xgboost/src/learner.cc

/*!
 * Copyright 2014-2021 by Contributors
 * \file learner.cc
 * \brief Implementation of learning algorithm.
 * \author Tianqi Chen
 */
#include <dmlc/io.h>
#include <dmlc/parameter.h>
#include <dmlc/thread_local.h>

#include <atomic>
#include <mutex>
#include <algorithm>
#include <iomanip>
#include <limits>
#include <memory>
#include <sstream>
#include <string>
#include <stack>
#include <utility>
#include <vector>

#include "dmlc/any.h"
#include "xgboost/base.h"
#include "xgboost/c_api.h"
#include "xgboost/data.h"
#include "xgboost/model.h"
#include "xgboost/predictor.h"
#include "xgboost/feature_map.h"
#include "xgboost/gbm.h"
#include "xgboost/generic_parameters.h"
#include "xgboost/host_device_vector.h"
#include "xgboost/json.h"
#include "xgboost/learner.h"
#include "xgboost/logging.h"
#include "xgboost/metric.h"
#include "xgboost/objective.h"
#include "xgboost/parameter.h"

#include "common/common.h"
#include "common/io.h"
#include "common/observer.h"
#include "common/random.h"
#include "common/timer.h"
#include "common/charconv.h"
#include "common/version.h"
#include "common/threading_utils.h"

namespace {

const char* kMaxDeltaStepDefaultValue = "0.7";
}  // anonymous namespace

namespace xgboost {

enum class DataSplitMode : int {
  kAuto = 0, kCol = 1, kRow = 2
};
}  // namespace xgboost

DECLARE_FIELD_ENUM_CLASS(xgboost::DataSplitMode);

namespace xgboost {
// implementation of base learner.
bool Learner::AllowLazyCheckPoint() const {
  return gbm_->AllowLazyCheckPoint();
}

Learner::~Learner() = default;

/*! \brief training parameter for regression
 *
 * Should be deprecated, but still used for being compatible with binary IO.
 * Once it's gone, `LearnerModelParam` should handle transforming `base_margin`
 * with objective by itself.
 */
struct LearnerModelParamLegacy : public dmlc::Parameter<LearnerModelParamLegacy> {
  /* \brief global bias */
  bst_float base_score;
  /* \brief number of features  */
  uint32_t num_feature;
  /* \brief number of classes, if it is multi-class classification  */
  int32_t num_class;
  /*! \brief Model contain additional properties */
  int32_t contain_extra_attrs;
  /*! \brief Model contain eval metrics */
  int32_t contain_eval_metrics;
  /*! \brief the version of XGBoost. */
  uint32_t major_version;
  uint32_t minor_version;
  /*! \brief reserved field */
  int reserved[27];
  /*! \brief constructor */
  LearnerModelParamLegacy() {
    std::memset(this, 0, sizeof(LearnerModelParamLegacy));
    base_score = 0.5f;
    major_version = std::get<0>(Version::Self());
    minor_version = std::get<1>(Version::Self());
    static_assert(sizeof(LearnerModelParamLegacy) == 136,
                  "Do not change the size of this struct, as it will break binary IO.");
  }
  // Skip other legacy fields.
  Json ToJson() const {
    Object obj;
    char floats[NumericLimits<float>::kToCharsSize];
    auto ret = to_chars(floats, floats + NumericLimits<float>::kToCharsSize, base_score);
    CHECK(ret.ec == std::errc());
    obj["base_score"] =
        std::string{floats, static_cast<size_t>(std::distance(floats, ret.ptr))};

    char integers[NumericLimits<int64_t>::kToCharsSize];
    ret = to_chars(integers, integers + NumericLimits<int64_t>::kToCharsSize,
                   static_cast<int64_t>(num_feature));
    CHECK(ret.ec == std::errc());
    obj["num_feature"] =
        std::string{integers, static_cast<size_t>(std::distance(integers, ret.ptr))};
    ret = to_chars(integers, integers + NumericLimits<int64_t>::kToCharsSize,
                   static_cast<int64_t>(num_class));
    CHECK(ret.ec == std::errc());
    obj["num_class"] =
        std::string{integers, static_cast<size_t>(std::distance(integers, ret.ptr))};
    return Json(std::move(obj));
  }
  void FromJson(Json const& obj) {
    auto const& j_param = get<Object const>(obj);
    std::map<std::string, std::string> m;
    m["num_feature"] = get<String const>(j_param.at("num_feature"));
    m["num_class"] = get<String const>(j_param.at("num_class"));
    this->Init(m);
    std::string str = get<String const>(j_param.at("base_score"));
    from_chars(str.c_str(), str.c_str() + str.size(), base_score);
  }
  inline LearnerModelParamLegacy ByteSwap() const {
    LearnerModelParamLegacy x = *this;
    dmlc::ByteSwap(&x.base_score, sizeof(x.base_score), 1);
    dmlc::ByteSwap(&x.num_feature, sizeof(x.num_feature), 1);
    dmlc::ByteSwap(&x.num_class, sizeof(x.num_class), 1);
    dmlc::ByteSwap(&x.contain_extra_attrs, sizeof(x.contain_extra_attrs), 1);
    dmlc::ByteSwap(&x.contain_eval_metrics, sizeof(x.contain_eval_metrics), 1);
    dmlc::ByteSwap(&x.major_version, sizeof(x.major_version), 1);
    dmlc::ByteSwap(&x.minor_version, sizeof(x.minor_version), 1);
    dmlc::ByteSwap(x.reserved, sizeof(x.reserved[0]), sizeof(x.reserved) / sizeof(x.reserved[0]));
    return x;
  }

  // declare parameters
  DMLC_DECLARE_PARAMETER(LearnerModelParamLegacy) {
    DMLC_DECLARE_FIELD(base_score)
        .set_default(0.5f)
        .describe("Global bias of the model.");
    DMLC_DECLARE_FIELD(num_feature)
        .set_default(0)
        .describe(
            "Number of features in training data,"
            " this parameter will be automatically detected by learner.");
    DMLC_DECLARE_FIELD(num_class).set_default(0).set_lower_bound(0).describe(
        "Number of class option for multi-class classifier. "
        " By default equals 0 and corresponds to binary classifier.");
  }
};

