-
Notifications
You must be signed in to change notification settings - Fork 3.9k
/
Copy pathdataset.cpp
1671 lines (1592 loc) · 64.5 KB
/
dataset.cpp
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 (c) 2016 Microsoft Corporation. All rights reserved.
* Licensed under the MIT License. See LICENSE file in the project root for
* license information.
*/
#include <LightGBM/dataset.h>
#include <LightGBM/feature_group.h>
#include <LightGBM/cuda/vector_cudahost.h>
#include <LightGBM/utils/array_args.h>
#include <LightGBM/utils/openmp_wrapper.h>
#include <LightGBM/utils/threading.h>
#include <chrono>
#include <cstdio>
#include <limits>
#include <sstream>
#include <unordered_map>
namespace LightGBM {
const char* Dataset::binary_file_token =
"______LightGBM_Binary_File_Token______\n";
Dataset::Dataset() {
data_filename_ = "noname";
num_data_ = 0;
is_finish_load_ = false;
wait_for_manual_finish_ = false;
has_raw_ = false;
}
Dataset::Dataset(data_size_t num_data) {
CHECK_GT(num_data, 0);
data_filename_ = "noname";
num_data_ = num_data;
metadata_.Init(num_data_, NO_SPECIFIC, NO_SPECIFIC);
is_finish_load_ = false;
wait_for_manual_finish_ = false;
group_bin_boundaries_.push_back(0);
has_raw_ = false;
}
Dataset::~Dataset() {}
std::vector<std::vector<int>> OneFeaturePerGroup(const std::vector<int>& used_features) {
std::vector<std::vector<int>> features_in_group;
features_in_group.resize(used_features.size());
for (size_t i = 0; i < used_features.size(); ++i) {
features_in_group[i].emplace_back(used_features[i]);
}
return features_in_group;
}
int GetConflictCount(const std::vector<bool>& mark, const int* indices,
int num_indices, data_size_t max_cnt) {
int ret = 0;
for (int i = 0; i < num_indices; ++i) {
if (mark[indices[i]]) {
++ret;
}
if (ret > max_cnt) {
return -1;
}
}
return ret;
}
void MarkUsed(std::vector<bool>* mark, const int* indices,
data_size_t num_indices) {
auto& ref_mark = *mark;
for (int i = 0; i < num_indices; ++i) {
ref_mark[indices[i]] = true;
}
}
std::vector<int> FixSampleIndices(const BinMapper* bin_mapper,
int num_total_samples, int num_indices,
const int* sample_indices,
const double* sample_values) {
std::vector<int> ret;
if (bin_mapper->GetDefaultBin() == bin_mapper->GetMostFreqBin()) {
return ret;
}
int i = 0, j = 0;
while (i < num_total_samples) {
if (j < num_indices && sample_indices[j] < i) {
++j;
} else if (j < num_indices && sample_indices[j] == i) {
if (bin_mapper->ValueToBin(sample_values[j]) !=
bin_mapper->GetMostFreqBin()) {
ret.push_back(i);
}
++i;
} else {
ret.push_back(i++);
}
}
return ret;
}
std::vector<std::vector<int>> FindGroups(
const std::vector<std::unique_ptr<BinMapper>>& bin_mappers,
const std::vector<int>& find_order, int** sample_indices,
const int* num_per_col, int num_sample_col, data_size_t total_sample_cnt,
data_size_t num_data, bool is_use_gpu, bool is_sparse,
std::vector<int8_t>* multi_val_group) {
const int max_search_group = 100;
const int max_bin_per_group = 256;
const data_size_t single_val_max_conflict_cnt =
static_cast<data_size_t>(total_sample_cnt / 10000);
multi_val_group->clear();
Random rand(num_data);
std::vector<std::vector<int>> features_in_group;
std::vector<std::vector<bool>> conflict_marks;
std::vector<data_size_t> group_used_row_cnt;
std::vector<data_size_t> group_total_data_cnt;
std::vector<int> group_num_bin;
// first round: fill the single val group
for (auto fidx : find_order) {
bool is_filtered_feature = fidx >= num_sample_col;
const data_size_t cur_non_zero_cnt =
is_filtered_feature ? 0 : num_per_col[fidx];
std::vector<int> available_groups;
for (int gid = 0; gid < static_cast<int>(features_in_group.size()); ++gid) {
auto cur_num_bin = group_num_bin[gid] + bin_mappers[fidx]->num_bin() +
(bin_mappers[fidx]->GetDefaultBin() == 0 ? -1 : 0);
if (group_total_data_cnt[gid] + cur_non_zero_cnt <=
total_sample_cnt + single_val_max_conflict_cnt) {
if (!is_use_gpu || cur_num_bin <= max_bin_per_group) {
available_groups.push_back(gid);
}
}
}
std::vector<int> search_groups;
if (!available_groups.empty()) {
int last = static_cast<int>(available_groups.size()) - 1;
auto indices = rand.Sample(last, std::min(last, max_search_group - 1));
// always push the last group
