/
adapter.h
1153 lines (1004 loc) · 35.5 KB
/
adapter.h
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) 2019~2021 by Contributors
* \file adapter.h
*/
#ifndef XGBOOST_DATA_ADAPTER_H_
#define XGBOOST_DATA_ADAPTER_H_
#include <dmlc/data.h>
#include <cstddef>
#include <functional>
#include <limits>
#include <memory>
#include <string>
#include <utility>
#include <vector>
#include <map>
#include <algorithm>
#include "xgboost/logging.h"
#include "xgboost/base.h"
#include "xgboost/data.h"
#include "xgboost/span.h"
#include "array_interface.h"
#include "../c_api/c_api_error.h"
#include "../common/math.h"
#include "arrow-cdi.h"
namespace xgboost {
namespace data {
/** External data formats should implement an adapter as below. The
* adapter provides a uniform access to data outside xgboost, allowing
* construction of DMatrix objects from a range of sources without duplicating
* code.
*
* The adapter object is an iterator that returns batches of data. Each batch
* contains a number of "lines". A line represents a set of elements from a
* sparse input matrix, normally a row in the case of a CSR matrix or a column
* for a CSC matrix. Typically in sparse matrix formats we can efficiently
* access subsets of elements at a time, but cannot efficiently lookups elements
* by random access, hence the "line" abstraction, allowing the sparse matrix to
* return subsets of elements efficiently. Individual elements are described by
* a COO tuple (row index, column index, value).
*
* This abstraction allows us to read through different sparse matrix formats
* using the same interface. In particular we can write a DMatrix constructor
* that uses the same code to construct itself from a CSR matrix, CSC matrix,
* dense matrix, CSV, LIBSVM file, or potentially other formats. To see why this
* is necessary, imagine we have 5 external matrix formats and 5 internal
* DMatrix types where each DMatrix needs a custom constructor for each possible
* input. The number of constructors is 5*5=25. Using an abstraction over the
* input data types the number of constructors is reduced to 5, as each DMatrix
* is oblivious to the external data format. Adding a new input source is simply
* a case of implementing an adapter.
*
* Most of the below adapters do not need more than one batch as the data
* originates from an in memory source. The file adapter does require batches to
* avoid loading the entire file in memory.
*
* An important detail is empty row/column handling. Files loaded from disk do
* not provide meta information about the number of rows/columns to expect, this
* needs to be inferred during construction. Other sparse formats may specify a
* number of rows/columns, but we can encounter entirely sparse rows or columns,
* leading to disagreement between the inferred number and the meta-info
* provided. To resolve this, adapters have methods specifying the number of
* rows/columns expected, these methods may return zero where these values must
* be inferred from data. A constructed DMatrix should agree with the input
* source on numbers of rows/columns, appending empty rows if necessary.
* */
/** \brief An adapter can return this value for number of rows or columns
* indicating that this value is currently unknown and should be inferred while
* passing over the data. */
constexpr size_t kAdapterUnknownSize = std::numeric_limits<size_t >::max();
struct COOTuple {
COOTuple() = default;
XGBOOST_DEVICE COOTuple(size_t row_idx, size_t column_idx, float value)
: row_idx(row_idx), column_idx(column_idx), value(value) {}
size_t row_idx{0};
size_t column_idx{0};
float value{0};
};
struct IsValidFunctor {
float missing;
XGBOOST_DEVICE explicit IsValidFunctor(float missing) : missing(missing) {}
XGBOOST_DEVICE bool operator()(float value) const {
return !(common::CheckNAN(value) || value == missing);
}
XGBOOST_DEVICE bool operator()(const data::COOTuple& e) const {
return !(common::CheckNAN(e.value) || e.value == missing);
}
XGBOOST_DEVICE bool operator()(const Entry& e) const {
return !(common::CheckNAN(e.fvalue) || e.fvalue == missing);
}
};
namespace detail {
/**
* \brief Simplifies the use of DataIter when there is only one batch.
