/
gradient_boosted_trees.cc
2794 lines (2483 loc) · 115 KB
/
gradient_boosted_trees.cc
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
/*
* Copyright 2021 Google LLC.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* https://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
#include "yggdrasil_decision_forests/learner/gradient_boosted_trees/gradient_boosted_trees.h"
#include <algorithm>
#include <cmath>
#include <cstddef>
#include <limits>
#include <memory>
#include <numeric>
#include <optional>
#include <random>
#include <string>
#include <utility>
#include <vector>
#include "absl/container/flat_hash_map.h"
#include "absl/container/flat_hash_set.h"
#include "absl/memory/memory.h"
#include "absl/status/status.h"
#include "absl/strings/str_cat.h"
#include "absl/strings/str_format.h"
#include "absl/strings/string_view.h"
#include "absl/time/clock.h"
#include "absl/time/time.h"
#include "yggdrasil_decision_forests/dataset/data_spec.h"
#include "yggdrasil_decision_forests/dataset/data_spec.pb.h"
#include "yggdrasil_decision_forests/dataset/formats.h"
#include "yggdrasil_decision_forests/dataset/vertical_dataset.h"
#include "yggdrasil_decision_forests/dataset/vertical_dataset_io.h"
#include "yggdrasil_decision_forests/dataset/weight.h"
#include "yggdrasil_decision_forests/learner/abstract_learner.h"
#include "yggdrasil_decision_forests/learner/abstract_learner.pb.h"
#include "yggdrasil_decision_forests/learner/decision_tree/decision_tree.pb.h"
#include "yggdrasil_decision_forests/learner/decision_tree/generic_parameters.h"
#include "yggdrasil_decision_forests/learner/decision_tree/training.h"
#include "yggdrasil_decision_forests/learner/gradient_boosted_trees/gradient_boosted_trees.pb.h"
#include "yggdrasil_decision_forests/learner/gradient_boosted_trees/gradient_boosted_trees_loss.h"
#include "yggdrasil_decision_forests/metric/ranking_ndcg.h"
#include "yggdrasil_decision_forests/model/abstract_model.h"
#include "yggdrasil_decision_forests/model/abstract_model.pb.h"
#include "yggdrasil_decision_forests/model/decision_tree/decision_tree.h"
#include "yggdrasil_decision_forests/model/gradient_boosted_trees/gradient_boosted_trees.h"
#include "yggdrasil_decision_forests/model/gradient_boosted_trees/gradient_boosted_trees.pb.h"
#include "yggdrasil_decision_forests/utils/adaptive_work.h"
#include "yggdrasil_decision_forests/utils/compatibility.h"
#include "yggdrasil_decision_forests/utils/csv.h"
#include "yggdrasil_decision_forests/utils/feature_importance.h"
#include "yggdrasil_decision_forests/utils/filesystem.h"
#include "yggdrasil_decision_forests/utils/hyper_parameters.h"
#include "yggdrasil_decision_forests/utils/logging.h"
#include "yggdrasil_decision_forests/utils/random.h"
#include "yggdrasil_decision_forests/utils/snapshot.h"
#include "yggdrasil_decision_forests/utils/status_macros.h"
#include "yggdrasil_decision_forests/utils/usage.h"
namespace yggdrasil_decision_forests {
namespace model {
namespace gradient_boosted_trees {
using row_t = dataset::VerticalDataset::row_t;
constexpr char GradientBoostedTreesLearner::kRegisteredName[];
// Generic hyper parameter names.
constexpr char GradientBoostedTreesLearner::kHParamNumTrees[];
constexpr char GradientBoostedTreesLearner::kHParamShrinkage[];
constexpr char GradientBoostedTreesLearner::kHParamL1Regularization[];
constexpr char GradientBoostedTreesLearner::kHParamL2Regularization[];
constexpr char
GradientBoostedTreesLearner::kHParamL2CategoricalRegularization[];
constexpr char GradientBoostedTreesLearner::kHParamLambdaLoss[];
constexpr char GradientBoostedTreesLearner::kHParamDartDropOut[];
constexpr char GradientBoostedTreesLearner::
kHParamAdaptSubsampleForMaximumTrainingDuration[];
constexpr char GradientBoostedTreesLearner::kHParamUseHessianGain[];
constexpr char GradientBoostedTreesLearner::kHParamSamplingMethod[];
constexpr char GradientBoostedTreesLearner::kSamplingMethodNone[];
constexpr char GradientBoostedTreesLearner::kSamplingMethodRandom[];
constexpr char GradientBoostedTreesLearner::kSamplingMethodGOSS[];
constexpr char GradientBoostedTreesLearner::kSamplingMethodSelGB[];
constexpr char GradientBoostedTreesLearner::kHParamGossAlpha[];
constexpr char GradientBoostedTreesLearner::kHParamGossBeta[];
constexpr char GradientBoostedTreesLearner::kHParamSelGBRatio[];
constexpr char GradientBoostedTreesLearner::kHParamSubsample[];
constexpr char GradientBoostedTreesLearner::kHParamForestExtraction[];
constexpr char GradientBoostedTreesLearner::kHParamForestExtractionMart[];
constexpr char GradientBoostedTreesLearner::kHParamForestExtractionDart[];
constexpr char GradientBoostedTreesLearner::kHParamValidationSetRatio[];
constexpr char GradientBoostedTreesLearner::kHParamEarlyStopping[];
constexpr char GradientBoostedTreesLearner::kHParamEarlyStoppingNone[];
constexpr char
GradientBoostedTreesLearner::kHParamEarlyStoppingMinLossFullModel[];
constexpr char GradientBoostedTreesLearner::kHParamEarlyStoppingLossIncrease[];
constexpr char
GradientBoostedTreesLearner::kHParamEarlyStoppingNumTreesLookAhead[];
constexpr char GradientBoostedTreesLearner::kHParamApplyLinkFunction[];
constexpr char
GradientBoostedTreesLearner::kHParamComputePermutationVariableImportance[];
using dataset::VerticalDataset;
using CategoricalColumn = VerticalDataset::CategoricalColumn;
constexpr double kAdaptativeWarmUpSeconds = 5.0;
namespace {
// During training, the training dataset is duplicated (shallow copy) and
// modified to train each individual tree. This modified dataset is called the
// "gradient dataset" (because the label is the gradient of the loss).