search_groups.push_back(available_groups.back());
for (auto idx : indices) {
search_groups.push_back(available_groups[idx]);
}
}
int best_gid = -1;
int best_conflict_cnt = -1;
for (auto gid : search_groups) {
const data_size_t rest_max_cnt = single_val_max_conflict_cnt -
group_total_data_cnt[gid] +
group_used_row_cnt[gid];
const data_size_t cnt =
is_filtered_feature
? 0
: GetConflictCount(conflict_marks[gid], sample_indices[fidx],
num_per_col[fidx], rest_max_cnt);
if (cnt >= 0 && cnt <= rest_max_cnt && cnt <= cur_non_zero_cnt / 2) {
best_gid = gid;
best_conflict_cnt = cnt;
break;
}
}
if (best_gid >= 0) {
features_in_group[best_gid].push_back(fidx);
group_total_data_cnt[best_gid] += cur_non_zero_cnt;
group_used_row_cnt[best_gid] += cur_non_zero_cnt - best_conflict_cnt;
if (!is_filtered_feature) {
MarkUsed(&conflict_marks[best_gid], sample_indices[fidx],
num_per_col[fidx]);
}
group_num_bin[best_gid] +=
bin_mappers[fidx]->num_bin() +
(bin_mappers[fidx]->GetDefaultBin() == 0 ? -1 : 0);
} else {
features_in_group.emplace_back();
features_in_group.back().push_back(fidx);
conflict_marks.emplace_back(total_sample_cnt, false);
if (!is_filtered_feature) {
MarkUsed(&(conflict_marks.back()), sample_indices[fidx],
num_per_col[fidx]);
}
group_total_data_cnt.emplace_back(cur_non_zero_cnt);
group_used_row_cnt.emplace_back(cur_non_zero_cnt);
group_num_bin.push_back(
1 + bin_mappers[fidx]->num_bin() +
(bin_mappers[fidx]->GetDefaultBin() == 0 ? -1 : 0));
}
}
if (!is_sparse) {
multi_val_group->resize(features_in_group.size(), false);
return features_in_group;
}
std::vector<int> second_round_features;
std::vector<std::vector<int>> features_in_group2;
std::vector<std::vector<bool>> conflict_marks2;
const double dense_threshold = 0.4;
for (int gid = 0; gid < static_cast<int>(features_in_group.size()); ++gid) {
const double dense_rate =
static_cast<double>(group_used_row_cnt[gid]) / total_sample_cnt;
if (dense_rate >= dense_threshold) {
features_in_group2.push_back(std::move(features_in_group[gid]));
conflict_marks2.push_back(std::move(conflict_marks[gid]));
} else {
for (auto fidx : features_in_group[gid]) {
second_round_features.push_back(fidx);
}
}
}
features_in_group = features_in_group2;
conflict_marks = conflict_marks2;
multi_val_group->resize(features_in_group.size(), false);
if (!second_round_features.empty()) {
features_in_group.emplace_back();
conflict_marks.emplace_back(total_sample_cnt, false);
bool is_multi_val = is_use_gpu ? true : false;
int conflict_cnt = 0;
for (auto fidx : second_round_features) {
features_in_group.back().push_back(fidx);
if (!is_multi_val) {
const int rest_max_cnt = single_val_max_conflict_cnt - conflict_cnt;
const auto cnt =
GetConflictCount(conflict_marks.back(), sample_indices[fidx],
num_per_col[fidx], rest_max_cnt);
conflict_cnt += cnt;
if (cnt < 0 || conflict_cnt > single_val_max_conflict_cnt) {
is_multi_val = true;
continue;
}
MarkUsed(&(conflict_marks.back()), sample_indices[fidx],
num_per_col[fidx]);
}
}
multi_val_group->push_back(is_multi_val);
}
return features_in_group;
}
std::vector<std::vector<int>> FastFeatureBundling(
const std::vector<std::unique_ptr<BinMapper>>& bin_mappers,
int** sample_indices, double** sample_values, const int* num_per_col,
int num_sample_col, data_size_t total_sample_cnt,
const std::vector<int>& used_features, data_size_t num_data,
bool is_use_gpu, bool is_sparse, std::vector<int8_t>* multi_val_group) {
Common::FunctionTimer fun_timer("Dataset::FastFeatureBundling", global_timer);
std::vector<size_t> feature_non_zero_cnt;
feature_non_zero_cnt.reserve(used_features.size());
// put dense feature first
for (auto fidx : used_features) {
if (fidx < num_sample_col) {
feature_non_zero_cnt.emplace_back(num_per_col[fidx]);
} else {
feature_non_zero_cnt.emplace_back(0);
}
}
// sort by non zero cnt
std::vector<int> sorted_idx;
sorted_idx.reserve(used_features.size());
for (int i = 0; i < static_cast<int>(used_features.size()); ++i) {
sorted_idx.emplace_back(i);
}
// sort by non zero cnt, bigger first
std::stable_sort(sorted_idx.begin(), sorted_idx.end(),
[&feature_non_zero_cnt](int a, int b) {
return feature_non_zero_cnt[a] > feature_non_zero_cnt[b];
});
