gbdt.cpp 27.3 KB
Newer Older
Guolin Ke's avatar
Guolin Ke committed
1
2
#include "gbdt.h"

3
#include <LightGBM/utils/openmp_wrapper.h>
4

Guolin Ke's avatar
Guolin Ke committed
5
6
7
8
9
10
11
12
13
14
15
#include <LightGBM/utils/common.h>

#include <LightGBM/objective_function.h>
#include <LightGBM/metric.h>

#include <ctime>

#include <sstream>
#include <chrono>
#include <string>
#include <vector>
16
#include <utility>
Guolin Ke's avatar
Guolin Ke committed
17
18
19

namespace LightGBM {

Guolin Ke's avatar
Guolin Ke committed
20
21
22
#ifdef TIMETAG
std::chrono::duration<double, std::milli> boosting_time;
std::chrono::duration<double, std::milli> train_score_time;
Guolin Ke's avatar
Guolin Ke committed
23
std::chrono::duration<double, std::milli> out_of_bag_score_time;
Guolin Ke's avatar
Guolin Ke committed
24
25
26
27
28
29
30
std::chrono::duration<double, std::milli> valid_score_time;
std::chrono::duration<double, std::milli> metric_time;
std::chrono::duration<double, std::milli> bagging_time;
std::chrono::duration<double, std::milli> sub_gradient_time;
std::chrono::duration<double, std::milli> tree_time;
#endif // TIMETAG

31
GBDT::GBDT()
32
  :iter_(0),
33
34
35
36
37
  train_data_(nullptr),
  object_function_(nullptr),
  early_stopping_round_(0),
  max_feature_idx_(0),
  num_class_(1),
38
  sigmoid_(-1.0f),
39
  num_iteration_for_pred_(0),
40
  shrinkage_rate_(0.1f),
wxchan's avatar
wxchan committed
41
  num_init_iteration_(0) {
42
43
44
45
46
#pragma omp parallel
#pragma omp master
    {
      num_threads_ = omp_get_num_threads();
    }
Guolin Ke's avatar
Guolin Ke committed
47
48
49
}

GBDT::~GBDT() {
Guolin Ke's avatar
Guolin Ke committed
50
51
52
#ifdef TIMETAG
  Log::Info("GBDT::boosting costs %f", boosting_time * 1e-3);
  Log::Info("GBDT::train_score costs %f", train_score_time * 1e-3);
Guolin Ke's avatar
Guolin Ke committed
53
  Log::Info("GBDT::out_of_bag_score costs %f", out_of_bag_score_time * 1e-3);
Guolin Ke's avatar
Guolin Ke committed
54
55
56
57
58
59
  Log::Info("GBDT::valid_score costs %f", valid_score_time * 1e-3);
  Log::Info("GBDT::metric costs %f", metric_time * 1e-3);
  Log::Info("GBDT::bagging costs %f", bagging_time * 1e-3);
  Log::Info("GBDT::sub_gradient costs %f", sub_gradient_time * 1e-3);
  Log::Info("GBDT::tree costs %f", tree_time * 1e-3);
#endif
Guolin Ke's avatar
Guolin Ke committed
60
61
}

62
63
64
void GBDT::Init(const BoostingConfig* config, const Dataset* train_data, const ObjectiveFunction* object_function,
     const std::vector<const Metric*>& training_metrics) {
  iter_ = 0;
wxchan's avatar
wxchan committed
65
  num_iteration_for_pred_ = 0;
66
  max_feature_idx_ = 0;
wxchan's avatar
wxchan committed
67
68
  num_class_ = config->num_class;
  train_data_ = nullptr;
Guolin Ke's avatar
Guolin Ke committed
69
  gbdt_config_ = nullptr;
70
  tree_learner_ = nullptr;
wxchan's avatar
wxchan committed
71
72
73
74
75
  ResetTrainingData(config, train_data, object_function, training_metrics);
}

void GBDT::ResetTrainingData(const BoostingConfig* config, const Dataset* train_data, const ObjectiveFunction* object_function,
  const std::vector<const Metric*>& training_metrics) {
Guolin Ke's avatar
Guolin Ke committed
76
  auto new_config = std::unique_ptr<BoostingConfig>(new BoostingConfig(*config));
wxchan's avatar
wxchan committed
77
78
79
  if (train_data_ != nullptr && !train_data_->CheckAlign(*train_data)) {
    Log::Fatal("cannot reset training data, since new training data has different bin mappers");
  }
Guolin Ke's avatar
Guolin Ke committed
80
81
82
  early_stopping_round_ = new_config->early_stopping_round;
  shrinkage_rate_ = new_config->learning_rate;

Guolin Ke's avatar
Guolin Ke committed
83
  object_function_ = object_function;
Guolin Ke's avatar
Guolin Ke committed
84

Guolin Ke's avatar
Guolin Ke committed
85
  sigmoid_ = -1.0f;
wxchan's avatar
wxchan committed
86
  if (object_function_ != nullptr
Guolin Ke's avatar
Guolin Ke committed
87
88
    && std::string(object_function_->GetName()) == std::string("binary")) {
    // only binary classification need sigmoid transform
Guolin Ke's avatar
Guolin Ke committed
89
    sigmoid_ = new_config->sigmoid;
90
  }
Guolin Ke's avatar
Guolin Ke committed
91

Guolin Ke's avatar
Guolin Ke committed
92
  if (train_data_ != train_data && train_data != nullptr) {
93
94
    if (tree_learner_ == nullptr) {
      tree_learner_ = std::unique_ptr<TreeLearner>(TreeLearner::CreateTreeLearner(new_config->tree_learner_type, &new_config->tree_config));
Guolin Ke's avatar
Guolin Ke committed
95
96
    }
    // init tree learner
97
    tree_learner_->Init(train_data);
Guolin Ke's avatar
Guolin Ke committed
98