// Base name of the gradient and hessian column in the gradient dataset.
// Note: The hessian column is only created if necessary (e.g. non-constant
// hessian and use_hessian_gain=true).
constexpr char kBaseGradientColumnName[] = "__gradient__";
constexpr char kBaseHessianColumnName[] = "__hessian__";
// Name of the gradient column in the gradient dataset.
std::string GradientColumnName(const int grad_idx) {
return absl::StrCat(kBaseGradientColumnName, grad_idx);
}
// Name of the hessian column in the gradient dataset.
std::string HessianColumnName(const int grad_idx) {
return absl::StrCat(kBaseHessianColumnName, grad_idx);
}
// Creates the training configurations to "learn the gradients".
// Note: Gradients are the target of the learning.
void ConfigureTrainingConfigForGradients(
const model::proto::TrainingConfig& base_config,
const model::proto::TrainingConfigLinking& base_config_link,
const model::proto::Task gradient_task,
const dataset::VerticalDataset& dataset,
std::vector<GradientData>* gradients) {
for (auto& gradient : *gradients) {
gradient.config = base_config;
gradient.config.set_label(gradient.gradient_column_name);
gradient.config.set_task(gradient_task);
gradient.config_link = base_config_link;
gradient.config_link.set_label(
dataset.ColumnNameToColumnIdx(gradient.gradient_column_name));
}
}
// Computes the loss best adapted to the problem.
utils::StatusOr<proto::Loss> DefaultLoss(
const model::proto::Task task, const dataset::proto::Column& label_spec) {
if (task == model::proto::Task::CLASSIFICATION &&
label_spec.type() == dataset::proto::ColumnType::CATEGORICAL) {
if (label_spec.categorical().number_of_unique_values() == 3) {
// Note: "number_of_unique_values() == 3" because of the reserved
// "out-of-dictionary" item.
return proto::Loss::BINOMIAL_LOG_LIKELIHOOD;
} else if (label_spec.categorical().number_of_unique_values() > 3) {
return proto::Loss::MULTINOMIAL_LOG_LIKELIHOOD;
}
}
if (task == model::proto::Task::REGRESSION &&
label_spec.type() == dataset::proto::ColumnType::NUMERICAL) {
return proto::Loss::SQUARED_ERROR;
}
if (task == model::proto::Task::RANKING &&
label_spec.type() == dataset::proto::ColumnType::NUMERICAL) {
return proto::Loss::LAMBDA_MART_NDCG5;
}
return absl::InvalidArgumentError(
"No defined default loss for this combination of label type and task");
}
// Returns the task used to train the individual decision trees. This task might
// be different from the task that the GBT model is trained to solve.
//
// For example, in case of loss=BINOMIAL_LOG_LIKELIHOOD (which implies a binary
// classification), the trees are "regression trees".
model::proto::Task SubTask(const proto::Loss loss) {
// GBT trees are always (so far) regression trees.
return model::proto::REGRESSION;
}
// Set the default value of non-specified hyper-parameters.
void SetDefaultHyperParameters(model::proto::TrainingConfig* config) {
auto* gbt_config = config->MutableExtension(
gradient_boosted_trees::proto::gradient_boosted_trees_config);
decision_tree::SetDefaultHyperParameters(gbt_config->mutable_decision_tree());
if (gbt_config->has_sample_with_shards()) {
gbt_config->mutable_decision_tree()
->mutable_internal()
->set_sorting_strategy(
decision_tree::proto::DecisionTreeTrainingConfig::Internal::
IN_NODE);
}
if (!gbt_config->decision_tree().has_max_depth()) {
if (gbt_config->decision_tree().has_growing_strategy_best_first_global()) {
gbt_config->mutable_decision_tree()->set_max_depth(-1);
} else {
gbt_config->mutable_decision_tree()->set_max_depth(6);
}
}
if (!gbt_config->decision_tree().has_num_candidate_attributes() &&
!gbt_config->decision_tree().has_num_candidate_attributes_ratio()) {
// The basic definition of GBT does not have any attribute sampling.
