *= max_iteration;
    }
  }
  return num_preb_in_one_row;
}

LIGHTGBM_C_EXPORT int LGBM_BoosterCalcNumPredict(BoosterHandle handle,
                                                 int num_row,
                                                 int predict_type,
                                                 int num_iteration,
                                                 int64_t* out_len) {
  API_BEGIN();
  Booster* ref_booster = reinterpret_cast<Booster*>(handle);
  *out_len = static_cast<int64_t>(num_row * GetNumPredOneRow(ref_booster, predict_type, num_iteration));
  API_END();
}

LIGHTGBM_C_EXPORT int LGBM_BoosterPredictForCSR(BoosterHandle handle,
                                                const void* indptr,
                                                int indptr_type,
                                                const int32_t* indices,
                                                const void* data,
                                                int data_type,
                                                int64_t nindptr,
                                                int64_t nelem,
                                                int64_t,
                                                int predict_type,
                                                int num_iteration,
                                                int64_t* out_len,
                                                double* out_result) {
  API_BEGIN();
  Booster* ref_booster = reinterpret_cast<Booster*>(handle);
  auto predictor = ref_booster->NewPredictor(static_cast<int>(num_iteration), predict_type);
  auto get_row_fun = RowFunctionFromCSR(indptr, indptr_type, indices, data, data_type, nindptr, nelem);
  int64_t num_preb_in_one_row = GetNumPredOneRow(ref_booster, predict_type, num_iteration);
  int nrow = static_cast<int>(nindptr - 1);
#pragma omp parallel for schedule(static)
  for (int i = 0; i < nrow; ++i) {
    auto one_row = get_row_fun(i);
    auto predicton_result = predictor.GetPredictFunction()(one_row);
    for (int j = 0; j < static_cast<int>(predicton_result.size()); ++j) {
      out_result[i * num_preb_in_one_row + j] = static_cast<double>(predicton_result[j]);
    }
  }
  *out_len = nrow * num_preb_in_one_row;
  API_END();
}

LIGHTGBM_C_EXPORT int LGBM_BoosterPredictForCSC(BoosterHandle handle,
                                                const void* col_ptr,
                                                int col_ptr_type,
                                                const int32_t* indices,
                                                const void* data,
                                                int data_type,
                                                int64_t ncol_ptr,
                                                int64_t nelem,
                                                int64_t num_row,
                                                int predict_type,
                                                int num_iteration,
                                                int64_t* out_len,
                                                double* out_result) {
  API_BEGIN();
  Booster* ref_booster = reinterpret_cast<Booster*>(handle);
  auto predictor = ref_booster->NewPredictor(static_cast<int>(num_iteration), predict_type);
  int64_t num_preb_in_one_row = GetNumPredOneRow(ref_booster, predict_type, num_iteration);
  int ncol = static_cast<int>(ncol_ptr - 1);

  Threading::For<int64_t>(0, num_row,
                          [&predictor, &out_result, num_preb_in_one_row, ncol, col_ptr, col_ptr_type, indices, data, data_type, ncol_ptr, nelem]
  (int, data_size_t start, data_size_t end) {
    std::vector<CSC_RowIterator> iterators;
    for (int j = 0; j < ncol; ++j) {
      iterators.emplace_back(col_ptr, col_ptr_type, indices, data, data_type, ncol_ptr, nelem, j);
    }
    std::vector<std::pair<int, double>> one_row;
    for (int64_t i = start; i < end; ++i) {
      one_row.clear();
      for (int j = 0; j < ncol; ++j) {
        auto val = iterators[j].Get(static_cast<int>(i));
        if (std::fabs(val) > kEpsilon) {
          one_row.emplace_back(j, val);
        }
      }
      auto predicton_result = predictor.GetPredictFunction()(one_row);
      for (int j = 0; j < static_cast<int>(predicton_result.size()); ++j) {
        out_result[i * num_preb_in_one_row + j] = static_cast<double>(predicton_result[j]);
      }
    }
  });
  *out_len = num_row * num_preb_in_one_row;
  API_END();
}