Guolin Ke's avatar
Guolin Ke committed
99
100
101
102
103
104
    // push training metrics
    training_metrics_.clear();
    for (const auto& metric : training_metrics) {
      training_metrics_.push_back(metric);
    }
    training_metrics_.shrink_to_fit();
wxchan's avatar
wxchan committed
105
106
107
108
109
110
111
112
113
114
115
116
117
    // not same training data, need reset score and others
    // create score tracker
    train_score_updater_.reset(new ScoreUpdater(train_data, num_class_));
    // update score
    for (int i = 0; i < iter_; ++i) {
      for (int curr_class = 0; curr_class < num_class_; ++curr_class) {
        auto curr_tree = (i + num_init_iteration_) * num_class_ + curr_class;
        train_score_updater_->AddScore(models_[curr_tree].get(), curr_class);
      }
    }
    num_data_ = train_data->num_data();
    // create buffer for gradients and hessians
    if (object_function_ != nullptr) {
118
119
120
      size_t total_size = static_cast<size_t>(num_data_) * num_class_;
      gradients_.resize(total_size);
      hessians_.resize(total_size);
wxchan's avatar
wxchan committed
121
122
123
124
125
    }
    // get max feature index
    max_feature_idx_ = train_data->num_total_features() - 1;
    // get label index
    label_idx_ = train_data->label_idx();
126
127
    // get feature names
    feature_names_ = train_data->feature_names();
Guolin Ke's avatar
Guolin Ke committed
128
129

    feature_infos_ = train_data->feature_infos();
Guolin Ke's avatar
Guolin Ke committed
130
131
  }

Guolin Ke's avatar
Guolin Ke committed
132
133
  if ((train_data_ != train_data && train_data != nullptr)
    || (gbdt_config_ != nullptr && gbdt_config_->bagging_fraction != new_config->bagging_fraction)) {
wxchan's avatar
wxchan committed
134
    // if need bagging, create buffer
Guolin Ke's avatar
Guolin Ke committed
135
    if (new_config->bagging_fraction < 1.0 && new_config->bagging_freq > 0) {
136
137
      bag_data_cnt_ =
        static_cast<data_size_t>(new_config->bagging_fraction * num_data_);
138
      bag_data_indices_.resize(num_data_);
139
140
141
142
143
144
      tmp_indices_.resize(num_data_);
      offsets_buf_.resize(num_threads_);
      left_cnts_buf_.resize(num_threads_);
      right_cnts_buf_.resize(num_threads_);
      left_write_pos_buf_.resize(num_threads_);
      right_write_pos_buf_.resize(num_threads_);
Guolin Ke's avatar
Guolin Ke committed
145
146
      double average_bag_rate = new_config->bagging_fraction / new_config->bagging_freq;
      is_use_subset_ = false;
147
      if (average_bag_rate <= 0.5) {
Guolin Ke's avatar
Guolin Ke committed
148
        tmp_subset_.reset(new Dataset(bag_data_cnt_));
149
        tmp_subset_->CopyFeatureMapperFrom(train_data);
Guolin Ke's avatar
Guolin Ke committed
150
151
152
        is_use_subset_ = true;
        Log::Debug("use subset for bagging");
      }
wxchan's avatar
wxchan committed
153
154
155
    } else {
      bag_data_cnt_ = num_data_;
      bag_data_indices_.clear();
156
      tmp_indices_.clear();
Guolin Ke's avatar
Guolin Ke committed
157
      is_use_subset_ = false;
wxchan's avatar
wxchan committed
158
    }
Guolin Ke's avatar
Guolin Ke committed
159
  }
wxchan's avatar
wxchan committed
160
  train_data_ = train_data;
Guolin Ke's avatar
Guolin Ke committed
161
162
  if (train_data_ != nullptr) {
    // reset config for tree learner
163
    tree_learner_->ResetConfig(&new_config->tree_config);
Guolin Ke's avatar
Guolin Ke committed
164
  }
Guolin Ke's avatar
Guolin Ke committed
165
  gbdt_config_.reset(new_config.release());
Guolin Ke's avatar
Guolin Ke committed
166
167
}

wxchan's avatar
wxchan committed
168
void GBDT::AddValidDataset(const Dataset* valid_data,
Guolin Ke's avatar
Guolin Ke committed
169
  const std::vector<const Metric*>& valid_metrics) {
wxchan's avatar
wxchan committed
170
171
  if (!train_data_->CheckAlign(*valid_data)) {
    Log::Fatal("cannot add validation data, since it has different bin mappers with training data");
172
  }
Guolin Ke's avatar
Guolin Ke committed
173
  // for a validation dataset, we need its score and metric
Guolin Ke's avatar
Guolin Ke committed
174
  auto new_score_updater = std::unique_ptr<ScoreUpdater>(new ScoreUpdater(valid_data, num_class_));
wxchan's avatar
wxchan committed
175
176
177
178
179
180
181
  // update score
  for (int i = 0; i < iter_; ++i) {
    for (int curr_class = 0; curr_class < num_class_; ++curr_class) {
      auto curr_tree = (i + num_init_iteration_) * num_class_ + curr_class;
      new_score_updater->AddScore(models_[curr_tree].get(), curr_class);
    }
  }
Guolin Ke's avatar
Guolin Ke committed
182
  valid_score_updater_.push_back(std::move(new_score_updater));
Guolin Ke's avatar
Guolin Ke committed
183
  valid_metrics_.emplace_back();
184
185
186
  if (early_stopping_round_ > 0) {
    best_iter_.emplace_back();
    best_score_.emplace_back();
Guolin Ke's avatar
Guolin Ke committed
187
    best_msg_.emplace_back();
188
  }
Guolin Ke's avatar
Guolin Ke committed
189
190
  for (const auto& metric : valid_metrics) {
    valid_metrics_.back().push_back(metric);
191
192
193
    if (early_stopping_round_ > 0) {
      best_iter_.back().push_back(0);
      best_score_.back().push_back(kMinScore);
Guolin Ke's avatar
Guolin Ke committed
194
      best_msg_.back().emplace_back();
195
    }
Guolin Ke's avatar
Guolin Ke committed
196
  }
Guolin Ke's avatar
Guolin Ke committed
197
  valid_metrics_.back().shrink_to_fit();
Guolin Ke's avatar
Guolin Ke committed
198
199
}