gbt_config->mutable_decision_tree()->set_num_candidate_attributes(-1);
}
if (!gbt_config->has_shrinkage()) {
if (gbt_config->forest_extraction_case() ==
proto::GradientBoostedTreesTrainingConfig::kDart) {
gbt_config->set_shrinkage(1.f);
}
}
if (gbt_config->has_use_goss()) {
if (gbt_config->has_gradient_one_side_sampling()) {
LOG(WARNING) << "Ignoring deprecated use_goss, goss_alpha, and goss_beta "
"values because `gradient_one_side_sampling` is already "
"present in "
"the train config.";
} else if ((gbt_config->has_subsample() && gbt_config->subsample() < 1) ||
gbt_config->sampling_methods_case() !=
proto::GradientBoostedTreesTrainingConfig::
SAMPLING_METHODS_NOT_SET) {
LOG(WARNING)
<< "Ignoring deprecated use_goss, goss_alpha, and goss_beta "
"values because another sampling method is already present in the "
"train config.";
} else {
gbt_config->mutable_gradient_one_side_sampling()->set_alpha(
gbt_config->goss_alpha());
gbt_config->mutable_gradient_one_side_sampling()->set_beta(
gbt_config->goss_beta());
}
// Clear deprecated fields.
gbt_config->clear_subsample();
gbt_config->clear_use_goss();
gbt_config->clear_goss_alpha();
gbt_config->clear_goss_beta();
}
if (gbt_config->has_subsample()) {
if (gbt_config->has_stochastic_gradient_boosting()) {
LOG(WARNING)
<< "Ignoring deprecated subsample value because "
"`stochastic_gradient_boosting` is already present in the config.";
} else if (gbt_config->sampling_methods_case() !=
proto::GradientBoostedTreesTrainingConfig::
SAMPLING_METHODS_NOT_SET) {
LOG(WARNING) << "Ignoring deprecated subsample value because another "
"sampling method is already present in the train config.";
} else {
gbt_config->mutable_stochastic_gradient_boosting()->set_ratio(
gbt_config->subsample());
}
// Clear deprecated fields.
gbt_config->clear_subsample();
}
if (gbt_config->early_stopping() !=
proto::GradientBoostedTreesTrainingConfig::NONE &&
gbt_config->validation_set_ratio() == 0) {
LOG(WARNING)
<< "early_stopping != \"NONE\" requires validation_set_ratio>0. "
"Setting early_stopping=\"NONE\" (was \""
<< proto::GradientBoostedTreesTrainingConfig::EarlyStopping_Name(
gbt_config->early_stopping())
<< "\") i.e. sabling early stopping.";
gbt_config->set_early_stopping(
proto::GradientBoostedTreesTrainingConfig::NONE);
}
}
// Splits the training shards between effective training and validation.
absl::Status SplitShards(std::vector<std::string> all,
const int num_validation_shards,
std::vector<std::string>* training,
std::vector<std::string>* validation,
utils::RandomEngine* rnd) {
if (all.size() < num_validation_shards) {
return absl::InternalError("Not enough shards");
}
training->clear();
validation->clear();
if (num_validation_shards == 0) {
// No validation.
*training = all;
return absl::OkStatus();
}
std::shuffle(all.begin(), all.end(), *rnd);
validation->insert(validation->end(), all.begin(),
all.begin() + num_validation_shards);
training->insert(training->end(), all.begin() + num_validation_shards,
all.end());
return absl::OkStatus();
}
// Sample a subset of the candidate shards without replacement.
std::vector<std::string> SampleTrainingShards(
const std::vector<std::string>& candidates, const int num_selected,
utils::RandomEngine* rnd) {
// Note: Could use std::sample in C++17.
std::vector<std::string> selected = candidates;
std::shuffle(selected.begin(), selected.end(), *rnd);
selected.resize(num_selected);
return selected;
}
// Truncate the model (if early stopping is enabled), update the validation loss
// and display the final snippet.