std::vector<int> feature_order_by_cnt;
feature_order_by_cnt.reserve(sorted_idx.size());
for (auto sidx : sorted_idx) {
feature_order_by_cnt.push_back(used_features[sidx]);
}
std::vector<std::vector<int>> tmp_indices;
std::vector<int> tmp_num_per_col(num_sample_col, 0);
for (auto fidx : used_features) {
if (fidx >= num_sample_col) {
continue;
}
auto ret = FixSampleIndices(
bin_mappers[fidx].get(), static_cast<int>(total_sample_cnt),
num_per_col[fidx], sample_indices[fidx], sample_values[fidx]);
if (!ret.empty()) {
tmp_indices.push_back(ret);
tmp_num_per_col[fidx] = static_cast<int>(ret.size());
sample_indices[fidx] = tmp_indices.back().data();
} else {
tmp_num_per_col[fidx] = num_per_col[fidx];
}
}
std::vector<int8_t> group_is_multi_val, group_is_multi_val2;
auto features_in_group =
FindGroups(bin_mappers, used_features, sample_indices,
tmp_num_per_col.data(), num_sample_col, total_sample_cnt,
num_data, is_use_gpu, is_sparse, &group_is_multi_val);
auto group2 =
FindGroups(bin_mappers, feature_order_by_cnt, sample_indices,
tmp_num_per_col.data(), num_sample_col, total_sample_cnt,
num_data, is_use_gpu, is_sparse, &group_is_multi_val2);
if (features_in_group.size() > group2.size()) {
features_in_group = group2;
group_is_multi_val = group_is_multi_val2;
}
// shuffle groups
int num_group = static_cast<int>(features_in_group.size());
Random tmp_rand(num_data);
for (int i = 0; i < num_group - 1; ++i) {
int j = tmp_rand.NextShort(i + 1, num_group);
std::swap(features_in_group[i], features_in_group[j]);
// Using std::swap for vector<bool> will cause the wrong result.
std::swap(group_is_multi_val[i], group_is_multi_val[j]);
}
*multi_val_group = group_is_multi_val;
return features_in_group;
}
void Dataset::Construct(std::vector<std::unique_ptr<BinMapper>>* bin_mappers,
int num_total_features,
const std::vector<std::vector<double>>& forced_bins,
int** sample_non_zero_indices,
double** sample_values,
const int* num_per_col,
int num_sample_col,
size_t total_sample_cnt,
const Config& io_config) {
num_total_features_ = num_total_features;
CHECK_EQ(num_total_features_, static_cast<int>(bin_mappers->size()));
// get num_features
std::vector<int> used_features;
auto& ref_bin_mappers = *bin_mappers;
for (int i = 0; i < static_cast<int>(bin_mappers->size()); ++i) {
if (ref_bin_mappers[i] != nullptr && !ref_bin_mappers[i]->is_trivial()) {
used_features.emplace_back(i);
}
}
if (used_features.empty()) {
Log::Warning(
"There are no meaningful features which satisfy the provided configuration. "
"Decreasing Dataset parameters min_data_in_bin or min_data_in_leaf and re-constructing "
"Dataset might resolve this warning.");
}
auto features_in_group = OneFeaturePerGroup(used_features);
auto is_sparse = io_config.is_enable_sparse;
if (io_config.device_type == std::string("cuda") || io_config.device_type == std::string("cuda_exp")) {
LGBM_config_::current_device = lgbm_device_cuda;
if ((io_config.device_type == std::string("cuda") || io_config.device_type == std::string("cuda_exp")) && is_sparse) {
Log::Warning("Using sparse features with CUDA is currently not supported.");
is_sparse = false;
}
}
std::vector<int8_t> group_is_multi_val(used_features.size(), 0);
if (io_config.enable_bundle && !used_features.empty()) {
bool lgbm_is_gpu_used = io_config.device_type == std::string("gpu") || io_config.device_type == std::string("cuda")
|| io_config.device_type == std::string("cuda_exp");
features_in_group = FastFeatureBundling(
*bin_mappers, sample_non_zero_indices, sample_values, num_per_col,
num_sample_col, static_cast<data_size_t>(total_sample_cnt),
used_features, num_data_, lgbm_is_gpu_used,
is_sparse, &group_is_multi_val);
}
num_features_ = 0;
for (const auto& fs : features_in_group) {
num_features_ += static_cast<int>(fs.size());
}
int cur_fidx = 0;
used_feature_map_ = std::vector<int>(num_total_features_, -1);
num_groups_ = static_cast<int>(features_in_group.size());
real_feature_idx_.resize(num_features_);
feature2group_.resize(num_features_);
feature2subfeature_.resize(num_features_);
feature_need_push_zeros_.clear();
group_bin_boundaries_.clear();
uint64_t num_total_bin = 0;
group_bin_boundaries_.push_back(num_total_bin);