Guolin Ke's avatar
Guolin Ke committed
200
data_size_t GBDT::BaggingHelper(Random& cur_rand, data_size_t start, data_size_t cnt, data_size_t* buffer){
201
202
203
  if (cnt <= 0) {
    return 0;
  }
204
205
206
207
  data_size_t bag_data_cnt =
    static_cast<data_size_t>(gbdt_config_->bagging_fraction * cnt);
  data_size_t cur_left_cnt = 0;
  data_size_t cur_right_cnt = 0;
Guolin Ke's avatar
Guolin Ke committed
208
  auto right_buffer = buffer + bag_data_cnt;
209
210
  // random bagging, minimal unit is one record
  for (data_size_t i = 0; i < cnt; ++i) {
Guolin Ke's avatar
Guolin Ke committed
211
212
213
    float prob =
      (bag_data_cnt - cur_left_cnt) / static_cast<float>(cnt - i);
    if (cur_rand.NextFloat() < prob) {
214
215
      buffer[cur_left_cnt++] = start + i;
    } else {
Guolin Ke's avatar
Guolin Ke committed
216
      right_buffer[cur_right_cnt++] = start + i;
217
218
219
220
221
    }
  }
  CHECK(cur_left_cnt == bag_data_cnt);
  return cur_left_cnt;
}
Guolin Ke's avatar
Guolin Ke committed
222

Guolin Ke's avatar
Guolin Ke committed
223
224


225
void GBDT::Bagging(int iter) {
Guolin Ke's avatar
Guolin Ke committed
226
  // if need bagging
227
  if (bag_data_cnt_ < num_data_ && iter % gbdt_config_->bagging_freq == 0) {
Guolin Ke's avatar
Guolin Ke committed
228
    const data_size_t min_inner_size = 1000;
229
230
231
    data_size_t inner_size = (num_data_ + num_threads_ - 1) / num_threads_;
    if (inner_size < min_inner_size) { inner_size = min_inner_size; }

Guolin Ke's avatar
Guolin Ke committed
232
#pragma omp parallel for schedule(static,1)
233
234
235
236
237
238
239
    for (int i = 0; i < num_threads_; ++i) {
      left_cnts_buf_[i] = 0;
      right_cnts_buf_[i] = 0;
      data_size_t cur_start = i * inner_size;
      if (cur_start > num_data_) { continue; }
      data_size_t cur_cnt = inner_size;
      if (cur_start + cur_cnt > num_data_) { cur_cnt = num_data_ - cur_start; }
Guolin Ke's avatar
Guolin Ke committed
240
241
      Random cur_rand(gbdt_config_->bagging_seed + iter * num_threads_ + i);
      data_size_t cur_left_count = BaggingHelper(cur_rand, cur_start, cur_cnt, tmp_indices_.data() + cur_start);
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
      offsets_buf_[i] = cur_start;
      left_cnts_buf_[i] = cur_left_count;
      right_cnts_buf_[i] = cur_cnt - cur_left_count;
    }
    data_size_t left_cnt = 0;
    left_write_pos_buf_[0] = 0;
    right_write_pos_buf_[0] = 0;
    for (int i = 1; i < num_threads_; ++i) {
      left_write_pos_buf_[i] = left_write_pos_buf_[i - 1] + left_cnts_buf_[i - 1];
      right_write_pos_buf_[i] = right_write_pos_buf_[i - 1] + right_cnts_buf_[i - 1];
    }
    left_cnt = left_write_pos_buf_[num_threads_ - 1] + left_cnts_buf_[num_threads_ - 1];

#pragma omp parallel for schedule(static, 1)
    for (int i = 0; i < num_threads_; ++i) {
      if (left_cnts_buf_[i] > 0) {
        std::memcpy(bag_data_indices_.data() + left_write_pos_buf_[i],
          tmp_indices_.data() + offsets_buf_[i], left_cnts_buf_[i] * sizeof(data_size_t));
Guolin Ke's avatar
Guolin Ke committed
260
      }
261
262
263
      if (right_cnts_buf_[i] > 0) {
        std::memcpy(bag_data_indices_.data() + left_cnt + right_write_pos_buf_[i],
          tmp_indices_.data() + offsets_buf_[i] + left_cnts_buf_[i], right_cnts_buf_[i] * sizeof(data_size_t));
Guolin Ke's avatar
Guolin Ke committed
264
265
      }
    }
Guolin Ke's avatar
Guolin Ke committed
266
267
    bag_data_cnt_ = left_cnt;
    CHECK(bag_data_indices_[bag_data_cnt_ - 1] > bag_data_indices_[bag_data_cnt_]);
Guolin Ke's avatar
Guolin Ke committed
268
    Log::Debug("Re-bagging, using %d data to train", bag_data_cnt_);
Guolin Ke's avatar
Guolin Ke committed
269
    // set bagging data to tree learner
Guolin Ke's avatar
Guolin Ke committed
270
271
272
273
    if (!is_use_subset_) {
      tree_learner_->SetBaggingData(bag_data_indices_.data(), bag_data_cnt_);
    } else {
      // get subset
Guolin Ke's avatar
Guolin Ke committed
274
275
      tmp_subset_->ReSize(bag_data_cnt_);
      tmp_subset_->CopySubset(train_data_, bag_data_indices_.data(), bag_data_cnt_, false);
Guolin Ke's avatar
Guolin Ke committed
276
277
      tree_learner_->ResetTrainingData(tmp_subset_.get());
    }
Guolin Ke's avatar
Guolin Ke committed
278
279
280
  }
}

