Repository navigation
Expand file tree
/
Copy pathsparse-matrix.cc
More file actions
1159 lines (1066 loc) · 36.3 KB
/
Copy pathsparse-matrix.cc
File metadata and controls
1159 lines (1066 loc) · 36.3 KB
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
// matrix/sparse-matrix.cc
// Copyright 2015 Johns Hopkins University (author: Daniel Povey)
// 2015 Guoguo Chen
// See ../../COPYING for clarification regarding multiple authors
//
// 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
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// THIS CODE IS PROVIDED *AS IS* BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
// KIND, EITHER EXPRESS OR IMPLIED, INCLUDING WITHOUT LIMITATION ANY IMPLIED
// WARRANTIES OR CONDITIONS OF TITLE, FITNESS FOR A PARTICULAR PURPOSE,
// MERCHANTABLITY OR NON-INFRINGEMENT.
// See the Apache 2 License for the specific language governing permissions and
// limitations under the License.
#include <algorithm>
#include <limits>
#include <string>
#include "matrix/sparse-matrix.h"
#include "matrix/kaldi-matrix.h"
namespace kaldi {
template <typename Real>
std::pair<MatrixIndexT, Real>* SparseVector<Real>::Data() {
if (pairs_.empty())
return NULL;
else
return &(pairs_[0]);
}
template <typename Real>
const std::pair<MatrixIndexT, Real>* SparseVector<Real>::Data() const {
if (pairs_.empty())
return NULL;
else
return &(pairs_[0]);
}
template <typename Real>
Real SparseVector<Real>::Sum() const {
Real sum = 0;
for (int32 i = 0; i < pairs_.size(); ++i) {
sum += pairs_[i].second;
}
return sum;
}
template <typename Real>
template <typename OtherReal>
void SparseVector<Real>::CopyToVec(VectorBase<OtherReal> *vec) const {
KALDI_ASSERT(vec->Dim() == this->dim_);
vec->SetZero();
OtherReal *other_data = vec->Data();
typename std::vector<std::pair<MatrixIndexT, Real> >::const_iterator
iter = pairs_.begin(), end = pairs_.end();
for (; iter != end; ++iter)
other_data[iter->first] = iter->second;
}
template void SparseVector<float>::CopyToVec(VectorBase<float> *vec) const;
template void SparseVector<float>::CopyToVec(VectorBase<double> *vec) const;
template void SparseVector<double>::CopyToVec(VectorBase<float> *vec) const;
template void SparseVector<double>::CopyToVec(VectorBase<double> *vec) const;
template <typename Real>
template <typename OtherReal>
void SparseVector<Real>::AddToVec(Real alpha,
VectorBase<OtherReal> *vec) const {
KALDI_ASSERT(vec->Dim() == dim_);
OtherReal *other_data = vec->Data();
typename std::vector<std::pair<MatrixIndexT, Real> >::const_iterator
iter = pairs_.begin(), end = pairs_.end();
if (alpha == 1.0) { // treat alpha==1.0 case specially.
for (; iter != end; ++iter)
other_data[iter->first] += iter->second;
} else {
for (; iter != end; ++iter)
other_data[iter->first] += alpha * iter->second;
}
}
template
void SparseVector<float>::AddToVec(float alpha, VectorBase<float> *vec) const;
template
void SparseVector<float>::AddToVec(float alpha, VectorBase<double> *vec) const;
template
void SparseVector<double>::AddToVec(double alpha, VectorBase<float> *vec) const;
template
void SparseVector<double>::AddToVec(double alpha,
VectorBase<double> *vec) const;
template <typename Real>
template <typename OtherReal>
void SparseVector<Real>::CopyFromSvec(const SparseVector<OtherReal> &other) {
dim_ = other.Dim();
pairs_.clear();
if (dim_ == 0) return;
for (int32 i = 0; i < other.NumElements(); ++i) {
pairs_.push_back(std::make_pair(
other.GetElement(i).first,
static_cast<Real>(other.GetElement(i).second)));
}
}
template
void SparseVector<float>::CopyFromSvec(const SparseVector<float> &svec);
template
void SparseVector<float>::CopyFromSvec(const SparseVector<double> &svec);
template
void SparseVector<double>::CopyFromSvec(const SparseVector<float> &svec);
template
void SparseVector<double>::CopyFromSvec(const SparseVector<double> &svec);
