forked from tensorflow/tensorflow
-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathsparse_ops.py
More file actions
1679 lines (1344 loc) · 62.4 KB
/
Copy pathsparse_ops.py
File metadata and controls
1679 lines (1344 loc) · 62.4 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
# Copyright 2015 The TensorFlow Authors. All Rights Reserved.
#
# 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
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ==============================================================================
# pylint: disable=g-short-docstring-punctuation
"""Sparse Tensor Representation. See the @{python/sparse_ops} guide.
@@SparseTensor
@@SparseTensorValue
@@sparse_to_dense
@@sparse_tensor_to_dense
@@sparse_to_indicator
@@sparse_merge
@@sparse_concat
@@sparse_reorder
@@sparse_reshape
@@sparse_split
@@sparse_retain
@@sparse_reset_shape
@@sparse_fill_empty_rows
@@sparse_transpose
@@sparse_reduce_sum
@@sparse_reduce_sum_sparse
@@sparse_add
@@sparse_softmax
@@sparse_tensor_dense_matmul
@@sparse_maximum
@@sparse_minimum
"""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import numpy as np
from tensorflow.python.framework import dtypes
from tensorflow.python.framework import ops
from tensorflow.python.framework import sparse_tensor
from tensorflow.python.ops import array_ops
from tensorflow.python.ops import check_ops
from tensorflow.python.ops import control_flow_ops
from tensorflow.python.ops import gen_sparse_ops
from tensorflow.python.ops import math_ops
# go/tf-wildcard-import
# pylint: disable=wildcard-import
from tensorflow.python.ops.gen_sparse_ops import *
# pylint: enable=wildcard-import
from tensorflow.python.util import deprecation
def _convert_to_sparse_tensor(sp_input):
"""Convert `sp_input` to `SparseTensor` and return it.
Args:
sp_input: `SparseTensor` or `SparseTensorValue`.
Returns:
`sp_input` converted to `SparseTensor`.
Raises:
ValueError: if `sp_input` is neither `SparseTensor` nor `SparseTensorValue`.
"""
if isinstance(sp_input, sparse_tensor.SparseTensorValue):
return sparse_tensor.SparseTensor.from_value(sp_input)
if not isinstance(sp_input, sparse_tensor.SparseTensor):
raise TypeError("Input must be a SparseTensor.")
return sp_input
def _convert_to_sparse_tensors(sp_inputs):
"""Convert `sp_inputs` to `SparseTensor` objects and return them.
Args:
sp_inputs: `list` or `tuple` of `SparseTensor` or `SparseTensorValue`
objects.
Returns:
`sp_inputs` converted to `SparseTensor` objects.
Raises:
ValueError: if any item in `sp_inputs` is neither `SparseTensor` nor
`SparseTensorValue`.
"""
if isinstance(sp_inputs, list):
return [_convert_to_sparse_tensor(sp_input) for sp_input in sp_inputs]
if isinstance(sp_inputs, tuple):
return (_convert_to_sparse_tensor(sp_input) for sp_input in sp_inputs)
raise TypeError("Inputs must be a list or tuple.")
# pylint: disable=protected-access
def sparse_concat(axis,
sp_inputs,
name=None,
expand_nonconcat_dim=False,
concat_dim=None):
"""Concatenates a list of `SparseTensor` along the specified dimension.
Concatenation is with respect to the dense versions of each sparse input.
It is assumed that each inputs is a `SparseTensor` whose elements are ordered
along increasing dimension number.
If expand_nonconcat_dim is False, all inputs' shapes must match, except for
the concat dimension. If expand_nonconcat_dim is True, then inputs' shapes are
allowed to vary among all inputs.
The `indices`, `values`, and `shapes` lists must have the same length.
If expand_nonconcat_dim is False, then the output shape is identical to the
inputs', except along the concat dimension, where it is the sum of the inputs'
sizes along that dimension.
If expand_nonconcat_dim is True, then the output shape along the non-concat
dimensions will be expand to be the largest among all inputs, and it is the
sum of the inputs sizes along the concat dimension.
The output elements will be resorted to preserve the sort order along
increasing dimension number.
This op runs in `O(M log M)` time, where `M` is the total number of non-empty
values across all inputs. This is due to the need for an internal sort in
order to concatenate efficiently across an arbitrary dimension.
