/
lookup_ops.py
2462 lines (2055 loc) · 91.4 KB
/
lookup_ops.py
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 2017 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.
# ==============================================================================
"""Lookup operations."""
# pylint: disable=g-bad-name
import collections
import functools
import uuid
from tensorflow.python.checkpoint import saveable_compat
from tensorflow.python.eager import context
from tensorflow.python.framework import constant_op
from tensorflow.python.framework import dtypes
from tensorflow.python.framework import ops
from tensorflow.python.framework import sparse_tensor
from tensorflow.python.framework import tensor as tensor_lib
from tensorflow.python.framework import tensor_shape
from tensorflow.python.framework import tensor_util
from tensorflow.python.ops import array_ops
from tensorflow.python.ops import control_flow_ops
from tensorflow.python.ops import gen_lookup_ops
# Ensure lookup gradients are registered
from tensorflow.python.ops import lookup_grad # pylint: disable=unused-import
from tensorflow.python.ops import math_ops
from tensorflow.python.ops import string_ops
# go/tf-wildcard-import
# pylint: disable=wildcard-import
from tensorflow.python.ops.gen_lookup_ops import *
from tensorflow.python.saved_model import registration
from tensorflow.python.trackable import asset
# pylint: enable=wildcard-import
from tensorflow.python.trackable import base as trackable_base
from tensorflow.python.trackable import resource
from tensorflow.python.training.saver import BaseSaverBuilder
from tensorflow.python.types import internal
from tensorflow.python.util import compat as compat_util
from tensorflow.python.util.deprecation import deprecated
from tensorflow.python.util.tf_export import tf_export
@tf_export(v1=["initialize_all_tables"])
@deprecated(None, "Use `tf.tables_initializer` instead.")
def initialize_all_tables(name="init_all_tables"):
"""Returns an Op that initializes all tables of the default graph.
Args:
name: Optional name for the initialization op.
Returns:
An Op that initializes all tables. Note that if there are
not tables the returned Op is a NoOp.
"""
return tables_initializer(name)
@tf_export(v1=["initializers.tables_initializer", "tables_initializer"])
def tables_initializer(name="init_all_tables"):
"""Returns an Op that initializes all tables of the default graph.
Args:
name: Optional name for the initialization op.
Returns:
An Op that initializes all tables. Note that if there are
not tables the returned Op is a NoOp.
@compatibility(TF2)
`tf.compat.v1.tables_initializer` is no longer needed with eager execution and
`tf.function`. In TF2, when creating an initializable table like a
`tf.lookup.StaticHashTable`, the table will automatically be initialized on
creation.
#### Before & After Usage Example
Before:
>>> with tf.compat.v1.Session():
... init = tf.compat.v1.lookup.KeyValueTensorInitializer(['a', 'b'], [1, 2])
... table = tf.compat.v1.lookup.StaticHashTable(init, default_value=-1)
... tf.compat.v1.tables_initializer().run()
... result = table.lookup(tf.constant(['a', 'c'])).eval()
>>> result
array([ 1, -1], dtype=int32)
After:
>>> init = tf.lookup.KeyValueTensorInitializer(['a', 'b'], [1, 2])
>>> table = tf.lookup.StaticHashTable(init, default_value=-1)
>>> table.lookup(tf.constant(['a', 'c'])).numpy()
array([ 1, -1], dtype=int32)
@end_compatibility
"""
initializers = ops.get_collection(ops.GraphKeys.TABLE_INITIALIZERS)
if initializers:
return control_flow_ops.group(*initializers, name=name)
return control_flow_ops.no_op(name=name)
def check_table_dtypes(table, key_dtype, value_dtype):
"""Check that the given key_dtype and value_dtype matches the table dtypes.
Args:
table: The table to check types against to.
key_dtype: The key data type to check.
value_dtype: The value data type to check.
Raises:
TypeError: when 'key_dtype' or 'value_dtype' doesn't match the table data
types.