template <typename Real>
SparseVector<Real>& SparseVector<Real>::operator = (
const SparseVector<Real> &other) {
this->CopyFromSvec(other);
dim_ = other.dim_;
pairs_ = other.pairs_;
return *this;
}
template <typename Real>
void SparseVector<Real>::Swap(SparseVector<Real> *other) {
pairs_.swap(other->pairs_);
std::swap(dim_, other->dim_);
}
template <typename Real>
void SparseVector<Real>::Write(std::ostream &os, bool binary) const {
if (binary) {
WriteToken(os, binary, "SV");
WriteBasicType(os, binary, dim_);
MatrixIndexT num_elems = pairs_.size();
WriteBasicType(os, binary, num_elems);
typename std::vector<std::pair<MatrixIndexT, Real> >::const_iterator
iter = pairs_.begin(), end = pairs_.end();
for (; iter != end; ++iter) {
WriteBasicType(os, binary, iter->first);
WriteBasicType(os, binary, iter->second);
}
} else {
// In text-mode, use a human-friendly, script-friendly format;
// format is "dim=5 [ dim=5 0 0.2 3 0.9 ] "
os << "dim=" << dim_ << " [ ";
typename std::vector<std::pair<MatrixIndexT, Real> >::const_iterator
iter = pairs_.begin(), end = pairs_.end();
for (; iter != end; ++iter)
os << iter->first << ' ' << iter->second << ' ';
os << "] ";
}
}
template <typename Real>
void SparseVector<Real>::Read(std::istream &is, bool binary) {
if (binary) {
ExpectToken(is, binary, "SV");
ReadBasicType(is, binary, &dim_);
KALDI_ASSERT(dim_ >= 0);
int32 num_elems;
ReadBasicType(is, binary, &num_elems);
KALDI_ASSERT(num_elems >= 0 && num_elems <= dim_);
pairs_.resize(num_elems);
typename std::vector<std::pair<MatrixIndexT, Real> >::iterator
iter = pairs_.begin(), end = pairs_.end();
for (; iter != end; ++iter) {
ReadBasicType(is, binary, &(iter->first));
ReadBasicType(is, binary, &(iter->second));
}
} else {
// In text-mode, format is "dim=5 [ 0 0.2 3 0.9 ]
std::string str;
is >> str;
if (str.substr(0, 4) != "dim=")
KALDI_ERR << "Reading sparse vector, expected 'dim=xxx', got " << str;
std::string dim_str = str.substr(4, std::string::npos);
std::istringstream dim_istr(dim_str);
int32 dim = -1;
dim_istr >> dim;
if (dim < 0 || dim_istr.fail()) {
KALDI_ERR << "Reading sparse vector, expected 'dim=[int]', got " << str;
}
dim_ = dim;
is >> std::ws;
is >> str;
if (str != "[")
KALDI_ERR << "Reading sparse vector, expected '[', got " << str;
pairs_.clear();
while (1) {
is >> std::ws;
if (is.peek() == ']') {
is.get();
break;
}
MatrixIndexT i;
BaseFloat p;
is >> i >> p;
if (is.fail())
KALDI_ERR << "Error reading sparse vector, expecting numbers.";
KALDI_ASSERT(i >= 0 && i < dim
&& (pairs_.empty() || i > pairs_.back().first));
pairs_.push_back(std::pair<MatrixIndexT, BaseFloat>(i, p));
}
}
}
namespace sparse_vector_utils {
template <typename Real>
struct CompareFirst {
inline bool operator() (const std::pair<MatrixIndexT, Real> &p1,
const std::pair<MatrixIndexT, Real> &p2) const {
return p1.first < p2.first;
}
};
}
template <typename Real>
SparseVector<Real>::SparseVector(
MatrixIndexT dim, const std::vector<std::pair<MatrixIndexT, Real> > &pairs):
dim_(dim),
pairs_(pairs) {
std::sort(pairs_.begin(), pairs_.end(),
sparse_vector_utils::CompareFirst<Real>());
typename std::vector<std::pair<MatrixIndexT, Real> >::iterator
out = pairs_.begin(), in = out, end = pairs_.end();
// special case: while there is nothing to be changed, skip over
// initial input (avoids unnecessary copying).
while (in + 1 < end && in[0].first != in[1].first && in[0].second != 0.0) {
in++;
out++;
}
while (in < end) {
// We reach this point only at the first element of
// each stretch of identical .first elements.
*out = *in;
++in;
while (in < end && in->first == out->first) {
out->second += in->second; // this is the merge operation.
++in;
}
if (out->second != Real(0.0)) // Don't keep zero elements.
out++;
}
pairs_.erase(out, end);
if (!pairs_.empty()) {
// range check.