LearnerModelParam::LearnerModelParam(
    LearnerModelParamLegacy const &user_param, float base_margin)
    : base_score{base_margin}, num_feature{user_param.num_feature},
      num_output_group{user_param.num_class == 0
                       ? 1
                       : static_cast<uint32_t>(user_param.num_class)}
{}

struct LearnerTrainParam : public XGBoostParameter<LearnerTrainParam> {
  // data split mode, can be row, col, or none.
  DataSplitMode dsplit {DataSplitMode::kAuto};
  // flag to disable default metric
  bool disable_default_eval_metric {false};
  // FIXME(trivialfis): The following parameters belong to model itself, but can be
  // specified by users.  Move them to model parameter once we can get rid of binary IO.
  std::string booster;
  std::string objective;

  // declare parameters
  DMLC_DECLARE_PARAMETER(LearnerTrainParam) {
    DMLC_DECLARE_FIELD(dsplit)
        .set_default(DataSplitMode::kAuto)
        .add_enum("auto", DataSplitMode::kAuto)
        .add_enum("col", DataSplitMode::kCol)
        .add_enum("row", DataSplitMode::kRow)
        .describe("Data split mode for distributed training.");
    DMLC_DECLARE_FIELD(disable_default_eval_metric)
        .set_default(false)
        .describe("Flag to disable default metric. Set to >0 to disable");
    DMLC_DECLARE_FIELD(booster)
        .set_default("gbtree")
        .describe("Gradient booster used for training.");
    DMLC_DECLARE_FIELD(objective)
        .set_default("reg:squarederror")
        .describe("Objective function used for obtaining gradient.");
  }
};


DMLC_REGISTER_PARAMETER(LearnerModelParamLegacy);
DMLC_REGISTER_PARAMETER(LearnerTrainParam);
DMLC_REGISTER_PARAMETER(GenericParameter);

int constexpr GenericParameter::kCpuId;
int64_t constexpr GenericParameter::kDefaultSeed;

void GenericParameter::ConfigureGpuId(bool require_gpu) {
#if defined(XGBOOST_USE_CUDA)
  if (gpu_id == kCpuId) {  // 0. User didn't specify the `gpu_id'
    if (require_gpu) {     // 1. `tree_method' or `predictor' or both are using
                           // GPU.
      // 2. Use device 0 as default.
      this->UpdateAllowUnknown(Args{{"gpu_id", "0"}});
    }
  }

  // 3. When booster is loaded from a memory image (Python pickle or R
  // raw model), number of available GPUs could be different.  Wrap around it.
  int32_t n_gpus = common::AllVisibleGPUs();
  if (n_gpus == 0) {
    if (gpu_id != kCpuId) {
      LOG(WARNING) << "No visible GPU is found, setting `gpu_id` to -1";
    }
    this->UpdateAllowUnknown(Args{{"gpu_id", std::to_string(kCpuId)}});
  } else if (fail_on_invalid_gpu_id) {
    CHECK(gpu_id == kCpuId || gpu_id < n_gpus)
      << "Only " << n_gpus << " GPUs are visible, gpu_id "
      << gpu_id << " is invalid.";
  } else if (gpu_id != kCpuId && gpu_id >= n_gpus) {
    LOG(WARNING) << "Only " << n_gpus
                 << " GPUs are visible, setting `gpu_id` to " << gpu_id % n_gpus;
    this->UpdateAllowUnknown(Args{{"gpu_id", std::to_string(gpu_id % n_gpus)}});
  }
#else
  // Just set it to CPU, don't think about it.
  this->UpdateAllowUnknown(Args{{"gpu_id", std::to_string(kCpuId)}});
#endif  // defined(XGBOOST_USE_CUDA)
}

using LearnerAPIThreadLocalStore =
    dmlc::ThreadLocalStore<std::map<Learner const *, XGBAPIThreadLocalEntry>>;

using ThreadLocalPredictionCache =
    dmlc::ThreadLocalStore<std::map<Learner const *, PredictionContainer>>;

class LearnerConfiguration : public Learner {
 private:
  std::mutex config_lock_;

 protected:
  static std::string const kEvalMetric;  // NOLINT

 protected:
  std::atomic<bool> need_configuration_;
  std::map<std::string, std::string> cfg_;
  // Stores information like best-iteration for early stopping.
  std::map<std::string, std::string> attributes_;
  // Name of each feature, usually set from DMatrix.
  std::vector<std::string> feature_names_;
  // Type of each feature, usually set from DMatrix.
  std::vector<std::string> feature_types_;

  common::Monitor monitor_;
  LearnerModelParamLegacy mparam_;
  LearnerModelParam learner_model_param_;
  LearnerTrainParam tparam_;
  std::vector<std::string> metric_names_;

 public:
  explicit LearnerConfiguration(std::vector<std::shared_ptr<DMatrix> > cache)
      : need_configuration_{true} {
    monitor_.Init("Learner");
    auto& local_cache = (*ThreadLocalPredictionCache::Get())[this];
    for (std::shared_ptr<DMatrix> const& d : cache) {
      local_cache.Cache(d, GenericParameter::kCpuId);
    }
  }
  ~LearnerConfiguration() override {
    auto local_cache = ThreadLocalPredictionCache::Get();
    if (local_cache->find(this) != local_cache->cend()) {
      local_cache->erase(this);
    }
  }

  // Configuration before data is known.
  void Configure() override {
    // Varient of double checked lock
    if (!this->need_configuration_) { return; }
    std::lock_guard<std::mutex> guard(config_lock_);
    if (!this->need_configuration_) { return; }

    monitor_.Start("Configure");
    auto old_tparam = tparam_;
    Args args = {cfg_.cbegin(), cfg_.cend()};

    tparam_.UpdateAllowUnknown(args);
    auto mparam_backup = mparam_;

    mparam_.UpdateAllowUnknown(args);

    auto initialized = generic_parameters_.GetInitialised();
    auto old_seed = generic_parameters_.seed;
    generic_parameters_.UpdateAllowUnknown(args);

    ConsoleLogger::Configure(args);
    common::OmpSetNumThreads(&generic_parameters_.nthread);