*/
template <typename DType>
class SingleBatchDataIter : dmlc::DataIter<DType> {
public:
void BeforeFirst() override { counter_ = 0; }
bool Next() override {
if (counter_ == 0) {
counter_++;
return true;
}
return false;
}
private:
int counter_{0};
};
/** \brief Indicates this data source cannot contain meta-info such as labels,
* weights or qid. */
class NoMetaInfo {
public:
const float* Labels() const { return nullptr; }
const float* Weights() const { return nullptr; }
const uint64_t* Qid() const { return nullptr; }
const float* BaseMargin() const { return nullptr; }
};
}; // namespace detail
class CSRAdapterBatch : public detail::NoMetaInfo {
public:
class Line {
public:
Line(size_t row_idx, size_t size, const unsigned* feature_idx,
const float* values)
: row_idx_(row_idx),
size_(size),
feature_idx_(feature_idx),
values_(values) {}
size_t Size() const { return size_; }
COOTuple GetElement(size_t idx) const {
return COOTuple{row_idx_, feature_idx_[idx], values_[idx]};
}
private:
size_t row_idx_;
size_t size_;
const unsigned* feature_idx_;
const float* values_;
};
CSRAdapterBatch(const size_t* row_ptr, const unsigned* feature_idx,
const float* values, size_t num_rows, size_t, size_t)
: row_ptr_(row_ptr),
feature_idx_(feature_idx),
values_(values),
num_rows_(num_rows) {}
const Line GetLine(size_t idx) const {
size_t begin_offset = row_ptr_[idx];
size_t end_offset = row_ptr_[idx + 1];
return Line(idx, end_offset - begin_offset, &feature_idx_[begin_offset],
&values_[begin_offset]);
}
size_t Size() const { return num_rows_; }
static constexpr bool kIsRowMajor = true;
private:
const size_t* row_ptr_;
const unsigned* feature_idx_;
const float* values_;
size_t num_rows_;
};
class CSRAdapter : public detail::SingleBatchDataIter<CSRAdapterBatch> {
public:
CSRAdapter(const size_t* row_ptr, const unsigned* feature_idx,
const float* values, size_t num_rows, size_t num_elements,
size_t num_features)
: batch_(row_ptr, feature_idx, values, num_rows, num_elements,
num_features),
num_rows_(num_rows),
num_columns_(num_features) {}
const CSRAdapterBatch& Value() const override { return batch_; }
size_t NumRows() const { return num_rows_; }
size_t NumColumns() const { return num_columns_; }
private:
CSRAdapterBatch batch_;
size_t num_rows_;
size_t num_columns_;
};
class DenseAdapterBatch : public detail::NoMetaInfo {
public:
DenseAdapterBatch(const float* values, size_t num_rows, size_t num_features)
: values_(values),
num_rows_(num_rows),
num_features_(num_features) {}
private:
class Line {
public:
Line(const float* values, size_t size, size_t row_idx)
: row_idx_(row_idx), size_(size), values_(values) {}
size_t Size() const { return size_; }
COOTuple GetElement(size_t idx) const {
return COOTuple{row_idx_, idx, values_[idx]};
}
private:
size_t row_idx_;
size_t size_;
const float* values_;
};
public:
size_t Size() const { return num_rows_; }
const Line GetLine(size_t idx) const {
return Line(values_ + idx * num_features_, num_features_, idx);
}
static constexpr bool kIsRowMajor = true;
private:
const float* values_;
size_t num_rows_;
size_t num_features_;
};
class DenseAdapter : public detail::SingleBatchDataIter<DenseAdapterBatch> {
public:
DenseAdapter(const float* values, size_t num_rows, size_t num_features)
: batch_(values, num_rows, num_features),
num_rows_(num_rows),
num_columns_(num_features) {}
const DenseAdapterBatch& Value() const override { return batch_; }
size_t NumRows() const { return num_rows_; }
size_t NumColumns() const { return num_columns_; }
private:
DenseAdapterBatch batch_;
size_t num_rows_;
size_t num_columns_;
};