For example, if `axis = 1` and the inputs are
sp_inputs[0]: shape = [2, 3]
[0, 2]: "a"
[1, 0]: "b"
[1, 1]: "c"
sp_inputs[1]: shape = [2, 4]
[0, 1]: "d"
[0, 2]: "e"
then the output will be
shape = [2, 7]
[0, 2]: "a"
[0, 4]: "d"
[0, 5]: "e"
[1, 0]: "b"
[1, 1]: "c"
Graphically this is equivalent to doing
[ a] concat [ d e ] = [ a d e ]
[b c ] [ ] [b c ]
Another example, if 'axis = 1' and the inputs are
sp_inputs[0]: shape = [3, 3]
[0, 2]: "a"
[1, 0]: "b"
[2, 1]: "c"
sp_inputs[1]: shape = [2, 4]
[0, 1]: "d"
[0, 2]: "e"
if expand_nonconcat_dim = False, this will result in an error. But if
expand_nonconcat_dim = True, this will result in:
shape = [3, 7]
[0, 2]: "a"
[0, 4]: "d"
[0, 5]: "e"
[1, 0]: "b"
[2, 1]: "c"
Graphically this is equivalent to doing
[ a] concat [ d e ] = [ a d e ]
[b ] [ ] [b ]
[ c ] [ c ]
Args:
axis: Dimension to concatenate along. Must be in range [-rank, rank),
where rank is the number of dimensions in each input `SparseTensor`.
sp_inputs: List of `SparseTensor` to concatenate.
name: A name prefix for the returned tensors (optional).
expand_nonconcat_dim: Whether to allow the expansion in the non-concat
dimensions. Defaulted to False.
concat_dim: The old (deprecated) name for axis.
Returns:
A `SparseTensor` with the concatenated output.
Raises:
TypeError: If `sp_inputs` is not a list of `SparseTensor`.
"""
axis = deprecation.deprecated_argument_lookup("axis", axis, "concat_dim",
concat_dim)
sp_inputs = _convert_to_sparse_tensors(sp_inputs)
if len(sp_inputs) == 1: # Degenerate case of one tensor.
return sp_inputs[0]
inds = [sp_input.indices for sp_input in sp_inputs]
vals = [sp_input.values for sp_input in sp_inputs]
shapes = [sp_input.dense_shape for sp_input in sp_inputs]
if expand_nonconcat_dim:
max_shape = math_ops.reduce_max(
array_ops.concat(
[array_ops.reshape(shape, [1, -1]) for shape in shapes], 0), 0)
shapes = [
array_ops.concat([
max_shape[:axis], shape[-1:] if axis == -1 else
shape[axis:axis + 1], [] if axis == -1 else max_shape[axis + 1:]
], 0) for shape in shapes
]
output_ind, output_val, output_shape = (gen_sparse_ops._sparse_concat(
inds, vals, shapes, axis, name=name))
return sparse_tensor.SparseTensor(output_ind, output_val, output_shape)
def sparse_add(a, b, thresh=0):
"""Adds two tensors, at least one of each is a `SparseTensor`.
If one `SparseTensor` and one `Tensor` are passed in, returns a `Tensor`. If
both arguments are `SparseTensor`s, this returns a `SparseTensor`. The order
of arguments does not matter. Use vanilla `tf.add()` for adding two dense
`Tensor`s.
The indices of any input `SparseTensor` are assumed ordered in standard
lexicographic order. If this is not the case, before this step run
`SparseReorder` to restore index ordering.
If both arguments are sparse, we perform "clipping" as follows. By default,
if two values sum to zero at some index, the output `SparseTensor` would still
include that particular location in its index, storing a zero in the
corresponding value slot. To override this, callers can specify `thresh`,
indicating that if the sum has a magnitude strictly smaller than `thresh`, its
corresponding value and index would then not be included. In particular,
`thresh == 0.0` (default) means everything is kept and actual thresholding
happens only for a positive value.
For example, suppose the logical sum of two sparse operands is (densified):
[ 2]
[.1 0]
[ 6 -.2]
Then,
* `thresh == 0` (the default): all 5 index/value pairs will be returned.
* `thresh == 0.11`: only .1 and 0 will vanish, and the remaining three
index/value pairs will be returned.
* `thresh == 0.21`: .1, 0, and -.2 will vanish.
Args:
a: The first operand; `SparseTensor` or `Tensor`.
b: The second operand; `SparseTensor` or `Tensor`. At least one operand
must be sparse.
thresh: A 0-D `Tensor`. The magnitude threshold that determines if an
output value/index pair takes space. Its dtype should match that of the
values if they are real; if the latter are complex64/complex128, then the
dtype should be float32/float64, correspondingly.