"""
if key_dtype.base_dtype != table.key_dtype:
raise TypeError(f"Invalid key dtype for table, expected {table.key_dtype} "
f"but got {key_dtype}.")
if value_dtype.base_dtype != table.value_dtype:
raise TypeError("Invalid value dtype for table, expected "
f"{table.value_dtype} but got {value_dtype}.")
class LookupInterface(resource.TrackableResource):
"""Represent a lookup table that persists across different steps."""
def __init__(self, key_dtype, value_dtype):
"""Construct a lookup table interface.
Args:
key_dtype: The table key type.
value_dtype: The table value type.
"""
self._key_dtype = dtypes.as_dtype(key_dtype)
self._value_dtype = dtypes.as_dtype(value_dtype)
super(LookupInterface, self).__init__()
def _create_resource(self):
raise NotImplementedError
@property
def key_dtype(self):
"""The table key dtype."""
return self._key_dtype
@property
def value_dtype(self):
"""The table value dtype."""
return self._value_dtype
@property
def name(self):
"""The name of the table."""
return NotImplementedError
def size(self, name=None):
"""Compute the number of elements in this table."""
raise NotImplementedError
def lookup(self, keys, name=None):
"""Looks up `keys` in a table, outputs the corresponding values."""
raise NotImplementedError
def __getitem__(self, keys):
"""Looks up `keys` in a table, outputs the corresponding values."""
return self.lookup(keys)
class InitializableLookupTableBase(LookupInterface):
"""Initializable lookup table interface.
An initializable lookup tables persist across different steps.
"""
def __init__(self, default_value, initializer):
"""Construct a table object from a table reference.
If requires a table initializer object (subclass of `TableInitializerBase`).
It provides the table key and value types, as well as the op to initialize
the table. The caller is responsible to execute the initialization op.
Args:
default_value: The value to use if a key is missing in the table.
initializer: The table initializer to use.
"""
super(InitializableLookupTableBase, self).__init__(initializer.key_dtype,
initializer.value_dtype)
self._default_value = ops.convert_to_tensor(
default_value, dtype=self._value_dtype)
self._default_value.get_shape().merge_with(tensor_shape.TensorShape([]))
if isinstance(initializer, trackable_base.Trackable):
self._initializer = self._track_trackable(initializer, "_initializer")
with ops.init_scope():
self._resource_handle = self._create_resource()
if (not context.executing_eagerly() and
ops.get_default_graph()._get_control_flow_context() is not None): # pylint: disable=protected-access
with ops.init_scope():
self._init_op = self._initialize()
else:
self._init_op = self._initialize()
def _initialize(self):
return self._initializer.initialize(self)
@property
def default_value(self):
"""The default value of the table."""
return self._default_value
def size(self, name=None):
"""Compute the number of elements in this table.
Args:
name: A name for the operation (optional).
Returns:
A scalar tensor containing the number of elements in this table.
"""
with ops.name_scope(name, "%s_Size" % self.name, [self.resource_handle]):
return gen_lookup_ops.lookup_table_size_v2(self.resource_handle)
def lookup(self, keys, name=None):
"""Looks up `keys` in a table, outputs the corresponding values.
The `default_value` is used for keys not present in the table.
Args:
keys: Keys to look up. May be either a `SparseTensor` or dense `Tensor`.
name: A name for the operation (optional).
Returns:
A `SparseTensor` if keys are sparse, a `RaggedTensor` if keys are ragged,
otherwise a dense `Tensor`.
Raises:
TypeError: when `keys` or `default_value` doesn't match the table data
types.
"""
key_tensor = keys
# TODO(b/296302236): Remove RaggedTensor check by adding ragged
# dispatching.
if isinstance(keys, (sparse_tensor.SparseTensor, internal.RaggedTensor)):
key_tensor = keys.values
if keys.dtype.base_dtype != self._key_dtype:
raise TypeError(f"Dtype of argument `keys` must be {self._key_dtype}, "
f"received: {keys.dtype}")
with ops.name_scope(
name, "%s_Lookup" % self.name,
(self.resource_handle, key_tensor, self._default_value)):
values = gen_lookup_ops.lookup_table_find_v2(self.resource_handle,
key_tensor,
self._default_value)
values.set_shape(key_tensor.get_shape())
if isinstance(keys, sparse_tensor.SparseTensor):
return sparse_tensor.SparseTensor(keys.indices, values, keys.dense_shape)