absl::Status FinalizeModelWithValidationDataset(
const internal::AllTrainingConfiguration& config,
const internal::EarlyStopping& early_stopping,
const dataset::VerticalDataset& validation_dataset,
GradientBoostedTreesModel* mdl) {
std::vector<float> final_secondary_metrics;
if (config.gbt_config->early_stopping() ==
proto::GradientBoostedTreesTrainingConfig::
MIN_VALIDATION_LOSS_ON_FULL_MODEL ||
config.gbt_config->early_stopping() ==
proto::GradientBoostedTreesTrainingConfig::VALIDATION_LOSS_INCREASE) {
LOG(INFO) << "Truncates the model to " << early_stopping.best_num_trees()
<< " tree(s) i.e. "
<< early_stopping.best_num_trees() / mdl->num_trees_per_iter()
<< " iteration(s).";
if (early_stopping.best_num_trees() < 0) {
return absl::InvalidArgumentError(
"The model should be evaluated once on the validation dataset.");
}
mdl->set_validation_loss(early_stopping.best_loss());
final_secondary_metrics = early_stopping.best_metrics();
mdl->mutable_decision_trees()->resize(early_stopping.best_num_trees());
} else {
mdl->set_validation_loss(early_stopping.last_loss());
final_secondary_metrics = early_stopping.last_metrics();
}
// Final snippet
std::string snippet;
absl::StrAppendFormat(
&snippet, "Final model num-trees:%d valid-loss:%f",
early_stopping.best_num_trees() / mdl->num_trees_per_iter(),
mdl->validation_loss());
if (!final_secondary_metrics.empty()) {
for (int secondary_metric_idx = 0;
secondary_metric_idx <
mdl->training_logs().secondary_metric_names().size();
secondary_metric_idx++) {
absl::StrAppendFormat(
&snippet, " valid-%s:%f",
mdl->training_logs().secondary_metric_names(secondary_metric_idx),
final_secondary_metrics[secondary_metric_idx]);
}
}
LOG(INFO) << snippet;
if (config.gbt_config->compute_permutation_variable_importance()) {
LOG(INFO) << "Compute permutation variable importances";
RETURN_IF_ERROR(
utils::ComputePermutationFeatureImportance(validation_dataset, mdl));
}
return absl::OkStatus();
}
absl::Status MaybeExportTrainingLogs(const absl::string_view log_directory,
GradientBoostedTreesModel* mdl) {
mdl->mutable_training_logs()->set_number_of_trees_in_final_model(
mdl->NumTrees() / mdl->num_trees_per_iter());
if (!log_directory.empty()) {
RETURN_IF_ERROR(
internal::ExportTrainingLogs(mdl->training_logs(), log_directory));
}
return absl::OkStatus();
}
absl::Status FinalizeModel(const absl::string_view log_directory,
GradientBoostedTreesModel* mdl) {
// Cache the structural variable importance in the model data.
RETURN_IF_ERROR(mdl->PrecomputeVariableImportances(
mdl->AvailableStructuralVariableImportances()));
return MaybeExportTrainingLogs(log_directory, mdl);
}
// Returns a non owning vector of tree pointers from a vector of tree
// unique_ptr.
std::vector<const decision_tree::DecisionTree*> RemoveUniquePtr(
const std::vector<std::unique_ptr<decision_tree::DecisionTree>>& src) {
std::vector<const decision_tree::DecisionTree*> dst;
dst.reserve(src.size());
for (const auto& tree : src) {
dst.push_back(tree.get());
}
return dst;
}
// Builds the internal (i.e. generally not accessible to user) configuration for
// the weak learner.
decision_tree::InternalTrainConfig BuildWeakLearnerInternalConfig(
const internal::AllTrainingConfiguration& config, const int num_threads,
const int grad_idx, const std::vector<GradientData>& gradients,
const std::vector<float>& predictions, const absl::Time& begin_training) {
// Timeout in the tree training.
absl::optional<absl::Time> timeout;
if (config.train_config.has_maximum_training_duration_seconds()) {
timeout =
begin_training +
absl::Seconds(config.train_config.maximum_training_duration_seconds());
}
decision_tree::InternalTrainConfig internal_config;
internal_config.set_leaf_value_functor = config.loss->SetLeafFunctor(
predictions, gradients, config.train_config_link.label());
internal_config.use_hessian_gain = config.gbt_config->use_hessian_gain();
internal_config.hessian_col_idx = gradients[grad_idx].hessian_col_idx;
internal_config.hessian_l1 = config.gbt_config->l1_regularization();
internal_config.hessian_l2_numerical = config.gbt_config->l2_regularization();
internal_config.hessian_l2_categorical =
config.gbt_config->l2_regularization_categorical();
internal_config.num_threads = num_threads;
internal_config.duplicated_selected_examples = false;
internal_config.timeout = timeout;
return internal_config;
}
} // namespace
GradientBoostedTreesLearner::GradientBoostedTreesLearner(
const model::proto::TrainingConfig& training_config)
: AbstractLearner(training_config) {}
absl::Status GradientBoostedTreesLearner::CheckConfiguration(
const dataset::proto::DataSpecification& data_spec,
const model::proto::TrainingConfig& config,
const model::proto::TrainingConfigLinking& config_link,
const proto::GradientBoostedTreesTrainingConfig& gbt_config,
const model::proto::DeploymentConfig& deployment) {
RETURN_IF_ERROR(AbstractLearner::CheckConfiguration(data_spec, config,
config_link, deployment));
if ((gbt_config.has_subsample() && gbt_config.subsample() < 1) &&
gbt_config.sampling_methods_case() !=
gbt_config.SAMPLING_METHODS_NOT_SET) {
LOG(WARNING) << "More than one sampling strategy is present.";
}
if (gbt_config.has_sample_with_shards()) {
if (config.task() == model::proto::RANKING) {
return absl::InvalidArgumentError(
"Ranking is not supported for per-shard sampling. Unset "
"sample_with_shards.");
}
if (gbt_config.has_dart()) {
return absl::InvalidArgumentError(
"Dart is not supported for per-shard sampling. Unset "
"sample_with_shards.");
}
if (gbt_config.adapt_subsample_for_maximum_training_duration()) {
return absl::InvalidArgumentError(
"Adaptive sub-sampling is not supported for per-shard sampling. "
"Unset sample_with_shards.");
}
}
return absl::OkStatus();
}
absl::Status GradientBoostedTreesLearner::BuildAllTrainingConfiguration(
const dataset::proto::DataSpecification& data_spec,
internal::AllTrainingConfiguration* all_config) const {
all_config->train_config = training_config();
SetDefaultHyperParameters(&all_config->train_config);
all_config->gbt_config = &all_config->train_config.GetExtension(
gradient_boosted_trees::proto::gradient_boosted_trees_config);
RETURN_IF_ERROR(AbstractLearner::LinkTrainingConfig(
all_config->train_config, data_spec, &all_config->train_config_link));
RETURN_IF_ERROR(CheckConfiguration(data_spec, all_config->train_config,
all_config->train_config_link,
*all_config->gbt_config, deployment()));
auto* mutable_gbt_config = all_config->train_config.MutableExtension(
gradient_boosted_trees::proto::gradient_boosted_trees_config);
// Select the loss function.