Returns:
A `SparseTensor` or a `Tensor`, representing the sum.
Raises:
TypeError: If both `a` and `b` are `Tensor`s. Use `tf.add()` instead.
"""
sparse_classes = (sparse_tensor.SparseTensor, sparse_tensor.SparseTensorValue)
if not any(isinstance(inp, sparse_classes) for inp in [a, b]):
raise TypeError("At least one input should be SparseTensor; do you mean to"
" use tf.add()?")
if all(isinstance(inp, sparse_classes) for inp in [a, b]):
a = _convert_to_sparse_tensor(a)
thresh = ops.convert_to_tensor(
thresh, dtype=a.values.dtype.real_dtype, name="thresh")
output_ind, output_val, output_shape = (gen_sparse_ops._sparse_add(
a.indices, a.values, a.dense_shape,
b.indices, b.values, b.dense_shape,
thresh))
return sparse_tensor.SparseTensor(output_ind, output_val, output_shape)
else:
# swap to make `a` the SparseTensor.
if isinstance(b, sparse_classes):
a, b = b, a
return gen_sparse_ops._sparse_tensor_dense_add(
a.indices, a.values, a.dense_shape, b)
def sparse_dense_cwise_add(sp_t, dense_t):
"""Adds up a SparseTensor and a dense Tensor, using these special rules:
(1) Broadcasts the dense side to have the same shape as the sparse side, if
eligible;
(2) Then, only the dense values pointed to by the indices of the SparseTensor
participate in the cwise addition.
By the rules, the result is a logical SparseTensor with exactly the same
indices and shape, but possibly with different non-zero values. The output of
this Op is the resultant non-zero values.
Args:
sp_t: the SparseTensor operand.
dense_t: the dense Tensor operand; must have the same dtype and a
broadcast-compatible shape as `sp_t`.
Returns:
output: the SparseTensor output.
"""
result = gen_sparse_ops.sparse_dense_cwise_add(sp_t.indices, sp_t.values,
sp_t.dense_shape, dense_t)
return sparse_tensor.SparseTensor(sp_t.indices, result, sp_t.dense_shape)
def sparse_reorder(sp_input, name=None):
"""Reorders a `SparseTensor` into the canonical, row-major ordering.
Note that by convention, all sparse ops preserve the canonical ordering
along increasing dimension number. The only time ordering can be violated
is during manual manipulation of the indices and values to add entries.
Reordering does not affect the shape of the `SparseTensor`.
For example, if `sp_input` has shape `[4, 5]` and `indices` / `values`:
[0, 3]: b
[0, 1]: a
[3, 1]: d
[2, 0]: c
then the output will be a `SparseTensor` of shape `[4, 5]` and
`indices` / `values`:
[0, 1]: a
[0, 3]: b
[2, 0]: c
[3, 1]: d
Args:
sp_input: The input `SparseTensor`.
name: A name prefix for the returned tensors (optional)
Returns:
A `SparseTensor` with the same shape and non-empty values, but in
canonical ordering.
Raises:
TypeError: If `sp_input` is not a `SparseTensor`.
"""
sp_input = _convert_to_sparse_tensor(sp_input)
reordered_ind, reordered_val = (gen_sparse_ops._sparse_reorder(
sp_input.indices, sp_input.values, sp_input.dense_shape, name=name))
return sparse_tensor.SparseTensor(reordered_ind, reordered_val,
array_ops.identity(sp_input.dense_shape))
def sparse_reshape(sp_input, shape, name=None):
"""Reshapes a `SparseTensor` to represent values in a new dense shape.
This operation has the same semantics as `reshape` on the represented dense
tensor. The indices of non-empty values in `sp_input` are recomputed based
on the new dense shape, and a new `SparseTensor` is returned containing the
new indices and new shape. The order of non-empty values in `sp_input` is
unchanged.
If one component of `shape` is the special value -1, the size of that
dimension is computed so that the total dense size remains constant. At
most one component of `shape` can be -1. The number of dense elements
implied by `shape` must be the same as the number of dense elements
originally represented by `sp_input`.