KALDI_ASSERT(pairs_.front().first >= 0 && pairs_.back().first < dim_);
}
}
template <typename Real>
void SparseVector<Real>::SetRandn(BaseFloat zero_prob) {
pairs_.clear();
KALDI_ASSERT(zero_prob >= 0 && zero_prob <= 1.0);
for (MatrixIndexT i = 0; i < dim_; i++)
if (WithProb(1.0 - zero_prob))
pairs_.push_back(std::pair<MatrixIndexT, Real>(i, RandGauss()));
}
template <typename Real>
void SparseVector<Real>::Resize(MatrixIndexT dim,
MatrixResizeType resize_type) {
if (resize_type != kCopyData || dim == 0)
pairs_.clear();
KALDI_ASSERT(dim >= 0);
if (dim < dim_ && resize_type == kCopyData)
while (!pairs_.empty() && pairs_.back().first >= dim)
pairs_.pop_back();
dim_ = dim;
}
template <typename Real>
MatrixIndexT SparseMatrix<Real>::NumRows() const {
return rows_.size();
}
template <typename Real>
MatrixIndexT SparseMatrix<Real>::NumCols() const {
if (rows_.empty())
return 0.0;
else
return rows_[0].Dim();
}
template <typename Real>
MatrixIndexT SparseMatrix<Real>::NumElements() const {
int32 num_elements = 0;
for (int32 i = 0; i < rows_.size(); ++i) {
num_elements += rows_[i].NumElements();
}
return num_elements;
}
template <typename Real>
SparseVector<Real>* SparseMatrix<Real>::Data() {
if (rows_.empty())
return NULL;
else
return rows_.data();
}
template <typename Real>
const SparseVector<Real>* SparseMatrix<Real>::Data() const {
if (rows_.empty())
return NULL;
else
return rows_.data();
}
template <typename Real>
Real SparseMatrix<Real>::Sum() const {
Real sum = 0;
for (int32 i = 0; i < rows_.size(); ++i) {
sum += rows_[i].Sum();
}
return sum;
}
template<typename Real>
Real SparseMatrix<Real>::FrobeniusNorm() const {
Real squared_sum = 0;
for (int32 i = 0; i < rows_.size(); ++i) {
const std::pair<MatrixIndexT, Real> *row_data = rows_[i].Data();
for (int32 j = 0; j < rows_[i].NumElements(); ++j) {
squared_sum += row_data[j].second * row_data[j].second;
}
}
return std::sqrt(squared_sum);
}
template<typename Real>
void SparseMatrix<Real>::Transpose() {
if(rows_.empty())
return;
std::vector<SparseVector<Real> > trans_rows
(NumCols(), SparseVector<Real>(NumRows()));
for (int32 i = 0; i < rows_.size(); ++i) {
const std::pair<MatrixIndexT, Real> *row_data = rows_[i].Data();
for (int32 j = 0; j < rows_[i].NumElements(); ++j) {
trans_rows[row_data[j].first].pairs_.push_back(std::make_pair(
i,
row_data[j].second));
}
}
rows_.swap(trans_rows);
}
template <typename Real>
template <typename OtherReal>
void SparseMatrix<Real>::CopyToMat(MatrixBase<OtherReal> *other,
MatrixTransposeType trans) const {
if (trans == kNoTrans) {
MatrixIndexT num_rows = rows_.size();
KALDI_ASSERT(other->NumRows() == num_rows);
for (MatrixIndexT i = 0; i < num_rows; i++) {
SubVector<OtherReal> vec(*other, i);
rows_[i].CopyToVec(&vec);
}
} else {
OtherReal *other_col_data = other->Data();
MatrixIndexT other_stride = other->Stride(),
num_rows = NumRows(), num_cols = NumCols();
KALDI_ASSERT(num_rows == other->NumCols() && num_cols == other->NumRows());
other->SetZero();
for (MatrixIndexT row = 0; row < num_rows; row++, other_col_data++) {
const SparseVector<Real> &svec = rows_[row];
MatrixIndexT num_elems = svec.NumElements();
const std::pair<MatrixIndexT, Real> *sdata = svec.Data();
for (MatrixIndexT e = 0; e < num_elems; e++)
other_col_data[sdata[e].first * other_stride] = sdata[e].second;
}
}
}
template
void SparseMatrix<float>::CopyToMat(MatrixBase<float> *other,