if (mutable_gbt_config->loss() == proto::Loss::DEFAULT) {
ASSIGN_OR_RETURN(
const auto default_loss,
DefaultLoss(all_config->train_config.task(),
data_spec.columns(all_config->train_config_link.label())));
mutable_gbt_config->set_loss(default_loss);
LOG(INFO) << "Default loss set to "
<< proto::Loss_Name(mutable_gbt_config->loss());
}
ASSIGN_OR_RETURN(
all_config->loss,
CreateLoss(all_config->gbt_config->loss(),
all_config->train_config.task(),
data_spec.columns(all_config->train_config_link.label()),
*all_config->gbt_config));
if (all_config->loss->RequireGroupingAttribute()) {
if (!all_config->gbt_config->validation_set_group_feature().empty()) {
return absl::InvalidArgumentError(
"\"validation_set_group_feature\" cannot be specified for "
"a ranking task. Instead, use \"ranking_group\".");
}
all_config->effective_validation_set_group =
all_config->train_config_link.ranking_group();
} else {
if (!all_config->gbt_config->validation_set_group_feature().empty()) {
RETURN_IF_ERROR(dataset::GetSingleColumnIdxFromName(
all_config->gbt_config->validation_set_group_feature(), data_spec,
&all_config->effective_validation_set_group));
}
}
return absl::OkStatus();
}
std::unique_ptr<GradientBoostedTreesModel>
GradientBoostedTreesLearner::InitializeModel(
const internal::AllTrainingConfiguration& config,
const dataset::proto::DataSpecification& data_spec) const {
auto mdl = absl::make_unique<GradientBoostedTreesModel>();
mdl->set_data_spec(data_spec);
internal::InitializeModelWithTrainingConfig(
config.train_config, config.train_config_link, mdl.get());
mdl->set_loss(config.gbt_config->loss());
const auto secondary_metric_names = config.loss->SecondaryMetricNames();
*mdl->training_logs_.mutable_secondary_metric_names() = {
secondary_metric_names.begin(), secondary_metric_names.end()};
if (mdl->task() == model::proto::Task::CLASSIFICATION &&
!config.gbt_config->apply_link_function()) {
// The model output might not be a probability.
mdl->set_classification_outputs_probabilities(false);
}
mdl->set_output_logits(!config.gbt_config->apply_link_function());
return mdl;
}
utils::StatusOr<std::unique_ptr<AbstractModel>>
GradientBoostedTreesLearner::TrainWithStatus(
const absl::string_view typed_path,
const dataset::proto::DataSpecification& data_spec) const {
const auto& gbt_config = training_config().GetExtension(
gradient_boosted_trees::proto::gradient_boosted_trees_config);
if (!gbt_config.has_sample_with_shards()) {
// Regular training.
return AbstractLearner::TrainWithStatus(typed_path, data_spec);
}
return ShardedSamplingTrain(typed_path, data_spec);
}
utils::StatusOr<std::unique_ptr<AbstractModel>>
GradientBoostedTreesLearner::ShardedSamplingTrain(
const absl::string_view typed_path,
const dataset::proto::DataSpecification& data_spec) const {
// The logic of this method is similar to "TrainWithStatus", with the
// exceptions:
// - The loss on the training dataset is computed.
// - Instead of using a dataset loaded in memory, each tree is trained on a
// dataset sampled using shards.
// - No support for the DART algorithm.
// - No support for Ranking.
// TODO(gbm): Splitting method.
const auto begin_training = absl::Now();
// Initialize the configuration.
internal::AllTrainingConfiguration config;
RETURN_IF_ERROR(BuildAllTrainingConfiguration(data_spec, &config));
utils::usage::OnTrainingStart(data_spec, config.train_config,
config.train_config_link,
/*num_examples=*/-1);
// Initialize the model.
auto mdl = InitializeModel(config, data_spec);
utils::RandomEngine random(config.train_config.random_seed());
// Get the dataset shards.