For example, if `sp_input` has shape `[2, 3, 6]` and `indices` / `values`:
[0, 0, 0]: a
[0, 0, 1]: b
[0, 1, 0]: c
[1, 0, 0]: d
[1, 2, 3]: e
and `shape` is `[9, -1]`, then the output will be a `SparseTensor` of
shape `[9, 4]` and `indices` / `values`:
[0, 0]: a
[0, 1]: b
[1, 2]: c
[4, 2]: d
[8, 1]: e
Args:
sp_input: The input `SparseTensor`.
shape: A 1-D (vector) int64 `Tensor` specifying the new dense shape of the
represented `SparseTensor`.
name: A name prefix for the returned tensors (optional)
Returns:
A `SparseTensor` with the same non-empty values but with indices calculated
by the new dense shape.
Raises:
TypeError: If `sp_input` is not a `SparseTensor`.
"""
sp_input = _convert_to_sparse_tensor(sp_input)
with ops.name_scope(name, "SparseReshape", [sp_input]) as name:
reshaped_ind, reshaped_shape = gen_sparse_ops._sparse_reshape(
sp_input.indices, sp_input.dense_shape, shape, name=name)
return sparse_tensor.SparseTensor(
reshaped_ind, array_ops.identity(sp_input.values),
reshaped_shape)
# TODO(aselle): Remove keyword required once for 1.0 final
class KeywordRequired(object):
def __repr__(self):
# This is needed to make documentation without fully qualified module paths
return "KeywordRequired()"
def sparse_split(keyword_required=KeywordRequired(),
sp_input=None, num_split=None, axis=None,
name=None, split_dim=None):
"""Split a `SparseTensor` into `num_split` tensors along `axis`.
If the `sp_input.dense_shape[axis]` is not an integer multiple of `num_split`
each slice starting from 0:`shape[axis] % num_split` gets extra one
dimension. For example, if `axis = 1` and `num_split = 2` and the
input is:
input_tensor = shape = [2, 7]
[ a d e ]
[b c ]
Graphically the output tensors are:
output_tensor[0] =
[ a ]
[b c ]
output_tensor[1] =
[ d e ]
[ ]
Args:
keyword_required: Python 2 standin for * (temporary for argument reorder)
sp_input: The `SparseTensor` to split.
num_split: A Python integer. The number of ways to split.
axis: A 0-D `int32` `Tensor`. The dimension along which to split.
name: A name for the operation (optional).
split_dim: Deprecated old name for axis.
Returns:
`num_split` `SparseTensor` objects resulting from splitting `value`.
Raises:
TypeError: If `sp_input` is not a `SparseTensor`.
ValueError: If the deprecated `split_dim` and `axis` are both non None.
"""
if not isinstance(keyword_required, KeywordRequired):
raise ValueError("Keyword arguments are required for this function.")
if sp_input is None:
raise ValueError("sp_input is required")
if num_split is None:
raise ValueError("num_split is required")
if axis is None:
raise ValueError("axis is required")
axis = deprecation.deprecated_argument_lookup("axis", axis, "split_dim",
split_dim)
sp_input = _convert_to_sparse_tensor(sp_input)
output_inds, output_vals, output_shapes = (gen_sparse_ops._sparse_split(
axis,
sp_input.indices,
sp_input.values,
sp_input.dense_shape,
num_split,
name=name))
sparse_tensors = []
for i in range(0, num_split):
sparse_tensors.append(
sparse_tensor.SparseTensor(
output_inds[i], output_vals[i], output_shapes[i]))
return sparse_tensors
def sparse_to_dense(sparse_indices,
output_shape,
sparse_values,
default_value=0,
validate_indices=True,
name=None):
"""Converts a sparse representation into a dense tensor.
Builds an array `dense` with shape `output_shape` such that
```python
# If sparse_indices is scalar
dense[i] = (i == sparse_indices ? sparse_values : default_value)
# If sparse_indices is a vector, then for each i
dense[sparse_indices[i]] = sparse_values[i]
# If sparse_indices is an n by d matrix, then for each i in [0, n)
dense[sparse_indices[i][0], ..., sparse_indices[i][d-1]] = sparse_values[i]
```
All other values in `dense` are set to `default_value`. If `sparse_values`
is a scalar, all sparse indices are set to this single value.
Indices should be sorted in lexicographic order, and indices must not
contain any repeats. If `validate_indices` is True, these properties
are checked during execution.
Args:
sparse_indices: A 0-D, 1-D, or 2-D `Tensor` of type `int32` or `int64`.