    // add additional parameters
    // These are cosntraints that need to be satisfied.
    if (tparam_.dsplit == DataSplitMode::kAuto && rabit::IsDistributed()) {
      tparam_.dsplit = DataSplitMode::kRow;
    }

    // set seed only before the model is initialized
    if (!initialized || generic_parameters_.seed != old_seed) {
      common::GlobalRandom().seed(generic_parameters_.seed);
    }

    // must precede configure gbm since num_features is required for gbm
    this->ConfigureNumFeatures();
    args = {cfg_.cbegin(), cfg_.cend()};  // renew
    this->ConfigureObjective(old_tparam, &args);

    // Before 1.0.0, we save `base_score` into binary as a transformed value by objective.
    // After 1.0.0 we save the value provided by user and keep it immutable instead.  To
    // keep the stability, we initialize it in binary LoadModel instead of configuration.
    // Under what condition should we omit the transformation:
    //
    // - base_score is loaded from old binary model.
    //
    // What are the other possible conditions:
    //
    // - model loaded from new binary or JSON.
    // - model is created from scratch.
    // - model is configured second time due to change of parameter
    if (!learner_model_param_.Initialized() || mparam_.base_score != mparam_backup.base_score) {
      learner_model_param_ = LearnerModelParam(mparam_,
                                               obj_->ProbToMargin(mparam_.base_score));
    }

    this->ConfigureGBM(old_tparam, args);
    generic_parameters_.ConfigureGpuId(this->gbm_->UseGPU());

    this->ConfigureMetrics(args);

    this->need_configuration_ = false;
    if (generic_parameters_.validate_parameters) {
      this->ValidateParameters();
    }

    // FIXME(trivialfis): Clear the cache once binary IO is gone.
    monitor_.Stop("Configure");
  }

  virtual PredictionContainer* GetPredictionCache() const {
    return &((*ThreadLocalPredictionCache::Get())[this]);
  }

  void LoadConfig(Json const& in) override {
    CHECK(IsA<Object>(in));
    Version::Load(in);

    auto const& learner_parameters = get<Object>(in["learner"]);
    FromJson(learner_parameters.at("learner_train_param"), &tparam_);

    auto const& gradient_booster = learner_parameters.at("gradient_booster");

    auto const& objective_fn = learner_parameters.at("objective");
    if (!obj_) {
      obj_.reset(ObjFunction::Create(tparam_.objective, &generic_parameters_));
    }
    obj_->LoadConfig(objective_fn);

    tparam_.booster = get<String>(gradient_booster["name"]);
    if (!gbm_) {
      gbm_.reset(GradientBooster::Create(tparam_.booster,
                                         &generic_parameters_, &learner_model_param_));
    }
    gbm_->LoadConfig(gradient_booster);

    auto const& j_metrics = learner_parameters.at("metrics");
    auto n_metrics = get<Array const>(j_metrics).size();
    metric_names_.resize(n_metrics);
    metrics_.resize(n_metrics);
    for (size_t i = 0; i < n_metrics; ++i) {
      metric_names_[i]= get<String>(j_metrics[i]);
      metrics_[i] = std::unique_ptr<Metric>(
          Metric::Create(metric_names_[i], &generic_parameters_));
    }

    FromJson(learner_parameters.at("generic_param"), &generic_parameters_);
    // make sure the GPU ID is valid in new environment before start running configure.
    generic_parameters_.ConfigureGpuId(false);

    this->need_configuration_ = true;
  }

  void SaveConfig(Json* p_out) const override {
    CHECK(!this->need_configuration_) << "Call Configure before saving model.";
    Version::Save(p_out);
    Json& out { *p_out };
    // parameters
    out["learner"] = Object();
    auto& learner_parameters = out["learner"];

    learner_parameters["learner_train_param"] = ToJson(tparam_);
    learner_parameters["learner_model_param"] = mparam_.ToJson();
    learner_parameters["gradient_booster"] = Object();
    auto& gradient_booster = learner_parameters["gradient_booster"];
    gbm_->SaveConfig(&gradient_booster);

    learner_parameters["objective"] = Object();
    auto& objective_fn = learner_parameters["objective"];
    obj_->SaveConfig(&objective_fn);

    std::vector<Json> metrics(metrics_.size());
    for (size_t i = 0; i < metrics_.size(); ++i) {
      metrics[i] = String(metrics_[i]->Name());
    }
    learner_parameters["metrics"] = Array(std::move(metrics));

    learner_parameters["generic_param"] = ToJson(generic_parameters_);
  }

  void SetParam(const std::string& key, const std::string& value) override {
    this->need_configuration_ = true;
    if (key == kEvalMetric) {
      if (std::find(metric_names_.cbegin(), metric_names_.cend(),
                    value) == metric_names_.cend()) {
        metric_names_.emplace_back(value);
      }
    } else {
      cfg_[key] = value;
    }
  }
  // Short hand for setting multiple parameters
  void SetParams(std::vector<std::pair<std::string, std::string>> const& args) override {
    for (auto const& kv : args) {
      this->SetParam(kv.first, kv.second);
    }
  }

  uint32_t GetNumFeature() const override {
    return learner_model_param_.num_feature;
  }

  void SetAttr(const std::string& key, const std::string& value) override {
    attributes_[key] = value;
    mparam_.contain_extra_attrs = 1;
  }

  bool GetAttr(const std::string& key, std::string* out) const override {
    auto it = attributes_.find(key);
    if (it == attributes_.end()) return false;
    *out = it->second;
    return true;
  }

  bool DelAttr(const std::string& key) override {
    auto it = attributes_.find(key);
    if (it == attributes_.end()) { return false; }
    attributes_.erase(it);
    return true;
  }

  void SetFeatureNames(std::vector<std::string> const& fn) override {
    feature_names_ = fn;
  }

  void GetFeatureNames(std::vector<std::string>* fn) const override {
    *fn = feature_names_;
  }

  void SetFeatureTypes(std::vector<std::string> const& ft) override {
    this->feature_types_ = ft;
  }

  void GetFeatureTypes(std::vector<std::string>* p_ft) const override {
    auto& ft = *p_ft;
    ft = this->feature_types_;
  }

  std::vector<std::string> GetAttrNames() const override {
    std::vector<std::string> out;
    for (auto const& kv : attributes_) {
      out.emplace_back(kv.first);
    }
    return out;
  }

  const std::map<std::string, std::string>& GetConfigurationArguments() const override {
    return cfg_;
  }