# TODO(b/296302236): Remove RaggedTensor check by adding ragged
# dispatching.
elif isinstance(keys, internal.RaggedTensor):
return keys.with_values(values)
else:
return values
class InitializableLookupTableBaseV1(InitializableLookupTableBase):
@property
def initializer(self):
return self._init_op
@registration.register_tf_serializable(
predicate=lambda obj: isinstance(obj, StaticHashTable))
@tf_export("lookup.StaticHashTable", v1=[])
class StaticHashTable(InitializableLookupTableBase):
"""A generic hash table that is immutable once initialized.
Example usage:
>>> keys_tensor = tf.constant(['a', 'b', 'c'])
>>> vals_tensor = tf.constant([7, 8, 9])
>>> input_tensor = tf.constant(['a', 'f'])
>>> table = tf.lookup.StaticHashTable(
... tf.lookup.KeyValueTensorInitializer(keys_tensor, vals_tensor),
... default_value=-1)
>>> table.lookup(input_tensor).numpy()
array([ 7, -1], dtype=int32)
Or for more pythonic code:
>>> table[input_tensor].numpy()
array([ 7, -1], dtype=int32)
The result of a lookup operation has the same shape as the argument:
>>> input_tensor = tf.constant([['a', 'b'], ['c', 'd']])
>>> table[input_tensor].numpy()
array([[ 7, 8],
[ 9, -1]], dtype=int32)
"""
def __init__(self,
initializer,
default_value,
name=None,
experimental_is_anonymous=False):
"""Creates a non-initialized `HashTable` object.
Creates a table, the type of its keys and values are specified by the
initializer.
Before using the table you will have to initialize it. After initialization
the table will be immutable.
Args:
initializer: The table initializer to use. See `HashTable` kernel for
supported key and value types.
default_value: The value to use if a key is missing in the table.
name: A name for the operation (optional).
experimental_is_anonymous: Whether to use anonymous mode for the
table (default is False). In anonymous mode, the table
resource can only be accessed via a resource handle. It can't
be looked up by a name. When all resource handles pointing to
that resource are gone, the resource will be deleted
automatically.
Returns:
A `HashTable` object.
"""
self._initializer = initializer
self._default_value = default_value
self._is_anonymous = experimental_is_anonymous
if not self._is_anonymous:
self._shared_name = self._initializer._shared_name # pylint: disable=protected-access
if not self._shared_name:
# Force using a shared name so that StaticHashTable resources can be
# shared across different kernels. If no "shared_name" is set and
# "use_node_name_sharing" is False, then each kernel gets its own local
# resource.
self._shared_name = "hash_table_%s" % (str(uuid.uuid4()),)
self._name = name or "hash_table"
self._table_name = None
super(StaticHashTable, self).__init__(default_value, initializer)
self._value_shape = self._default_value.get_shape()
def _create_resource(self):
if self._is_anonymous:
table_ref = gen_lookup_ops.anonymous_hash_table(
key_dtype=self._initializer.key_dtype,
value_dtype=self._initializer.value_dtype,
name=self._name)
else:
table_ref = gen_lookup_ops.hash_table_v2(
shared_name=self._shared_name,
key_dtype=self._initializer.key_dtype,
value_dtype=self._initializer.value_dtype,
name=self._name)
if context.executing_eagerly():
self._table_name = None
else:
self._table_name = table_ref.op.name.split("/")[-1]
return table_ref
@property
def name(self):
return self._table_name
def export(self, name=None):
"""Returns tensors of all keys and values in the table.
Args:
name: A name for the operation (optional).
Returns:
A pair of tensors with the first tensor containing all keys and the
second tensors containing all values in the table.