std::string dataset_path, dataset_prefix;
ASSIGN_OR_RETURN(std::tie(dataset_prefix, dataset_path),
dataset::SplitTypeAndPath(typed_path));
std::vector<std::string> all_shards;
RETURN_IF_ERROR(utils::ExpandInputShards(dataset_path, &all_shards));
LOG(INFO) << "Training gradient boosted tree on " << all_shards.size()
<< " shard(s) and " << config.train_config_link.features().size()
<< " feature(s).";
if (all_shards.size() < 10) {
return absl::InvalidArgumentError(absl::Substitute(
"The number of shards in $0 is too small $1<10. For best "
"performances, sampling in the shards should be approximately similar, "
"for the model training, as sampling the examples e.g. >100.",
typed_path, all_shards.size()));
}
// Split the shards between train and validation.
std::vector<std::string> training_shards;
std::vector<std::string> validation_shards;
int num_validation_shards = std::lround(
all_shards.size() * config.gbt_config->validation_set_ratio());
const bool has_validation_dataset = num_validation_shards > 0;
if (config.gbt_config->validation_set_ratio() > 0.f &&
num_validation_shards == 0) {
num_validation_shards = 1;
}
RETURN_IF_ERROR(SplitShards(all_shards, num_validation_shards,
&training_shards, &validation_shards, &random));
// Load and prepare the validation dataset.
std::unique_ptr<internal::CompleteTrainingDatasetForWeakLearner> validation;
if (has_validation_dataset) {
const auto begin_load_validation = absl::Now();
LOG(INFO) << "Loading validation dataset from " << validation_shards.size()
<< " shards";
ASSIGN_OR_RETURN(validation,
internal::LoadCompleteDatasetForWeakLearner(
validation_shards, dataset_prefix, data_spec, config,
/*allocate_gradient=*/false, mdl.get()));
LOG(INFO) << validation->dataset.nrow()
<< " examples loaded in the validation dataset in "
<< (absl::Now() - begin_load_validation);
}
internal::EarlyStopping early_stopping(
config.gbt_config->early_stopping_num_trees_look_ahead());
// Load the first sample of training dataset.
int num_sample_train_shards =
std::lround(training_shards.size() *
config.gbt_config->stochastic_gradient_boosting().ratio());
if (num_sample_train_shards == 0) {
num_sample_train_shards = 1;
}
std::unique_ptr<internal::CompleteTrainingDatasetForWeakLearner>
current_train_dataset, next_train_dataset;
LOG(INFO) << "Loading first training sample dataset from "
<< num_sample_train_shards << " shards";
const auto begin_load_first_sample = absl::Now();
ASSIGN_OR_RETURN(current_train_dataset,
internal::LoadCompleteDatasetForWeakLearner(
SampleTrainingShards(training_shards,
num_sample_train_shards, &random),
dataset_prefix, data_spec, config,
/*allocate_gradient=*/true, mdl.get()));
LOG(INFO) << current_train_dataset->dataset.nrow()
<< " examples loaded in the first training sample in "
<< (absl::Now() - begin_load_first_sample);
// Timer accumulators.
struct {
// Amount of time spent waiting for the preparation thread (IO + parsing +
// preprocess).
absl::Duration sum_duration_wait_prepare;
absl::Mutex mutex_sum_duration;
// Amount of time in the shard loading logic (IO + parsing).
absl::Duration sum_duration_load ABSL_GUARDED_BY(mutex_sum_duration);
// Amount of time in the shard preprocess logic (running the previously
// learned trees).
absl::Duration sum_duration_preprocess ABSL_GUARDED_BY(mutex_sum_duration);
} time_accumulators;
// Fast version of the model. The fast engine is cheaper to run but more
// expensive to construct.
std::unique_ptr<serving::FastEngine> last_engine;
int num_trees_in_last_engine = 0;
// Load a random sample of training data, prepare it for the weak learner
// training, and compute the cached predictions with "trees".
utils::RandomEngine shard_random(random());
absl::Mutex shard_random_mutex;
auto load_and_prepare_next_sample =
[&training_shards, num_sample_train_shards, &shard_random,
&shard_random_mutex, &dataset_prefix, &data_spec, &config, &mdl,
&time_accumulators, &last_engine, &num_trees_in_last_engine](
const std::vector<decision_tree::DecisionTree*>& trees)
-> utils::StatusOr<
std::unique_ptr<internal::CompleteTrainingDatasetForWeakLearner>> {
auto time_begin_load = absl::Now();
std::vector<std::string> selected_shards;
{
absl::MutexLock lock(&shard_random_mutex);
selected_shards = SampleTrainingShards(
training_shards, num_sample_train_shards, &shard_random);
}
ASSIGN_OR_RETURN(auto dataset,
internal::LoadCompleteDatasetForWeakLearner(
selected_shards, dataset_prefix, data_spec, config,
/*allocate_gradient=*/true, mdl.get()));
auto time_begin_predict = absl::Now();
RETURN_IF_ERROR(internal::ComputePredictions(
mdl.get(), last_engine.get(), trees, config, dataset->gradient_dataset,
&dataset->predictions));
dataset->predictions_from_num_trees =
num_trees_in_last_engine + trees.size();
auto time_end_all = absl::Now();
{
absl::MutexLock results_lock(&time_accumulators.mutex_sum_duration);
time_accumulators.sum_duration_load +=
time_begin_predict - time_begin_load;
time_accumulators.sum_duration_preprocess +=
time_end_all - time_begin_predict;
}
return dataset;
};
// List of selected examples. Always contains all the training examples.
std::vector<row_t> selected_examples;
// Thread loading the sample of shard for the next tree.