`sparse_indices[i]` contains the complete index where `sparse_values[i]`
will be placed.
output_shape: A 1-D `Tensor` of the same type as `sparse_indices`. Shape
of the dense output tensor.
sparse_values: A 0-D or 1-D `Tensor`. Values corresponding to each row of
`sparse_indices`, or a scalar value to be used for all sparse indices.
default_value: A 0-D `Tensor` of the same type as `sparse_values`. Value
to set for indices not specified in `sparse_indices`. Defaults to zero.
validate_indices: A boolean value. If True, indices are checked to make
sure they are sorted in lexicographic order and that there are no repeats.
name: A name for the operation (optional).
Returns:
Dense `Tensor` of shape `output_shape`. Has the same type as
`sparse_values`.
"""
return gen_sparse_ops._sparse_to_dense(
sparse_indices,
output_shape,
sparse_values,
default_value=default_value,
validate_indices=validate_indices,
name=name)
def sparse_reduce_sum(sp_input, axis=None, keep_dims=False,
reduction_axes=None):
"""Computes the sum of elements across dimensions of a SparseTensor.
This Op takes a SparseTensor and is the sparse counterpart to
`tf.reduce_sum()`. In particular, this Op also returns a dense `Tensor`
instead of a sparse one.
Reduces `sp_input` along the dimensions given in `reduction_axes`. Unless
`keep_dims` is true, the rank of the tensor is reduced by 1 for each entry in
`reduction_axes`. If `keep_dims` is true, the reduced dimensions are retained
with length 1.
If `reduction_axes` has no entries, all dimensions are reduced, and a tensor
with a single element is returned. Additionally, the axes can be negative,
similar to the indexing rules in Python.
For example:
```python
# 'x' represents [[1, ?, 1]
# [?, 1, ?]]
# where ? is implicitly-zero.
tf.sparse_reduce_sum(x) ==> 3
tf.sparse_reduce_sum(x, 0) ==> [1, 1, 1]
tf.sparse_reduce_sum(x, 1) ==> [2, 1] # Can also use -1 as the axis.
tf.sparse_reduce_sum(x, 1, keep_dims=True) ==> [[2], [1]]
tf.sparse_reduce_sum(x, [0, 1]) ==> 3
```
Args:
sp_input: The SparseTensor to reduce. Should have numeric type.
axis: The dimensions to reduce; list or scalar. If `None` (the
default), reduces all dimensions.
keep_dims: If true, retain reduced dimensions with length 1.
reduction_axes: Deprecated name of axis.
Returns:
The reduced Tensor.
"""
return gen_sparse_ops.sparse_reduce_sum(
sp_input.indices, sp_input.values,
sp_input.dense_shape,
math_ops._ReductionDims(sp_input, axis, reduction_axes),
keep_dims)
def sparse_reduce_sum_sparse(sp_input, axis=None, keep_dims=False,
reduction_axes=None):
"""Computes the sum of elements across dimensions of a SparseTensor.
This Op takes a SparseTensor and is the sparse counterpart to
`tf.reduce_sum()`. In contrast to SparseReduceSum, this Op returns a
SparseTensor.
Reduces `sp_input` along the dimensions given in `reduction_axes`. Unless
`keep_dims` is true, the rank of the tensor is reduced by 1 for each entry in
`reduction_axes`. If `keep_dims` is true, the reduced dimensions are retained
with length 1.
If `reduction_axes` has no entries, all dimensions are reduced, and a tensor
with a single element is returned. Additionally, the axes can be negative,
which are interpreted according to the indexing rules in Python.
Args:
sp_input: The SparseTensor to reduce. Should have numeric type.
axis: The dimensions to reduce; list or scalar. If `None` (the
default), reduces all dimensions.
keep_dims: If true, retain reduced dimensions with length 1.
reduction_axes: Deprecated name of axis
Returns:
The reduced SparseTensor.
"""
output_ind, output_val, output_shape = (
gen_sparse_ops.sparse_reduce_sum_sparse(
sp_input.indices, sp_input.values,
sp_input.dense_shape, math_ops._ReductionDims(sp_input, axis,
reduction_axes),
keep_dims))
return sparse_tensor.SparseTensor(output_ind, output_val, output_shape)
def sparse_tensor_to_dense(sp_input,
default_value=0,
validate_indices=True,
name=None):
"""Converts a `SparseTensor` into a dense tensor.
This op is a convenience wrapper around `sparse_to_dense` for `SparseTensor`s.