281
void GBDT::UpdateScoreOutOfBag(const Tree* tree, const int curr_class) {
Guolin Ke's avatar
Guolin Ke committed
282
283
284
#ifdef TIMETAG
  auto start_time = std::chrono::steady_clock::now();
#endif
Hui Xue's avatar
Hui Xue committed
285
  // we need to predict out-of-bag socres of data for boosting
Guolin Ke's avatar
Guolin Ke committed
286
  if (num_data_ - bag_data_cnt_ > 0 && !is_use_subset_) {
287
    train_score_updater_->AddScore(tree, bag_data_indices_.data() + bag_data_cnt_, num_data_ - bag_data_cnt_, curr_class);
Guolin Ke's avatar
Guolin Ke committed
288
  }
Guolin Ke's avatar
Guolin Ke committed
289
#ifdef TIMETAG
Guolin Ke's avatar
Guolin Ke committed
290
  out_of_bag_score_time += std::chrono::steady_clock::now() - start_time;
Guolin Ke's avatar
Guolin Ke committed
291
#endif
Guolin Ke's avatar
Guolin Ke committed
292
293
}

294
bool GBDT::TrainOneIter(const score_t* gradient, const score_t* hessian, bool is_eval) {
Guolin Ke's avatar
Guolin Ke committed
295
296
  // boosting first
  if (gradient == nullptr || hessian == nullptr) {
Guolin Ke's avatar
Guolin Ke committed
297
298
299
#ifdef TIMETAG
    auto start_time = std::chrono::steady_clock::now();
#endif
Guolin Ke's avatar
Guolin Ke committed
300
301
302
    Boosting();
    gradient = gradients_.data();
    hessian = hessians_.data();
Guolin Ke's avatar
Guolin Ke committed
303
304
305
#ifdef TIMETAG
    boosting_time += std::chrono::steady_clock::now() - start_time;
#endif
Guolin Ke's avatar
Guolin Ke committed
306
  }
Guolin Ke's avatar
Guolin Ke committed
307
308
309
#ifdef TIMETAG
  auto start_time = std::chrono::steady_clock::now();
#endif
310
311
  // bagging logic
  Bagging(iter_);
Guolin Ke's avatar
Guolin Ke committed
312
313
314
#ifdef TIMETAG
  bagging_time += std::chrono::steady_clock::now() - start_time;
#endif
Guolin Ke's avatar
Guolin Ke committed
315
  if (is_use_subset_ && bag_data_cnt_ < num_data_) {
Guolin Ke's avatar
Guolin Ke committed
316
317
318
#ifdef TIMETAG
    start_time = std::chrono::steady_clock::now();
#endif
Guolin Ke's avatar
Guolin Ke committed
319
320
321
322
323
324
325
326
    if (gradients_.empty()) {
      size_t total_size = static_cast<size_t>(num_data_) * num_class_;
      gradients_.resize(total_size);
      hessians_.resize(total_size);
    }    
    // get sub gradients
    for (int curr_class = 0; curr_class < num_class_; ++curr_class) {
      auto bias = curr_class * num_data_;
Guolin Ke's avatar
Guolin Ke committed
327
      // cannot multi-threding
Guolin Ke's avatar
Guolin Ke committed
328
329
330
331
332
333
334
      for (int i = 0; i < bag_data_cnt_; ++i) {
        gradients_[bias + i] = gradient[bias + bag_data_indices_[i]];
        hessians_[bias + i] = hessian[bias + bag_data_indices_[i]];
      }
    }
    gradient = gradients_.data();
    hessian = hessians_.data();
Guolin Ke's avatar
Guolin Ke committed
335
336
337
#ifdef TIMETAG
    sub_gradient_time += std::chrono::steady_clock::now() - start_time;
#endif
Guolin Ke's avatar
Guolin Ke committed
338
  }
Guolin Ke's avatar
Guolin Ke committed
339
  for (int curr_class = 0; curr_class < num_class_; ++curr_class) {
Guolin Ke's avatar
Guolin Ke committed
340
341
342
#ifdef TIMETAG
    start_time = std::chrono::steady_clock::now();
#endif
Guolin Ke's avatar
Guolin Ke committed
343
    // train a new tree
344
    std::unique_ptr<Tree> new_tree(tree_learner_->Train(gradient + curr_class * num_data_, hessian + curr_class * num_data_));
Guolin Ke's avatar
Guolin Ke committed
345
346
347
#ifdef TIMETAG
    tree_time += std::chrono::steady_clock::now() - start_time;
#endif
Guolin Ke's avatar
Guolin Ke committed
348
349
    // if cannot learn a new tree, then stop
    if (new_tree->num_leaves() <= 1) {
Guolin Ke's avatar
Guolin Ke committed
350
      Log::Info("Stopped training because there are no more leaves that meet the split requirements.");
Guolin Ke's avatar
Guolin Ke committed
351
      return true;
352
    }
353

Guolin Ke's avatar
Guolin Ke committed
354
355
356
357
358
    // shrinkage by learning rate
    new_tree->Shrinkage(shrinkage_rate_);
    // update score
    UpdateScore(new_tree.get(), curr_class);
    UpdateScoreOutOfBag(new_tree.get(), curr_class);
359

Guolin Ke's avatar
Guolin Ke committed
360
361
362
363
364
365
366
367
368
    // add model
    models_.push_back(std::move(new_tree));
  }
  ++iter_;
  if (is_eval) {
    return EvalAndCheckEarlyStopping();
  } else {
    return false;
  }
369

Guolin Ke's avatar
Guolin Ke committed
370
}
371

wxchan's avatar
wxchan committed
372
void GBDT::RollbackOneIter() {
373
  if (iter_ <= 0) { return; }
wxchan's avatar
wxchan committed
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
  int cur_iter = iter_ + num_init_iteration_ - 1;
  // reset score
  for (int curr_class = 0; curr_class < num_class_; ++curr_class) {
    auto curr_tree = cur_iter * num_class_ + curr_class;
    models_[curr_tree]->Shrinkage(-1.0);
    train_score_updater_->AddScore(models_[curr_tree].get(), curr_class);
    for (auto& score_updater : valid_score_updater_) {
      score_updater->AddScore(models_[curr_tree].get(), curr_class);
    }
  }
  // remove model
  for (int curr_class = 0; curr_class < num_class_; ++curr_class) {
    models_.pop_back();
  }
  --iter_;
}