class ArrayAdapterBatch : public detail::NoMetaInfo {
public:
static constexpr bool kIsRowMajor = true;
private:
ArrayInterface<2> array_interface_;
class Line {
ArrayInterface<2> array_interface_;
size_t ridx_;
public:
Line(ArrayInterface<2> array_interface, size_t ridx)
: array_interface_{std::move(array_interface)}, ridx_{ridx} {}
size_t Size() const { return array_interface_.Shape(1); }
COOTuple GetElement(size_t idx) const {
return {ridx_, idx, array_interface_(ridx_, idx)};
}
};
public:
ArrayAdapterBatch() = default;
Line const GetLine(size_t idx) const {
return Line{array_interface_, idx};
}
size_t NumRows() const { return array_interface_.Shape(0); }
size_t NumCols() const { return array_interface_.Shape(1); }
size_t Size() const { return this->NumRows(); }
explicit ArrayAdapterBatch(ArrayInterface<2> array_interface)
: array_interface_{std::move(array_interface)} {}
};
/**
* Adapter for dense array on host, in Python that's `numpy.ndarray`. This is similar to
* `DenseAdapter`, but supports __array_interface__ instead of raw pointers. An
* advantage is this can handle various data type without making a copy.
*/
class ArrayAdapter : public detail::SingleBatchDataIter<ArrayAdapterBatch> {
public:
explicit ArrayAdapter(StringView array_interface) {
auto j = Json::Load(array_interface);
array_interface_ = ArrayInterface<2>(get<Object const>(j));
batch_ = ArrayAdapterBatch{array_interface_};
}
ArrayAdapterBatch const& Value() const override { return batch_; }
size_t NumRows() const { return array_interface_.Shape(0); }
size_t NumColumns() const { return array_interface_.Shape(1); }
private:
ArrayAdapterBatch batch_;
ArrayInterface<2> array_interface_;
};
class CSRArrayAdapterBatch : public detail::NoMetaInfo {
ArrayInterface<1> indptr_;
ArrayInterface<1> indices_;
ArrayInterface<1> values_;
bst_feature_t n_features_;
class Line {
ArrayInterface<1> indices_;
ArrayInterface<1> values_;
size_t ridx_;
size_t offset_;
public:
Line(ArrayInterface<1> indices, ArrayInterface<1> values, size_t ridx,
size_t offset)
: indices_{std::move(indices)}, values_{std::move(values)}, ridx_{ridx},
offset_{offset} {}
COOTuple GetElement(size_t idx) const {
return {ridx_, TypedIndex<size_t, 1>{indices_}(offset_ + idx), values_(offset_ + idx)};
}
size_t Size() const {
return values_.Shape(0);
}
};
public:
static constexpr bool kIsRowMajor = true;
public:
CSRArrayAdapterBatch() = default;
CSRArrayAdapterBatch(ArrayInterface<1> indptr, ArrayInterface<1> indices,
ArrayInterface<1> values, bst_feature_t n_features)
: indptr_{std::move(indptr)},
indices_{std::move(indices)},
values_{std::move(values)},
n_features_{n_features} {
}
size_t NumRows() const {
size_t size = indptr_.Shape(0);
size = size == 0 ? 0 : size - 1;
return size;
}
size_t NumCols() const { return n_features_; }
size_t Size() const { return this->NumRows(); }
Line const GetLine(size_t idx) const {
auto begin_no_stride = TypedIndex<size_t, 1>{indptr_}(idx);
auto end_no_stride = TypedIndex<size_t, 1>{indptr_}(idx + 1);
auto indices = indices_;
auto values = values_;
// Slice indices and values, stride remains unchanged since this is slicing by
// specific index.
auto offset = indices.strides[0] * begin_no_stride;
indices.shape[0] = end_no_stride - begin_no_stride;
values.shape[0] = end_no_stride - begin_no_stride;
return Line{indices, values, idx, offset};
}
};
/**
* Adapter for CSR array on host, in Python that's `scipy.sparse.csr_matrix`. This is
* similar to `CSRAdapter`, but supports __array_interface__ instead of raw pointers. An
* advantage is this can handle various data type without making a copy.