group_feature_start_.resize(num_groups_);
group_feature_cnt_.resize(num_groups_);
for (int i = 0; i < num_groups_; ++i) {
auto cur_features = features_in_group[i];
int cur_cnt_features = static_cast<int>(cur_features.size());
group_feature_start_[i] = cur_fidx;
group_feature_cnt_[i] = cur_cnt_features;
// get bin_mappers
std::vector<std::unique_ptr<BinMapper>> cur_bin_mappers;
for (int j = 0; j < cur_cnt_features; ++j) {
int real_fidx = cur_features[j];
used_feature_map_[real_fidx] = cur_fidx;
real_feature_idx_[cur_fidx] = real_fidx;
feature2group_[cur_fidx] = i;
feature2subfeature_[cur_fidx] = j;
cur_bin_mappers.emplace_back(ref_bin_mappers[real_fidx].release());
if (cur_bin_mappers.back()->GetDefaultBin() !=
cur_bin_mappers.back()->GetMostFreqBin()) {
feature_need_push_zeros_.push_back(cur_fidx);
}
++cur_fidx;
}
feature_groups_.emplace_back(std::unique_ptr<FeatureGroup>(
new FeatureGroup(cur_cnt_features, group_is_multi_val[i], &cur_bin_mappers, num_data_, i)));
num_total_bin += feature_groups_[i]->num_total_bin_;
group_bin_boundaries_.push_back(num_total_bin);
}
if (!io_config.max_bin_by_feature.empty()) {
CHECK_EQ(static_cast<size_t>(num_total_features_),
io_config.max_bin_by_feature.size());
CHECK_GT(*(std::min_element(io_config.max_bin_by_feature.begin(),
io_config.max_bin_by_feature.end())), 1);
max_bin_by_feature_.resize(num_total_features_);
max_bin_by_feature_.assign(io_config.max_bin_by_feature.begin(),
io_config.max_bin_by_feature.end());
}
forced_bin_bounds_ = forced_bins;
max_bin_ = io_config.max_bin;
min_data_in_bin_ = io_config.min_data_in_bin;
bin_construct_sample_cnt_ = io_config.bin_construct_sample_cnt;
use_missing_ = io_config.use_missing;
zero_as_missing_ = io_config.zero_as_missing;
has_raw_ = false;
if (io_config.linear_tree) {
has_raw_ = true;
}
numeric_feature_map_ = std::vector<int>(num_features_, -1);
num_numeric_features_ = 0;
for (int i = 0; i < num_features_; ++i) {
if (FeatureBinMapper(i)->bin_type() == BinType::NumericalBin) {
numeric_feature_map_[i] = num_numeric_features_;
++num_numeric_features_;
}
}
device_type_ = io_config.device_type;
gpu_device_id_ = io_config.gpu_device_id;
}
void Dataset::FinishLoad() {
if (is_finish_load_) {
return;
}
if (num_groups_ > 0) {
for (int i = 0; i < num_groups_; ++i) {
feature_groups_[i]->FinishLoad();
}
}
metadata_.FinishLoad();
#ifdef USE_CUDA_EXP
if (device_type_ == std::string("cuda_exp")) {
CreateCUDAColumnData();
metadata_.CreateCUDAMetadata(gpu_device_id_);
} else {
cuda_column_data_.reset(nullptr);
}
#endif // USE_CUDA_EXP
is_finish_load_ = true;
}
void PushDataToMultiValBin(
data_size_t num_data, const std::vector<uint32_t> most_freq_bins,
const std::vector<uint32_t> offsets,
std::vector<std::vector<std::unique_ptr<BinIterator>>>* iters,
MultiValBin* ret) {
Common::FunctionTimer fun_time("Dataset::PushDataToMultiValBin",
global_timer);
if (ret->IsSparse()) {
Threading::For<data_size_t>(
0, num_data, 1024, [&](int tid, data_size_t start, data_size_t end) {
std::vector<uint32_t> cur_data;
cur_data.reserve(most_freq_bins.size());
for (size_t j = 0; j < most_freq_bins.size(); ++j) {
(*iters)[tid][j]->Reset(start);
}
for (data_size_t i = start; i < end; ++i) {
cur_data.clear();
for (size_t j = 0; j < most_freq_bins.size(); ++j) {
// for sparse multi value bin, we store the feature bin values with offset added
auto cur_bin = (*iters)[tid][j]->Get(i);
if (cur_bin == most_freq_bins[j]) {
continue;
}
cur_bin += offsets[j];
if (most_freq_bins[j] == 0) {
cur_bin -= 1;
}
cur_data.push_back(cur_bin);
}
ret->PushOneRow(tid, i, cur_data);
}
});
} else {
Threading::For<data_size_t>(
0, num_data, 1024, [&](int tid, data_size_t start, data_size_t end) {
std::vector<uint32_t> cur_data(most_freq_bins.size(), 0);
for (size_t j = 0; j < most_freq_bins.size(); ++j) {
(*iters)[tid][j]->Reset(start);
}
for (data_size_t i = start; i < end; ++i) {
for (size_t j = 0; j < most_freq_bins.size(); ++j) {
// for dense multi value bin, the feature bin values without offsets are used
auto cur_bin = (*iters)[tid][j]->Get(i);
cur_data[j] = cur_bin;
}
ret->PushOneRow(tid, i, cur_data);
}
});
}
}
MultiValBin* Dataset::GetMultiBinFromSparseFeatures(const std::vector<uint32_t>& offsets) const {