Guolin Ke's avatar
Guolin Ke committed
391
bool GBDT::EvalAndCheckEarlyStopping() {
392
  bool is_met_early_stopping = false;
Guolin Ke's avatar
Guolin Ke committed
393
394
395
#ifdef TIMETAG
  auto start_time = std::chrono::steady_clock::now();
#endif
396
  // print message for metric
Guolin Ke's avatar
Guolin Ke committed
397
  auto best_msg = OutputMetric(iter_);
Guolin Ke's avatar
Guolin Ke committed
398
399
400
#ifdef TIMETAG
  metric_time += std::chrono::steady_clock::now() - start_time;
#endif
Guolin Ke's avatar
Guolin Ke committed
401
  is_met_early_stopping = !best_msg.empty();
402
403
404
  if (is_met_early_stopping) {
    Log::Info("Early stopping at iteration %d, the best iteration round is %d",
      iter_, iter_ - early_stopping_round_);
Guolin Ke's avatar
Guolin Ke committed
405
    Log::Info("Output of best iteration round:\n%s", best_msg.c_str());
406
    // pop last early_stopping_round_ models
407
    for (int i = 0; i < early_stopping_round_ * num_class_; ++i) {
408
409
410
411
      models_.pop_back();
    }
  }
  return is_met_early_stopping;
Guolin Ke's avatar
Guolin Ke committed
412
413
}

414
void GBDT::UpdateScore(const Tree* tree, const int curr_class) {
Guolin Ke's avatar
Guolin Ke committed
415
416
417
#ifdef TIMETAG
  auto start_time = std::chrono::steady_clock::now();
#endif
Guolin Ke's avatar
Guolin Ke committed
418
  // update training score
Guolin Ke's avatar
Guolin Ke committed
419
420
421
422
423
  if (!is_use_subset_) {
    train_score_updater_->AddScore(tree_learner_.get(), curr_class);
  } else {
    train_score_updater_->AddScore(tree, curr_class);
  }
Guolin Ke's avatar
Guolin Ke committed
424
425
426
427
428
429
#ifdef TIMETAG
  train_score_time += std::chrono::steady_clock::now() - start_time;
#endif
#ifdef TIMETAG
  start_time = std::chrono::steady_clock::now();
#endif
Guolin Ke's avatar
Guolin Ke committed
430
  // update validation score
Guolin Ke's avatar
Guolin Ke committed
431
432
  for (auto& score_updater : valid_score_updater_) {
    score_updater->AddScore(tree, curr_class);
Guolin Ke's avatar
Guolin Ke committed
433
  }
Guolin Ke's avatar
Guolin Ke committed
434
435
436
#ifdef TIMETAG
  valid_score_time += std::chrono::steady_clock::now() - start_time;
#endif
Guolin Ke's avatar
Guolin Ke committed
437
438
}

Guolin Ke's avatar
Guolin Ke committed
439
440
441
442
std::string GBDT::OutputMetric(int iter) {
  bool need_output = (iter % gbdt_config_->output_freq) == 0;
  std::string ret = "";
  std::stringstream msg_buf;
443
  std::vector<std::pair<size_t, size_t>> meet_early_stopping_pairs;
Guolin Ke's avatar
Guolin Ke committed
444
  // print training metric
Guolin Ke's avatar
Guolin Ke committed
445
  if (need_output) {
446
447
448
    for (auto& sub_metric : training_metrics_) {
      auto name = sub_metric->GetName();
      auto scores = sub_metric->Eval(train_score_updater_->score());
Guolin Ke's avatar
Guolin Ke committed
449
      for (size_t k = 0; k < name.size(); ++k) {
Guolin Ke's avatar
Guolin Ke committed
450
451
452
453
454
455
456
457
        std::stringstream tmp_buf;
        tmp_buf << "Iteration:" << iter
          << ", training " << name[k]
          << " : " << scores[k];
        Log::Info(tmp_buf.str().c_str());
        if (early_stopping_round_ > 0) {
          msg_buf << tmp_buf.str() << std::endl;
        }
458
      }
459
    }
Guolin Ke's avatar
Guolin Ke committed
460
461
  }
  // print validation metric
Guolin Ke's avatar
Guolin Ke committed
462
  if (need_output || early_stopping_round_ > 0) {
463
464
465
    for (size_t i = 0; i < valid_metrics_.size(); ++i) {
      for (size_t j = 0; j < valid_metrics_[i].size(); ++j) {
        auto test_scores = valid_metrics_[i][j]->Eval(valid_score_updater_[i]->score());
Guolin Ke's avatar
Guolin Ke committed
466
467
468
469
470
471
472
473
474
475
476
        auto name = valid_metrics_[i][j]->GetName();
        for (size_t k = 0; k < name.size(); ++k) {
          std::stringstream tmp_buf;
          tmp_buf << "Iteration:" << iter
            << ", valid_" << i + 1 << " " << name[k]
            << " : " << test_scores[k];
          if (need_output) {
            Log::Info(tmp_buf.str().c_str());
          }
          if (early_stopping_round_ > 0) {
            msg_buf << tmp_buf.str() << std::endl;
477
          }
wxchan's avatar
wxchan committed
478
        }
Guolin Ke's avatar
Guolin Ke committed
479
        if (ret.empty() && early_stopping_round_ > 0) {
480
481
482
          auto cur_score = valid_metrics_[i][j]->factor_to_bigger_better() * test_scores.back();
          if (cur_score > best_score_[i][j]) {
            best_score_[i][j] = cur_score;
483
            best_iter_[i][j] = iter;
Guolin Ke's avatar
Guolin Ke committed
484
            meet_early_stopping_pairs.emplace_back(i, j);
485
          } else {
Guolin Ke's avatar
Guolin Ke committed
486
            if (iter - best_iter_[i][j] >= early_stopping_round_) { ret = best_msg_[i][j]; }
487
          }
wxchan's avatar
wxchan committed
488
489
        }
      }
Guolin Ke's avatar
Guolin Ke committed
490
491
    }
  }
Guolin Ke's avatar
Guolin Ke committed
492
493
494
  for (auto& pair : meet_early_stopping_pairs) {
    best_msg_[pair.first][pair.second] = msg_buf.str();
  }
wxchan's avatar
wxchan committed
495
  return ret;
Guolin Ke's avatar
Guolin Ke committed
496
497
}