"""
with ops.name_scope(name, "%s_Export" % self.name, [self.resource_handle]):
exported_keys, exported_values = gen_lookup_ops.lookup_table_export_v2(
self.resource_handle, self._key_dtype, self._value_dtype)
exported_values.set_shape(exported_keys.get_shape().concatenate(
self._value_shape))
return exported_keys, exported_values
def _serialize_to_proto(self, **unused_kwargs):
return None
def _add_trackable_child(self, name, value):
setattr(self, name, value)
if isinstance(value, trackable_base.Trackable):
self._track_trackable(value, name) # pylint:disable=protected-access
@classmethod
def _deserialize_from_proto(cls, **kwargs):
class _RestoredStaticHashTable(resource.RestoredResource): # pylint: disable=protected-access
@classmethod
def _resource_type(cls):
return "RestoredStaticHashTable"
return _RestoredStaticHashTable._deserialize_from_proto(**kwargs) # pylint: disable=protected-access
@tf_export(v1=["lookup.StaticHashTable"])
class StaticHashTableV1(StaticHashTable):
"""A generic hash table that is immutable once initialized.
When running in graph mode, you must evaluate the tensor returned by
`tf.tables_initializer()` before evaluating the tensor returned by
this class's `lookup()` method. Example usage in graph mode:
```python
keys_tensor = tf.constant([1, 2])
vals_tensor = tf.constant([3, 4])
input_tensor = tf.constant([1, 5])
table = tf.lookup.StaticHashTable(
tf.lookup.KeyValueTensorInitializer(keys_tensor, vals_tensor), -1)
out = table.lookup(input_tensor)
with tf.Session() as sess:
sess.run(tf.tables_initializer())
print(sess.run(out))
```
Note that in graph mode if you set `experimental_is_anonymous` to
`True`, you should only call `Session.run` once, otherwise each
`Session.run` will create (and destroy) a new table unrelated to
each other, leading to errors such as "Table not initialized".
You can do so like this:
```python
keys_tensor = tf.constant([1, 2])
vals_tensor = tf.constant([3, 4])
input_tensor = tf.constant([1, 5])
table = tf.lookup.StaticHashTable(
tf.lookup.KeyValueTensorInitializer(keys_tensor, vals_tensor), -1,
experimental_is_anonymous=True)
with tf.control_dependencies([tf.tables_initializer()]):
out = table.lookup(input_tensor)
with tf.Session() as sess:
print(sess.run(out))
```
In eager mode, no special code is needed to initialize the table.
Example usage in eager mode:
```python
tf.enable_eager_execution()
keys_tensor = tf.constant([1, 2])
vals_tensor = tf.constant([3, 4])
input_tensor = tf.constant([1, 5])
table = tf.lookup.StaticHashTable(
tf.lookup.KeyValueTensorInitializer(keys_tensor, vals_tensor), -1)
print(table.lookup(input_tensor))
```
"""
@property
def initializer(self):
return self._init_op
# For backwards compatibility. This will be removed in TF 2.0.
class HashTable(StaticHashTableV1):
@property
def init(self):
return self.initializer
class TableInitializerBase(trackable_base.Trackable):
"""Base class for lookup table initializers."""
def __init__(self, key_dtype, value_dtype):
"""Construct a table initializer object.
Args:
key_dtype: Type of the table keys.
value_dtype: Type of the table values.
"""
self._key_dtype = dtypes.as_dtype(key_dtype)
self._value_dtype = dtypes.as_dtype(value_dtype)
@property
def key_dtype(self):
"""The expected table key dtype."""
return self._key_dtype
@property
def value_dtype(self):
"""The expected table value dtype."""
return self._value_dtype
def initialize(self, table):
"""Returns the table initialization op."""
raise NotImplementedError
@property
def _shared_name(self):
"""Returns a shared name to be used by the table."""
shared_name = ""
if context.executing_eagerly():
# Ensure a unique name when eager execution is enabled to avoid spurious
# sharing issues.
# TODO(rohanj): Use context.anonymous_name() instead.
shared_name += str(ops.uid())
return shared_name
@tf_export("lookup.KeyValueTensorInitializer")
class KeyValueTensorInitializer(TableInitializerBase):
"""Table initializers given `keys` and `values` tensors.