Common::FunctionTimer fun_time("Dataset::GetMultiBinFromSparseFeatures",
global_timer);
int multi_group_id = -1;
for (int i = 0; i < num_groups_; ++i) {
if (feature_groups_[i]->is_multi_val_) {
if (multi_group_id < 0) {
multi_group_id = i;
} else {
Log::Fatal("Bug. There should be only one multi-val group.");
}
}
}
if (multi_group_id < 0) {
return nullptr;
}
const int num_feature = feature_groups_[multi_group_id]->num_feature_;
int num_threads = OMP_NUM_THREADS();
std::vector<std::vector<std::unique_ptr<BinIterator>>> iters(num_threads);
std::vector<uint32_t> most_freq_bins;
double sum_sparse_rate = 0;
for (int i = 0; i < num_feature; ++i) {
#pragma omp parallel for schedule(static, 1)
for (int tid = 0; tid < num_threads; ++tid) {
iters[tid].emplace_back(
feature_groups_[multi_group_id]->SubFeatureIterator(i));
}
most_freq_bins.push_back(
feature_groups_[multi_group_id]->bin_mappers_[i]->GetMostFreqBin());
sum_sparse_rate +=
feature_groups_[multi_group_id]->bin_mappers_[i]->sparse_rate();
}
sum_sparse_rate /= num_feature;
Log::Debug("Dataset::GetMultiBinFromSparseFeatures: sparse rate %f",
sum_sparse_rate);
std::unique_ptr<MultiValBin> ret;
ret.reset(MultiValBin::CreateMultiValBin(num_data_, offsets.back(),
num_feature, sum_sparse_rate, offsets));
PushDataToMultiValBin(num_data_, most_freq_bins, offsets, &iters, ret.get());
ret->FinishLoad();
return ret.release();
}
MultiValBin* Dataset::GetMultiBinFromAllFeatures(const std::vector<uint32_t>& offsets) const {
Common::FunctionTimer fun_time("Dataset::GetMultiBinFromAllFeatures",
global_timer);
int num_threads = OMP_NUM_THREADS();
double sum_dense_ratio = 0;
std::unique_ptr<MultiValBin> ret;
std::vector<std::vector<std::unique_ptr<BinIterator>>> iters(num_threads);
std::vector<uint32_t> most_freq_bins;
int ncol = 0;
for (int gid = 0; gid < num_groups_; ++gid) {
if (feature_groups_[gid]->is_multi_val_) {
ncol += feature_groups_[gid]->num_feature_;
} else {
++ncol;
}
for (int fid = 0; fid < feature_groups_[gid]->num_feature_; ++fid) {
const auto& bin_mapper = feature_groups_[gid]->bin_mappers_[fid];
sum_dense_ratio += 1.0f - bin_mapper->sparse_rate();
}
}
sum_dense_ratio /= ncol;
for (int gid = 0; gid < num_groups_; ++gid) {
if (feature_groups_[gid]->is_multi_val_) {
for (int fid = 0; fid < feature_groups_[gid]->num_feature_; ++fid) {
const auto& bin_mapper = feature_groups_[gid]->bin_mappers_[fid];
most_freq_bins.push_back(bin_mapper->GetMostFreqBin());
#pragma omp parallel for schedule(static, 1)
for (int tid = 0; tid < num_threads; ++tid) {
iters[tid].emplace_back(
feature_groups_[gid]->SubFeatureIterator(fid));
}
}
} else {
most_freq_bins.push_back(0);
for (int tid = 0; tid < num_threads; ++tid) {
iters[tid].emplace_back(feature_groups_[gid]->FeatureGroupIterator());
}
}
}
CHECK(static_cast<int>(most_freq_bins.size()) == ncol);
Log::Debug("Dataset::GetMultiBinFromAllFeatures: sparse rate %f",
1.0 - sum_dense_ratio);
ret.reset(MultiValBin::CreateMultiValBin(
num_data_, offsets.back(), static_cast<int>(most_freq_bins.size()),
1.0 - sum_dense_ratio, offsets));
PushDataToMultiValBin(num_data_, most_freq_bins, offsets, &iters, ret.get());
ret->FinishLoad();
return ret.release();
}
TrainingShareStates* Dataset::GetShareStates(
score_t* gradients, score_t* hessians,
const std::vector<int8_t>& is_feature_used, bool is_constant_hessian,
bool force_col_wise, bool force_row_wise) const {
Common::FunctionTimer fun_timer("Dataset::TestMultiThreadingMethod",
global_timer);
if (force_col_wise && force_row_wise) {
Log::Fatal(
"Cannot set both of `force_col_wise` and `force_row_wise` to `true` at "
"the same time");
}
if (num_groups_ <= 0) {
TrainingShareStates* share_state = new TrainingShareStates();
share_state->is_col_wise = true;
share_state->is_constant_hessian = is_constant_hessian;
return share_state;
}
if (force_col_wise) {
TrainingShareStates* share_state = new TrainingShareStates();
std::vector<uint32_t> offsets;
share_state->CalcBinOffsets(
feature_groups_, &offsets, true);
share_state->SetMultiValBin(GetMultiBinFromSparseFeatures(offsets),
num_data_, feature_groups_, false, true);
share_state->is_col_wise = true;
share_state->is_constant_hessian = is_constant_hessian;
return share_state;
} else if (force_row_wise) {
TrainingShareStates* share_state = new TrainingShareStates();
std::vector<uint32_t> offsets;
share_state->CalcBinOffsets(
feature_groups_, &offsets, false);
share_state->SetMultiValBin(GetMultiBinFromAllFeatures(offsets), num_data_,
feature_groups_, false, false);
share_state->is_col_wise = false;
share_state->is_constant_hessian = is_constant_hessian;
return share_state;
} else {
std::unique_ptr<MultiValBin> sparse_bin;
std::unique_ptr<MultiValBin> all_bin;
std::unique_ptr<TrainingShareStates> col_wise_state;
std::unique_ptr<TrainingShareStates> row_wise_state;
col_wise_state.reset(new TrainingShareStates());
row_wise_state.reset(new TrainingShareStates());
std::chrono::duration<double, std::milli> col_wise_init_time, row_wise_init_time;
auto start_time = std::chrono::steady_clock::now();
std::vector<uint32_t> col_wise_offsets;
col_wise_state->CalcBinOffsets(feature_groups_, &col_wise_offsets, true);
col_wise_state->SetMultiValBin(GetMultiBinFromSparseFeatures(col_wise_offsets), num_data_,
feature_groups_, false, true);
col_wise_init_time = std::chrono::steady_clock::now() - start_time;
start_time = std::chrono::steady_clock::now();
std::vector<uint32_t> row_wise_offsets;
row_wise_state->CalcBinOffsets(feature_groups_, &row_wise_offsets, false);
row_wise_state->SetMultiValBin(GetMultiBinFromAllFeatures(row_wise_offsets), num_data_,
feature_groups_, false, false);
row_wise_init_time = std::chrono::steady_clock::now() - start_time;
uint64_t max_total_bin = std::max<uint64_t>(row_wise_state->num_hist_total_bin(),
col_wise_state->num_hist_total_bin());
std::vector<hist_t, Common::AlignmentAllocator<hist_t, kAlignedSize>>
hist_data(max_total_bin * 2);
Log::Debug(
"init for col-wise cost %f seconds, init for row-wise cost %f seconds",
col_wise_init_time * 1e-3, row_wise_init_time * 1e-3);
col_wise_state->is_col_wise = true;
col_wise_state->is_constant_hessian = is_constant_hessian;
InitTrain(is_feature_used, col_wise_state.get());
row_wise_state->is_col_wise = false;
row_wise_state->is_constant_hessian = is_constant_hessian;
InitTrain(is_feature_used, row_wise_state.get());
std::chrono::duration<double, std::milli> col_wise_time, row_wise_time;
start_time = std::chrono::steady_clock::now();
ConstructHistograms(is_feature_used, nullptr, num_data_, gradients,
hessians, gradients, hessians, col_wise_state.get(),
hist_data.data());
col_wise_time = std::chrono::steady_clock::now() - start_time;
start_time = std::chrono::steady_clock::now();
ConstructHistograms(is_feature_used, nullptr, num_data_, gradients,
hessians, gradients, hessians, row_wise_state.get(),
hist_data.data());
row_wise_time = std::chrono::steady_clock::now() - start_time;
if (col_wise_time < row_wise_time) {
auto overhead_cost = row_wise_init_time + row_wise_time + col_wise_time;
Log::Warning(
"Auto-choosing col-wise multi-threading, the overhead of testing was "
"%f seconds.\n"
"You can set `force_col_wise=true` to remove the overhead.",
overhead_cost * 1e-3);
return col_wise_state.release();
} else {
auto overhead_cost = col_wise_init_time + row_wise_time + col_wise_time;
Log::Warning(
"Auto-choosing row-wise multi-threading, the overhead of testing was "
"%f seconds.\n"
"You can set `force_row_wise=true` to remove the overhead.\n"
"And if memory is not enough, you can set `force_col_wise=true`.",
overhead_cost * 1e-3);
if (row_wise_state->IsSparseRowwise()) {
Log::Debug("Using Sparse Multi-Val Bin");
} else {
Log::Debug("Using Dense Multi-Val Bin");
}
return row_wise_state.release();
}
}
}
void Dataset::CopyFeatureMapperFrom(const Dataset* dataset) {
feature_groups_.clear();
num_features_ = dataset->num_features_;
num_groups_ = dataset->num_groups_;
has_raw_ = dataset->has_raw();
// copy feature bin mapper data
for (int i = 0; i < num_groups_; ++i) {
feature_groups_.emplace_back(
new FeatureGroup(*dataset->feature_groups_[i], num_data_));
}
feature_groups_.shrink_to_fit();
used_feature_map_ = dataset->used_feature_map_;
num_total_features_ = dataset->num_total_features_;
feature_names_ = dataset->feature_names_;
label_idx_ = dataset->label_idx_;
real_feature_idx_ = dataset->real_feature_idx_;