498
/*! \brief Get eval result */
499
std::vector<double> GBDT::GetEvalAt(int data_idx) const {
Guolin Ke's avatar
Guolin Ke committed
500
  CHECK(data_idx >= 0 && data_idx <= static_cast<int>(valid_score_updater_.size()));
501
502
  std::vector<double> ret;
  if (data_idx == 0) {
503
504
    for (auto& sub_metric : training_metrics_) {
      auto scores = sub_metric->Eval(train_score_updater_->score());
505
506
507
      for (auto score : scores) {
        ret.push_back(score);
      }
508
509
    }
  }
510
511
512
513
514
515
516
  else {
    auto used_idx = data_idx - 1;
    for (size_t j = 0; j < valid_metrics_[used_idx].size(); ++j) {
      auto test_scores = valid_metrics_[used_idx][j]->Eval(valid_score_updater_[used_idx]->score());
      for (auto score : test_scores) {
        ret.push_back(score);
      }
517
518
519
520
521
    }
  }
  return ret;
}

Guolin Ke's avatar
Guolin Ke committed
522
/*! \brief Get training scores result */
523
const double* GBDT::GetTrainingScore(int64_t* out_len) {
524
  *out_len = static_cast<int64_t>(train_score_updater_->num_data()) * num_class_;
Guolin Ke's avatar
Guolin Ke committed
525
  return train_score_updater_->score();
526
527
}

Guolin Ke's avatar
Guolin Ke committed
528
529
void GBDT::GetPredictAt(int data_idx, double* out_result, int64_t* out_len) {
  CHECK(data_idx >= 0 && data_idx <= static_cast<int>(valid_score_updater_.size()));
Guolin Ke's avatar
Guolin Ke committed
530

531
  const double* raw_scores = nullptr;
Guolin Ke's avatar
Guolin Ke committed
532
533
  data_size_t num_data = 0;
  if (data_idx == 0) {
wxchan's avatar
wxchan committed
534
    raw_scores = GetTrainingScore(out_len);
Guolin Ke's avatar
Guolin Ke committed
535
536
537
538
539
    num_data = train_score_updater_->num_data();
  } else {
    auto used_idx = data_idx - 1;
    raw_scores = valid_score_updater_[used_idx]->score();
    num_data = valid_score_updater_[used_idx]->num_data();
540
    *out_len = static_cast<int64_t>(num_data) * num_class_;
Guolin Ke's avatar
Guolin Ke committed
541
542
  }
  if (num_class_ > 1) {
wxchan's avatar
wxchan committed
543
#pragma omp parallel for schedule(static)
Guolin Ke's avatar
Guolin Ke committed
544
    for (data_size_t i = 0; i < num_data; ++i) {
545
      std::vector<double> tmp_result(num_class_);
Guolin Ke's avatar
Guolin Ke committed
546
      for (int j = 0; j < num_class_; ++j) {
547
        tmp_result[j] = raw_scores[j * num_data + i];
Guolin Ke's avatar
Guolin Ke committed
548
549
550
      }
      Common::Softmax(&tmp_result);
      for (int j = 0; j < num_class_; ++j) {
Guolin Ke's avatar
Guolin Ke committed
551
        out_result[j * num_data + i] = static_cast<double>(tmp_result[j]);
Guolin Ke's avatar
Guolin Ke committed
552
553
      }
    }
Guolin Ke's avatar
Guolin Ke committed
554
  } else if(sigmoid_ > 0.0f){
wxchan's avatar
wxchan committed
555
#pragma omp parallel for schedule(static)
Guolin Ke's avatar
Guolin Ke committed
556
    for (data_size_t i = 0; i < num_data; ++i) {
557
      out_result[i] = static_cast<double>(1.0f / (1.0f + std::exp(- sigmoid_ * raw_scores[i])));
Guolin Ke's avatar
Guolin Ke committed
558
559
    }
  } else {
wxchan's avatar
wxchan committed
560
#pragma omp parallel for schedule(static)
Guolin Ke's avatar
Guolin Ke committed
561
    for (data_size_t i = 0; i < num_data; ++i) {
Guolin Ke's avatar
Guolin Ke committed
562
      out_result[i] = static_cast<double>(raw_scores[i]);
Guolin Ke's avatar
Guolin Ke committed
563
564
565
566
567
    }
  }

}

Guolin Ke's avatar
Guolin Ke committed
568
void GBDT::Boosting() {
569
570
571
  if (object_function_ == nullptr) {
    Log::Fatal("No object function provided");
  }
Hui Xue's avatar
Hui Xue committed
572
  // objective function will calculate gradients and hessians
573
  int64_t num_score = 0;
Guolin Ke's avatar
Guolin Ke committed
574
  object_function_->
Guolin Ke's avatar
Guolin Ke committed
575
    GetGradients(GetTrainingScore(&num_score), gradients_.data(), hessians_.data());
Guolin Ke's avatar
Guolin Ke committed
576
577
}

578
std::string GBDT::DumpModel(int num_iteration) const {
Guolin Ke's avatar
Guolin Ke committed
579
  std::stringstream str_buf;
wxchan's avatar
wxchan committed
580

Guolin Ke's avatar
Guolin Ke committed
581
  str_buf << "{";
Guolin Ke's avatar
Guolin Ke committed
582
  str_buf << "\"name\":\"" << SubModelName() << "\"," << std::endl;
Guolin Ke's avatar
Guolin Ke committed
583
584
585
586
  str_buf << "\"num_class\":" << num_class_ << "," << std::endl;
  str_buf << "\"label_index\":" << label_idx_ << "," << std::endl;
  str_buf << "\"max_feature_idx\":" << max_feature_idx_ << "," << std::endl;
  str_buf << "\"sigmoid\":" << sigmoid_ << "," << std::endl;
wxchan's avatar
wxchan committed
587