MatrixTransposeType trans) const;
template
void SparseMatrix<float>::CopyToMat(MatrixBase<double> *other,
MatrixTransposeType trans) const;
template
void SparseMatrix<double>::CopyToMat(MatrixBase<float> *other,
MatrixTransposeType trans) const;
template
void SparseMatrix<double>::CopyToMat(MatrixBase<double> *other,
MatrixTransposeType trans) const;
template <typename Real>
template <typename OtherReal>
void SparseMatrix<Real>::CopyFromMat(const MatrixBase<OtherReal> &other) {
rows_.resize(other.NumRows());
if (rows_.size() == 0) return;
for (int32 r = 0; r < rows_.size(); r++) {
SparseVector<Real> row(other.NumCols());
for (int32 c = 0; c < other.NumCols(); c++)
if (other(r, c) != 0.0)
row.pairs_.push_back(std::make_pair(c, other(r, c)));
rows_[r].Swap(&row);
}
}
template void SparseMatrix<float>::CopyFromMat(const MatrixBase<float> &other);
template void SparseMatrix<float>::CopyFromMat(const MatrixBase<double> &other);
template void SparseMatrix<double>::CopyFromMat(const MatrixBase<float> &other);
template void SparseMatrix<double>::CopyFromMat(const MatrixBase<double> &other);
template <typename Real>
template <typename OtherReal>
void SparseMatrix<Real>::CopyToVec(VectorBase<OtherReal> *other) const {
KALDI_ASSERT(other->Dim() == NumElements());
OtherReal *dst_data = other->Data();
int32 dst_index = 0;
for (int32 i = 0; i < rows_.size(); ++i) {
for (int32 j = 0; j < rows_[i].NumElements(); ++j) {
dst_data[dst_index] =
static_cast<OtherReal>(rows_[i].GetElement(j).second);
dst_index++;
}
}
}
template void SparseMatrix<float>::CopyToVec(VectorBase<float> *other) const;
template void SparseMatrix<float>::CopyToVec(VectorBase<double> *other) const;
template void SparseMatrix<double>::CopyToVec(VectorBase<float> *other) const;
template void SparseMatrix<double>::CopyToVec(VectorBase<double> *other) const;
template <typename Real>
template <typename OtherReal>
void SparseMatrix<Real>::CopyFromSmat(const SparseMatrix<OtherReal> &other) {
rows_.resize(other.NumRows());
if (rows_.size() == 0) return;
for (int32 r = 0; r < rows_.size(); ++r) {
rows_[r].CopyFromSvec(other.Row(r));
}
}
template
void SparseMatrix<float>::CopyFromSmat(const SparseMatrix<float> &other);
template
void SparseMatrix<float>::CopyFromSmat(const SparseMatrix<double> &other);
template
void SparseMatrix<double>::CopyFromSmat(const SparseMatrix<float> &other);
template
void SparseMatrix<double>::CopyFromSmat(const SparseMatrix<double> &other);
template <typename Real>
void SparseMatrix<Real>::Write(std::ostream &os, bool binary) const {
if (binary) {
// Note: we can use the same marker for float and double SparseMatrix,
// because internally we use WriteBasicType and ReadBasicType to read the
// floats and doubles, and this will automatically take care of type
// conversion.
WriteToken(os, binary, "SM");
int32 num_rows = rows_.size();
WriteBasicType(os, binary, num_rows);
for (int32 row = 0; row < num_rows; row++)
rows_[row].Write(os, binary);
} else {
// The format is "rows=10 dim=20 [ 1 0.4 9 1.2 ] dim=20 [ 3 1.7 19 0.6 ] ..
// not 100% efficient, but easy to work with, and we can re-use the
// read/write code from SparseVector.
int32 num_rows = rows_.size();
os << "rows=" << num_rows << " ";
for (int32 row = 0; row < num_rows; row++)
rows_[row].Write(os, binary);
os << "\n"; // Might make it a little more readable.