feature2group_ = dataset->feature2group_;
feature2subfeature_ = dataset->feature2subfeature_;
group_bin_boundaries_ = dataset->group_bin_boundaries_;
group_feature_start_ = dataset->group_feature_start_;
group_feature_cnt_ = dataset->group_feature_cnt_;
forced_bin_bounds_ = dataset->forced_bin_bounds_;
feature_need_push_zeros_ = dataset->feature_need_push_zeros_;
}
void Dataset::CreateValid(const Dataset* dataset) {
feature_groups_.clear();
num_features_ = dataset->num_features_;
num_groups_ = num_features_;
max_bin_ = dataset->max_bin_;
min_data_in_bin_ = dataset->min_data_in_bin_;
bin_construct_sample_cnt_ = dataset->bin_construct_sample_cnt_;
use_missing_ = dataset->use_missing_;
zero_as_missing_ = dataset->zero_as_missing_;
feature2group_.clear();
feature2subfeature_.clear();
has_raw_ = dataset->has_raw();
numeric_feature_map_ = dataset->numeric_feature_map_;
num_numeric_features_ = dataset->num_numeric_features_;
// copy feature bin mapper data
feature_need_push_zeros_.clear();
group_bin_boundaries_.clear();
uint64_t num_total_bin = 0;
group_bin_boundaries_.push_back(num_total_bin);
group_feature_start_.resize(num_groups_);
group_feature_cnt_.resize(num_groups_);
for (int i = 0; i < num_features_; ++i) {
std::vector<std::unique_ptr<BinMapper>> bin_mappers;
bin_mappers.emplace_back(new BinMapper(*(dataset->FeatureBinMapper(i))));
if (bin_mappers.back()->GetDefaultBin() !=
bin_mappers.back()->GetMostFreqBin()) {
feature_need_push_zeros_.push_back(i);
}
feature_groups_.emplace_back(new FeatureGroup(&bin_mappers, num_data_));
feature2group_.push_back(i);
feature2subfeature_.push_back(0);
num_total_bin += feature_groups_[i]->num_total_bin_;
group_bin_boundaries_.push_back(num_total_bin);
group_feature_start_[i] = i;
group_feature_cnt_[i] = 1;
}
feature_groups_.shrink_to_fit();
used_feature_map_ = dataset->used_feature_map_;
num_total_features_ = dataset->num_total_features_;
feature_names_ = dataset->feature_names_;
label_idx_ = dataset->label_idx_;
real_feature_idx_ = dataset->real_feature_idx_;
forced_bin_bounds_ = dataset->forced_bin_bounds_;
device_type_ = dataset->device_type_;
gpu_device_id_ = dataset->gpu_device_id_;
}
void Dataset::ReSize(data_size_t num_data) {
if (num_data_ != num_data) {
num_data_ = num_data;
OMP_INIT_EX();
#pragma omp parallel for schedule(static)
for (int group = 0; group < num_groups_; ++group) {
OMP_LOOP_EX_BEGIN();
feature_groups_[group]->ReSize(num_data_);
OMP_LOOP_EX_END();
}
OMP_THROW_EX();
}
}
void Dataset::CopySubrow(const Dataset* fullset,
const data_size_t* used_indices,
data_size_t num_used_indices, bool need_meta_data) {
CHECK_EQ(num_used_indices, num_data_);
std::vector<int> group_ids, subfeature_ids;
group_ids.reserve(num_features_);
subfeature_ids.reserve(num_features_);
for (int group = 0; group < num_groups_; ++group) {
if (fullset->IsMultiGroup(group)) {
for (int sub_feature = 0; sub_feature <
fullset->feature_groups_[group]->num_feature_; ++sub_feature) {
group_ids.emplace_back(group);
subfeature_ids.emplace_back(sub_feature);
}
} else {
group_ids.emplace_back(group);
subfeature_ids.emplace_back(-1);
}
}
int num_copy_tasks = static_cast<int>(group_ids.size());
OMP_INIT_EX();
#pragma omp parallel for schedule(dynamic)
for (int task_id = 0; task_id < num_copy_tasks; ++task_id) {
OMP_LOOP_EX_BEGIN();
int group = group_ids[task_id];
int subfeature = subfeature_ids[task_id];
feature_groups_[group]->CopySubrowByCol(fullset->feature_groups_[group].get(),
used_indices, num_used_indices, subfeature);
OMP_LOOP_EX_END();
}
OMP_THROW_EX();
if (need_meta_data) {
metadata_.Init(fullset->metadata_, used_indices, num_used_indices);
}
is_finish_load_ = true;
numeric_feature_map_ = fullset->numeric_feature_map_;
num_numeric_features_ = fullset->num_numeric_features_;
if (has_raw_) {
ResizeRaw(num_used_indices);
#pragma omp parallel for schedule(static)
for (int i = 0; i < num_used_indices; ++i) {
for (int j = 0; j < num_numeric_features_; ++j) {
raw_data_[j][i] = fullset->raw_data_[j][used_indices[i]];
}
}
}
// update CUDA storage for column data and metadata
device_type_ = fullset->device_type_;
gpu_device_id_ = fullset->gpu_device_id_;
#ifdef USE_CUDA_EXP