*/
class CSRArrayAdapter : public detail::SingleBatchDataIter<CSRArrayAdapterBatch> {
public:
CSRArrayAdapter(StringView indptr, StringView indices, StringView values,
size_t num_cols)
: indptr_{indptr}, indices_{indices}, values_{values}, num_cols_{num_cols} {
batch_ = CSRArrayAdapterBatch{indptr_, indices_, values_,
static_cast<bst_feature_t>(num_cols_)};
}
CSRArrayAdapterBatch const& Value() const override {
return batch_;
}
size_t NumRows() const {
size_t size = indptr_.Shape(0);
size = size == 0 ? 0 : size - 1;
return size;
}
size_t NumColumns() const { return num_cols_; }
private:
CSRArrayAdapterBatch batch_;
ArrayInterface<1> indptr_;
ArrayInterface<1> indices_;
ArrayInterface<1> values_;
size_t num_cols_;
};
class CSCAdapterBatch : public detail::NoMetaInfo {
public:
CSCAdapterBatch(const size_t* col_ptr, const unsigned* row_idx,
const float* values, size_t num_features)
: col_ptr_(col_ptr),
row_idx_(row_idx),
values_(values),
num_features_(num_features) {}
private:
class Line {
public:
Line(size_t col_idx, size_t size, const unsigned* row_idx,
const float* values)
: col_idx_(col_idx), size_(size), row_idx_(row_idx), values_(values) {}
size_t Size() const { return size_; }
COOTuple GetElement(size_t idx) const {
return COOTuple{row_idx_[idx], col_idx_, values_[idx]};
}
private:
size_t col_idx_;
size_t size_;
const unsigned* row_idx_;
const float* values_;
};
public:
size_t Size() const { return num_features_; }
const Line GetLine(size_t idx) const {
size_t begin_offset = col_ptr_[idx];
size_t end_offset = col_ptr_[idx + 1];
return Line(idx, end_offset - begin_offset, &row_idx_[begin_offset],
&values_[begin_offset]);
}
static constexpr bool kIsRowMajor = false;
private:
const size_t* col_ptr_;
const unsigned* row_idx_;
const float* values_;
size_t num_features_;
};
class CSCAdapter : public detail::SingleBatchDataIter<CSCAdapterBatch> {
public:
CSCAdapter(const size_t* col_ptr, const unsigned* row_idx,
const float* values, size_t num_features, size_t num_rows)
: batch_(col_ptr, row_idx, values, num_features),
num_rows_(num_rows),
num_columns_(num_features) {}
const CSCAdapterBatch& Value() const override { return batch_; }
// JVM package sends 0 as unknown
size_t NumRows() const {
return num_rows_ == 0 ? kAdapterUnknownSize : num_rows_;
}
size_t NumColumns() const { return num_columns_; }
private:
CSCAdapterBatch batch_;
size_t num_rows_;
size_t num_columns_;
};
class DataTableAdapterBatch : public detail::NoMetaInfo {
public:
DataTableAdapterBatch(void** data, const char** feature_stypes,
size_t num_rows, size_t num_features)
: data_(data),
feature_stypes_(feature_stypes),
num_features_(num_features),
num_rows_(num_rows) {}
private:
enum class DTType : uint8_t {
kFloat32 = 0,
kFloat64 = 1,
kBool8 = 2,
kInt32 = 3,
kInt8 = 4,
kInt16 = 5,
kInt64 = 6,
kUnknown = 7
};
DTType DTGetType(std::string type_string) const {
if (type_string == "float32") {
return DTType::kFloat32;
} else if (type_string == "float64") {
return DTType::kFloat64;
} else if (type_string == "bool8") {
return DTType::kBool8;
} else if (type_string == "int32") {
return DTType::kInt32;
} else if (type_string == "int8") {
return DTType::kInt8;
} else if (type_string == "int16") {
return DTType::kInt16;
} else if (type_string == "int64") {
return DTType::kInt64;
} else {
LOG(FATAL) << "Unknown data table type.";
return DTType::kUnknown;
}
}
class Line {
float DTGetValue(const void* column, DTType dt_type, size_t ridx) const {