>>> keys_tensor = tf.constant(['a', 'b', 'c'])
>>> vals_tensor = tf.constant([7, 8, 9])
>>> input_tensor = tf.constant(['a', 'f'])
>>> init = tf.lookup.KeyValueTensorInitializer(keys_tensor, vals_tensor)
>>> table = tf.lookup.StaticHashTable(
... init,
... default_value=-1)
>>> table.lookup(input_tensor).numpy()
array([ 7, -1], dtype=int32)
"""
def __init__(self, keys, values, key_dtype=None, value_dtype=None, name=None):
"""Constructs a table initializer object based on keys and values tensors.
Args:
keys: The tensor for the keys.
values: The tensor for the values.
key_dtype: The `keys` data type. Used when `keys` is a python array.
value_dtype: The `values` data type. Used when `values` is a python array.
name: A name for the operation (optional).
"""
if (not context.executing_eagerly() and
ops.get_default_graph()._get_control_flow_context() is not None): # pylint: disable=protected-access
with ops.init_scope():
self._keys = ops.convert_to_tensor(keys, dtype=key_dtype, name="keys")
self._values = ops.convert_to_tensor(
values, dtype=value_dtype, name="values")
else:
self._keys = ops.convert_to_tensor(keys, dtype=key_dtype, name="keys")
self._values = ops.convert_to_tensor(
values, dtype=value_dtype, name="values")
self._name = name if name is not None else "key_value_init"
if context.executing_eagerly():
# Ensure a unique name when eager execution is enabled to avoid spurious
# sharing issues.
# TODO(rohanj): Use context.anonymous_name() instead.
self._name += str(ops.uid())
super(KeyValueTensorInitializer, self).__init__(self._keys.dtype,
self._values.dtype)
def initialize(self, table):
"""Initializes the given `table` with `keys` and `values` tensors.
Args:
table: The table to initialize.
Returns:
The operation that initializes the table.
Raises:
TypeError: when the keys and values data types do not match the table
key and value data types.
"""
check_table_dtypes(table, self._keys.dtype, self._values.dtype)
with ops.name_scope(
self._name, values=(table.resource_handle, self._keys, self._values)):
init_op = gen_lookup_ops.lookup_table_import_v2(table.resource_handle,
self._keys, self._values)
ops.add_to_collection(ops.GraphKeys.TABLE_INITIALIZERS, init_op)
return init_op
@tf_export("lookup.TextFileIndex")
class TextFileIndex:
"""The key and value content to get from each line.
This class defines the key and value used for `tf.lookup.TextFileInitializer`.
The key and value content to get from each line is specified either
by the following, or a value `>=0`.
* `TextFileIndex.LINE_NUMBER` means use the line number starting from zero,
expects data type int64.
* `TextFileIndex.WHOLE_LINE` means use the whole line content, expects data
type string.
A value `>=0` means use the index (starting at zero) of the split line based
on `delimiter`.
"""
WHOLE_LINE = -2
LINE_NUMBER = -1
@tf_export("lookup.TextFileInitializer")
class TextFileInitializer(TableInitializerBase):
r"""Table initializers from a text file.
This initializer assigns one entry in the table for each line in the file.
The key and value type of the table to initialize is given by `key_dtype` and
`value_dtype`.
The key and value content to get from each line is specified by
the `key_index` and `value_index`.
* `TextFileIndex.LINE_NUMBER` means use the line number starting from zero,
expects data type int64.
* `TextFileIndex.WHOLE_LINE` means use the whole line content, expects data
type string.
* A value `>=0` means use the index (starting at zero) of the split line based
on `delimiter`.
For example if we have a file with the following content:
>>> import tempfile
>>> f = tempfile.NamedTemporaryFile(delete=False)
>>> content='\n'.join(["emerson 10", "lake 20", "palmer 30",])
>>> f.file.write(content.encode('utf-8'))
>>> f.file.close()
The following snippet initializes a table with the first column as keys and
second column as values:
* `emerson -> 10`
* `lake -> 20`
* `palmer -> 30`
>>> init= tf.lookup.TextFileInitializer(
... filename=f.name,
... key_dtype=tf.string, key_index=0,
... value_dtype=tf.int64, value_index=1,
... delimiter=" ")
>>> table = tf.lookup.StaticHashTable(init, default_value=-1)
>>> table.lookup(tf.constant(['palmer','lake','tarkus'])).numpy()
Similarly to initialize the whole line as keys and the line number as values.