if (device_type_ == std::string("cuda_exp")) {
if (cuda_column_data_ == nullptr) {
cuda_column_data_.reset(new CUDAColumnData(fullset->num_data(), gpu_device_id_));
metadata_.CreateCUDAMetadata(gpu_device_id_);
}
cuda_column_data_->CopySubrow(fullset->cuda_column_data(), used_indices, num_used_indices);
}
#endif // USE_CUDA_EXP
}
bool Dataset::SetFloatField(const char* field_name, const float* field_data,
data_size_t num_element) {
std::string name(field_name);
name = Common::Trim(name);
if (name == std::string("label") || name == std::string("target")) {
#ifdef LABEL_T_USE_DOUBLE
Log::Fatal("Don't support LABEL_T_USE_DOUBLE");
#else
metadata_.SetLabel(field_data, num_element);
#endif
} else if (name == std::string("weight") || name == std::string("weights")) {
#ifdef LABEL_T_USE_DOUBLE
Log::Fatal("Don't support LABEL_T_USE_DOUBLE");
#else
metadata_.SetWeights(field_data, num_element);
#endif
} else {
return false;
}
return true;
}
bool Dataset::SetDoubleField(const char* field_name, const double* field_data,
data_size_t num_element) {
std::string name(field_name);
name = Common::Trim(name);
if (name == std::string("init_score")) {
metadata_.SetInitScore(field_data, num_element);
} else {
return false;
}
return true;
}
bool Dataset::SetIntField(const char* field_name, const int* field_data,
data_size_t num_element) {
std::string name(field_name);
name = Common::Trim(name);
if (name == std::string("query") || name == std::string("group")) {
metadata_.SetQuery(field_data, num_element);
} else {
return false;
}
return true;
}
bool Dataset::GetFloatField(const char* field_name, data_size_t* out_len,
const float** out_ptr) {
std::string name(field_name);
name = Common::Trim(name);
if (name == std::string("label") || name == std::string("target")) {
#ifdef LABEL_T_USE_DOUBLE
Log::Fatal("Don't support LABEL_T_USE_DOUBLE");
#else
*out_ptr = metadata_.label();
*out_len = num_data_;
#endif
} else if (name == std::string("weight") || name == std::string("weights")) {
#ifdef LABEL_T_USE_DOUBLE
Log::Fatal("Don't support LABEL_T_USE_DOUBLE");
#else
*out_ptr = metadata_.weights();
*out_len = num_data_;
#endif
} else {
return false;
}
return true;
}
bool Dataset::GetDoubleField(const char* field_name, data_size_t* out_len,
const double** out_ptr) {
std::string name(field_name);
name = Common::Trim(name);
if (name == std::string("init_score")) {
*out_ptr = metadata_.init_score();
*out_len = static_cast<data_size_t>(metadata_.num_init_score());
} else {
return false;
}
return true;
}
bool Dataset::GetIntField(const char* field_name, data_size_t* out_len,
const int** out_ptr) {
std::string name(field_name);
name = Common::Trim(name);
if (name == std::string("query") || name == std::string("group")) {
*out_ptr = metadata_.query_boundaries();
*out_len = metadata_.num_queries() + 1;
} else {
return false;
}
return true;
}
void Dataset::SaveBinaryFile(const char* bin_filename) {
if (bin_filename != nullptr && std::string(bin_filename) == data_filename_) {
Log::Warning("Binary file %s already exists", bin_filename);
return;
}
// if not pass a filename, just append ".bin" of original file
std::string bin_filename_str(data_filename_);
if (bin_filename == nullptr || bin_filename[0] == '\0') {
bin_filename_str.append(".bin");
bin_filename = bin_filename_str.c_str();
}
bool is_file_existed = false;
if (VirtualFileWriter::Exists(bin_filename)) {
is_file_existed = true;
Log::Warning("File %s exists, cannot save binary to it", bin_filename);
}
if (!is_file_existed) {
auto writer = VirtualFileWriter::Make(bin_filename);
if (!writer->Init()) {
Log::Fatal("Cannot write binary data to %s ", bin_filename);
}
Log::Info("Saving data to binary file %s", bin_filename);
size_t size_of_token = std::strlen(binary_file_token);
writer->AlignedWrite(binary_file_token, size_of_token);
// get size of header
size_t size_of_header =
VirtualFileWriter::AlignedSize(sizeof(num_data_)) +
VirtualFileWriter::AlignedSize(sizeof(num_features_)) +
VirtualFileWriter::AlignedSize(sizeof(num_total_features_)) +
VirtualFileWriter::AlignedSize(sizeof(int) * num_total_features_) +
VirtualFileWriter::AlignedSize(sizeof(label_idx_)) +
VirtualFileWriter::AlignedSize(sizeof(num_groups_)) +