// Note: The shard loaded in multi-threaded by the vertical dataset IO lib.
std::unique_ptr<utils::concurrency::Thread> thread_load_next_shards;
// Begin time of the training, excluding the model preparation. Used to
// compute the IO bottle neck.
const auto begin_training_loop = absl::Now();
// Gets the fraction of time spend waiting for the loader thread (and not
// training).
const auto get_ratio_waiting_for_loader = [&]() {
const auto denominator =
absl::ToDoubleSeconds(absl::Now() - begin_training_loop);
if (denominator == 0.0) {
return 0.;
}
return absl::ToDoubleSeconds(time_accumulators.sum_duration_wait_prepare) /
denominator;
};
// Amount of time spend in preprocessing in the preparation of the shards.
const auto get_ratio_prepare_in_shard_preparation = [&]() {
absl::MutexLock results_lock(&time_accumulators.mutex_sum_duration);
const auto denominator =
absl::ToDoubleSeconds(time_accumulators.sum_duration_preprocess +
time_accumulators.sum_duration_load);
if (denominator == 0.0) {
return 0.;
}
return absl::ToDoubleSeconds(time_accumulators.sum_duration_preprocess) /
denominator;
};
for (int iter_idx = 0; iter_idx < config.gbt_config->num_trees();
iter_idx++) {
// If true, the sample in "current_train_dataset" will be re-used (instead
// of discarded and replaced by "next_train_dataset").
const bool recycle_current =
(iter_idx %
(1 + config.gbt_config->sample_with_shards().num_recycling())) != 0;
// Same as "recycle_current", but for the next iteration.
const bool recycle_next =
((iter_idx + 1) %
(1 + config.gbt_config->sample_with_shards().num_recycling())) != 0;
if (!recycle_current) {
// Retrieve the set of sharded being loaded.
if (iter_idx > 0) {
// Wait for the loading thread.
const auto begin_wait_loader = absl::Now();
thread_load_next_shards->Join();
time_accumulators.sum_duration_wait_prepare +=
absl::Now() - begin_wait_loader;
thread_load_next_shards = {};
if (!next_train_dataset) {
return absl::InternalError("Missing next sample");
}
// Note: At this point, the pre-computed predictions do not take into
// account the trees added in the last iteration.
// Add the predictions of the trees learned in the last iteration(s).
const int num_redo_iters =
1 + config.gbt_config->sample_with_shards().num_recycling();
DCHECK_EQ(mdl->NumTrees(),
next_train_dataset->predictions_from_num_trees +
num_redo_iters * mdl->num_trees_per_iter());
for (int redo_iter_idx = 0; redo_iter_idx < num_redo_iters;
redo_iter_idx++) {
std::vector<const decision_tree::DecisionTree*> last_trees;
last_trees.reserve(mdl->num_trees_per_iter());
const auto begin_tree_idx =
mdl->NumTrees() -
(num_redo_iters - redo_iter_idx) * mdl->num_trees_per_iter();
for (int tree_idx_in_iter = 0;
tree_idx_in_iter < mdl->num_trees_per_iter();
tree_idx_in_iter++) {
last_trees.push_back(
mdl->decision_trees()[begin_tree_idx + tree_idx_in_iter].get());
}
// Caches the predictions of the trees.
RETURN_IF_ERROR(config.loss->UpdatePredictions(
last_trees, next_train_dataset->gradient_dataset,
&next_train_dataset->predictions,
/*mean_abs_prediction=*/nullptr));
next_train_dataset->predictions_from_num_trees +=
mdl->num_trees_per_iter();
}
current_train_dataset = std::move(next_train_dataset);
}
// Start the loading of the next training sample.
//
// Note: We don't need to do it for the last tree.
if (iter_idx < config.gbt_config->num_trees() - 1) {
// Compile the trees into an engine.
mdl->set_output_logits(true);
auto engine_or = mdl->BuildFastEngine();
mdl->set_output_logits(false);
if (engine_or.ok()) {
last_engine = std::move(engine_or.value());
num_trees_in_last_engine = mdl->NumTrees();
}
// Extract the trees of the current model.
std::vector<decision_tree::DecisionTree*> trees;
for (int tree_idx = num_trees_in_last_engine;
tree_idx < mdl->NumTrees(); tree_idx++) {
trees.push_back(&*mdl->decision_trees()[tree_idx]);
}
thread_load_next_shards = absl::make_unique<utils::concurrency::Thread>(
[&load_and_prepare_next_sample, &next_train_dataset, trees]() {
next_train_dataset = load_and_prepare_next_sample(trees).value();
});
}
}
if (current_train_dataset->dataset.nrow() != selected_examples.size()) {
// Select all the training examples in the sample.
selected_examples.resize(current_train_dataset->dataset.nrow());
std::iota(selected_examples.begin(), selected_examples.end(), 0);
}
if (iter_idx == 0) {
// The first dataset is used to compute the initial predictions (a.k.a
// bias).
mdl->num_trees_per_iter_ = current_train_dataset->gradients.size();
early_stopping.set_trees_per_iterations(mdl->num_trees_per_iter_);
ASSIGN_OR_RETURN(
const auto initial_predictions,
config.loss->InitialPredictions(current_train_dataset->dataset,
config.train_config_link.label(),
current_train_dataset->weights));
mdl->set_initial_predictions(initial_predictions);
internal::SetInitialPredictions(mdl->initial_predictions(),
current_train_dataset->dataset.nrow(),
¤t_train_dataset->predictions);
if (has_validation_dataset) {
internal::SetInitialPredictions(mdl->initial_predictions(),
validation->dataset.nrow(),
&validation->predictions);
}
}
// Compute the gradient.