Guolin Ke's avatar
Guolin Ke committed
588
  str_buf << "\"feature_names\":[\"" 
589
     << Common::Join(feature_names_, "\",\"") << "\"]," 
Guolin Ke's avatar
Guolin Ke committed
590
591
     << std::endl;

Guolin Ke's avatar
Guolin Ke committed
592
  str_buf << "\"tree_info\":[";
593
594
595
596
597
  int num_used_model = static_cast<int>(models_.size());
  if (num_iteration > 0) {
    num_used_model = std::min(num_iteration * num_class_, num_used_model);
  } 
  for (int i = 0; i < num_used_model; ++i) {
wxchan's avatar
wxchan committed
598
    if (i > 0) {
Guolin Ke's avatar
Guolin Ke committed
599
      str_buf << ",";
wxchan's avatar
wxchan committed
600
    }
Guolin Ke's avatar
Guolin Ke committed
601
602
603
604
    str_buf << "{";
    str_buf << "\"tree_index\":" << i << ",";
    str_buf << models_[i]->ToJSON();
    str_buf << "}";
wxchan's avatar
wxchan committed
605
  }
Guolin Ke's avatar
Guolin Ke committed
606
  str_buf << "]" << std::endl;
wxchan's avatar
wxchan committed
607

Guolin Ke's avatar
Guolin Ke committed
608
  str_buf << "}" << std::endl;
wxchan's avatar
wxchan committed
609

Guolin Ke's avatar
Guolin Ke committed
610
  return str_buf.str();
wxchan's avatar
wxchan committed
611
612
}

613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
std::string GBDT::SaveModelToString(int num_iterations) const {
    std::stringstream ss;

    // output model type
    ss << SubModelName() << std::endl;
    // output number of class
    ss << "num_class=" << num_class_ << std::endl;
    // output label index
    ss << "label_index=" << label_idx_ << std::endl;
    // output max_feature_idx
    ss << "max_feature_idx=" << max_feature_idx_ << std::endl;
    // output objective name
    if (object_function_ != nullptr) {
      ss << "objective=" << object_function_->GetName() << std::endl;
    }
    // output sigmoid parameter
    ss << "sigmoid=" << sigmoid_ << std::endl;

    ss << "feature_names=" << Common::Join(feature_names_, " ") << std::endl;

Guolin Ke's avatar
Guolin Ke committed
633
634
    ss << "feature_infos=" << Common::Join(feature_infos_, " ") << std::endl;

635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
    ss << std::endl;
    int num_used_model = static_cast<int>(models_.size());
    if (num_iterations > 0) {
      num_used_model = std::min(num_iterations * num_class_, num_used_model);
    }
    // output tree models
    for (int i = 0; i < num_used_model; ++i) {
      ss << "Tree=" << i << std::endl;
      ss << models_[i]->ToString() << std::endl;
    }

    std::vector<std::pair<size_t, std::string>> pairs = FeatureImportance();
    ss << std::endl << "feature importances:" << std::endl;
    for (size_t i = 0; i < pairs.size(); ++i) {
      ss << pairs[i].second << "=" << std::to_string(pairs[i].first) << std::endl;
    }

    return ss.str();
}

655
bool GBDT::SaveModelToFile(int num_iteration, const char* filename) const {
wxchan's avatar
wxchan committed
656
657
658
  /*! \brief File to write models */
  std::ofstream output_file;
  output_file.open(filename);
659

660
  output_file << SaveModelToString(num_iteration);
661

wxchan's avatar
wxchan committed
662
  output_file.close();
663
664

  return (bool)output_file;
Guolin Ke's avatar
Guolin Ke committed
665
666
}

667
bool GBDT::LoadModelFromString(const std::string& model_str) {
Guolin Ke's avatar
Guolin Ke committed
668
669
670
  // use serialized string to restore this object
  models_.clear();
  std::vector<std::string> lines = Common::Split(model_str.c_str(), '\n');
671
672

  // get number of classes
673
674
675
676
  auto line = Common::FindFromLines(lines, "num_class=");
  if (line.size() > 0) {
    Common::Atoi(Common::Split(line.c_str(), '=')[1].c_str(), &num_class_);
  } else {
677
    Log::Fatal("Model file doesn't specify the number of classes");
678
    return false;
679
  }
Guolin Ke's avatar
Guolin Ke committed
680
  // get index of label
681
682
683
684
  line = Common::FindFromLines(lines, "label_index=");
  if (line.size() > 0) {
    Common::Atoi(Common::Split(line.c_str(), '=')[1].c_str(), &label_idx_);
  } else {
685
    Log::Fatal("Model file doesn't specify the label index");
686
    return false;
Guolin Ke's avatar
Guolin Ke committed
687
  }
Guolin Ke's avatar
Guolin Ke committed
688
  // get max_feature_idx first
689
690
691
692
  line = Common::FindFromLines(lines, "max_feature_idx=");
  if (line.size() > 0) {
    Common::Atoi(Common::Split(line.c_str(), '=')[1].c_str(), &max_feature_idx_);
  } else {
693
    Log::Fatal("Model file doesn't specify max_feature_idx");
694
    return false;
Guolin Ke's avatar
Guolin Ke committed
695
696
  }
  // get sigmoid parameter
697
698
699
700
  line = Common::FindFromLines(lines, "sigmoid=");
  if (line.size() > 0) {
    Common::Atof(Common::Split(line.c_str(), '=')[1].c_str(), &sigmoid_);
  } else {
701
    sigmoid_ = -1.0f;
Guolin Ke's avatar
Guolin Ke committed
702
  }
Guolin Ke's avatar
Guolin Ke committed
703
704
705
  // get feature names
  line = Common::FindFromLines(lines, "feature_names=");
  if (line.size() > 0) {
Guolin Ke's avatar
Guolin Ke committed
706
    feature_names_ = Common::Split(line.substr(std::strlen("feature_names=")).c_str(), " ");
Guolin Ke's avatar
Guolin Ke committed
707
708
    if (feature_names_.size() != static_cast<size_t>(max_feature_idx_ + 1)) {
      Log::Fatal("Wrong size of feature_names");
709
      return false;
Guolin Ke's avatar
Guolin Ke committed
710
    }
Guolin Ke's avatar
Guolin Ke committed
711
712
  }
  else {
Guolin Ke's avatar
Guolin Ke committed
713
    Log::Fatal("Model file doesn't contain feature names");
714
    return false;
Guolin Ke's avatar
Guolin Ke committed
715
716
  }