  GenericParameter const& GetGenericParameter() const override {
    return generic_parameters_;
  }

 private:
  void ValidateParameters() {
    Json config { Object() };
    this->SaveConfig(&config);
    std::stack<Json> stack;
    stack.push(config);
    std::string const postfix{"_param"};

    auto is_parameter = [&postfix](std::string const &key) {
      return key.size() > postfix.size() &&
             std::equal(postfix.rbegin(), postfix.rend(), key.rbegin());
    };

    // Extract all parameters
    std::vector<std::string> keys;
    // First global parameters
    Json const global_config{ToJson(*GlobalConfigThreadLocalStore::Get())};
    for (auto const& items : get<Object const>(global_config)) {
      keys.emplace_back(items.first);
    }
    // Parameters in various xgboost components.
    while (!stack.empty()) {
      auto j_obj = stack.top();
      stack.pop();
      auto const &obj = get<Object const>(j_obj);

      for (auto const &kv : obj) {
        if (is_parameter(kv.first)) {
          auto parameter = get<Object const>(kv.second);
          std::transform(parameter.begin(), parameter.end(), std::back_inserter(keys),
                         [](std::pair<std::string const&, Json const&> const& kv) {
                           return kv.first;
                         });
        } else if (IsA<Object>(kv.second)) {
          stack.push(kv.second);
        }
      }
    }

    // FIXME(trivialfis): Make eval_metric a training parameter.
    keys.emplace_back(kEvalMetric);
    keys.emplace_back("num_output_group");

    std::sort(keys.begin(), keys.end());

    std::vector<std::string> provided;
    for (auto const &kv : cfg_) {
      if (std::any_of(kv.first.cbegin(), kv.first.cend(),
                      [](char ch) { return std::isspace(ch); })) {
        LOG(FATAL) << "Invalid parameter \"" << kv.first << "\" contains whitespace.";
      }
      provided.push_back(kv.first);
    }
    std::sort(provided.begin(), provided.end());

    std::vector<std::string> diff;
    std::set_difference(provided.begin(), provided.end(), keys.begin(),
                        keys.end(), std::back_inserter(diff));
    if (diff.size() != 0) {
      std::stringstream ss;
      ss << "\nParameters: { ";
      for (size_t i = 0; i < diff.size() - 1; ++i) {
        ss << "\"" << diff[i] << "\", ";
      }
      ss << "\"" << diff.back() << "\"";
      ss << R"W( } might not be used.

  This may not be accurate due to some parameters are only used in language bindings but
  passed down to XGBoost core.  Or some parameters are not used but slip through this
  verification. Please open an issue if you find above cases.

)W";
      LOG(WARNING) << ss.str();
    }
  }

  void ConfigureNumFeatures() {
    // Compute number of global features if parameter not already set
    if (mparam_.num_feature == 0) {
      // TODO(hcho3): Change num_feature to 64-bit integer
      unsigned num_feature = 0;
      auto local_cache = this->GetPredictionCache();
      for (auto& matrix : local_cache->Container()) {
        CHECK(matrix.first);
        CHECK(!matrix.second.ref.expired());
        const uint64_t num_col = matrix.first->Info().num_col_;
        CHECK_LE(num_col,
                 static_cast<uint64_t>(std::numeric_limits<unsigned>::max()))
            << "Unfortunately, XGBoost does not support data matrices with "
            << std::numeric_limits<unsigned>::max() << " features or greater";
        num_feature = std::max(num_feature, static_cast<uint32_t>(num_col));
      }

      rabit::Allreduce<rabit::op::Max>(&num_feature, 1);
      if (num_feature > mparam_.num_feature) {
        mparam_.num_feature = num_feature;
      }
    }
    CHECK_NE(mparam_.num_feature, 0)
        << "0 feature is supplied.  Are you using raw Booster interface?";
    // Remove these once binary IO is gone.
    cfg_["num_feature"] = common::ToString(mparam_.num_feature);
    cfg_["num_class"] = common::ToString(mparam_.num_class);
  }

  void ConfigureGBM(LearnerTrainParam const& old, Args const& args) {
    if (gbm_ == nullptr || old.booster != tparam_.booster) {
      gbm_.reset(GradientBooster::Create(tparam_.booster, &generic_parameters_,
                                         &learner_model_param_));
    }
    gbm_->Configure(args);
  }

  void ConfigureObjective(LearnerTrainParam const& old, Args* p_args) {
    // Once binary IO is gone, NONE of these config is useful.
    if (cfg_.find("num_class") != cfg_.cend() && cfg_.at("num_class") != "0" &&
        tparam_.objective != "multi:softprob") {
      cfg_["num_output_group"] = cfg_["num_class"];
      if (atoi(cfg_["num_class"].c_str()) > 1 && cfg_.count("objective") == 0) {
        tparam_.objective = "multi:softmax";
      }
    }

    if (cfg_.find("max_delta_step") == cfg_.cend() &&
        cfg_.find("objective") != cfg_.cend() &&
        tparam_.objective == "count:poisson") {
      // max_delta_step is a duplicated parameter in Poisson regression and tree param.
      // Rename one of them once binary IO is gone.
      cfg_["max_delta_step"] = kMaxDeltaStepDefaultValue;
    }
    if (obj_ == nullptr || tparam_.objective != old.objective) {
      obj_.reset(ObjFunction::Create(tparam_.objective, &generic_parameters_));
    }
    auto& args = *p_args;
    args = {cfg_.cbegin(), cfg_.cend()};  // renew
    obj_->Configure(args);
  }

  void ConfigureMetrics(Args const& args) {
    for (auto const& name : metric_names_) {
      auto DupCheck = [&name](std::unique_ptr<Metric> const& m) {
                        return m->Name() != name;
                      };
      if (std::all_of(metrics_.begin(), metrics_.end(), DupCheck)) {
        metrics_.emplace_back(std::unique_ptr<Metric>(Metric::Create(name, &generic_parameters_)));
        mparam_.contain_eval_metrics = 1;
      }
    }
    for (auto& p_metric : metrics_) {
      p_metric->Configure(args);
    }
  }
};