LIGHTGBM_C_EXPORT int LGBM_BoosterPredictForMat(BoosterHandle handle,
                                                const void* data,
                                                int data_type,
                                                int32_t nrow,
                                                int32_t ncol,
                                                int is_row_major,
                                                int predict_type,
                                                int num_iteration,
                                                int64_t* out_len,
                                                double* out_result) {
  API_BEGIN();
  Booster* ref_booster = reinterpret_cast<Booster*>(handle);
  auto predictor = ref_booster->NewPredictor(static_cast<int>(num_iteration), predict_type);
  auto get_row_fun = RowPairFunctionFromDenseMatric(data, nrow, ncol, data_type, is_row_major);
  int64_t num_preb_in_one_row = GetNumPredOneRow(ref_booster, predict_type, num_iteration);
#pragma omp parallel for schedule(static)
  for (int i = 0; i < nrow; ++i) {
    auto one_row = get_row_fun(i);
    auto predicton_result = predictor.GetPredictFunction()(one_row);
    for (int j = 0; j < static_cast<int>(predicton_result.size()); ++j) {
      out_result[i * num_preb_in_one_row + j] = static_cast<double>(predicton_result[j]);
    }
  }
  *out_len = nrow * num_preb_in_one_row;
  API_END();
}

LIGHTGBM_C_EXPORT int LGBM_BoosterSaveModel(BoosterHandle handle,
                                            int num_iteration,
                                            const char* filename) {
  API_BEGIN();
  Booster* ref_booster = reinterpret_cast<Booster*>(handle);
  ref_booster->SaveModelToFile(num_iteration, filename);
  API_END();
}

LIGHTGBM_C_EXPORT int LGBM_BoosterSaveModelToString(BoosterHandle handle,
                                                    int num_iteration,
                                                    int buffer_len,
                                                    int* out_len,
                                                    char* out_str) {
  API_BEGIN();
  Booster* ref_booster = reinterpret_cast<Booster*>(handle);
  std::string model = ref_booster->SaveModelToString(num_iteration);
  *out_len = static_cast<int>(model.size()) + 1;
  if (*out_len <= buffer_len) {
    std::strcpy(out_str, model.c_str());
  }
  API_END();
}

LIGHTGBM_C_EXPORT int LGBM_BoosterDumpModel(BoosterHandle handle,
                                            int num_iteration,
                                            int buffer_len,
                                            int* out_len,
                                            char* out_str) {
  API_BEGIN();
  Booster* ref_booster = reinterpret_cast<Booster*>(handle);
  std::string model = ref_booster->DumpModel(num_iteration);
  *out_len = static_cast<int>(model.size()) + 1;
  if (*out_len <= buffer_len) {
    std::strcpy(out_str, model.c_str());
  }
  API_END();
}

LIGHTGBM_C_EXPORT int LGBM_BoosterGetLeafValue(BoosterHandle handle,
                                               int tree_idx,
                                               int leaf_idx,
                                               double* out_val) {
  API_BEGIN();
  Booster* ref_booster = reinterpret_cast<Booster*>(handle);
  *out_val = static_cast<double>(ref_booster->GetLeafValue(tree_idx, leaf_idx));
  API_END();
}

LIGHTGBM_C_EXPORT int LGBM_BoosterSetLeafValue(BoosterHandle handle,
                                               int tree_idx,
                                               int leaf_idx,
                                               double val) {
  API_BEGIN();
  Booster* ref_booster = reinterpret_cast<Booster*>(handle);
  ref_booster->SetLeafValue(tree_idx, leaf_idx, val);
  API_END();
}