Guolin Ke's avatar
Guolin Ke committed
717
718
719
720
721
722
723
724
725
726
727
728
  line = Common::FindFromLines(lines, "feature_infos=");
  if (line.size() > 0) {
    feature_infos_ = Common::Split(line.substr(std::strlen("feature_infos=")).c_str(), " ");
    if (feature_infos_.size() != static_cast<size_t>(max_feature_idx_ + 1)) {
      Log::Fatal("Wrong size of feature_infos");
      return false;
    }
  } else {
    Log::Fatal("Model file doesn't contain feature infos");
    return false;
  }

Guolin Ke's avatar
Guolin Ke committed
729
  // get tree models
730
  size_t i = 0;
Guolin Ke's avatar
Guolin Ke committed
731
732
733
734
735
736
737
  while (i < lines.size()) {
    size_t find_pos = lines[i].find("Tree=");
    if (find_pos != std::string::npos) {
      ++i;
      int start = static_cast<int>(i);
      while (i < lines.size() && lines[i].find("Tree=") == std::string::npos) { ++i; }
      int end = static_cast<int>(i);
Guolin Ke's avatar
Guolin Ke committed
738
      std::string tree_str = Common::Join<std::string>(lines, start, end, "\n");
Guolin Ke's avatar
Guolin Ke committed
739
740
      auto new_tree = std::unique_ptr<Tree>(new Tree(tree_str));
      models_.push_back(std::move(new_tree));
Guolin Ke's avatar
Guolin Ke committed
741
742
743
744
    } else {
      ++i;
    }
  }
745
  Log::Info("Finished loading %d models", models_.size());
wxchan's avatar
wxchan committed
746
747
  num_iteration_for_pred_ = static_cast<int>(models_.size()) / num_class_;
  num_init_iteration_ = num_iteration_for_pred_;
748
  iter_ = 0;
749
750

  return true;
Guolin Ke's avatar
Guolin Ke committed
751
752
}

wxchan's avatar
wxchan committed
753
std::vector<std::pair<size_t, std::string>> GBDT::FeatureImportance() const {
754

755
  std::vector<size_t> feature_importances(max_feature_idx_ + 1, 0);
756
    for (size_t iter = 0; iter < models_.size(); ++iter) {
757
        for (int split_idx = 0; split_idx < models_[iter]->num_leaves() - 1; ++split_idx) {
Guolin Ke's avatar
Guolin Ke committed
758
            ++feature_importances[models_[iter]->split_feature(split_idx)];
wxchan's avatar
wxchan committed
759
760
        }
    }
761
762
763
    // store the importance first
    std::vector<std::pair<size_t, std::string>> pairs;
    for (size_t i = 0; i < feature_importances.size(); ++i) {
Guolin Ke's avatar
Guolin Ke committed
764
      if (feature_importances[i] > 0) {
765
        pairs.emplace_back(feature_importances[i], feature_names_[i]);
Guolin Ke's avatar
Guolin Ke committed
766
      }
767
768
769
770
771
    }
    // sort the importance
    std::sort(pairs.begin(), pairs.end(),
      [](const std::pair<size_t, std::string>& lhs,
        const std::pair<size_t, std::string>& rhs) {
772
      return lhs.first > rhs.first;
773
    });
wxchan's avatar
wxchan committed
774
    return pairs;
wxchan's avatar
wxchan committed
775
776
}

777
778
std::vector<double> GBDT::PredictRaw(const double* value) const {
  std::vector<double> ret(num_class_, 0.0f);
wxchan's avatar
wxchan committed
779
  for (int i = 0; i < num_iteration_for_pred_; ++i) {
780
781
782
    for (int j = 0; j < num_class_; ++j) {
      ret[j] += models_[i * num_class_ + j]->Predict(value);
    }
Guolin Ke's avatar
Guolin Ke committed
783
784
785
786
  }
  return ret;
}

787
std::vector<double> GBDT::Predict(const double* value) const {
788
  std::vector<double> ret(num_class_, 0.0f);
wxchan's avatar
wxchan committed
789
  for (int i = 0; i < num_iteration_for_pred_; ++i) {
790
791
    for (int j = 0; j < num_class_; ++j) {
      ret[j] += models_[i * num_class_ + j]->Predict(value);
792
793
    }
  }
794
795
  // if need sigmoid transform
  if (sigmoid_ > 0 && num_class_ == 1) {
796
    ret[0] = 1.0f / (1.0f + std::exp(-sigmoid_ * ret[0]));
797
798
799
  } else if (num_class_ > 1) {
    Common::Softmax(&ret);
  }
800
801
802
  return ret;
}

803
std::vector<int> GBDT::PredictLeafIndex(const double* value) const {
wxchan's avatar
wxchan committed
804
  std::vector<int> ret;
wxchan's avatar
wxchan committed
805
  for (int i = 0; i < num_iteration_for_pred_; ++i) {
806
807
808
    for (int j = 0; j < num_class_; ++j) {
      ret.push_back(models_[i * num_class_ + j]->PredictLeafIndex(value));
    }
wxchan's avatar
wxchan committed
809
810
811
812
  }
  return ret;
}

Guolin Ke's avatar
Guolin Ke committed
813
}  // namespace LightGBM