// Compute the gradient of the residual relative to the examples.
RETURN_IF_ERROR(config.loss->UpdateGradients(
current_train_dataset->gradient_dataset,
config.train_config_link.label(), current_train_dataset->predictions,
nullptr, ¤t_train_dataset->gradients, &random));
// Train a tree on the gradient.
DCHECK_EQ(current_train_dataset->predictions_from_num_trees,
mdl->NumTrees());
std::vector<std::unique_ptr<decision_tree::DecisionTree>> new_trees;
new_trees.reserve(mdl->num_trees_per_iter());
for (int grad_idx = 0; grad_idx < mdl->num_trees_per_iter(); grad_idx++) {
auto tree = absl::make_unique<decision_tree::DecisionTree>();
const auto internal_config = BuildWeakLearnerInternalConfig(
config, deployment().num_threads(), grad_idx,
current_train_dataset->gradients, current_train_dataset->predictions,
begin_training);
RETURN_IF_ERROR(decision_tree::Train(
current_train_dataset->gradient_dataset, selected_examples,
current_train_dataset->gradients[grad_idx].config,
current_train_dataset->gradients[grad_idx].config_link,
config.gbt_config->decision_tree(), deployment(),
current_train_dataset->weights, &random, tree.get(),
internal_config));
new_trees.push_back(std::move(tree));
}
if (has_validation_dataset) {
// Update the predictions on the validation dataset.
RETURN_IF_ERROR(config.loss->UpdatePredictions(
RemoveUniquePtr(new_trees), validation->gradient_dataset,
&validation->predictions,
/*mean_abs_prediction=*/nullptr));
validation->predictions_from_num_trees += new_trees.size();
}
if (recycle_next) {
// Update the predictions on the sample because it will be recycled.
RETURN_IF_ERROR(config.loss->UpdatePredictions(
RemoveUniquePtr(new_trees), current_train_dataset->gradient_dataset,
¤t_train_dataset->predictions,
/*mean_abs_prediction=*/nullptr));
current_train_dataset->predictions_from_num_trees += new_trees.size();
}
// Add the tree to the model.
for (auto& tree : new_trees) {
mdl->AddTree(std::move(tree));
}
// Validation & training logs
if (((iter_idx + 1) % config.gbt_config->validation_interval_in_trees()) ==
0) {
float training_loss;
std::vector<float> train_secondary_metrics;
DCHECK_EQ(validation->predictions_from_num_trees, mdl->NumTrees());
RETURN_IF_ERROR(config.loss->Loss(
current_train_dataset->gradient_dataset,
config.train_config_link.label(), current_train_dataset->predictions,
current_train_dataset->weights, nullptr, &training_loss,
&train_secondary_metrics));
auto* log_entry = mdl->training_logs_.mutable_entries()->Add();
log_entry->set_number_of_trees(iter_idx + 1);
log_entry->set_training_loss(training_loss);
*log_entry->mutable_training_secondary_metrics() = {
train_secondary_metrics.begin(), train_secondary_metrics.end()};
std::string snippet = absl::StrFormat("\tnum-trees:%d train-loss:%f",
iter_idx + 1, training_loss);
for (int secondary_metric_idx = 0;
secondary_metric_idx <
mdl->training_logs_.secondary_metric_names().size();
secondary_metric_idx++) {
absl::StrAppendFormat(
&snippet, " train-%s:%f",
mdl->training_logs_.secondary_metric_names(secondary_metric_idx),
train_secondary_metrics[secondary_metric_idx]);
}
if (has_validation_dataset) {
float validation_loss;
std::vector<float> validation_secondary_metrics;
RETURN_IF_ERROR(config.loss->Loss(
validation->gradient_dataset, config.train_config_link.label(),
validation->predictions, validation->weights, nullptr,
&validation_loss, &validation_secondary_metrics));
log_entry->set_validation_loss(validation_loss);
*log_entry->mutable_validation_secondary_metrics() = {
validation_secondary_metrics.begin(),
validation_secondary_metrics.end()};
absl::StrAppendFormat(&snippet, " valid-loss:%f", validation_loss);
for (int secondary_metric_idx = 0;
secondary_metric_idx <
mdl->training_logs_.secondary_metric_names().size();
secondary_metric_idx++) {
absl::StrAppendFormat(
&snippet, " valid-%s:%f",
mdl->training_logs_.secondary_metric_names(secondary_metric_idx),
validation_secondary_metrics[secondary_metric_idx]);
}