}
}
template <typename Real>
void SparseMatrix<Real>::Read(std::istream &is, bool binary) {
if (binary) {
ExpectToken(is, binary, "SM");
int32 num_rows;
ReadBasicType(is, binary, &num_rows);
KALDI_ASSERT(num_rows >= 0 && num_rows < 10000000);
rows_.resize(num_rows);
for (int32 row = 0; row < num_rows; row++)
rows_[row].Read(is, binary);
} else {
std::string str;
is >> str;
if (str.substr(0, 5) != "rows=")
KALDI_ERR << "Reading sparse matrix, expected 'rows=xxx', got " << str;
std::string rows_str = str.substr(5, std::string::npos);
std::istringstream rows_istr(rows_str);
int32 num_rows = -1;
rows_istr >> num_rows;
if (num_rows < 0 || rows_istr.fail()) {
KALDI_ERR << "Reading sparse vector, expected 'rows=[int]', got " << str;
}
rows_.resize(num_rows);
for (int32 row = 0; row < num_rows; row++)
rows_[row].Read(is, binary);
}
}
template <typename Real>
void SparseMatrix<Real>::AddToMat(BaseFloat alpha,
MatrixBase<Real> *other,
MatrixTransposeType trans) const {
if (trans == kNoTrans) {
MatrixIndexT num_rows = rows_.size();
KALDI_ASSERT(other->NumRows() == num_rows);
for (MatrixIndexT i = 0; i < num_rows; i++) {
SubVector<Real> vec(*other, i);
rows_[i].AddToVec(alpha, &vec);
}
} else {
Real *other_col_data = other->Data();
MatrixIndexT other_stride = other->Stride(),
num_rows = NumRows(), num_cols = NumCols();
KALDI_ASSERT(num_rows == other->NumCols() && num_cols == other->NumRows());
for (MatrixIndexT row = 0; row < num_rows; row++, other_col_data++) {
const SparseVector<Real> &svec = rows_[row];
MatrixIndexT num_elems = svec.NumElements();
const std::pair<MatrixIndexT, Real> *sdata = svec.Data();
for (MatrixIndexT e = 0; e < num_elems; e++)
other_col_data[sdata[e].first * other_stride] +=
alpha * sdata[e].second;
}
}
}
template <typename Real>
Real VecSvec(const VectorBase<Real> &vec,
const SparseVector<Real> &svec) {
KALDI_ASSERT(vec.Dim() == svec.Dim());
MatrixIndexT n = svec.NumElements();
const std::pair<MatrixIndexT, Real> *sdata = svec.Data();
const Real *data = vec.Data();
Real ans = 0.0;
for (MatrixIndexT i = 0; i < n; i++)
ans += data[sdata[i].first] * sdata[i].second;
return ans;
}
template
float VecSvec(const VectorBase<float> &vec,
const SparseVector<float> &svec);
template
double VecSvec(const VectorBase<double> &vec,
const SparseVector<double> &svec);
template <typename Real>
const SparseVector<Real> &SparseMatrix<Real>::Row(MatrixIndexT r) const {
KALDI_ASSERT(static_cast<size_t>(r) < rows_.size());
return rows_[r];
}
template <typename Real>
void SparseMatrix<Real>::SetRow(int32 r, const SparseVector<Real> &vec) {
KALDI_ASSERT(static_cast<size_t>(r) < rows_.size() &&
vec.Dim() == rows_[0].Dim());
rows_[r] = vec;
}
template <typename Real>
SparseMatrix<Real>& SparseMatrix<Real>::operator = (
const SparseMatrix<Real> &other) {
rows_ = other.rows_;
return *this;
}
template <typename Real>
void SparseMatrix<Real>::Swap(SparseMatrix<Real> *other) {
rows_.swap(other->rows_);
}
template<typename Real>
SparseMatrix<Real>::SparseMatrix(
MatrixIndexT dim,
const std::vector<std::vector<std::pair<MatrixIndexT, Real> > > &pairs):
rows_(pairs.size()) {
MatrixIndexT num_rows = pairs.size();
for (MatrixIndexT row = 0; row < num_rows; row++) {
SparseVector<Real> svec(dim, pairs[row]);
rows_[row].Swap(&svec);
}
}
template<typename Real>
void SparseMatrix<Real>::CopyFromPosterior(
MatrixIndexT dim,
const std::vector<std::vector<std::pair<MatrixIndexT, Real> > > &pairs) {
rows_.clear();
rows_.resize(pairs.size());
MatrixIndexT num_rows = pairs.size();
for (MatrixIndexT row = 0; row < num_rows; row++) {
SparseVector<Real> svec(dim);
svec.pairs_ = pairs[row];
rows_[row].Swap(&svec);
}
}
template <typename Real>
void SparseMatrix<Real>::SetRandn(BaseFloat zero_prob) {
MatrixIndexT num_rows = rows_.size();
for (MatrixIndexT row = 0; row < num_rows; row++)
rows_[row].SetRandn(zero_prob);
}
template <typename Real>
void SparseMatrix<Real>::Resize(MatrixIndexT num_rows,
MatrixIndexT num_cols,
MatrixResizeType resize_type) {
KALDI_ASSERT(num_rows >= 0 && num_cols >= 0);
if (resize_type == kSetZero || resize_type == kUndefined) {
rows_.clear();
Resize(num_rows, num_cols, kCopyData);
} else {
// Assume resize_type == kCopyData from here.