* `emerson 10 -> 0`
* `lake 20 -> 1`
* `palmer 30 -> 2`
>>> init = tf.lookup.TextFileInitializer(
... filename=f.name,
... key_dtype=tf.string, key_index=tf.lookup.TextFileIndex.WHOLE_LINE,
... value_dtype=tf.int64, value_index=tf.lookup.TextFileIndex.LINE_NUMBER)
>>> table = tf.lookup.StaticHashTable(init, -1)
>>> table.lookup(tf.constant('palmer 30')).numpy()
2
"""
def __init__(self,
filename,
key_dtype,
key_index,
value_dtype,
value_index,
vocab_size=None,
delimiter="\t",
name=None,
value_index_offset=0):
"""Constructs a table initializer object to populate from a text file.
It generates one key-value pair per line. The type of table key and
value are specified by `key_dtype` and `value_dtype`, respectively.
Similarly the content of the key and value are specified by the key_index
and value_index.
- TextFileIndex.LINE_NUMBER means use the line number starting from zero,
expects data type int64.
- TextFileIndex.WHOLE_LINE means use the whole line content, expects data
type string or int64.
- A value >=0 means use the index (starting at zero) of the split line based
on `delimiter`.
Args:
filename: The filename of the text file to be used for initialization. The
path must be accessible from wherever the graph is initialized (eg.
trainer or eval workers). The filename may be a scalar `Tensor`.
key_dtype: The `key` data type.
key_index: the index that represents information of a line to get the
table 'key' values from.
value_dtype: The `value` data type.
value_index: the index that represents information of a line to get the
table 'value' values from.'
vocab_size: The number of elements in the file, if known.
delimiter: The delimiter to separate fields in a line.
name: A name for the operation (optional).
value_index_offset: A number to add to all indices extracted from the file
This is useful for cases where a user would like to reserve one or more
low index values for control characters. For instance, if you would
like to ensure that no vocabulary item is mapped to index 0 (so you can
reserve 0 for a masking value), you can set value_index_offset to 1;
this will mean that the first vocabulary element is mapped to 1
instead of 0.
Raises:
ValueError: when the filename is empty, or when the table key and value
data types do not match the expected data types.
"""
if not isinstance(filename, tensor_lib.Tensor) and not filename:
raise ValueError("`filename` argument required for tf.lookup.TextFileInitializer")
self._filename_arg = filename
key_dtype = dtypes.as_dtype(key_dtype)
value_dtype = dtypes.as_dtype(value_dtype)
if key_index < -2:
raise ValueError(f"`key_index` should be >= -2, received: {key_index}.")
if key_index == TextFileIndex.LINE_NUMBER and key_dtype != dtypes.int64:
raise ValueError("`key_dtype` must be int64 if `key_index` is "
f"{TextFileIndex.LINE_NUMBER}, received: {key_dtype}")
if ((key_index == TextFileIndex.WHOLE_LINE) and
(not key_dtype.is_integer) and (key_dtype != dtypes.string)):
raise ValueError(
"`key_dtype` should be either integer or string for `key_index` "
f"{TextFileIndex.WHOLE_LINE}, received: {key_dtype}")
if value_index < -2:
raise ValueError("`value_index` should be >= -2, received: "
f"{value_index}")
if value_index == TextFileIndex.LINE_NUMBER and value_dtype != dtypes.int64:
raise ValueError("`value_dtype` must be int64 for `value_index` "
f"{TextFileIndex.LINE_NUMBER}, received: {value_dtype}")
if ((value_index == TextFileIndex.WHOLE_LINE) and
(not value_dtype.is_integer) and (value_dtype != dtypes.string)):
raise ValueError(
"`value_dtype` should be either integer or string for `value_index` "
f"{TextFileIndex.WHOLE_LINE}, received: {value_dtype}")
if (vocab_size is not None) and (vocab_size <= 0):
raise ValueError(f"`vocab_size` should be > 0, received: {vocab_size}")
self._key_index = key_index
self._value_index = value_index
self._vocab_size = vocab_size
self._delimiter = delimiter
self._name = name
self._filename = self._track_trackable(
asset.Asset(filename), "_filename")
self._offset = value_index_offset
super(TextFileInitializer, self).__init__(key_dtype, value_dtype)
def initialize(self, table):
"""Initializes the table from a text file.