For example, if `sp_input` has shape `[3, 5]` and non-empty string values:
[0, 1]: a
[0, 3]: b
[2, 0]: c
and `default_value` is `x`, then the output will be a dense `[3, 5]`
string tensor with values:
[[x a x b x]
[x x x x x]
[c x x x x]]
Indices must be without repeats. This is only
tested if validate_indices is True.
Args:
sp_input: The input `SparseTensor`.
default_value: Scalar value to set for indices not specified in
`sp_input`. Defaults to zero.
validate_indices: A boolean value. If `True`, indices are checked to make
sure they are sorted in lexicographic order and that there are no repeats.
name: A name prefix for the returned tensors (optional).
Returns:
A dense tensor with shape `sp_input.dense_shape` and values specified by
the non-empty values in `sp_input`. Indices not in `sp_input` are assigned
`default_value`.
Raises:
TypeError: If `sp_input` is not a `SparseTensor`.
"""
sp_input = _convert_to_sparse_tensor(sp_input)
return sparse_to_dense(
sp_input.indices,
sp_input.dense_shape,
sp_input.values,
default_value=default_value,
validate_indices=validate_indices,
name=name)
def sparse_to_indicator(sp_input, vocab_size, name=None):
"""Converts a `SparseTensor` of ids into a dense bool indicator tensor.
The last dimension of `sp_input.indices` is discarded and replaced with
the values of `sp_input`. If `sp_input.dense_shape = [D0, D1, ..., Dn, K]`,
then `output.shape = [D0, D1, ..., Dn, vocab_size]`, where
output[d_0, d_1, ..., d_n, sp_input[d_0, d_1, ..., d_n, k]] = True
and False elsewhere in `output`.
For example, if `sp_input.dense_shape = [2, 3, 4]` with non-empty values:
[0, 0, 0]: 0
[0, 1, 0]: 10
[1, 0, 3]: 103
[1, 1, 2]: 150
[1, 1, 3]: 149
[1, 1, 4]: 150
[1, 2, 1]: 121
and `vocab_size = 200`, then the output will be a `[2, 3, 200]` dense bool
tensor with False everywhere except at positions
(0, 0, 0), (0, 1, 10), (1, 0, 103), (1, 1, 149), (1, 1, 150),
(1, 2, 121).
Note that repeats are allowed in the input SparseTensor.
This op is useful for converting `SparseTensor`s into dense formats for
compatibility with ops that expect dense tensors.
The input `SparseTensor` must be in row-major order.
Args:
sp_input: A `SparseTensor` with `values` property of type `int32` or
`int64`.
vocab_size: A scalar int64 Tensor (or Python int) containing the new size
of the last dimension, `all(0 <= sp_input.values < vocab_size)`.
name: A name prefix for the returned tensors (optional)
Returns:
A dense bool indicator tensor representing the indices with specified value.
Raises:
TypeError: If `sp_input` is not a `SparseTensor`.
"""
sp_input = _convert_to_sparse_tensor(sp_input)
with ops.name_scope(name, "SparseToIndicator", [sp_input]) as name:
num_entries = array_ops.shape(sp_input.indices)[0]
new_values = array_ops.fill(array_ops.expand_dims(num_entries, 0), True)
sp_values = sparse_tensor.SparseTensor(
sp_input.indices, new_values, sp_input.dense_shape)
sp_new = sparse_merge(sp_input, sp_values, vocab_size, name)
# validate_indices may be False because we allow duplicates in new_indices:
# repeated indices are allowed when creating an indicator matrix.
return sparse_tensor_to_dense(
sp_new, default_value=False, validate_indices=False, name=name)
def sparse_merge(sp_ids, sp_values, vocab_size, name=None,
already_sorted=False):
"""Combines a batch of feature ids and values into a single `SparseTensor`.
The most common use case for this function occurs when feature ids and
their corresponding values are stored in `Example` protos on disk.
`parse_example` will return a batch of ids and a batch of values, and this
function joins them into a single logical `SparseTensor` for use in
functions such as `sparse_tensor_dense_matmul`, `sparse_to_dense`, etc.
The `SparseTensor` returned by this function has the following properties:
- `indices` is equivalent to `sp_ids.indices` with the last
dimension discarded and replaced with `sp_ids.values`.
- `values` is simply `sp_values.values`.
- If `sp_ids.dense_shape = [D0, D1, ..., Dn, K]`, then
`output.shape = [D0, D1, ..., Dn, vocab_size]`.
