/
variable_scope.py
1927 lines (1700 loc) · 81 KB
/
variable_scope.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 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.
# ==============================================================================
"""A class to store named variables and a scope operator to manage sharing."""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import collections as collections_lib
import copy
import enum # pylint: disable=g-bad-import-order
import functools
import traceback
import six
from six.moves import xrange # pylint: disable=redefined-builtin
from tensorflow.python.eager import context
from tensorflow.python.estimator import util as estimator_util
from tensorflow.python.framework import dtypes
from tensorflow.python.framework import ops
from tensorflow.python.framework import tensor_shape
from tensorflow.python.ops import array_ops
from tensorflow.python.ops import init_ops
from tensorflow.python.ops import resource_variable_ops
from tensorflow.python.ops import variables
from tensorflow.python.platform import tf_logging as logging
from tensorflow.python.util import tf_contextlib
__all__ = ["AUTO_REUSE", "VariableScope", "get_variable_scope",
"get_variable", "get_local_variable", "variable_scope",
"variable_op_scope", "no_regularizer"]
class _PartitionInfo(object):
"""Holds partition info used by initializer functions.
"""
def __init__(self, full_shape, var_offset):
"""Constructor.
Args:
full_shape: Tuple or list of `int` indicating the full combined shape
of the partitioned variables.
var_offset: Tuple or list of `int` specifying offset of this partition
with respect to the full variable for each dimension.
Raises:
TypeError: If `full_shape` or `var_offset` is not a sequence.
ValueError: If `full_shape` or `var_offset` differ in length. If
`var_offset` exceeds `full_shape` in any dimension.
"""
if not isinstance(full_shape, collections_lib.Sequence) or isinstance(
full_shape, six.string_types):
raise TypeError(
"`full_shape` must be a sequence (like tuple or list) instead of " +
type(full_shape).__name__)
if not isinstance(var_offset, collections_lib.Sequence) or isinstance(
var_offset, six.string_types):
raise TypeError(
"`var_offset` must be a sequence (like tuple or list) instead of " +
type(var_offset).__name__)
if len(var_offset) != len(full_shape):
raise ValueError(
"Expected equal length, but `var_offset` is of length {} while "
"full_shape is of length {}.".format(
len(var_offset), len(full_shape)))
for i in xrange(len(full_shape)):
offset = var_offset[i]
shape = full_shape[i]
if offset < 0 or offset >= shape:
raise ValueError(
"Expected 0 <= offset < shape but found offset={}, shape={} for "
"var_offset={}, full_shape={}".format(offset, shape, var_offset,
full_shape))
self._full_shape = full_shape
self._var_offset = var_offset
@property
def full_shape(self):
return self._full_shape
@property
def var_offset(self):
return self._var_offset
def single_offset(self, shape):
"""Returns the offset when the variable is partitioned in at most one dim.
Args:
shape: Tuple or list of `int` indicating the shape of one specific
variable partition.
Returns:
`int` representing the offset in the dimension along which the variable is
partitioned. Returns 0 if the variable is not being partitioned.
Raises:
ValueError: Depending on self.single_slice_dim().
"""
single_slice_dim = self.single_slice_dim(shape)
# If this variable is not being partitioned at all, single_slice_dim() could
# return None.
if single_slice_dim is None:
return 0
return self.var_offset[single_slice_dim]
def single_slice_dim(self, shape):
"""Returns the slice dim when the variable is partitioned only in one dim.
Args:
shape: Tuple or list of `int` indicating the shape of one specific
variable partition.
Returns:
`int` representing the dimension that the variable is partitioned in, or
`None` if the variable doesn't seem to be partitioned at all.
Raises:
TypeError: If `shape` is not a sequence.
ValueError: If `shape` is not the same length as `self.full_shape`. If
the variable is partitioned in more than one dimension.
"""
if not isinstance(shape, collections_lib.Sequence) or isinstance(
shape, six.string_types):
raise TypeError(
"`shape` must be a sequence (like tuple or list) instead of " +
type(shape).__name__)
if len(shape) != len(self.full_shape):
raise ValueError(
"Expected equal length, but received shape={} of length {} while "
"self.full_shape={} is of length {}.".format(shape, len(
shape), self.full_shape, len(self.full_shape)))
for i in xrange(len(shape)):
if self.var_offset[i] + shape[i] > self.full_shape[i]:
raise ValueError(
"With self.var_offset={}, a partition of shape={} would exceed "
"self.full_shape={} in dimension {}.".format(
self.var_offset, shape, self.full_shape, i))
slice_dim = None
for i in xrange(len(shape)):
if shape[i] == self.full_shape[i]:
continue
if slice_dim is not None:
raise ValueError(
"Cannot use single_slice_dim() with shape={} and "
"self.full_shape={} since slice dim could be either dimension {} "
"or {}.".format(shape, self.full_shape, i, slice_dim))
slice_dim = i
return slice_dim
class _ReuseMode(enum.Enum):
"""Mode for variable access within a variable scope."""
# Indicates that variables are to be fetched if they already exist or
# otherwise created.
AUTO_REUSE = 1
# TODO(alive): For TensorFlow 2.0, Deprecate True/False/None API in favor of
# enum values.
# REUSE_FALSE = 2
# REUSE_TRUE = 3
AUTO_REUSE = _ReuseMode.AUTO_REUSE
AUTO_REUSE.__doc__ = """
When passed in as the value for the `reuse` flag, AUTO_REUSE indicates that
get_variable() should create the requested variable if it doesn't exist or, if
it does exist, simply return it.
"""
class _VariableStore(object):
"""Variable store that carries a number of named Variables.
New variable names and new variables can be created; all stored
variables are initialized with the initializer passed to __init__.
Attributes:
vars: a dictionary with string names (same as passed in GetVar) as keys
and the corresponding TensorFlow Variables as values.
"""
def __init__(self):
"""Create a variable store."""
self._vars = {} # A dictionary of the stored TensorFlow variables.
self._partitioned_vars = {} # A dict of the stored PartitionedVariables.
self.variable_scopes_count = {} # Count re-used variable scopes.
def open_variable_scope(self, scope_name):
if scope_name in self.variable_scopes_count:
self.variable_scopes_count[scope_name] += 1
else:
self.variable_scopes_count[scope_name] = 1
def close_variable_subscopes(self, scope_name):
for k in self.variable_scopes_count:
if not scope_name or k.startswith(scope_name + "/"):
self.variable_scopes_count[k] = 0
def variable_scope_count(self, scope_name):
return self.variable_scopes_count.get(scope_name, 0)
def get_variable(self, name, shape=None, dtype=dtypes.float32,
initializer=None, regularizer=None, reuse=None,
trainable=True, collections=None, caching_device=None,
partitioner=None, validate_shape=True, use_resource=None,
custom_getter=None, constraint=None):
"""Gets an existing variable with these parameters or create a new one.
If a variable with the given name is already stored, we return the stored
variable. Otherwise, we create a new one.
Set `reuse` to `True` when you only want to reuse existing Variables.
Set `reuse` to `False` when you only want to create new Variables.
Set `reuse` to None (the default) or tf.AUTO_REUSE when you want
variables to be created if they don't exist or returned if they do.
If initializer is `None` (the default), the default initializer passed in
the constructor is used. If that one is `None` too, we use a new
`glorot_uniform_initializer`. If initializer is a Tensor, we use
it as a value and derive the shape from the initializer.
If a partitioner is provided, a `PartitionedVariable` is returned.
Accessing this object as a `Tensor` returns the shards concatenated along
the partition axis.
Some useful partitioners are available. See, e.g.,
`variable_axis_size_partitioner` and `min_max_variable_partitioner`.
Args:
name: The name of the new or existing variable.
shape: Shape of the new or existing variable.
dtype: Type of the new or existing variable (defaults to `DT_FLOAT`).
initializer: Initializer for the variable.
regularizer: A (Tensor -> Tensor or None) function; the result of
applying it on a newly created variable will be added to the collection
GraphKeys.REGULARIZATION_LOSSES and can be used for regularization.
reuse: a Boolean, None, or tf.AUTO_REUSE. Controls reuse or creation
of variables. In Eager mode, this argument is always forced to be
tf.AUTO_REUSE.
trainable: If `True` also add the variable to the graph collection
`GraphKeys.TRAINABLE_VARIABLES` (see `tf.Variable`).
collections: List of graph collections keys to add the `Variable` to.
Defaults to `[GraphKeys.GLOBAL_VARIABLES]` (see `tf.Variable`).
caching_device: Optional device string or function describing where the
Variable should be cached for reading. Defaults to the Variable's
device. If not `None`, caches on another device. Typical use is to
cache on the device where the Ops using the `Variable` reside, to
deduplicate copying through `Switch` and other conditional statements.
partitioner: Optional callable that accepts a fully defined `TensorShape`
and dtype of the `Variable` to be created, and returns a list of
partitions for each axis (currently only one axis can be partitioned).
validate_shape: If False, allows the variable to be initialized with a
value of unknown shape. If True, the default, the shape of initial_value
must be known.
use_resource: If False, creates a regular Variable. If True, creates
instead an experimental ResourceVariable which has well-defined
semantics. Defaults to False (will later change to True).
In Eager mode, this argument is always forced to be true.
custom_getter: Callable that takes as a first argument the true getter,
and allows overwriting the internal get_variable method.
The signature of `custom_getter` should match that of this method,
but the most future-proof version will allow for changes:
`def custom_getter(getter, *args, **kwargs)`. Direct access to
all `get_variable` parameters is also allowed:
`def custom_getter(getter, name, *args, **kwargs)`. A simple identity
custom getter that simply creates variables with modified names is:
```python
def custom_getter(getter, name, *args, **kwargs):
return getter(name + '_suffix', *args, **kwargs)
```
constraint: An optional projection function to be applied to the variable
after being updated by an `Optimizer` (e.g. used to implement norm
constraints or value constraints for layer weights). The function must
take as input the unprojected Tensor representing the value of the
variable and return the Tensor for the projected value
(which must have the same shape). Constraints are not safe to
use when doing asynchronous distributed training.
Returns:
The created or existing `Variable` (or `PartitionedVariable`, if a
partitioner was used).
Raises:
ValueError: when creating a new variable and shape is not declared,
when reusing a variable and specifying a conflicting shape,
or when violating reuse during variable creation.
"""
if custom_getter is not None and not callable(custom_getter):
raise ValueError(
"Passed a custom_getter which is not callable: %s" % custom_getter)
if context.in_eager_mode():
reuse = AUTO_REUSE
use_resource = True
# If a *_ref type is passed in an error would be triggered further down the
# stack. We prevent this using base_dtype to get a non-ref version of the
# type, before doing anything else. When _ref types are removed in favor of
# resources, this line can be removed.
try:
dtype = dtype.base_dtype
except AttributeError:
# .base_dtype not existing means that we will try and use the raw dtype
# which was passed in - this might be a NumPy type which is valid.
pass
# This is the main logic of get_variable. However, custom_getter
# may override this logic. So we save it as a callable and pass
# it to custom_getter.
# Note: the parameters of _true_getter, and their documentation, match
# *exactly* item-for-item with the docstring of this method.
def _true_getter(name, shape=None, dtype=dtypes.float32, # pylint: disable=missing-docstring
initializer=None, regularizer=None, reuse=None,
trainable=True, collections=None, caching_device=None,
partitioner=None, validate_shape=True, use_resource=None,
constraint=None):
is_scalar = (shape is not None
and isinstance(shape, collections_lib.Sequence)
and not shape)
# Partitioned variable case
if partitioner is not None and not is_scalar:
if not callable(partitioner):
raise ValueError(
"Partitioner must be callable, but received: %s" % partitioner)
with ops.name_scope(None):
return self._get_partitioned_variable(name=name,
shape=shape,
dtype=dtype,
initializer=initializer,
regularizer=regularizer,
reuse=reuse,
trainable=trainable,
collections=collections,
caching_device=caching_device,
partitioner=partitioner,
validate_shape=validate_shape,
use_resource=use_resource,
constraint=constraint)
# Special case for partitioned variable to allow reuse without having to
# specify partitioner.
if (reuse is True and partitioner is None
and name in self._partitioned_vars):
return self._get_partitioned_variable(name=name,
shape=shape,
dtype=dtype,
initializer=initializer,
regularizer=regularizer,
reuse=reuse,
trainable=trainable,
collections=collections,
caching_device=caching_device,
partitioner=None,
validate_shape=validate_shape,
use_resource=use_resource,
constraint=constraint)
# Single variable case
if "%s/part_0" % name in self._vars:
raise ValueError(
"No partitioner was provided, but a partitioned version of the "
"variable was found: %s/part_0. Perhaps a variable of the same "
"name was already created with partitioning?" % name)
return self._get_single_variable(
name=name, shape=shape, dtype=dtype,
initializer=initializer, regularizer=regularizer, reuse=reuse,
trainable=trainable, collections=collections,
caching_device=caching_device, validate_shape=validate_shape,
use_resource=use_resource, constraint=constraint)
if custom_getter is not None:
# Handle backwards compatibility with getter arguments that were added
# to the API after users started writing custom getters.
custom_getter_kwargs = {
"getter": _true_getter,
"name": name,
"shape": shape,
"dtype": dtype,
"initializer": initializer,
"regularizer": regularizer,
"reuse": reuse,
"trainable": trainable,
"collections": collections,
"caching_device": caching_device,
"partitioner": partitioner,
"validate_shape": validate_shape,
"use_resource": use_resource,
}
# `fn_args` can handle functions, `functools.partial`, `lambda`.
if "constraint" in estimator_util.fn_args(custom_getter):
custom_getter_kwargs["constraint"] = constraint
return custom_getter(**custom_getter_kwargs)
else:
return _true_getter(
name, shape=shape, dtype=dtype,
initializer=initializer, regularizer=regularizer,
reuse=reuse, trainable=trainable, collections=collections,
caching_device=caching_device, partitioner=partitioner,
validate_shape=validate_shape, use_resource=use_resource,
constraint=constraint)
def _get_partitioned_variable(
self, name, partitioner, shape=None, dtype=dtypes.float32,
initializer=None, regularizer=None, reuse=None,
trainable=True, collections=None, caching_device=None,
validate_shape=True, use_resource=None, constraint=None):
"""Gets or creates a sharded variable list with these parameters.
The `partitioner` must be a callable that accepts a fully defined
`TensorShape` and returns a sequence of integers (the `partitions`).
These integers describe how to partition the given sharded `Variable`
along the given dimension. That is, `partitions[1] = 3` means split
the `Variable` into 3 shards along dimension 1. Currently, sharding along
only one axis is supported.
If the list of variables with the given name (prefix) is already stored,
we return the stored variables. Otherwise, we create a new one.
Set `reuse` to `True` when you only want to reuse existing Variables.
Set `reuse` to `False` when you only want to create new Variables.
Set `reuse` to None (the default) or tf.AUTO_REUSE when you want
variables to be created if they don't exist or returned if they do.
If initializer is `None` (the default), the default initializer passed in
the constructor is used. If that one is `None` too, we use a new
`glorot_uniform_initializer`. If initializer is a Tensor, we use
it as a value and derive the shape from the initializer.
If the initializer is a callable, then it will be called for each
shard. Otherwise the initializer should match the shape of the entire
sharded Variable, and it will be sliced accordingly for each shard.
Some useful partitioners are available. See, e.g.,
`variable_axis_size_partitioner` and `min_max_variable_partitioner`.
Args:
name: the name of the new or existing sharded variable.
partitioner: Optional callable that accepts a fully defined `TensorShape`
and `dtype` of the Variable to be created, and returns a list of
partitions for each axis (currently only one axis can be partitioned).
shape: shape of the new or existing sharded variable.
dtype: type of the new or existing sharded variable
(defaults to `DT_FLOAT`).
initializer: initializer for the sharded variable.
regularizer: a (Tensor -> Tensor or None) function; the result of
applying it on a newly created variable will be added to the collection
GraphKeys.REGULARIZATION_LOSSES and can be used for regularization.
reuse: a Boolean, None, or tf.AUTO_REUSE. Controls reuse or creation
of variables.
trainable: If `True` also add the variable to the graph collection
`GraphKeys.TRAINABLE_VARIABLES` (see `tf.Variable`).
collections: List of graph collections keys to add the Variable to.
Defaults to `[GraphKeys.GLOBAL_VARIABLES]` (see `tf.Variable`).
caching_device: Optional device string or function describing where the
Variable should be cached for reading. Defaults to the Variable's
device. If not `None`, caches on another device. Typical use is to
cache on the device where the Ops using the Variable reside, to
deduplicate copying through `Switch` and other conditional statements.
validate_shape: If False, allows the variable to be initialized with a
value of unknown shape. If True, the default, the shape of initial_value
must be known.
use_resource: If False, creates a regular Variable. If True, creates an
experimental ResourceVariable which has well-defined semantics. Defaults
to False (will later change to True).
constraint: An optional projection function to be applied to the variable
after being updated by an `Optimizer` (e.g. used to implement norm
constraints or value constraints for layer weights). The function must
take as input the unprojected Tensor representing the value of the
variable and return the Tensor for the projected value
(which must have the same shape). Constraints are not safe to
use when doing asynchronous distributed training.
Returns:
A `PartitionedVariable` object.
Raises:
ValueError: when creating a new variable and shape is not declared,
when reusing a variable and specifying a conflicting shape,
when violating reuse during variable creation, or if an existing
sharded variable exists for the given name but with different sharding.
"""
if context.in_eager_mode():
raise NotImplementedError("Partitioned variables are not yet supported "
"in Eager mode.")
initializing_from_value = initializer is not None and isinstance(
initializer, ops.Tensor)
reuse_without_partition = reuse and not partitioner
if name in self._vars:
raise ValueError(
"A partitioner was provided, but an unpartitioned version of the "
"variable was found: %s. Perhaps a variable of the same name was "
"already created without partitioning?" % name)
shape = tensor_shape.as_shape(shape)
if initializing_from_value:
shape = shape.merge_with(initializer.get_shape())
if not reuse_without_partition:
if not shape.is_fully_defined():
raise ValueError("Shape of a new partitioned variable (%s) must be "
"fully defined, but instead was %s." % (name, shape))
if shape.ndims < 1:
raise ValueError("A partitioned Variable must have rank at least 1, "
"shape: %s" % shape)
partitions = partitioner(shape=shape, dtype=dtype)
if not isinstance(partitions, collections_lib.Sequence):
raise ValueError("Partitioner must return a sequence, but saw: %s"
% partitions)
if len(partitions) != shape.ndims:
raise ValueError(
"Partitioner returned a partition list that does not match the "
"Variable's rank: %s vs. %s" % (partitions, shape))
if any([p < 1 for p in partitions]):
raise ValueError(
"Partitioner returned zero partitions for some axes: %s" %
partitions)
if name in self._partitioned_vars:
if reuse is False:
raise ValueError(
"Partitioned variable with name %s already exists. Did you mean to "
"set reuse=True or reuse=tf.AUTO_REUSE in VarScope?"
% name)
existing_var = self._partitioned_vars[name]
if not shape.is_compatible_with(existing_var.get_shape()):
raise ValueError(
"Trying to reuse partitioned variable %s, but specified shape %s "
"and found shape %s."
% (name, shape, existing_var.get_shape()))
if not dtype.is_compatible_with(existing_var.dtype):
raise ValueError(
"Trying to reuse partitioned variable %s, but specified dtype %s "
"and found dtype %s."
% (name, dtype.name, existing_var.dtype.name))
# pylint: disable=protected-access
if (not reuse_without_partition and
existing_var._get_partitions() != partitions):
raise ValueError(
"Trying to reuse partitioned variable %s, but specified partitions "
"%s and found partitions %s." %
(name, partitions, existing_var._get_partitions()))
# pylint: enable=protected-access
return existing_var
if reuse is True:
raise ValueError("PartitionedVariable %s does not exist, or was not "
"created with tf.get_variable(). Did you mean to set "
"reuse=None in VarScope?" % name)
slice_dim, slice_shape = _compute_slice_dim_and_shape(
shape.as_list(), partitions)
vs = []
num_slices = partitions[slice_dim]
num_slices_with_excess = shape[slice_dim].value % num_slices
slice_offset = [0] * shape.ndims
if "%s/part_0" % name in self._vars:
if "%s/part_%d" % (name, num_slices - 1) not in self._vars:
raise ValueError(
"Partitioner returned a different partitioning than what was "
"already found. Partitioner returned %d shards, and shard "
"%s/part_0 was found, but %s/part_%d was not."
% (num_slices, name, name, num_slices - 1))
if "%s/part_%d" % (name, num_slices) in self._vars:
raise ValueError(
"Partitioner returned a different partitioning than what was "
"already found. Partitioner returned %d shards, and shard "
"%s/part_0 was found, but so was the extra shard %s/part_%d."
% (num_slices, name, name, num_slices))
for i in xrange(num_slices):
var_shape = slice_shape[:]
var_offset = slice_offset[:]
partition_info = _PartitionInfo(
full_shape=shape.as_list(), var_offset=var_offset)
if i < num_slices_with_excess:
var_shape[slice_dim] += 1
slice_offset[slice_dim] += var_shape[slice_dim]
var_full_name = "%s/part_%d" % (name, i)
with ops.name_scope(var_full_name + "/PartitionedInitializer"):
# Create the tensor to initialize the variable with default value.
if initializer is None:
init, initializing_from_value = self._get_default_initializer(
name=name, shape=shape, dtype=dtype)
if initializing_from_value:
init_shape = None
else:
init_shape = var_shape
elif callable(initializer):
init = initializer
init_shape = var_shape
elif isinstance(initializer, ops.Tensor):
init = array_ops.slice(initializer, var_offset, var_shape)
# Use the dtype of the given tensor.
dtype = init.dtype.base_dtype
init_shape = None
else:
init = ops.convert_to_tensor(initializer, dtype=dtype)
init = array_ops.slice(init, var_offset, var_shape)
init_shape = None
with ops.name_scope(None):
var = self._get_single_variable(
name=var_full_name,
shape=init_shape,
dtype=dtype,
initializer=init,
partition_info=partition_info,
regularizer=regularizer,
reuse=reuse,
trainable=trainable,
collections=collections,
caching_device=caching_device,
validate_shape=validate_shape,
use_resource=use_resource,
constraint=constraint)
# pylint: disable=protected-access
var._set_save_slice_info(variables.Variable.SaveSliceInfo(
name, shape.as_list(), var_offset, var_shape))
vs.append(var)
# pylint: enable=protected-access
# pylint: disable=protected-access
partitioned_var = variables.PartitionedVariable(name=name,
shape=shape,
dtype=dtype,
variable_list=vs,
partitions=partitions)
# pylint: enable=protected-access
self._partitioned_vars[name] = partitioned_var
return partitioned_var
def _get_single_variable(self,
name,
shape=None,
dtype=dtypes.float32,
initializer=None,
regularizer=None,
partition_info=None,
reuse=None,
trainable=True,
collections=None,
caching_device=None,
validate_shape=True,
use_resource=None,
constraint=None):
"""Get or create a single Variable (e.g. a shard or entire variable).
See the documentation of get_variable above (ignore partitioning components)
for details.
Args:
name: see get_variable.
shape: see get_variable.
dtype: see get_variable.
initializer: see get_variable.
regularizer: see get_variable.
partition_info: _PartitionInfo object.
reuse: see get_variable.
trainable: see get_variable.
collections: see get_variable.
caching_device: see get_variable.
validate_shape: see get_variable.
use_resource: see get_variable.
constraint: see get_variable.
Returns:
A Variable. See documentation of get_variable above.
Raises:
ValueError: See documentation of get_variable above.
"""
# Fast-path for get_variable in eager mode when the variable already
# exists. Note this skips error validation code, so mismatched shapes and
# dtypes will be caught when the variable is used instead of when the call
# to get_variable happens.
if context.in_eager_mode():
v = self._vars.get(name, None)
if v is not None:
return v
# Set to true if initializer is a constant.
initializing_from_value = False
if initializer is not None and not callable(initializer):
initializing_from_value = True
if shape is not None and initializing_from_value:
raise ValueError("If initializer is a constant, do not specify shape.")
dtype = dtypes.as_dtype(dtype)
shape = tensor_shape.as_shape(shape)
if name in self._vars:
# Here we handle the case when returning an existing variable.
if reuse is False:
tb = self._vars[name].op.traceback[::-1]
# Throw away internal tf entries and only take a few lines.
tb = [x for x in tb if "tensorflow/python" not in x[0]][:3]
raise ValueError("Variable %s already exists, disallowed."
" Did you mean to set reuse=True or "
"reuse=tf.AUTO_REUSE in VarScope? "
"Originally defined at:\n\n%s" % (
name, "".join(traceback.format_list(tb))))
found_var = self._vars[name]
if not shape.is_compatible_with(found_var.get_shape()):
raise ValueError("Trying to share variable %s, but specified shape %s"
" and found shape %s." % (name, shape,
found_var.get_shape()))
if not dtype.is_compatible_with(found_var.dtype):
dtype_str = dtype.name
found_type_str = found_var.dtype.name
raise ValueError("Trying to share variable %s, but specified dtype %s"
" and found dtype %s." % (name, dtype_str,
found_type_str))
return found_var
# The code below handles only the case of creating a new variable.
if reuse is True:
raise ValueError("Variable %s does not exist, or was not created with "
"tf.get_variable(). Did you mean to set "
"reuse=tf.AUTO_REUSE in VarScope?" % name)
if not shape.is_fully_defined() and not initializing_from_value:
raise ValueError("Shape of a new variable (%s) must be fully defined, "
"but instead was %s." % (name, shape))
# Create the tensor to initialize the variable with default value.
if initializer is None:
initializer, initializing_from_value = self._get_default_initializer(
name=name, shape=shape, dtype=dtype)
# Clear control dependencies while creating the initializer.
with ops.control_dependencies(None):
if initializing_from_value:
init_val = initializer
variable_dtype = None
else:
# Instantiate initializer if provided initializer is a type object.
if isinstance(initializer, type(init_ops.Initializer)):
initializer = initializer(dtype=dtype)
init_val = lambda: initializer( # pylint: disable=g-long-lambda
shape.as_list(), dtype=dtype, partition_info=partition_info)
variable_dtype = dtype.base_dtype
# Create the variable.
if use_resource is None:
# Set the default value if unspecified.
use_resource = False
if use_resource:
v = resource_variable_ops.ResourceVariable(
initial_value=init_val,
name=name,
trainable=trainable,
collections=collections,
caching_device=caching_device,
dtype=variable_dtype,
validate_shape=validate_shape,
constraint=constraint)
else:
v = variables.Variable(
initial_value=init_val,
name=name,
trainable=trainable,
collections=collections,
caching_device=caching_device,
dtype=variable_dtype,
validate_shape=validate_shape,
constraint=constraint)
self._vars[name] = v
logging.vlog(1, "Created variable %s with shape %s and init %s", v.name,
format(shape), initializer)
# Run the regularizer if requested and save the resulting loss.
if regularizer:
with ops.colocate_with(v):
with ops.name_scope(name + "/Regularizer/"):
loss = regularizer(v)
if loss is not None:
if context.in_graph_mode():
v_name = v.name
loss_name = loss.name
else:
v_name = "v_%s" % type(v)
loss_name = "loss_%s" % type(loss)
logging.vlog(1, "Applied regularizer to %s and added the result %s "
"to REGULARIZATION_LOSSES.", v_name, loss_name)
ops.add_to_collection(ops.GraphKeys.REGULARIZATION_LOSSES, loss)
return v
# Initialize variable when no initializer provided
def _get_default_initializer(self, name, shape=None, dtype=dtypes.float32):
"""Provide a default initializer and a corresponding value.
Args:
name: see get_variable.
shape: see get_variable.
dtype: see get_variable.
Returns:
initializer and initializing_from_value. See get_variable above.
Raises:
ValueError: When giving unsupported dtype.
"""
# If dtype is DT_FLOAT, provide a uniform unit scaling initializer
if dtype.is_floating:
initializer = init_ops.glorot_uniform_initializer()
initializing_from_value = False
# If dtype is DT_INT/DT_UINT, provide a default value `zero`
# If dtype is DT_BOOL, provide a default value `FALSE`
elif dtype.is_integer or dtype.is_unsigned or dtype.is_bool:
initializer = init_ops.zeros_initializer()(
shape=shape, dtype=dtype.base_dtype)
initializing_from_value = True
# NOTES:Do we need to support for handling DT_STRING and DT_COMPLEX here?
else:
raise ValueError("An initializer for variable %s of %s is required"
% (name, dtype.base_dtype))
return initializer, initializing_from_value
# To stop regularization, use this regularizer
def no_regularizer(_):
"""Use this function to prevent regularization of variables."""
return None
# TODO(alive): support caching devices and partitioned variables in Eager mode.
class VariableScope(object):
"""Variable scope object to carry defaults to provide to `get_variable`.
Many of the arguments we need for `get_variable` in a variable store are most
easily handled with a context. This object is used for the defaults.
Attributes:
name: name of the current scope, used as prefix in get_variable.
initializer: default initializer passed to get_variable.
regularizer: default regularizer passed to get_variable.
reuse: Boolean, None, or tf.AUTO_REUSE, setting the reuse in
get_variable. In Eager mode, this argument is always forced to be
tf.AUTO_REUSE.
caching_device: string, callable, or None: the caching device passed to
get_variable.
partitioner: callable or `None`: the partitioner passed to `get_variable`.
custom_getter: default custom getter passed to get_variable.
name_scope: The name passed to `tf.name_scope`.
dtype: default type passed to get_variable (defaults to DT_FLOAT).
use_resource: if False, create a normal Variable; if True create an
experimental ResourceVariable with well-defined semantics. Defaults
to False (will later change to True). In Eager mode, this argument is
always forced to be True.
constraint: An optional projection function to be applied to the variable
after being updated by an `Optimizer` (e.g. used to implement norm
constraints or value constraints for layer weights). The function must
take as input the unprojected Tensor representing the value of the
variable and return the Tensor for the projected value
(which must have the same shape). Constraints are not safe to
use when doing asynchronous distributed training.
"""
def __init__(self,
reuse,
name="",
initializer=None,
regularizer=None,
caching_device=None,
partitioner=None,
custom_getter=None,
name_scope="",
dtype=dtypes.float32,
use_resource=None,
constraint=None):
"""Creates a new VariableScope with the given properties."""
self._name = name
self._initializer = initializer
self._regularizer = regularizer
self._reuse = reuse
self._caching_device = caching_device
self._partitioner = partitioner
self._custom_getter = custom_getter
self._name_scope = name_scope
self._dtype = dtype
self._use_resource = use_resource
self._constraint = constraint
if context.in_eager_mode():
if self._caching_device is not None:
raise NotImplementedError("Caching devices is not yet supported "
"in Eager mode.")
if self._partitioner is not None:
raise NotImplementedError("Partitioned variables are not yet supported "
"in Eager mode.")
self._reuse = AUTO_REUSE
self._use_resource = True
@property
def name(self):
return self._name
@property
def original_name_scope(self):
return self._name_scope
@property
def reuse(self):
return self._reuse
@property
def initializer(self):
return self._initializer
@property
def dtype(self):
return self._dtype
@property
def use_resource(self):
return self._use_resource
@property
def regularizer(self):
return self._regularizer
@property
def caching_device(self):
return self._caching_device
@property
def partitioner(self):
return self._partitioner
@property
def custom_getter(self):
return self._custom_getter
@property
def constraint(self):
return self._constraint
def reuse_variables(self):
"""Reuse variables in this scope."""
self._reuse = True
def set_initializer(self, initializer):
"""Set initializer for this scope."""
self._initializer = initializer
def set_dtype(self, dtype):
"""Set data type for this scope."""
self._dtype = dtype
def set_use_resource(self, use_resource):
"""Sets whether to use ResourceVariables for this scope."""
if context.in_eager_mode() and not use_resource:
raise ValueError("In eager mode, use_resource cannot be set to false.")
self._use_resource = use_resource
def set_regularizer(self, regularizer):
"""Set regularizer for this scope."""
self._regularizer = regularizer
def set_caching_device(self, caching_device):
"""Set caching_device for this scope."""