int32 old_num_rows = rows_.size(), old_num_cols = NumCols();
SparseVector<Real> initializer(num_cols);
rows_.resize(num_rows, initializer);
if (num_cols != old_num_cols)
for (int32 row = 0; row < old_num_rows; row++)
rows_[row].Resize(num_cols, kCopyData);
}
}
template <typename Real>
void SparseMatrix<Real>::AppendSparseMatrixRows(
std::vector<SparseMatrix<Real> > *inputs) {
rows_.clear();
size_t num_rows = 0;
typename std::vector<SparseMatrix<Real> >::iterator
input_iter = inputs->begin(),
input_end = inputs->end();
for (; input_iter != input_end; ++input_iter)
num_rows += input_iter->rows_.size();
rows_.resize(num_rows);
typename std::vector<SparseVector<Real> >::iterator
row_iter = rows_.begin(),
row_end = rows_.end();
for (input_iter = inputs->begin(); input_iter != input_end; ++input_iter) {
typename std::vector<SparseVector<Real> >::iterator
input_row_iter = input_iter->rows_.begin(),
input_row_end = input_iter->rows_.end();
for (; input_row_iter != input_row_end; ++input_row_iter, ++row_iter)
row_iter->Swap(&(*input_row_iter));
}
KALDI_ASSERT(row_iter == row_end);
int32 num_cols = NumCols();
for (row_iter = rows_.begin(); row_iter != row_end; ++row_iter) {
if (row_iter->Dim() != num_cols)
KALDI_ERR << "Appending rows with inconsistent dimensions, "
<< row_iter->Dim() << " vs. " << num_cols;
}
inputs->clear();
}
template<typename Real>
Real TraceMatSmat(const MatrixBase<Real> &A,
const SparseMatrix<Real> &B,
MatrixTransposeType trans) {
Real sum = 0.0;
if (trans == kTrans) {
MatrixIndexT num_rows = A.NumRows();
KALDI_ASSERT(B.NumRows() == num_rows);
for (MatrixIndexT r = 0; r < num_rows; r++)
sum += VecSvec(A.Row(r), B.Row(r));
} else {
const Real *A_col_data = A.Data();
MatrixIndexT Astride = A.Stride(), Acols = A.NumCols(), Arows = A.NumRows();
KALDI_ASSERT(Arows == B.NumCols() && Acols == B.NumRows());
sum = 0.0;
for (MatrixIndexT i = 0; i < Acols; i++, A_col_data++) {
Real col_sum = 0.0;
const SparseVector<Real> &svec = B.Row(i);
MatrixIndexT num_elems = svec.NumElements();
const std::pair<MatrixIndexT, Real> *sdata = svec.Data();
for (MatrixIndexT e = 0; e < num_elems; e++)
col_sum += A_col_data[Astride * sdata[e].first] * sdata[e].second;
sum += col_sum;
}
}
return sum;
}
template
float TraceMatSmat(const MatrixBase<float> &A,
const SparseMatrix<float> &B,
MatrixTransposeType trans);
template
double TraceMatSmat(const MatrixBase<double> &A,
const SparseMatrix<double> &B,
MatrixTransposeType trans);
void GeneralMatrix::Clear() {
mat_.Resize(0, 0);
cmat_.Clear();
smat_.Resize(0, 0);
}
GeneralMatrix& GeneralMatrix::operator= (const MatrixBase<BaseFloat> &mat) {
Clear();
mat_ = mat;
return *this;
}
GeneralMatrix& GeneralMatrix::operator= (const CompressedMatrix &cmat) {
Clear();
cmat_ = cmat;
return *this;
}
GeneralMatrix& GeneralMatrix::operator= (const SparseMatrix<BaseFloat> &smat) {
Clear();
smat_ = smat;
return *this;
}
GeneralMatrix& GeneralMatrix::operator= (const GeneralMatrix &gmat) {
mat_ = gmat.mat_;
smat_ = gmat.smat_;
cmat_ = gmat.cmat_;
return *this;
}
GeneralMatrixType GeneralMatrix::Type() const {
if (smat_.NumRows() != 0)
return kSparseMatrix;
else if (cmat_.NumRows() != 0)