For example, consider the following feature vectors:
```python
vector1 = [-3, 0, 0, 0, 0, 0]
vector2 = [ 0, 1, 0, 4, 1, 0]
vector3 = [ 5, 0, 0, 9, 0, 0]
```
These might be stored sparsely in the following Example protos by storing
only the feature ids (column number if the vectors are treated as a matrix)
of the non-zero elements and the corresponding values:
```python
examples = [Example(features={
"ids": Feature(int64_list=Int64List(value=[0])),
"values": Feature(float_list=FloatList(value=[-3]))}),
Example(features={
"ids": Feature(int64_list=Int64List(value=[1, 4, 3])),
"values": Feature(float_list=FloatList(value=[1, 1, 4]))}),
Example(features={
"ids": Feature(int64_list=Int64List(value=[0, 3])),
"values": Feature(float_list=FloatList(value=[5, 9]))})]
```
The result of calling parse_example on these examples will produce a
dictionary with entries for "ids" and "values". Passing those two objects
to this function along with vocab_size=6, will produce a `SparseTensor` that
sparsely represents all three instances. Namely, the `indices` property will
contain the coordinates of the non-zero entries in the feature matrix (the
first dimension is the row number in the matrix, i.e., the index within the
batch, and the second dimension is the column number, i.e., the feature id);
`values` will contain the actual values. `shape` will be the shape of the
original matrix, i.e., (3, 6). For our example above, the output will be
equal to:
```python
SparseTensor(indices=[[0, 0], [1, 1], [1, 3], [1, 4], [2, 0], [2, 3]],
values=[-3, 1, 4, 1, 5, 9],
dense_shape=[3, 6])
```
Args:
sp_ids: A `SparseTensor` with `values` property of type `int32`
or `int64`.
sp_values: A`SparseTensor` of any type.
vocab_size: A scalar `int64` Tensor (or Python int) containing the new size
of the last dimension, `all(0 <= sp_ids.values < vocab_size)`.
name: A name prefix for the returned tensors (optional)
already_sorted: A boolean to specify whether the per-batch values in
`sp_values` are already sorted. If so skip sorting, False by default
(optional).
Returns:
A `SparseTensor` compactly representing a batch of feature ids and values,
useful for passing to functions that expect such a `SparseTensor`.
Raises:
TypeError: If `sp_ids` or `sp_values` are not a `SparseTensor`.
"""
sp_ids = _convert_to_sparse_tensor(sp_ids)
sp_values = _convert_to_sparse_tensor(sp_values)
with ops.name_scope(name, "SparseMerge", [sp_ids, sp_values]):
indices_shape = array_ops.shape(sp_ids.indices)
rank = indices_shape[1]
ids = sp_ids.values
if ids.dtype != dtypes.int64:
ids = math_ops.cast(ids, dtypes.int64)
# Slice off the last dimension of indices, then tack on the ids
indices_columns_to_preserve = array_ops.slice(
sp_ids.indices, [0, 0], array_ops.stack([-1, rank - 1]))
new_indices = array_ops.concat(
[indices_columns_to_preserve, array_ops.reshape(ids, [-1, 1])], 1)
new_values = sp_values.values
new_shape = array_ops.concat([
array_ops.slice(sp_ids.dense_shape, [0],
array_ops.expand_dims(rank - 1, 0)),
math_ops.cast(array_ops.stack([vocab_size]), dtypes.int64)
], 0)
result = sparse_tensor.SparseTensor(new_indices, new_values, new_shape)
return result if already_sorted else sparse_reorder(result)
def sparse_retain(sp_input, to_retain):
"""Retains specified non-empty values within a `SparseTensor`.
For example, if `sp_input` has shape `[4, 5]` and 4 non-empty string values:
[0, 1]: a
[0, 3]: b
[2, 0]: c
[3, 1]: d
and `to_retain = [True, False, False, True]`, then the output will
be a `SparseTensor` of shape `[4, 5]` with 2 non-empty values:
[0, 1]: a
[3, 1]: d
Args:
sp_input: The input `SparseTensor` with `N` non-empty elements.
to_retain: A bool vector of length `N` with `M` true values.
Returns:
A `SparseTensor` with the same shape as the input and `M` non-empty
elements corresponding to the true positions in `to_retain`.
Raises:
TypeError: If `sp_input` is not a `SparseTensor`.