LIGHTGBM_C_EXPORT int LGBM_AllocateArray(int64_t len, int type, ArrayHandle* out) {
  API_BEGIN();
  if (type == C_API_DTYPE_FLOAT32) {
    *out = new float[len];
  } else if (type == C_API_DTYPE_FLOAT64) {
    *out = new double[len];
  } else if (type == C_API_DTYPE_INT32) {
    *out = new int32_t[len];
  } else if (type == C_API_DTYPE_INT64) {
    *out = new int64_t[len];
  }
  API_END();
}

template<typename T>
void Copy(T* dst, const T* src, int64_t len) {
  std::memcpy(dst, src, sizeof(T) * len);
}

LIGHTGBM_C_EXPORT int LGBM_CopyToArray(ArrayHandle arr, int type, int64_t start_idx, const void* src, int64_t len) {
  API_BEGIN();
  if (type == C_API_DTYPE_FLOAT32) {
    Copy<float>(static_cast<float*>(arr) + start_idx, static_cast<const float*>(src), len);
  } else if (type == C_API_DTYPE_FLOAT64) {
    Copy<double>(static_cast<double*>(arr) + start_idx, static_cast<const double*>(src), len);
  } else if (type == C_API_DTYPE_INT32) {
    Copy<int32_t>(static_cast<int32_t*>(arr) + start_idx, static_cast<const int32_t*>(src), len);
  } else if (type == C_API_DTYPE_INT64) {
    Copy<int64_t>(static_cast<int64_t*>(arr) + start_idx, static_cast<const int64_t*>(src), len);
  }
  API_END();
}

LIGHTGBM_C_EXPORT int LGBM_FreeArray(ArrayHandle arr, int type) {
  API_BEGIN();
  if (type == C_API_DTYPE_FLOAT32) {
    delete[] static_cast<float*>(arr);
  } else if (type == C_API_DTYPE_FLOAT64) {
    delete[] static_cast<double*>(arr);
  } else if (type == C_API_DTYPE_INT32) {
    delete[] static_cast<int32_t*>(arr);
  } else if (type == C_API_DTYPE_INT64) {
    delete[] static_cast<int64_t*>(arr);
  }
  API_END();
}

// ---- start of some help functions

std::function<std::vector<double>(int row_idx)>
RowFunctionFromDenseMatric(const void* data, int num_row, int num_col, int data_type, int is_row_major) {
  if (data_type == C_API_DTYPE_FLOAT32) {
    const float* data_ptr = reinterpret_cast<const float*>(data);
    if (is_row_major) {
      return [data_ptr, num_col, num_row] (int row_idx) {
        std::vector<double> ret(num_col);
        auto tmp_ptr = data_ptr + num_col * row_idx;
        for (int i = 0; i < num_col; ++i) {
          ret[i] = static_cast<double>(*(tmp_ptr + i));
        }
        return ret;
      };
    } else {
      return [data_ptr, num_col, num_row] (int row_idx) {
        std::vector<double> ret(num_col);
        for (int i = 0; i < num_col; ++i) {
          ret[i] = static_cast<double>(*(data_ptr + num_row * i + row_idx));
        }
        return ret;
      };
    }
  } else if (data_type == C_API_DTYPE_FLOAT64) {
    const double* data_ptr = reinterpret_cast<const double*>(data);
    if (is_row_major) {
      return [data_ptr, num_col, num_row] (int row_idx) {
        std::vector<double> ret(num_col);
        auto tmp_ptr = data_ptr + num_col * row_idx;
        for (int i = 0; i < num_col; ++i) {
          ret[i] = static_cast<double>(*(tmp_ptr + i));
        }
        return ret;
      };
    } else {
      return [data_ptr, num_col, num_row] (int row_idx) {
        std::vector<double> ret(num_col);
        for (int i = 0; i < num_col; ++i) {
          ret[i] = static_cast<double>(*(data_ptr + num_row * i + row_idx));
        }
        return ret;
      };
    }
  }
  throw std::runtime_error("unknown data type in RowFunctionFromDenseMatric");
}