Args:
table: The table to be initialized.
Returns:
The operation that initializes the table.
Raises:
TypeError: when the keys and values data types do not match the table
key and value data types.
"""
check_table_dtypes(table, self.key_dtype, self.value_dtype)
with ops.name_scope(self._name, "text_file_init", (table.resource_handle,)):
filename = ops.convert_to_tensor(
self._filename, dtypes.string, name="asset_filepath")
init_op = gen_lookup_ops.initialize_table_from_text_file_v2(
table.resource_handle, filename, self._key_index, self._value_index,
-1 if self._vocab_size is None else self._vocab_size, self._delimiter,
self._offset)
ops.add_to_collection(ops.GraphKeys.TABLE_INITIALIZERS, init_op)
# If the filename tensor is anything other than a string constant (e.g.,
# if it is a placeholder) then it does not make sense to track it as an
# asset.
if not context.executing_eagerly() and constant_op.is_constant(filename):
ops.add_to_collection(ops.GraphKeys.ASSET_FILEPATHS, filename)
return init_op
@property
def _shared_name(self):
if self._vocab_size:
# Keep the shared_name:
# <table_type>_<filename>_<vocab_size>_<key_index>_<value_index>_<offset>
if self._offset:
shared_name = "hash_table_%s_%d_%s_%s_%s" % (
self._filename_arg, self._vocab_size, self._key_index,
self._value_index, self._offset)
else:
shared_name = "hash_table_%s_%d_%s_%s" % (
self._filename_arg, self._vocab_size, self._key_index,
self._value_index)
else:
# Keep the shared_name
# <table_type>_<filename>_<key_index>_<value_index>_<offset>
if self._offset:
shared_name = "hash_table_%s_%s_%s_%s" % (
self._filename_arg, self._key_index, self._value_index,
self._offset)
else:
shared_name = "hash_table_%s_%s_%s" % (
self._filename_arg, self._key_index, self._value_index)
return shared_name
class TextFileStringTableInitializer(TextFileInitializer):
"""Table initializer for `int64` IDs to string tables from a text file."""
def __init__(self,
filename,
key_column_index=TextFileIndex.LINE_NUMBER,
value_column_index=TextFileIndex.WHOLE_LINE,
vocab_size=None,
delimiter="\t",
name="text_file_string_table_init"):
"""Constructs an initializer for an id-to-string table from a text file.
It populates a table that its key and value types are int64 and string,
respectively. It generates one key-value pair per line.
The content of the key and value are specified by `key_column_index`
and `value_column_index`.
- TextFileIndex.LINE_NUMBER means use the line number starting from zero,
expects data type int64.
- TextFileIndex.WHOLE_LINE means use the whole line content, expects data
type string or int64.
- A value >=0 means use the index (starting at zero) of the split line based
on `delimiter`.
Args:
filename: The filename of the text file to be used for initialization. The
path must be accessible from wherever the graph is initialized (eg.
trainer or eval workers). The filename may be a scalar `Tensor`.
key_column_index: The column index from the text file to get the keys
from. The default is to use the line number, starting from zero.
value_column_index: The column index from the text file to get the values
from. The default is to use the whole line content.
vocab_size: The number of elements in the file, if known.
delimiter: The delimiter to separate fields in a line.
name: Optional name for the op.
Raises:
TypeError: when the filename is empty, or when the table key and value
data types do not match the expected data types.
"""
super(TextFileStringTableInitializer, self).__init__(
filename,
dtypes.int64,
key_column_index,
dtypes.string,
value_column_index,
vocab_size=vocab_size,
delimiter=delimiter,
name=name)
class TextFileIdTableInitializer(TextFileInitializer):
"""Table initializer for string to `int64` IDs tables from a text file."""
def __init__(self,
filename,
key_column_index=TextFileIndex.WHOLE_LINE,
value_column_index=TextFileIndex.LINE_NUMBER,
vocab_size=None,
delimiter="\t",
name="text_file_id_table_init",
key_dtype=dtypes.string):
"""Constructs an initializer for an string-to-id table from a text file.
It populates a table that its key and value types are string and int64,
respectively. It generates one key-value pair per line.
The content of the key and value are specified by the key_index
and value_index.
- TextFileIndex.LINE_NUMBER means use the line number starting from zero,
expects data type int64.
- TextFileIndex.WHOLE_LINE means use the whole line content, expects data
type string.
- A value >=0 means use the index (starting at zero) of the split line based
on `delimiter`.
Args:
filename: The filename of the text file to be used for initialization. The
path must be accessible from wherever the graph is initialized (eg.
trainer or eval workers). The filename may be a scalar `Tensor`.
key_column_index: The column index from the text file to get the `key`
values from. The default is to use the whole line content.
value_column_index: The column index from the text file to get the `value`
values from. The default is to use the line number, starting from zero.
vocab_size: The number of elements in the file, if known.
delimiter: The delimiter to separate fields in a line.
name: Optional name for the op.
key_dtype: The `key` data type.
Raises:
TypeError: when the filename is empty, or when the table key and value
data types do not match the expected data types.
"""
super(TextFileIdTableInitializer, self).__init__(
filename,
key_dtype,
key_column_index,
dtypes.int64,
value_column_index,
vocab_size=vocab_size,
delimiter=delimiter,
name=name)
class HasherSpec(collections.namedtuple("HasherSpec", ["hasher", "key"])):
"""A structure for the spec of the hashing function to use for hash buckets.
`hasher` is the name of the hashing function to use (eg. "fasthash",
"stronghash").
`key` is optional and specify the key to use for the hash function if
supported, currently only used by a strong hash.
Fields:
hasher: The hasher name to use.
key: The key to be used by the hashing function, if required.
"""
__slots__ = ()
FastHashSpec = HasherSpec("fasthash", None) # pylint: disable=invalid-name
class StrongHashSpec(HasherSpec):
"""A structure to specify a key of the strong keyed hash spec.
The strong hash requires a `key`, which is a list of 2 unsigned integer
numbers. These should be non-zero; random numbers generated from random.org
would be a fine choice.
Fields:
key: The key to be used by the keyed hashing function.
"""
__slots__ = ()
def __new__(cls, key):
if len(key) != 2:
raise ValueError(f"`key` must have size 2, received {len(key)}")
if not isinstance(key[0], compat_util.integral_types) or not isinstance(
key[1], compat_util.integral_types):
raise TypeError("Invalid key %s. Must be unsigned integer values." % key)
return super(cls, StrongHashSpec).__new__(cls, "stronghash", key)
def _as_string(tensor):
if dtypes.string == tensor.dtype.base_dtype:
return tensor
return string_ops.as_string(tensor)
class IdTableWithHashBuckets(LookupInterface):
r"""String to Id table wrapper that assigns out-of-vocabulary keys to buckets.
For example, if an instance of `IdTableWithHashBuckets` is initialized with a
string-to-id table that maps:
* `emerson -> 0`
* `lake -> 1`
* `palmer -> 2`
The `IdTableWithHashBuckets` object will performs the following mapping:
* `emerson -> 0`
* `lake -> 1`
* `palmer -> 2`
* `<other term> -> bucket_id`, where bucket_id will be between `3` and
`3 + num_oov_buckets - 1`, calculated by:
`hash(<term>) % num_oov_buckets + vocab_size`
If input_tensor is `["emerson", "lake", "palmer", "king", "crimson"]`,
the lookup result is `[0, 1, 2, 4, 7]`.
If `table` is None, only out-of-vocabulary buckets are used.
Example usage:
```python
num_oov_buckets = 3
input_tensor = tf.constant(["emerson", "lake", "palmer", "king", "crimnson"])
table = tf.IdTableWithHashBuckets(
tf.StaticHashTable(
tf.lookup.TextFileInitializer(
filename,
key_dtype=tf.string,