return kCompressedMatrix;
else
return kFullMatrix;
}
MatrixIndexT GeneralMatrix::NumRows() const {
MatrixIndexT r = smat_.NumRows();
if (r != 0)
return r;
r = cmat_.NumRows();
if (r != 0)
return r;
return mat_.NumRows();
}
MatrixIndexT GeneralMatrix::NumCols() const {
MatrixIndexT r = smat_.NumCols();
if (r != 0)
return r;
r = cmat_.NumCols();
if (r != 0)
return r;
return mat_.NumCols();
}
void GeneralMatrix::Compress() {
if (mat_.NumRows() != 0) {
cmat_.CopyFromMat(mat_);
mat_.Resize(0, 0);
}
}
void GeneralMatrix::Uncompress() {
if (cmat_.NumRows() != 0) {
cmat_.CopyToMat(&mat_);
cmat_.Clear();
}
}
void GeneralMatrix::GetMatrix(Matrix<BaseFloat> *mat) const {
if (mat_.NumRows() !=0) {
*mat = mat_;
} else if (cmat_.NumRows() != 0) {
mat->Resize(cmat_.NumRows(), cmat_.NumCols(), kUndefined);
cmat_.CopyToMat(mat);
} else if (smat_.NumRows() != 0) {
mat->Resize(smat_.NumRows(), smat_.NumCols(), kUndefined);
smat_.CopyToMat(mat);
} else {
mat->Resize(0, 0);
}
}
void GeneralMatrix::CopyToMat(MatrixBase<BaseFloat> *mat,
MatrixTransposeType trans) const {
if (mat_.NumRows() !=0) {
mat->CopyFromMat(mat_, trans);
} else if (cmat_.NumRows() != 0) {
cmat_.CopyToMat(mat, trans);
} else if (smat_.NumRows() != 0) {
smat_.CopyToMat(mat, trans);
} else {
KALDI_ASSERT(mat->NumRows() == 0);
}
}
const SparseMatrix<BaseFloat>& GeneralMatrix::GetSparseMatrix() const {
if (mat_.NumRows() != 0 || cmat_.NumRows() != 0)
KALDI_ERR << "GetSparseMatrix called on GeneralMatrix of wrong type.";
return smat_;
}
void GeneralMatrix::SwapSparseMatrix(SparseMatrix<BaseFloat> *smat) {
if (mat_.NumRows() != 0 || cmat_.NumRows() != 0)
KALDI_ERR << "GetSparseMatrix called on GeneralMatrix of wrong type.";
smat->Swap(&smat_);
}
const CompressedMatrix &GeneralMatrix::GetCompressedMatrix() const {
if (mat_.NumRows() != 0 || smat_.NumRows() != 0)
KALDI_ERR << "GetCompressedMatrix called on GeneralMatrix of wrong type.";
return cmat_;
}
const Matrix<BaseFloat> &GeneralMatrix::GetFullMatrix() const {
if (smat_.NumRows() != 0 || cmat_.NumRows() != 0)
KALDI_ERR << "GetFullMatrix called on GeneralMatrix of wrong type.";
return mat_;
}
void GeneralMatrix::SwapFullMatrix(Matrix<BaseFloat> *mat) {
if (cmat_.NumRows() != 0 || smat_.NumRows() != 0)
KALDI_ERR << "SwapMatrix called on GeneralMatrix of wrong type.";
mat->Swap(&mat_);
}
void GeneralMatrix::Write(std::ostream &os, bool binary) const {
if (smat_.NumRows() != 0) {
smat_.Write(os, binary);
} else if (cmat_.NumRows() != 0) {
cmat_.Write(os, binary);
} else {
mat_.Write(os, binary);
}
}
void GeneralMatrix::Read(std::istream &is, bool binary) {
Clear();
if (binary) {
int peekval = is.peek();
if (peekval == 'C') {
// Token CM for compressed matrix
cmat_.Read(is, binary);
} else if (peekval == 'S') {
// Token SM for sparse matrix
smat_.Read(is, binary);
} else {
mat_.Read(is, binary);
}
} else {
// note: in text mode we will only ever read regular
// or sparse matrices, because the compressed-matrix format just
// gets written as a regular matrix in text mode.
is >> std::ws; // Eat up white space.
int peekval = is.peek();
if (peekval == 'r') { // sparse format starts rows=[int].