float missing = std::numeric_limits<float>::quiet_NaN();
switch (dt_type) {
case DTType::kFloat32: {
float val = reinterpret_cast<const float*>(column)[ridx];
return std::isfinite(val) ? val : missing;
}
case DTType::kFloat64: {
double val = reinterpret_cast<const double*>(column)[ridx];
return std::isfinite(val) ? static_cast<float>(val) : missing;
}
case DTType::kBool8: {
bool val = reinterpret_cast<const bool*>(column)[ridx];
return static_cast<float>(val);
}
case DTType::kInt32: {
int32_t val = reinterpret_cast<const int32_t*>(column)[ridx];
return val != (-2147483647 - 1) ? static_cast<float>(val) : missing;
}
case DTType::kInt8: {
int8_t val = reinterpret_cast<const int8_t*>(column)[ridx];
return val != -128 ? static_cast<float>(val) : missing;
}
case DTType::kInt16: {
int16_t val = reinterpret_cast<const int16_t*>(column)[ridx];
return val != -32768 ? static_cast<float>(val) : missing;
}
case DTType::kInt64: {
int64_t val = reinterpret_cast<const int64_t*>(column)[ridx];
return val != -9223372036854775807 - 1 ? static_cast<float>(val)
: missing;
}
default: {
LOG(FATAL) << "Unknown data table type.";
return 0.0f;
}
}
}
public:
Line(DTType type, size_t size, size_t column_idx, const void* column)
: type_(type), size_(size), column_idx_(column_idx), column_(column) {}
size_t Size() const { return size_; }
COOTuple GetElement(size_t idx) const {
return COOTuple{idx, column_idx_, DTGetValue(column_, type_, idx)};
}
private:
DTType type_;
size_t size_;
size_t column_idx_;
const void* column_;
};
public:
size_t Size() const { return num_features_; }
const Line GetLine(size_t idx) const {
return Line(DTGetType(feature_stypes_[idx]), num_rows_, idx, data_[idx]);
}
static constexpr bool kIsRowMajor = false;
private:
void** data_;
const char** feature_stypes_;
size_t num_features_;
size_t num_rows_;
};
class DataTableAdapter
: public detail::SingleBatchDataIter<DataTableAdapterBatch> {
public:
DataTableAdapter(void** data, const char** feature_stypes, size_t num_rows,
size_t num_features)
: batch_(data, feature_stypes, num_rows, num_features),
num_rows_(num_rows),
num_columns_(num_features) {}
const DataTableAdapterBatch& Value() const override { return batch_; }
size_t NumRows() const { return num_rows_; }
size_t NumColumns() const { return num_columns_; }
private:
DataTableAdapterBatch batch_;
size_t num_rows_;
size_t num_columns_;
};
class FileAdapterBatch {
public:
class Line {
public:
Line(size_t row_idx, const uint32_t *feature_idx, const float *value,
size_t size)
: row_idx_(row_idx),
feature_idx_(feature_idx),
value_(value),
size_(size) {}
size_t Size() { return size_; }
COOTuple GetElement(size_t idx) {
float fvalue = value_ == nullptr ? 1.0f : value_[idx];
return COOTuple{row_idx_, feature_idx_[idx], fvalue};
}
private:
size_t row_idx_;
const uint32_t* feature_idx_;
const float* value_;
size_t size_;
};
FileAdapterBatch(const dmlc::RowBlock<uint32_t>* block, size_t row_offset)
: block_(block), row_offset_(row_offset) {}
Line GetLine(size_t idx) const {
auto begin = block_->offset[idx];
auto end = block_->offset[idx + 1];
return Line{idx + row_offset_, &block_->index[begin], &block_->value[begin],
end - begin};
}
const float* Labels() const { return block_->label; }
const float* Weights() const { return block_->weight; }
const uint64_t* Qid() const { return block_->qid; }
const float* BaseMargin() const { return nullptr; }
size_t Size() const { return block_->size; }
static constexpr bool kIsRowMajor = true;
private:
const dmlc::RowBlock<uint32_t>* block_;
size_t row_offset_;
};
/** \brief FileAdapter wraps dmlc::parser to read files and provide access in a
* common interface. */
class FileAdapter : dmlc::DataIter<FileAdapterBatch> {
public:
explicit FileAdapter(dmlc::Parser<uint32_t>* parser) : parser_(parser) {}
const FileAdapterBatch& Value() const override { return *batch_.get(); }
void BeforeFirst() override {
batch_.reset();
parser_->BeforeFirst();
row_offset_ = 0;
}
bool Next() override {
bool next = parser_->Next();
batch_.reset(new FileAdapterBatch(&parser_->Value(), row_offset_));
row_offset_ += parser_->Value().size;
return next;
}
// Indicates a number of rows/columns must be inferred
size_t NumRows() const { return kAdapterUnknownSize; }
size_t NumColumns() const { return kAdapterUnknownSize; }
private:
size_t row_offset_{0};
std::unique_ptr<FileAdapterBatch> batch_;
dmlc::Parser<uint32_t>* parser_;
};
/*! \brief Data iterator that takes callback to return data, used in JVM package for
* accepting data iterator. */
template <typename DataIterHandle, typename XGBCallbackDataIterNext, typename XGBoostBatchCSR>
class IteratorAdapter : public dmlc::DataIter<FileAdapterBatch> {
public:
IteratorAdapter(DataIterHandle data_handle, XGBCallbackDataIterNext* next_callback)
: columns_{data::kAdapterUnknownSize},
data_handle_(data_handle),
next_callback_(next_callback) {}
// override functions
void BeforeFirst() override {
CHECK(at_first_) << "Cannot reset IteratorAdapter";
}
bool Next() override {
if ((*next_callback_)(
data_handle_,
[](void *handle, XGBoostBatchCSR batch) -> int {
API_BEGIN();
static_cast<IteratorAdapter *>(handle)->SetData(batch);
API_END();
},
this) != 0) {
at_first_ = false;
return true;
} else {
return false;
}
}
FileAdapterBatch const& Value() const override {
return *batch_.get();
}
// callback to set the data
void SetData(const XGBoostBatchCSR& batch) {
offset_.clear();
label_.clear();
weight_.clear();
index_.clear();
value_.clear();
offset_.insert(offset_.end(), batch.offset, batch.offset + batch.size + 1);
if (batch.label != nullptr) {
label_.insert(label_.end(), batch.label, batch.label + batch.size);
}
if (batch.weight != nullptr) {
weight_.insert(weight_.end(), batch.weight, batch.weight + batch.size);
}
if (batch.index != nullptr) {
index_.insert(index_.end(), batch.index + offset_[0],
batch.index + offset_.back());
}
if (batch.value != nullptr) {
value_.insert(value_.end(), batch.value + offset_[0],
batch.value + offset_.back());
}
if (offset_[0] != 0) {
size_t base = offset_[0];
for (size_t &item : offset_) {
item -= base;
}
}
CHECK(columns_ == data::kAdapterUnknownSize || columns_ == batch.columns)
<< "Number of columns between batches changed from " << columns_
<< " to " << batch.columns;
columns_ = batch.columns;
block_.size = batch.size;
block_.offset = dmlc::BeginPtr(offset_);
block_.label = dmlc::BeginPtr(label_);
block_.weight = dmlc::BeginPtr(weight_);
block_.qid = nullptr;
block_.field = nullptr;
block_.index = dmlc::BeginPtr(index_);
block_.value = dmlc::BeginPtr(value_);
batch_.reset(new FileAdapterBatch(&block_, row_offset_));
row_offset_ += offset_.size() - 1;
}
size_t NumColumns() const { return columns_; }
size_t NumRows() const { return kAdapterUnknownSize; }
private:
std::vector<size_t> offset_;
std::vector<dmlc::real_t> label_;
std::vector<dmlc::real_t> weight_;
std::vector<uint32_t> index_;
std::vector<dmlc::real_t> value_;
size_t columns_;
size_t row_offset_{0};
// at the beginning.