std::function<std::vector<std::pair<int, double>>(int row_idx)>
RowPairFunctionFromDenseMatric(const void* data, int num_row, int num_col, int data_type, int is_row_major) {
  auto inner_function = RowFunctionFromDenseMatric(data, num_row, num_col, data_type, is_row_major);
  if (inner_function != nullptr) {
    return [inner_function] (int row_idx) {
      auto raw_values = inner_function(row_idx);
      std::vector<std::pair<int, double>> ret;
      for (int i = 0; i < static_cast<int>(raw_values.size()); ++i) {
        if (std::fabs(raw_values[i]) > 1e-15) {
          ret.emplace_back(i, raw_values[i]);
        }
      }
      return ret;
    };
  }
  return nullptr;
}

std::function<std::vector<std::pair<int, double>>(int idx)>
RowFunctionFromCSR(const void* indptr, int indptr_type, const int32_t* indices, const void* data, int data_type, int64_t nindptr, int64_t nelem) {
  if (data_type == C_API_DTYPE_FLOAT32) {
    const float* data_ptr = reinterpret_cast<const float*>(data);
    if (indptr_type == C_API_DTYPE_INT32) {
      const int32_t* ptr_indptr = reinterpret_cast<const int32_t*>(indptr);
      return [ptr_indptr, indices, data_ptr, nindptr, nelem] (int idx) {
        std::vector<std::pair<int, double>> ret;
        int64_t start = ptr_indptr[idx];
        int64_t end = ptr_indptr[idx + 1];
        for (int64_t i = start; i < end; ++i) {
          ret.emplace_back(indices[i], data_ptr[i]);
        }
        return ret;
      };
    } else if (indptr_type == C_API_DTYPE_INT64) {
      const int64_t* ptr_indptr = reinterpret_cast<const int64_t*>(indptr);
      return [ptr_indptr, indices, data_ptr, nindptr, nelem] (int idx) {
        std::vector<std::pair<int, double>> ret;
        int64_t start = ptr_indptr[idx];
        int64_t end = ptr_indptr[idx + 1];
        for (int64_t i = start; i < end; ++i) {
          ret.emplace_back(indices[i], data_ptr[i]);
        }
        return ret;
      };
    }
  } else if (data_type == C_API_DTYPE_FLOAT64) {
    const double* data_ptr = reinterpret_cast<const double*>(data);
    if (indptr_type == C_API_DTYPE_INT32) {
      const int32_t* ptr_indptr = reinterpret_cast<const int32_t*>(indptr);
      return [ptr_indptr, indices, data_ptr, nindptr, nelem] (int idx) {
        std::vector<std::pair<int, double>> ret;
        int64_t start = ptr_indptr[idx];
        int64_t end = ptr_indptr[idx + 1];
        for (int64_t i = start; i < end; ++i) {
          ret.emplace_back(indices[i], data_ptr[i]);
        }
        return ret;
      };
    } else if (indptr_type == C_API_DTYPE_INT64) {
      const int64_t* ptr_indptr = reinterpret_cast<const int64_t*>(indptr);
      return [ptr_indptr, indices, data_ptr, nindptr, nelem] (int idx) {
        std::vector<std::pair<int, double>> ret;
        int64_t start = ptr_indptr[idx];
        int64_t end = ptr_indptr[idx + 1];
        for (int64_t i = start; i < end; ++i) {
          ret.emplace_back(indices[i], data_ptr[i]);
        }
        return ret;
      };
    }
  }
  throw std::runtime_error("unknown data type in RowFunctionFromCSR");
}