smat_.Read(is, binary);
} else {
mat_.Read(is, binary);
}
}
}
void AppendGeneralMatrixRows(const std::vector<const GeneralMatrix *> &src,
GeneralMatrix *mat) {
mat->Clear();
int32 size = src.size();
if (size == 0)
return;
bool all_sparse = true;
for (int32 i = 0; i < size; i++) {
if (src[i]->Type() != kSparseMatrix && src[i]->NumRows() != 0) {
all_sparse = false;
break;
}
}
if (all_sparse) {
std::vector<SparseMatrix<BaseFloat> > sparse_mats(size);
for (int32 i = 0; i < size; i++)
sparse_mats[i] = src[i]->GetSparseMatrix();
SparseMatrix<BaseFloat> appended_mat;
appended_mat.AppendSparseMatrixRows(&sparse_mats);
mat->SwapSparseMatrix(&appended_mat);
} else {
int32 tot_rows = 0, num_cols = -1;
for (int32 i = 0; i < size; i++) {
const GeneralMatrix &src_mat = *(src[i]);
int32 src_rows = src_mat.NumRows(), src_cols = src_mat.NumCols();
if (src_rows != 0) {
tot_rows += src_rows;
if (num_cols == -1) num_cols = src_cols;
else if (num_cols != src_cols)
KALDI_ERR << "Appending rows of matrices with inconsistent num-cols: "
<< num_cols << " vs. " << src_cols;
}
}
Matrix<BaseFloat> appended_mat(tot_rows, num_cols, kUndefined);
int32 row_offset = 0;
for (int32 i = 0; i < size; i++) {
const GeneralMatrix &src_mat = *(src[i]);
int32 src_rows = src_mat.NumRows();
if (src_rows != 0) {
SubMatrix<BaseFloat> dest_submat(appended_mat, row_offset, src_rows,
0, num_cols);
src_mat.CopyToMat(&dest_submat);
row_offset += src_rows;
}
}
KALDI_ASSERT(row_offset == tot_rows);
mat->SwapFullMatrix(&appended_mat);
}
}
void FilterCompressedMatrixRows(const CompressedMatrix &in,
const std::vector<bool> &keep_rows,
Matrix<BaseFloat> *out) {
KALDI_ASSERT(keep_rows.size() == static_cast<size_t>(in.NumRows()));
int32 num_kept_rows = 0;
std::vector<bool>::const_iterator iter = keep_rows.begin(),
end = keep_rows.end();
for (; iter != end; ++iter)
if (*iter)
num_kept_rows++;
if (num_kept_rows == 0)
KALDI_ERR << "No kept rows";
if (num_kept_rows == static_cast<int32>(keep_rows.size())) {
out->Resize(in.NumRows(), in.NumCols(), kUndefined);
in.CopyToMat(out);
return;
}
const BaseFloat heuristic = 0.33;
// should be > 0 and < 1.0. represents the performance hit we get from
// iterating row-wise versus column-wise in compressed-matrix uncompression.
if (num_kept_rows > heuristic * in.NumRows()) {
// if quite a few of the the rows are kept, it may be more efficient
// to uncompress the entire compressed matrix, since per-column operation
// is more efficient.
Matrix<BaseFloat> full_mat(in);
FilterMatrixRows(full_mat, keep_rows, out);
} else {
out->Resize(num_kept_rows, in.NumCols(), kUndefined);
iter = keep_rows.begin();
int32 out_row = 0;
for (int32 in_row = 0; iter != end; ++iter, ++in_row) {
if (*iter) {
SubVector<BaseFloat> dest(*out, out_row);
in.CopyRowToVec(in_row, &dest);
out_row++;
}
}
KALDI_ASSERT(out_row == num_kept_rows);
}
}
template <typename Real>
void FilterMatrixRows(const Matrix<Real> &in,
const std::vector<bool> &keep_rows,
Matrix<Real> *out) {
KALDI_ASSERT(keep_rows.size() == static_cast<size_t>(in.NumRows()));
int32 num_kept_rows = 0;
std::vector<bool>::const_iterator iter = keep_rows.begin(),
end = keep_rows.end();
for (; iter != end; ++iter)
if (*iter)
num_kept_rows++;
if (num_kept_rows == 0)
KALDI_ERR << "No kept rows";
if (num_kept_rows == static_cast<int32>(keep_rows.size())) {
*out = in;
return;
}
out->Resize(num_kept_rows, in.NumCols(), kUndefined);
iter = keep_rows.begin();
int32 out_row = 0;
for (int32 in_row = 0; iter != end; ++iter, ++in_row) {
if (*iter) {
SubVector<Real> src(in, in_row);
SubVector<Real> dest(*out, out_row);
dest.CopyFromVec(src);
out_row++;
}
}
KALDI_ASSERT(out_row == num_kept_rows);
}
template
void FilterMatrixRows(const Matrix<float> &in,
const std::vector<bool> &keep_rows,
Matrix<float> *out);
template
void FilterMatrixRows(const Matrix<double> &in,