bool at_first_{true};
// handle to the iterator,
DataIterHandle data_handle_;
// call back to get the data.
XGBCallbackDataIterNext *next_callback_;
// internal Rowblock
dmlc::RowBlock<uint32_t> block_;
std::unique_ptr<FileAdapterBatch> batch_;
};
enum ColumnDType : uint8_t {
kUnknown,
kInt8,
kUInt8,
kInt16,
kUInt16,
kInt32,
kUInt32,
kInt64,
kUInt64,
kFloat,
kDouble
};
class Column {
public:
Column() = default;
Column(size_t col_idx, size_t length, size_t null_count, const uint8_t* bitmap)
: col_idx_{col_idx}, length_{length}, null_count_{null_count}, bitmap_{bitmap} {}
virtual ~Column() = default;
Column(const Column&) = delete;
Column& operator=(const Column&) = delete;
Column(Column&&) = delete;
Column& operator=(Column&&) = delete;
// whether the valid bit is set for this element
bool IsValid(size_t row_idx) const {
return (!bitmap_ || (bitmap_[row_idx/8] & (1 << (row_idx%8))));
}
virtual COOTuple GetElement(size_t row_idx) const = 0;
virtual bool IsValidElement(size_t row_idx) const = 0;
virtual std::vector<float> AsFloatVector() const = 0;
virtual std::vector<uint64_t> AsUint64Vector() const = 0;
size_t Length() const { return length_; }
protected:
size_t col_idx_;
size_t length_;
size_t null_count_;
const uint8_t* bitmap_;
};
// Only columns of primitive types are supported. An ArrowColumnarBatch is a
// collection of std::shared_ptr<PrimitiveColumn>. These columns can be of different data types.
// Hence, PrimitiveColumn is a class template; and all concrete PrimitiveColumns
// derive from the abstract class Column.
template <typename T>
class PrimitiveColumn : public Column {
static constexpr float kNaN = std::numeric_limits<float>::quiet_NaN();
public:
PrimitiveColumn(size_t idx, size_t length, size_t null_count,
const uint8_t* bitmap, const T* data, float missing)
: Column{idx, length, null_count, bitmap}, data_{data}, missing_{missing} {}
COOTuple GetElement(size_t row_idx) const override {
CHECK(data_ && row_idx < length_) << "Column is empty or out-of-bound index of the column";
return { row_idx, col_idx_, IsValidElement(row_idx) ?