std::function<std::pair<int, double>(int idx)>
IterateFunctionFromCSC(const void* col_ptr, int col_ptr_type, const int32_t* indices, const void* data, int data_type, int64_t ncol_ptr, int64_t nelem, int col_idx) {
  CHECK(col_idx < ncol_ptr && col_idx >= 0);
  if (data_type == C_API_DTYPE_FLOAT32) {
    const float* data_ptr = reinterpret_cast<const float*>(data);
    if (col_ptr_type == C_API_DTYPE_INT32) {
      const int32_t* ptr_col_ptr = reinterpret_cast<const int32_t*>(col_ptr);
      int64_t start = ptr_col_ptr[col_idx];
      int64_t end = ptr_col_ptr[col_idx + 1];
      return [ptr_col_ptr, indices, data_ptr, ncol_ptr, nelem, start, end] (int bias) {
        int64_t i = static_cast<int64_t>(start + bias);
        if (i >= end) {
          return std::make_pair(-1, 0.0);
        }
        int idx = static_cast<int>(indices[i]);
        double val = static_cast<double>(data_ptr[i]);
        return std::make_pair(idx, val);
      };
    } else if (col_ptr_type == C_API_DTYPE_INT64) {
      const int64_t* ptr_col_ptr = reinterpret_cast<const int64_t*>(col_ptr);
      int64_t start = ptr_col_ptr[col_idx];
      int64_t end = ptr_col_ptr[col_idx + 1];
      return [ptr_col_ptr, indices, data_ptr, ncol_ptr, nelem, start, end] (int bias) {
        int64_t i = static_cast<int64_t>(start + bias);
        if (i >= end) {
          return std::make_pair(-1, 0.0);
        }
        int idx = static_cast<int>(indices[i]);
        double val = static_cast<double>(data_ptr[i]);
        return std::make_pair(idx, val);
      };
    }
  } else if (data_type == C_API_DTYPE_FLOAT64) {
    const double* data_ptr = reinterpret_cast<const double*>(data);
    if (col_ptr_type == C_API_DTYPE_INT32) {
      const int32_t* ptr_col_ptr = reinterpret_cast<const int32_t*>(col_ptr);
      int64_t start = ptr_col_ptr[col_idx];
      int64_t end = ptr_col_ptr[col_idx + 1];
      return [ptr_col_ptr, indices, data_ptr, ncol_ptr, nelem, start, end] (int bias) {
        int64_t i = static_cast<int64_t>(start + bias);
        if (i >= end) {
          return std::make_pair(-1, 0.0);
        }
        int idx = static_cast<int>(indices[i]);
        double val = static_cast<double>(data_ptr[i]);
        return std::make_pair(idx, val);
      };
    } else if (col_ptr_type == C_API_DTYPE_INT64) {
      const int64_t* ptr_col_ptr = reinterpret_cast<const int64_t*>(col_ptr);
      int64_t start = ptr_col_ptr[col_idx];
      int64_t end = ptr_col_ptr[col_idx + 1];
      return [ptr_col_ptr, indices, data_ptr, ncol_ptr, nelem, start, end] (int bias) {
        int64_t i = static_cast<int64_t>(start + bias);
        if (i >= end) {
          return std::make_pair(-1, 0.0);
        }
        int idx = static_cast<int>(indices[i]);
        double val = static_cast<double>(data_ptr[i]);
        return std::make_pair(idx, val);
      };
    }
  }
  throw std::runtime_error("unknown data type in CSC matrix");
}

CSC_RowIterator::CSC_RowIterator(const void* col_ptr, int col_ptr_type, const int32_t* indices,
                                 const void* data, int data_type, int64_t ncol_ptr, int64_t nelem, int col_idx) {
  iter_fun_ = IterateFunctionFromCSC(col_ptr, col_ptr_type, indices, data, data_type, ncol_ptr, nelem, col_idx);
}

double CSC_RowIterator::Get(int idx) {
  while (idx > cur_idx_ && !is_end_) {
    auto ret = iter_fun_(nonzero_idx_);
    if (ret.first < 0) {
      is_end_ = true;
      break;
    }
    cur_idx_ = ret.first;
    cur_val_ = ret.second;
    ++nonzero_idx_;
  }
  if (idx == cur_idx_) {
    return cur_val_;
  } else {
    return 0.0f;
  }
}

std::pair<int, double> CSC_RowIterator::NextNonZero() {
  if (!is_end_) {
    auto ret = iter_fun_(nonzero_idx_);
    ++nonzero_idx_;
    if (ret.first < 0) {
      is_end_ = true;
    }
    return ret;
  } else {
    return std::make_pair(-1, 0.0);
  }
}