std::string const LearnerConfiguration::kEvalMetric {"eval_metric"};  // NOLINT

class LearnerIO : public LearnerConfiguration {
 private:
  std::set<std::string> saved_configs_ = {"num_round"};
  // Used to identify the offset of JSON string when
  // Will be removed once JSON takes over.  Right now we still loads some RDS files from R.
  std::string const serialisation_header_ { u8"CONFIG-offset:" };

 public:
  explicit LearnerIO(std::vector<std::shared_ptr<DMatrix> > cache) :
      LearnerConfiguration{cache} {}

  void LoadModel(Json const& in) override {
    CHECK(IsA<Object>(in));
    Version::Load(in);
    auto const& learner = get<Object>(in["learner"]);
    mparam_.FromJson(learner.at("learner_model_param"));

    auto const& objective_fn = learner.at("objective");

    std::string name = get<String>(objective_fn["name"]);
    tparam_.UpdateAllowUnknown(Args{{"objective", name}});
    obj_.reset(ObjFunction::Create(name, &generic_parameters_));
    obj_->LoadConfig(objective_fn);

    auto const& gradient_booster = learner.at("gradient_booster");
    name = get<String>(gradient_booster["name"]);
    tparam_.UpdateAllowUnknown(Args{{"booster", name}});
    gbm_.reset(GradientBooster::Create(tparam_.booster,
                                       &generic_parameters_, &learner_model_param_));
    gbm_->LoadModel(gradient_booster);

    auto const& j_attributes = get<Object const>(learner.at("attributes"));
    attributes_.clear();
    for (auto const& kv : j_attributes) {
      attributes_[kv.first] = get<String const>(kv.second);
    }

    // feature names and types are saved in xgboost 1.4
    auto it = learner.find("feature_names");
    if (it != learner.cend()) {
      auto const &feature_names = get<Array const>(it->second);
      feature_names_.clear();
      for (auto const &name : feature_names) {
        feature_names_.emplace_back(get<String const>(name));
      }
    }
    it = learner.find("feature_types");
    if (it != learner.cend()) {
      auto const &feature_types = get<Array const>(it->second);
      feature_types_.clear();
      for (auto const &name : feature_types) {
        auto type = get<String const>(name);
        feature_types_.emplace_back(type);
      }
    }

    this->need_configuration_ = true;
  }

  void SaveModel(Json* p_out) const override {
    CHECK(!this->need_configuration_) << "Call Configure before saving model.";

    Version::Save(p_out);
    Json& out { *p_out };

    out["learner"] = Object();
    auto& learner = out["learner"];

    learner["learner_model_param"] = mparam_.ToJson();
    learner["gradient_booster"] = Object();
    auto& gradient_booster = learner["gradient_booster"];
    gbm_->SaveModel(&gradient_booster);

    learner["objective"] = Object();
    auto& objective_fn = learner["objective"];
    obj_->SaveConfig(&objective_fn);

    learner["attributes"] = Object();
    for (auto const& kv : attributes_) {
      learner["attributes"][kv.first] = String(kv.second);
    }

    learner["feature_names"] = Array();
    auto& feature_names = get<Array>(learner["feature_names"]);
    for (auto const& name : feature_names_) {
      feature_names.emplace_back(name);
    }
    learner["feature_types"] = Array();
    auto& feature_types = get<Array>(learner["feature_types"]);
    for (auto const& type : feature_types_) {
      feature_types.emplace_back(type);
    }
  }
  // About to be deprecated by JSON format
  void LoadModel(dmlc::Stream* fi) override {
    generic_parameters_.UpdateAllowUnknown(Args{});
    tparam_.Init(std::vector<std::pair<std::string, std::string>>{});
    // TODO(tqchen) mark deprecation of old format.
    common::PeekableInStream fp(fi);

    // backward compatible header check.
    std::string header;
    header.resize(4);
    if (fp.PeekRead(&header[0], 4) == 4) {
      CHECK_NE(header, "bs64")
          << "Base64 format is no longer supported in brick.";
      if (header == "binf") {
        CHECK_EQ(fp.Read(&header[0], 4), 4U);
      }
    }

    if (header[0] == '{') {
      // Dispatch to JSON
      auto json_stream = common::FixedSizeStream(&fp);
      std::string buffer;
      json_stream.Take(&buffer);
      auto model = Json::Load({buffer.c_str(), buffer.size()});
      this->LoadModel(model);
      return;
    }
    // use the peekable reader.
    fi = &fp;
    // read parameter
    CHECK_EQ(fi->Read(&mparam_, sizeof(mparam_)), sizeof(mparam_))
        << "BoostLearner: wrong model format";
    if (!DMLC_IO_NO_ENDIAN_SWAP) {
      mparam_ = mparam_.ByteSwap();
    }
    CHECK(fi->Read(&tparam_.objective)) << "BoostLearner: wrong model format";
    CHECK(fi->Read(&tparam_.booster)) << "BoostLearner: wrong model format";

    obj_.reset(ObjFunction::Create(tparam_.objective, &generic_parameters_));
    gbm_.reset(GradientBooster::Create(tparam_.booster, &generic_parameters_,
                                       &learner_model_param_));
    gbm_->Load(fi);
    if (mparam_.contain_extra_attrs != 0) {
      std::vector<std::pair<std::string, std::string> > attr;
      fi->Read(&attr);
      for (auto& kv : attr) {
        const std::string prefix = "SAVED_PARAM_";
        if (kv.first.find(prefix) == 0) {
          const std::string saved_param = kv.first.substr(prefix.length());
          if (saved_configs_.find(saved_param) != saved_configs_.end()) {
            cfg_[saved_param] = kv.second;
          }
        }
      }
      attributes_ = std::map<std::string, std::string>(attr.begin(), attr.end());
    }
    bool warn_old_model { false };
    if (attributes_.find("count_poisson_max_delta_step") != attributes_.cend()) {
      // Loading model from < 1.0.0, objective is not saved.
      cfg_["max_delta_step"] = attributes_.at("count_poisson_max_delta_step");
      attributes_.erase("count_poisson_max_delta_step");
      warn_old_model = true;
    } else {
      warn_old_model = false;
    }

    if (mparam_.major_version < 1) {
      // Before 1.0.0, base_score is saved as a transformed value, and there's no version
      // attribute (saved a 0) in the saved model.
      std::string multi{"multi:"};
      if (!std::equal(multi.cbegin(), multi.cend(), tparam_.objective.cbegin())) {
        HostDeviceVector<float> t;
        t.HostVector().resize(1);
        t.HostVector().at(0) = mparam_.base_score;
        this->obj_->PredTransform(&t);
        auto base_score = t.HostVector().at(0);
        mparam_.base_score = base_score;
      }
      warn_old_model = true;
    }

    learner_model_param_ =
        LearnerModelParam(mparam_, obj_->ProbToMargin(mparam_.base_score));
    if (attributes_.find("objective") != attributes_.cend()) {
      auto obj_str = attributes_.at("objective");
      auto j_obj = Json::Load({obj_str.c_str(), obj_str.size()});
      obj_->LoadConfig(j_obj);
      attributes_.erase("objective");
    } else {
      warn_old_model = true;
    }
    if (attributes_.find("metrics") != attributes_.cend()) {
      auto metrics_str = attributes_.at("metrics");
      std::vector<std::string> names { common::Split(metrics_str, ';') };
      attributes_.erase("metrics");
      for (auto const& n : names) {
        this->SetParam(kEvalMetric, n);
      }
    }

    if (warn_old_model) {
      LOG(WARNING) << "Loading model from XGBoost < 1.0.0, consider saving it "
                      "again for improved compatibility";
    }

    // Renew the version.
    mparam_.major_version = std::get<0>(Version::Self());
    mparam_.minor_version = std::get<1>(Version::Self());

    cfg_["num_class"] = common::ToString(mparam_.num_class);
    cfg_["num_feature"] = common::ToString(mparam_.num_feature);

    auto n = tparam_.__DICT__();
    cfg_.insert(n.cbegin(), n.cend());

    // copy dsplit from config since it will not run again during restore
    if (tparam_.dsplit == DataSplitMode::kAuto && rabit::IsDistributed()) {
      tparam_.dsplit = DataSplitMode::kRow;
    }

    this->need_configuration_ = true;
  }

  // Save model into binary format.  The code is about to be deprecated by more robust
  // JSON serialization format.
  void SaveModel(dmlc::Stream* fo) const override {
    LearnerModelParamLegacy mparam = mparam_;  // make a copy to potentially modify
    std::vector<std::pair<std::string, std::string> > extra_attr;
    mparam.contain_extra_attrs = 1;

    {
      std::vector<std::string> saved_params;
      for (const auto& key : saved_params) {
        auto it = cfg_.find(key);
        if (it != cfg_.end()) {
          mparam.contain_extra_attrs = 1;
          extra_attr.emplace_back("SAVED_PARAM_" + key, it->second);
        }
      }
    }
    {
      // Similar to JSON model IO, we save the objective.
      Json j_obj { Object() };
      obj_->SaveConfig(&j_obj);
      std::string obj_doc;
      Json::Dump(j_obj, &obj_doc);
      extra_attr.emplace_back("objective", obj_doc);
    }
    // As of 1.0.0, JVM Package and R Package uses Save/Load model for serialization.
    // Remove this part once they are ported to use actual serialization methods.
    if (mparam.contain_eval_metrics != 0) {
      std::stringstream os;
      for (auto& ev : metrics_) {
        os << ev->Name() << ";";
      }
      extra_attr.emplace_back("metrics", os.str());
    }
    std::string header {"binf"};
    fo->Write(header.data(), 4);
    if (DMLC_IO_NO_ENDIAN_SWAP) {
      fo->Write(&mparam, sizeof(LearnerModelParamLegacy));
    } else {
      LearnerModelParamLegacy x = mparam.ByteSwap();
      fo->Write(&x, sizeof(LearnerModelParamLegacy));
    }
    fo->Write(tparam_.objective);
    fo->Write(tparam_.booster);
    gbm_->Save(fo);
    if (mparam.contain_extra_attrs != 0) {
      std::map<std::string, std::string> attr(attributes_);
      for (const auto& kv : extra_attr) {
        attr[kv.first] = kv.second;
      }
      fo->Write(std::vector<std::pair<std::string, std::string>>(
          attr.begin(), attr.end()));
    }
  }

  void Save(dmlc::Stream* fo) const override {
    Json memory_snapshot{Object()};
    memory_snapshot["Model"] = Object();
    auto &model = memory_snapshot["Model"];
    this->SaveModel(&model);
    memory_snapshot["Config"] = Object();
    auto &config = memory_snapshot["Config"];
    this->SaveConfig(&config);
    std::string out_str;
    Json::Dump(memory_snapshot, &out_str);
    fo->Write(out_str.c_str(), out_str.size());
  }

  void Load(dmlc::Stream* fi) override {
    common::PeekableInStream fp(fi);
    char c {0};
    fp.PeekRead(&c, 1);
    if (c == '{') {
      std::string buffer;
      common::FixedSizeStream{&fp}.Take(&buffer);
      auto memory_snapshot = Json::Load({buffer.c_str(), buffer.size()});
      this->LoadModel(memory_snapshot["Model"]);
      this->LoadConfig(memory_snapshot["Config"]);
    } else {
      std::string header;
      header.resize(serialisation_header_.size());
      CHECK_EQ(fp.Read(&header[0], header.size()), serialisation_header_.size());
      // Avoid printing the content in loaded header, which might be random binary code.
      CHECK(header == serialisation_header_)  // NOLINT
          << R"doc(

  If you are loading a serialized model (like pickle in Python) generated by older
  XGBoost, please export the model by calling `Booster.save_model` from that version
  first, then load it back in current version.  There's a simple script for helping
  the process. See:

    https://xgboost.readthedocs.io/en/latest/tutorials/saving_model.html

  for reference to the script, and more details about differences between saving model and
  serializing.

)doc";
      int64_t sz {-1};
      CHECK_EQ(fp.Read(&sz, sizeof(sz)), sizeof(sz));
      if (!DMLC_IO_NO_ENDIAN_SWAP) {
        dmlc::ByteSwap(&sz, sizeof(sz), 1);
      }
      CHECK_GT(sz, 0);
      size_t json_offset = static_cast<size_t>(sz);
      std::string buffer;
      common::FixedSizeStream{&fp}.Take(&buffer);

      common::MemoryFixSizeBuffer binary_buf(&buffer[0], json_offset);
      this->LoadModel(&binary_buf);

      auto config = Json::Load({buffer.c_str() + json_offset, buffer.size() - json_offset});
      this->LoadConfig(config);
    }
  }
};

/*!
 * \brief learner that performs gradient boosting for a specific objective
 * function. It does training and prediction.
 */
class LearnerImpl : public LearnerIO {
 public:
  explicit LearnerImpl(std::vector<std::shared_ptr<DMatrix> > cache)
      : LearnerIO{cache} {}
  ~LearnerImpl() override {
    auto local_map = LearnerAPIThreadLocalStore::Get();
    if (local_map->find(this) != local_map->cend()) {
      local_map->erase(this);
    }
  }
  // Configuration before data is known.
  void CheckDataSplitMode() {
    if (rabit::IsDistributed()) {
      CHECK(tparam_.dsplit != DataSplitMode::kAuto)
        << "Precondition violated; dsplit cannot be 'auto' in distributed mode";
      if (tparam_.dsplit == DataSplitMode::kCol) {
        LOG(FATAL) << "Column-wise data split is currently not supported.";
      }
    }
  }

  std::vector<std::string> DumpModel(const FeatureMap& fmap,
                                     bool with_stats,
                                     std::string format) override {
    this->Configure();
    return gbm_->DumpModel(fmap, with_stats, format);
  }

  Learner *Slice(int32_t begin_layer, int32_t end_layer, int32_t step,
                 bool *out_of_bound) override {
    this->Configure();
    CHECK_GE(begin_layer, 0);
    auto *out_impl = new LearnerImpl({});
    auto gbm = std::unique_ptr<GradientBooster>(GradientBooster::Create(
        this->tparam_.booster, &this->generic_parameters_,
        &this->learner_model_param_));
    this->gbm_->Slice(begin_layer, end_layer, step, gbm.get(), out_of_bound);
    out_impl->gbm_ = std::move(gbm);
    Json config { Object() };
    this->SaveConfig(&config);
    out_impl->mparam_ = this->mparam_;
    out_impl->attributes_ = this->attributes_;
    out_impl->learner_model_param_ = this->learner_model_param_;
    out_impl->LoadConfig(config);
    out_impl->Configure();
    return out_impl;
  }

  void UpdateOneIter(int iter, std::shared_ptr<DMatrix> train) override {
    monitor_.Start("UpdateOneIter");
    TrainingObserver::Instance().Update(iter);
    this->Configure();
    if (generic_parameters_.seed_per_iteration || rabit::IsDistributed()) {
      common::GlobalRandom().seed(generic_parameters_.seed * kRandSeedMagic + iter);
    }
    this->CheckDataSplitMode();
    this->ValidateDMatrix(train.get(), true);

    auto local_cache = this->GetPredictionCache();
    auto& predt = local_cache->Cache(train, generic_parameters_.gpu_id);

    monitor_.Start("PredictRaw");
    this->PredictRaw(train.get(), &predt, true, 0, 0);
    TrainingObserver::Instance().Observe(predt.predictions, "Predictions");
    monitor_.Stop("PredictRaw");

    monitor_.Start("GetGradient");
    obj_->GetGradient(predt.predictions, train->Info(), iter, &gpair_);
    monitor_.Stop("GetGradient");
    TrainingObserver::Instance().Observe(gpair_, "Gradients");

    gbm_->DoBoost(train.get(), &gpair_, &predt);
    monitor_.Stop("UpdateOneIter");
  }

  void BoostOneIter(int iter, std::shared_ptr<DMatrix> train,
                    HostDeviceVector<GradientPair>* in_gpair) override {
    monitor_.Start("BoostOneIter");
    this->Configure();
    if (generic_parameters_.seed_per_iteration || rabit::IsDistributed()) {
      common::GlobalRandom().seed(generic_parameters_.seed * kRandSeedMagic + iter);
    }
    this->CheckDataSplitMode();
    this->ValidateDMatrix(train.get(), true);
    auto local_cache = this->GetPredictionCache();
    local_cache->Cache(train, generic_parameters_.gpu_id);

    gbm_->DoBoost(train.get(), in_gpair, &local_cache->Entry(train.get()));
    monitor_.Stop("BoostOneIter");
  }

  std::string EvalOneIter(int iter,
                          const std::vector<std::shared_ptr<DMatrix>>& data_sets,
                          const std::vector<std::string>& data_names) override {
    monitor_.Start("EvalOneIter");
    this->Configure();

    std::ostringstream os;
    os << '[' << iter << ']' << std::setiosflags(std::ios::fixed);
    if (metrics_.size() == 0 && tparam_.disable_default_eval_metric <= 0) {
      auto warn_default_eval_metric = [](const std::string& objective, const std::string& before,
                                         const std::string& after, const std::string& version) {
        LOG(WARNING) << "Starting in XGBoost " << version << ", the default evaluation metric "
                     << "used with the objective '" << objective << "' was changed from '"
                     << before << "' to '" << after << "'. Explicitly set eval_metric if you'd "
                     << "like to restore the old behavior.";
      };
      if (tparam_.objective == "binary:logistic") {
        warn_default_eval_metric(tparam_.objective, "error", "logloss", "1.3.0");
      } else if (tparam_.objective == "binary:logitraw") {
        warn_default_eval_metric(tparam_.objective, "auc", "logloss", "1.4.0");
      } else if ((tparam_.objective == "multi:softmax" || tparam_.objective == "multi:softprob")) {
        warn_default_eval_metric(tparam_.objective, "merror", "mlogloss", "1.3.0");
      }
      metrics_.emplace_back(Metric::Create(obj_->DefaultEvalMetric(), &generic_parameters_));
      metrics_.back()->Configure({cfg_.begin(), cfg_.end()});
    }

    auto local_cache = this->GetPredictionCache();
    for (size_t i = 0; i < data_sets.size(); ++i) {
      std::shared_ptr<DMatrix> m = data_sets[i];
      auto &predt = local_cache->Cache(m, generic_parameters_.gpu_id);
      this->ValidateDMatrix(m.get(), false);
      this->PredictRaw(m.get(), &predt, false, 0, 0);

      auto &out = output_predictions_.Cache(m, generic_parameters_.gpu_id).predictions;
      out.Resize(predt.predictions.Size());
      out.Copy(predt.predictions);

      obj_->EvalTransform(&out);
      for (auto& ev : metrics_) {
        os << '\t' << data_names[i] << '-' << ev->Name() << ':'
           << ev->Eval(out, m->Info(), tparam_.dsplit == DataSplitMode::kRow);
      }
    }

    monitor_.Stop("EvalOneIter");
    return os.str();
  }

  void Predict(std::shared_ptr<DMatrix> data, bool output_margin,
               HostDeviceVector<bst_float> *out_preds, unsigned layer_begin,
               unsigned layer_end, bool training,
               bool pred_leaf, bool pred_contribs, bool approx_contribs,
               bool pred_interactions) override {
    int multiple_predictions = static_cast<int>(pred_leaf) +
                               static_cast<int>(pred_interactions) +
                               static_cast<int>(pred_contribs);
    this->Configure();
    CHECK_LE(multiple_predictions, 1) << "Perform one kind of prediction at a time.";
    if (pred_contribs) {
      gbm_->PredictContribution(data.get(), out_preds, layer_begin, layer_end, approx_contribs);
    } else if (pred_interactions) {
      gbm_->PredictInteractionContributions(data.get(), out_preds, layer_begin, layer_end,
                                            approx_contribs);
    } else if (pred_leaf) {
      gbm_->PredictLeaf(data.get(), out_preds, layer_begin, layer_end);
    } else {
      auto local_cache = this->GetPredictionCache();
      auto& prediction = local_cache->Cache(data, generic_parameters_.gpu_id);
      this->PredictRaw(data.get(), &prediction, training, layer_begin, layer_end);
      // Copy the prediction cache to output prediction. out_preds comes from C API
      out_preds->SetDevice(generic_parameters_.gpu_id);
      out_preds->Resize(prediction.predictions.Size());
      out_preds->Copy(prediction.predictions);
      if (!output_margin) {
        obj_->PredTransform(out_preds);
      }
    }
  }

  int32_t BoostedRounds() const override {
    if (!this->gbm_) { return 0; }  // haven't call train or LoadModel.
    CHECK(!this->need_configuration_);
    return this->gbm_->BoostedRounds();
  }
  uint32_t Groups() const override {
    CHECK(!this->need_configuration_);
    return this->learner_model_param_.num_output_group;
  }

  XGBAPIThreadLocalEntry& GetThreadLocal() const override {
    return (*LearnerAPIThreadLocalStore::Get())[this];
  }

  void InplacePredict(dmlc::any const &x, std::shared_ptr<DMatrix> p_m,
                      PredictionType type, float missing,
                      HostDeviceVector<bst_float> **out_preds,
                      uint32_t iteration_begin,
                      uint32_t iteration_end) override {
    this->Configure();
    auto& out_predictions = this->GetThreadLocal().prediction_entry;
    this->gbm_->InplacePredict(x, p_m, missing, &out_predictions,
                               iteration_begin, iteration_end);
    if (type == PredictionType::kValue) {
      obj_->PredTransform(&out_predictions.predictions);
    } else if (type == PredictionType::kMargin) {
      // do nothing
    } else {
      LOG(FATAL) << "Unsupported prediction type:" << static_cast<int>(type);
    }
    *out_preds = &out_predictions.predictions;
  }

  void CalcFeatureScore(std::string const &importance_type,
                        std::vector<bst_feature_t> *features,
                        std::vector<float> *scores) override {
    this->Configure();
    gbm_->FeatureScore(importance_type, features, scores);
  }

  const std::map<std::string, std::string>& GetConfigurationArguments() const override {
    return cfg_;
  }

 protected:
  /*!
   * \brief get un-transformed prediction
   * \param data training data matrix
   * \param out_preds output vector that stores the prediction
   * \param ntree_limit limit number of trees used for boosted tree
   *   predictor, when it equals 0, this means we are using all the trees
   * \param training allow dropout when the DART booster is being used
   */
  void PredictRaw(DMatrix *data, PredictionCacheEntry *out_preds, bool training,
                  unsigned layer_begin, unsigned layer_end) const {
    CHECK(gbm_ != nullptr) << "Predict must happen after Load or configuration";
    this->ValidateDMatrix(data, false);
    gbm_->PredictBatch(data, out_preds, training, layer_begin, layer_end);
  }

  void ValidateDMatrix(DMatrix* p_fmat, bool is_training) const {
    MetaInfo const& info = p_fmat->Info();
    info.Validate(generic_parameters_.gpu_id);

    auto const row_based_split = [this]() {
      return tparam_.dsplit == DataSplitMode::kRow ||
             tparam_.dsplit == DataSplitMode::kAuto;
    };
    if (row_based_split()) {
      if (is_training) {
        CHECK_EQ(learner_model_param_.num_feature, p_fmat->Info().num_col_)
            << "Number of columns does not match number of features in "
               "booster.";
      } else {
        CHECK_GE(learner_model_param_.num_feature, p_fmat->Info().num_col_)
            << "Number of columns does not match number of features in "
               "booster.";
      }
    }

    if (p_fmat->Info().num_row_ == 0) {
      LOG(WARNING) << "Empty dataset at worker: " << rabit::GetRank();
    }
  }

 private:
  /*! \brief random number transformation seed. */
  static int32_t constexpr kRandSeedMagic = 127;
  // gradient pairs
  HostDeviceVector<GradientPair> gpair_;
  /*! \brief Temporary storage to prediction.  Useful for storing data transformed by
   *  objective function */
  PredictionContainer output_predictions_;
};

constexpr int32_t LearnerImpl::kRandSeedMagic;

Learner* Learner::Create(
    const std::vector<std::shared_ptr<DMatrix> >& cache_data) {
  return new LearnerImpl(cache_data);
}
}  // namespace xgboost