static_cast<float>(data_[row_idx]) : kNaN };
}
bool IsValidElement(size_t row_idx) const override {
// std::isfinite needs to cast to double to prevent msvc report error
return IsValid(row_idx)
&& std::isfinite(static_cast<double>(data_[row_idx]))
&& static_cast<float>(data_[row_idx]) != missing_;
}
std::vector<float> AsFloatVector() const override {
CHECK(data_) << "Column is empty";
std::vector<float> fv(length_);
std::transform(data_, data_ + length_, fv.begin(),
[](T v) { return static_cast<float>(v); });
return fv;
}
std::vector<uint64_t> AsUint64Vector() const override {
CHECK(data_) << "Column is empty";
std::vector<uint64_t> iv(length_);
std::transform(data_, data_ + length_, iv.begin(),
[](T v) { return static_cast<uint64_t>(v); });
return iv;
}
private:
const T* data_;
float missing_; // user specified missing value
};
struct ColumnarMetaInfo {
// data type of the column
ColumnDType type{ColumnDType::kUnknown};
// location of the column in an Arrow record batch
int64_t loc{-1};
};
struct ArrowSchemaImporter {
std::vector<ColumnarMetaInfo> columns;
// map Arrow format strings to types
static ColumnDType FormatMap(char const* format_str) {
CHECK(format_str) << "Format string cannot be empty";
switch (format_str[0]) {
case 'c':
return ColumnDType::kInt8;
case 'C':
return ColumnDType::kUInt8;
case 's':
return ColumnDType::kInt16;
case 'S':
return ColumnDType::kUInt16;
case 'i':
return ColumnDType::kInt32;
case 'I':
return ColumnDType::kUInt32;
case 'l':
return ColumnDType::kInt64;
case 'L':
return ColumnDType::kUInt64;
case 'f':
return ColumnDType::kFloat;
case 'g':
return ColumnDType::kDouble;
default:
CHECK(false) << "Column data type not supported by XGBoost";
return ColumnDType::kUnknown;
}
}
void Import(struct ArrowSchema *schema) {
if (schema) {
CHECK(std::string(schema->format) == "+s"); // NOLINT
CHECK(columns.empty());
for (auto i = 0; i < schema->n_children; ++i) {
std::string name{schema->children[i]->name};
ColumnDType type = FormatMap(schema->children[i]->format);
ColumnarMetaInfo col_info{type, i};
columns.push_back(col_info);
}
if (schema->release) {
schema->release(schema);
}
}
}
};
class ArrowColumnarBatch {
public:
ArrowColumnarBatch(struct ArrowArray *rb, struct ArrowSchemaImporter* schema)
: rb_{rb}, schema_{schema} {
CHECK(rb_) << "Cannot import non-existent record batch";
CHECK(!schema_->columns.empty()) << "Cannot import record batch without a schema";
}
size_t Import(float missing) {
auto& infov = schema_->columns;
for (size_t i = 0; i < infov.size(); ++i) {
columns_.push_back(CreateColumn(i, infov[i], missing));
}
// Compute the starting location for every row in this batch
auto batch_size = rb_->length;
auto num_columns = columns_.size();
row_offsets_.resize(batch_size + 1, 0);
for (auto i = 0; i < batch_size; ++i) {
row_offsets_[i+1] = row_offsets_[i];
for (size_t j = 0; j < num_columns; ++j) {
if (GetColumn(j).IsValidElement(i)) {
row_offsets_[i+1]++;
}
}
}
// return number of elements in the batch
return row_offsets_.back();
}
ArrowColumnarBatch(const ArrowColumnarBatch&) = delete;
ArrowColumnarBatch& operator=(const ArrowColumnarBatch&) = delete;
ArrowColumnarBatch(ArrowColumnarBatch&&) = delete;
ArrowColumnarBatch& operator=(ArrowColumnarBatch&&) = delete;
virtual ~ArrowColumnarBatch() {
if (rb_ && rb_->release) {
rb_->release(rb_);
rb_ = nullptr;
}
columns_.clear();
}
size_t Size() const { return rb_ ? rb_->length : 0; }
size_t NumColumns() const { return columns_.size(); }
size_t NumElements() const { return row_offsets_.back(); }
const Column& GetColumn(size_t col_idx) const {
return *columns_[col_idx];
}
void ShiftRowOffsets(size_t batch_offset) {
std::transform(row_offsets_.begin(), row_offsets_.end(), row_offsets_.begin(),
[=](size_t c) { return c + batch_offset; });
}
const std::vector<size_t>& RowOffsets() const { return row_offsets_; }
private:
std::shared_ptr<Column> CreateColumn(size_t idx,
ColumnarMetaInfo info,
float missing) const {