"""
sp_input = _convert_to_sparse_tensor(sp_input)
to_retain = ops.convert_to_tensor(to_retain)
# Shape checking, if shape is known at graph construction time
retain_shape = to_retain.get_shape()
retain_shape.assert_has_rank(1)
sp_input.values.get_shape()[0].merge_with(retain_shape[0])
where_true = array_ops.reshape(array_ops.where(to_retain), [-1])
new_indices = array_ops.gather(sp_input.indices, where_true)
new_values = array_ops.gather(sp_input.values, where_true)
return sparse_tensor.SparseTensor(new_indices, new_values,
array_ops.identity(sp_input.dense_shape))
def sparse_reset_shape(sp_input, new_shape=None):
"""Resets the shape of a `SparseTensor` with indices and values unchanged.
If `new_shape` is None, returns a copy of `sp_input` with its shape reset
to the tight bounding box of `sp_input`.
If `new_shape` is provided, then it must be larger or equal in all dimensions
compared to the shape of `sp_input`. When this condition is met, the returned
SparseTensor will have its shape reset to `new_shape` and its indices and
values unchanged from that of `sp_input.`
For example:
Consider a `sp_input` with shape [2, 3, 5]:
[0, 0, 1]: a
[0, 1, 0]: b
[0, 2, 2]: c
[1, 0, 3]: d
- It is an error to set `new_shape` as [3, 7] since this represents a
rank-2 tensor while `sp_input` is rank-3. This is either a ValueError
during graph construction (if both shapes are known) or an OpError during
run time.
- Setting `new_shape` as [2, 3, 6] will be fine as this shape is larger or
equal in every dimension compared to the original shape [2, 3, 5].
- On the other hand, setting new_shape as [2, 3, 4] is also an error: The
third dimension is smaller than the original shape [2, 3, 5] (and an
`InvalidArgumentError` will be raised).
- If `new_shape` is None, the returned SparseTensor will have a shape
[2, 3, 4], which is the tight bounding box of `sp_input`.
Args:
sp_input: The input `SparseTensor`.
new_shape: None or a vector representing the new shape for the returned
`SparseTensor`.
Returns:
A `SparseTensor` indices and values unchanged from `input_sp`. Its shape is
`new_shape` if that is set. Otherwise it is the tight bounding box of
`input_sp`
Raises:
TypeError: If `sp_input` is not a `SparseTensor`.
ValueError: If `new_shape` represents a tensor with a different rank from
that of `sp_input` (if shapes are known when graph is constructed).
OpError:
- If `new_shape` has dimension sizes that are too small.
- If shapes are not known during graph construction time, and during run
time it is found out that the ranks do not match.
"""
sp_input = _convert_to_sparse_tensor(sp_input)
in_indices = array_ops.identity(sp_input.indices)
in_values = array_ops.identity(sp_input.values)
in_shape = array_ops.identity(sp_input.dense_shape)
if new_shape is None:
dim_low_bound = math_ops.reduce_max(in_indices, 0)
output_shape_tensor = math_ops.add(dim_low_bound,
array_ops.ones_like(in_shape))
else:
output_shape_tensor = ops.convert_to_tensor(new_shape)
output_shape_tensor.get_shape().assert_has_rank(1)
output_shape_tensor = math_ops.cast(output_shape_tensor, dtypes.int64)
# For cases when shape is known during graph construction, this catches the
# error before the sparse_tensor.SparseTensor catches it.
output_shape_tensor.get_shape()[0].merge_with(in_shape.get_shape()[0])
# For cases where shape is not known during graph construction.
output_shape_tensor = control_flow_ops.with_dependencies(
[check_ops.assert_equal(
array_ops.shape(in_shape), array_ops.shape(output_shape_tensor))],
output_shape_tensor)
output_shape_tensor = control_flow_ops.with_dependencies(
[check_ops.assert_less_equal(in_shape, output_shape_tensor)],
output_shape_tensor)
return sparse_tensor.SparseTensor(in_indices, in_values, output_shape_tensor)
def sparse_fill_empty_rows(sp_input, default_value, name=None):
"""Fills empty rows in the input 2-D `SparseTensor` with a default value.
This op adds entries with the specified `default_value` at index
`[row, 0]` for any row in the input that does not already have a value.
For example, suppose `sp_input` has shape `[5, 6]` and non-empty values:
[0, 1]: a
[0, 3]: b
[2, 0]: c
[3, 1]: d
Rows 1 and 4 are empty, so the output will be of shape `[5, 6]` with values: