-
Notifications
You must be signed in to change notification settings - Fork 74k
/
core.py
1206 lines (1009 loc) · 43.7 KB
/
core.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.
# ==============================================================================
"""Core Keras layers.
"""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import copy
import sys
import textwrap
import types as python_types
import warnings
import numpy as np
from tensorflow.python.eager import backprop
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 tensor_shape
from tensorflow.python.keras import activations
from tensorflow.python.keras import backend as K
from tensorflow.python.keras import constraints
from tensorflow.python.keras import initializers
from tensorflow.python.keras import regularizers
from tensorflow.python.keras.engine.base_layer import Layer
from tensorflow.python.keras.engine.input_spec import InputSpec
from tensorflow.python.keras.utils import conv_utils
from tensorflow.python.keras.utils import generic_utils
from tensorflow.python.keras.utils import tf_utils
from tensorflow.python.ops import array_ops
from tensorflow.python.ops import gen_math_ops
from tensorflow.python.ops import math_ops
from tensorflow.python.ops import nn
from tensorflow.python.ops import sparse_ops
from tensorflow.python.ops import standard_ops
from tensorflow.python.ops import variable_scope
from tensorflow.python.platform import tf_logging
from tensorflow.python.training.tracking import base as trackable
from tensorflow.python.util import nest
from tensorflow.python.util import tf_inspect
from tensorflow.python.util.tf_export import keras_export
@keras_export('keras.layers.Masking')
class Masking(Layer):
"""Masks a sequence by using a mask value to skip timesteps.
For each timestep in the input tensor (dimension #1 in the tensor),
if all values in the input tensor at that timestep
are equal to `mask_value`, then the timestep will be masked (skipped)
in all downstream layers (as long as they support masking).
If any downstream layer does not support masking yet receives such
an input mask, an exception will be raised.
Example:
Consider a Numpy data array `x` of shape `(samples, timesteps, features)`,
to be fed to an LSTM layer. You want to mask timestep #3 and #5 because you
lack data for these timesteps. You can:
- Set `x[:, 3, :] = 0.` and `x[:, 5, :] = 0.`
- Insert a `Masking` layer with `mask_value=0.` before the LSTM layer:
```python
samples, timesteps, features = 32, 10, 8
inputs = np.random.random([samples, timesteps, features]).astype(np.float32)
inputs[:, 3, :] = 0.
inputs[:, 5, :] = 0.
model = tf.keras.models.Sequential()
model.add(tf.keras.layers.Masking(mask_value=0.,
input_shape=(timesteps, features)))
model.add(tf.keras.layers.LSTM(32))
output = model(inputs)
# The time step 3 and 5 will be skipped from LSTM calculation.
```
See [the masking and padding
guide](https://www.tensorflow.org/guide/keras/masking_and_padding)
for more details.
"""
def __init__(self, mask_value=0., **kwargs):
super(Masking, self).__init__(**kwargs)
self.supports_masking = True
self.mask_value = mask_value
self._compute_output_and_mask_jointly = True
def compute_mask(self, inputs, mask=None):
return K.any(math_ops.not_equal(inputs, self.mask_value), axis=-1)
def call(self, inputs):
boolean_mask = K.any(
math_ops.not_equal(inputs, self.mask_value), axis=-1, keepdims=True)
outputs = inputs * math_ops.cast(boolean_mask, inputs.dtype)
# Compute the mask and outputs simultaneously.
outputs._keras_mask = array_ops.squeeze(boolean_mask, axis=-1) # pylint: disable=protected-access
return outputs
def compute_output_shape(self, input_shape):
return input_shape
def get_config(self):
config = {'mask_value': self.mask_value}
base_config = super(Masking, self).get_config()
return dict(list(base_config.items()) + list(config.items()))
@keras_export('keras.layers.Dropout')
class Dropout(Layer):
"""Applies Dropout to the input.
Dropout consists in randomly setting
a fraction `rate` of input units to 0 at each update during training time,
which helps prevent overfitting.
Arguments:
rate: Float between 0 and 1. Fraction of the input units to drop.
noise_shape: 1D integer tensor representing the shape of the
binary dropout mask that will be multiplied with the input.
For instance, if your inputs have shape
`(batch_size, timesteps, features)` and
you want the dropout mask to be the same for all timesteps,
you can use `noise_shape=(batch_size, 1, features)`.
seed: A Python integer to use as random seed.
Call arguments:
inputs: Input tensor (of any rank).
training: Python boolean indicating whether the layer should behave in
training mode (adding dropout) or in inference mode (doing nothing).
"""
def __init__(self, rate, noise_shape=None, seed=None, **kwargs):
super(Dropout, self).__init__(**kwargs)
self.rate = rate
self.noise_shape = noise_shape
self.seed = seed
self.supports_masking = True
def _get_noise_shape(self, inputs):
# Subclasses of `Dropout` may implement `_get_noise_shape(self, inputs)`,
# which will override `self.noise_shape`, and allows for custom noise
# shapes with dynamically sized inputs.
if self.noise_shape is None:
return None
concrete_inputs_shape = array_ops.shape(inputs)
noise_shape = []
for i, value in enumerate(self.noise_shape):
noise_shape.append(concrete_inputs_shape[i] if value is None else value)
return ops.convert_to_tensor(noise_shape)
def call(self, inputs, training=None):
if training is None:
training = K.learning_phase()
def dropped_inputs():
return nn.dropout(
inputs,
noise_shape=self._get_noise_shape(inputs),
seed=self.seed,
rate=self.rate)
output = tf_utils.smart_cond(training,
dropped_inputs,
lambda: array_ops.identity(inputs))
return output
def compute_output_shape(self, input_shape):
return input_shape
def get_config(self):
config = {
'rate': self.rate,
'noise_shape': self.noise_shape,
'seed': self.seed
}
base_config = super(Dropout, self).get_config()
return dict(list(base_config.items()) + list(config.items()))
@keras_export('keras.layers.SpatialDropout1D')
class SpatialDropout1D(Dropout):
"""Spatial 1D version of Dropout.
This version performs the same function as Dropout, however it drops
entire 1D feature maps instead of individual elements. If adjacent frames
within feature maps are strongly correlated (as is normally the case in
early convolution layers) then regular dropout will not regularize the
activations and will otherwise just result in an effective learning rate
decrease. In this case, SpatialDropout1D will help promote independence
between feature maps and should be used instead.
Arguments:
rate: Float between 0 and 1. Fraction of the input units to drop.
Call arguments:
inputs: A 3D tensor.
training: Python boolean indicating whether the layer should behave in
training mode (adding dropout) or in inference mode (doing nothing).
Input shape:
3D tensor with shape:
`(samples, timesteps, channels)`
Output shape:
Same as input.
References:
- [Efficient Object Localization Using Convolutional
Networks](https://arxiv.org/abs/1411.4280)
"""
def __init__(self, rate, **kwargs):
super(SpatialDropout1D, self).__init__(rate, **kwargs)
self.input_spec = InputSpec(ndim=3)
def _get_noise_shape(self, inputs):
input_shape = array_ops.shape(inputs)
noise_shape = (input_shape[0], 1, input_shape[2])
return noise_shape
@keras_export('keras.layers.SpatialDropout2D')
class SpatialDropout2D(Dropout):
"""Spatial 2D version of Dropout.
This version performs the same function as Dropout, however it drops
entire 2D feature maps instead of individual elements. If adjacent pixels
within feature maps are strongly correlated (as is normally the case in
early convolution layers) then regular dropout will not regularize the
activations and will otherwise just result in an effective learning rate
decrease. In this case, SpatialDropout2D will help promote independence
between feature maps and should be used instead.
Arguments:
rate: Float between 0 and 1. Fraction of the input units to drop.
data_format: 'channels_first' or 'channels_last'.
In 'channels_first' mode, the channels dimension
(the depth) is at index 1,
in 'channels_last' mode is it at index 3.
It defaults to the `image_data_format` value found in your
Keras config file at `~/.keras/keras.json`.
If you never set it, then it will be "channels_last".
Call arguments:
inputs: A 4D tensor.
training: Python boolean indicating whether the layer should behave in
training mode (adding dropout) or in inference mode (doing nothing).
Input shape:
4D tensor with shape:
`(samples, channels, rows, cols)` if data_format='channels_first'
or 4D tensor with shape:
`(samples, rows, cols, channels)` if data_format='channels_last'.
Output shape:
Same as input.
References:
- [Efficient Object Localization Using Convolutional
Networks](https://arxiv.org/abs/1411.4280)
"""
def __init__(self, rate, data_format=None, **kwargs):
super(SpatialDropout2D, self).__init__(rate, **kwargs)
if data_format is None:
data_format = K.image_data_format()
if data_format not in {'channels_last', 'channels_first'}:
raise ValueError('data_format must be in '
'{"channels_last", "channels_first"}')
self.data_format = data_format
self.input_spec = InputSpec(ndim=4)
def _get_noise_shape(self, inputs):
input_shape = array_ops.shape(inputs)
if self.data_format == 'channels_first':
return (input_shape[0], input_shape[1], 1, 1)
elif self.data_format == 'channels_last':
return (input_shape[0], 1, 1, input_shape[3])
@keras_export('keras.layers.SpatialDropout3D')
class SpatialDropout3D(Dropout):
"""Spatial 3D version of Dropout.
This version performs the same function as Dropout, however it drops
entire 3D feature maps instead of individual elements. If adjacent voxels
within feature maps are strongly correlated (as is normally the case in
early convolution layers) then regular dropout will not regularize the
activations and will otherwise just result in an effective learning rate
decrease. In this case, SpatialDropout3D will help promote independence
between feature maps and should be used instead.
Arguments:
rate: Float between 0 and 1. Fraction of the input units to drop.
data_format: 'channels_first' or 'channels_last'.
In 'channels_first' mode, the channels dimension (the depth)
is at index 1, in 'channels_last' mode is it at index 4.
It defaults to the `image_data_format` value found in your
Keras config file at `~/.keras/keras.json`.
If you never set it, then it will be "channels_last".
Call arguments:
inputs: A 5D tensor.
training: Python boolean indicating whether the layer should behave in
training mode (adding dropout) or in inference mode (doing nothing).
Input shape:
5D tensor with shape:
`(samples, channels, dim1, dim2, dim3)` if data_format='channels_first'
or 5D tensor with shape:
`(samples, dim1, dim2, dim3, channels)` if data_format='channels_last'.
Output shape:
Same as input.
References:
- [Efficient Object Localization Using Convolutional
Networks](https://arxiv.org/abs/1411.4280)
"""
def __init__(self, rate, data_format=None, **kwargs):
super(SpatialDropout3D, self).__init__(rate, **kwargs)
if data_format is None:
data_format = K.image_data_format()
if data_format not in {'channels_last', 'channels_first'}:
raise ValueError('data_format must be in '
'{"channels_last", "channels_first"}')
self.data_format = data_format
self.input_spec = InputSpec(ndim=5)
def _get_noise_shape(self, inputs):
input_shape = array_ops.shape(inputs)
if self.data_format == 'channels_first':
return (input_shape[0], input_shape[1], 1, 1, 1)
elif self.data_format == 'channels_last':
return (input_shape[0], 1, 1, 1, input_shape[4])
@keras_export('keras.layers.Activation')
class Activation(Layer):
"""Applies an activation function to an output.
Arguments:
activation: Activation function, such as `tf.nn.relu`, or string name of
built-in activation function, such as "relu".
Input shape:
Arbitrary. Use the keyword argument `input_shape`
(tuple of integers, does not include the samples axis)
when using this layer as the first layer in a model.
Output shape:
Same shape as input.
"""
def __init__(self, activation, **kwargs):
super(Activation, self).__init__(**kwargs)
self.supports_masking = True
self.activation = activations.get(activation)
def call(self, inputs):
return self.activation(inputs)
def compute_output_shape(self, input_shape):
return input_shape
def get_config(self):
config = {'activation': activations.serialize(self.activation)}
base_config = super(Activation, self).get_config()
return dict(list(base_config.items()) + list(config.items()))
@keras_export('keras.layers.Reshape')
class Reshape(Layer):
"""Reshapes an output to a certain shape.
Arguments:
target_shape: Target shape. Tuple of integers,
does not include the samples dimension (batch size).
Input shape:
Arbitrary, although all dimensions in the input shaped must be fixed.
Use the keyword argument `input_shape`
(tuple of integers, does not include the samples axis)
when using this layer as the first layer in a model.
Output shape:
`(batch_size,) + target_shape`
Example:
```python
# as first layer in a Sequential model
model = Sequential()
model.add(Reshape((3, 4), input_shape=(12,)))
# now: model.output_shape == (None, 3, 4)
# note: `None` is the batch dimension
# as intermediate layer in a Sequential model
model.add(Reshape((6, 2)))
# now: model.output_shape == (None, 6, 2)
# also supports shape inference using `-1` as dimension
model.add(Reshape((-1, 2, 2)))
# now: model.output_shape == (None, None, 2, 2)
```
"""
def __init__(self, target_shape, **kwargs):
super(Reshape, self).__init__(**kwargs)
self.target_shape = tuple(target_shape)
def _fix_unknown_dimension(self, input_shape, output_shape):
"""Find and replace a missing dimension in an output shape.
This is a near direct port of the internal Numpy function
`_fix_unknown_dimension` in `numpy/core/src/multiarray/shape.c`
Arguments:
input_shape: Shape of array being reshaped
output_shape: Desired shape of the array with at most
a single -1 which indicates a dimension that should be
derived from the input shape.
Returns:
The new output shape with a -1 replaced with its computed value.
Raises:
ValueError: If the total array size of the output_shape is
different than the input_shape, or more than one unknown dimension
is specified.
"""
output_shape = list(output_shape)
msg = 'total size of new array must be unchanged'
known, unknown = 1, None
for index, dim in enumerate(output_shape):
if dim < 0:
if unknown is None:
unknown = index
else:
raise ValueError('Can only specify one unknown dimension.')
else:
known *= dim
original = np.prod(input_shape, dtype=int)
if unknown is not None:
if known == 0 or original % known != 0:
raise ValueError(msg)
output_shape[unknown] = original // known
elif original != known:
raise ValueError(msg)
return output_shape
def compute_output_shape(self, input_shape):
input_shape = tensor_shape.TensorShape(input_shape).as_list()
if None in input_shape[1:]:
output_shape = [input_shape[0]]
# input shape (partially) unknown? replace -1's with None's
output_shape += tuple(s if s != -1 else None for s in self.target_shape)
else:
output_shape = [input_shape[0]]
output_shape += self._fix_unknown_dimension(input_shape[1:],
self.target_shape)
return tensor_shape.TensorShape(output_shape)
def call(self, inputs):
return array_ops.reshape(inputs,
(array_ops.shape(inputs)[0],) + self.target_shape)
def get_config(self):
config = {'target_shape': self.target_shape}
base_config = super(Reshape, self).get_config()
return dict(list(base_config.items()) + list(config.items()))
@keras_export('keras.layers.Permute')
class Permute(Layer):
"""Permutes the dimensions of the input according to a given pattern.
Useful for e.g. connecting RNNs and convnets together.
Example:
```python
model = Sequential()
model.add(Permute((2, 1), input_shape=(10, 64)))
# now: model.output_shape == (None, 64, 10)
# note: `None` is the batch dimension
```
Arguments:
dims: Tuple of integers. Permutation pattern, does not include the
samples dimension. Indexing starts at 1.
For instance, `(2, 1)` permutes the first and second dimensions
of the input.
Input shape:
Arbitrary. Use the keyword argument `input_shape`
(tuple of integers, does not include the samples axis)
when using this layer as the first layer in a model.
Output shape:
Same as the input shape, but with the dimensions re-ordered according
to the specified pattern.
"""
def __init__(self, dims, **kwargs):
super(Permute, self).__init__(**kwargs)
self.dims = tuple(dims)
if sorted(dims) != list(range(1, len(dims) + 1)):
raise ValueError(
'Invalid permutation `dims` for Permute Layer: %s. '
'The set of indices in `dims` must be consecutive and start from 1.' %
(dims,))
self.input_spec = InputSpec(ndim=len(self.dims) + 1)
def compute_output_shape(self, input_shape):
input_shape = tensor_shape.TensorShape(input_shape).as_list()
output_shape = copy.copy(input_shape)
for i, dim in enumerate(self.dims):
target_dim = input_shape[dim]
output_shape[i + 1] = target_dim
return tensor_shape.TensorShape(output_shape)
def call(self, inputs):
return array_ops.transpose(inputs, perm=(0,) + self.dims)
def get_config(self):
config = {'dims': self.dims}
base_config = super(Permute, self).get_config()
return dict(list(base_config.items()) + list(config.items()))
@keras_export('keras.layers.Flatten')
class Flatten(Layer):
"""Flattens the input. Does not affect the batch size.
If inputs are shaped `(batch,)` without a channel dimension, then flattening
adds an extra channel dimension and output shapes are `(batch, 1)`.
Arguments:
data_format: A string,
one of `channels_last` (default) or `channels_first`.
The ordering of the dimensions in the inputs.
`channels_last` corresponds to inputs with shape
`(batch, ..., channels)` while `channels_first` corresponds to
inputs with shape `(batch, channels, ...)`.
It defaults to the `image_data_format` value found in your
Keras config file at `~/.keras/keras.json`.
If you never set it, then it will be "channels_last".
Example:
```python
model = Sequential()
model.add(Convolution2D(64, 3, 3,
border_mode='same',
input_shape=(3, 32, 32)))
# now: model.output_shape == (None, 64, 32, 32)
model.add(Flatten())
# now: model.output_shape == (None, 65536)
```
"""
def __init__(self, data_format=None, **kwargs):
super(Flatten, self).__init__(**kwargs)
self.data_format = conv_utils.normalize_data_format(data_format)
self.input_spec = InputSpec(min_ndim=1)
def call(self, inputs):
if (self.data_format == 'channels_first'
and K.ndim(inputs) is not None and K.ndim(inputs) > 1):
permutation = [0]
permutation.extend([i for i in
range(2, K.ndim(inputs))])
permutation.append(1)
inputs = array_ops.transpose(inputs, perm=permutation)
input_shape = inputs.shape
if input_shape[1:].is_fully_defined():
flattened_dim = tensor_shape.dimension_value(
np.prod(input_shape[1:], dtype=int))
# Temporary fix for integer overflow issue.
if flattened_dim > np.iinfo(np.int32).max:
shape_dtype = dtypes.int64
else:
shape_dtype = dtypes.int32
outputs = array_ops.reshape(
inputs, constant_op.constant((-1, flattened_dim), dtype=shape_dtype))
else:
batch_size = tensor_shape.dimension_value(inputs.shape[0])
if batch_size:
# Temporary fix for integer overflow issue.
if batch_size > np.iinfo(np.int32).max:
shape_dtype = dtypes.int64
else:
shape_dtype = dtypes.int32
outputs = array_ops.reshape(
inputs, constant_op.constant((batch_size, -1), dtype=shape_dtype))
else:
outputs = array_ops.reshape(inputs, (array_ops.shape(inputs)[0], -1))
if not context.executing_eagerly():
outputs.set_shape(self.compute_output_shape(inputs.shape))
return outputs
def compute_output_shape(self, input_shape):
input_shape = tensor_shape.as_shape(input_shape).as_list()
if not input_shape:
output_shape = tensor_shape.TensorShape([1])
else:
output_shape = [input_shape[0]]
if all(input_shape[1:]):
output_shape += [np.prod(input_shape[1:], dtype=int)]
else:
output_shape += [None]
return tensor_shape.TensorShape(output_shape)
def get_config(self):
config = {'data_format': self.data_format}
base_config = super(Flatten, self).get_config()
return dict(list(base_config.items()) + list(config.items()))
@keras_export('keras.layers.RepeatVector')
class RepeatVector(Layer):
"""Repeats the input n times.
Example:
```python
model = Sequential()
model.add(Dense(32, input_dim=32))
# now: model.output_shape == (None, 32)
# note: `None` is the batch dimension
model.add(RepeatVector(3))
# now: model.output_shape == (None, 3, 32)
```
Arguments:
n: Integer, repetition factor.
Input shape:
2D tensor of shape `(num_samples, features)`.
Output shape:
3D tensor of shape `(num_samples, n, features)`.
"""
def __init__(self, n, **kwargs):
super(RepeatVector, self).__init__(**kwargs)
self.n = n
self.input_spec = InputSpec(ndim=2)
def compute_output_shape(self, input_shape):
input_shape = tensor_shape.TensorShape(input_shape).as_list()
return tensor_shape.TensorShape([input_shape[0], self.n, input_shape[1]])
def call(self, inputs):
return K.repeat(inputs, self.n)
def get_config(self):
config = {'n': self.n}
base_config = super(RepeatVector, self).get_config()
return dict(list(base_config.items()) + list(config.items()))
@keras_export('keras.layers.Lambda')
class Lambda(Layer):
"""Wraps arbitrary expressions as a `Layer` object.
The `Lambda` layer exists so that arbitrary TensorFlow functions
can be used when constructing `Sequential` and Functional API
models. `Lambda` layers are best suited for simple operations or
quick experimentation. For more advanced usecases, follow
[this guide](https://www.tensorflow.org/guide/keras/custom_layers_and_models)
for subclassing `tf.keras.layers.Layer`.
The main reason to subclass `tf.keras.layers.Layer` instead of using a
`Lambda` layer is saving and inspecting a Model. `Lambda` layers
are saved by serializing the Python bytecode, whereas subclassed
Layers can be saved via overriding their `get_config` method. Overriding
`get_config` improves the portability of Models. Models that rely on
subclassed Layers are also often easier to visualize and reason about.
Examples:
```python
# add a x -> x^2 layer
model.add(Lambda(lambda x: x ** 2))
```
```python
# add a layer that returns the concatenation
# of the positive part of the input and
# the opposite of the negative part
def antirectifier(x):
x -= K.mean(x, axis=1, keepdims=True)
x = K.l2_normalize(x, axis=1)
pos = K.relu(x)
neg = K.relu(-x)
return K.concatenate([pos, neg], axis=1)
model.add(Lambda(antirectifier))
```
Variables:
While it is possible to use Variables with Lambda layers, this practice is
discouraged as it can easily lead to bugs. For instance, consider the
following layer:
```python
scale = tf.Variable(1.)
scale_layer = tf.keras.layers.Lambda(lambda x: x * scale)
```
Because scale_layer does not directly track the `scale` variable, it will
not appear in `scale_layer.trainable_weights` and will therefore not be
trained if `scale_layer` is used in a Model.
A better pattern is to write a subclassed Layer:
```python
class ScaleLayer(tf.keras.layers.Layer):
def __init__(self):
super(ScaleLayer, self).__init__()
self.scale = tf.Variable(1.)
def call(self, inputs):
return inputs * self.scale
```
In general, Lambda layers can be convenient for simple stateless
computation, but anything more complex should use a subclass Layer instead.
Arguments:
function: The function to be evaluated. Takes input tensor as first
argument.
output_shape: Expected output shape from function. This argument can be
inferred if not explicitly provided. Can be a tuple or function. If a
tuple, it only specifies the first dimension onward;
sample dimension is assumed either the same as the input: `output_shape =
(input_shape[0], ) + output_shape` or, the input is `None` and
the sample dimension is also `None`: `output_shape = (None, ) +
output_shape` If a function, it specifies the entire shape as a function
of the
input shape: `output_shape = f(input_shape)`
mask: Either None (indicating no masking) or a callable with the same
signature as the `compute_mask` layer method, or a tensor that will be
returned as output mask regardless what the input is.
arguments: Optional dictionary of keyword arguments to be passed to the
function.
Input shape: Arbitrary. Use the keyword argument input_shape (tuple of
integers, does not include the samples axis) when using this layer as the
first layer in a model.
Output shape: Specified by `output_shape` argument
"""
@trackable.no_automatic_dependency_tracking
def __init__(self, function, output_shape=None, mask=None, arguments=None,
**kwargs):
super(Lambda, self).__init__(**kwargs)
self.arguments = arguments or {}
self.function = function
if mask is not None:
self.supports_masking = True
self.mask = mask
self._supports_ragged_inputs = True
self._output_shape = output_shape
# Warning on every invocation will be quite irksome in Eager mode.
self._already_warned = False
function_args = tf_inspect.getfullargspec(function).args
self._fn_expects_training_arg = 'training' in function_args
self._fn_expects_mask_arg = 'mask' in function_args
@tf_utils.shape_type_conversion
def compute_output_shape(self, input_shape):
if self._output_shape is None:
# Make use of existing autocomputation but provide Lambda-specific
# error message. This is always safe to run even when the outer context
# is Graph mode because Lambda layers don't have side effects such as
# `add_loss`.
with context.eager_mode():
try:
return super(Lambda, self).compute_output_shape(input_shape)
except NotImplementedError:
raise NotImplementedError(
'We could not automatically infer the shape of the Lambda\'s '
'output. Please specify `output_shape` for this Lambda.')
if callable(self._output_shape):
output_shapes = self._output_shape(input_shape)
return tf_utils.convert_shapes(output_shapes, to_tuples=False)
# Output shapes are passed directly and don't include batch dimension.
input_tensor_shape = tf_utils.convert_shapes(input_shape, to_tuples=False)
batch_size = nest.flatten(input_tensor_shape)[0][0] if input_shape else None
def _add_batch(shape):
return tensor_shape.TensorShape([batch_size] + shape.as_list())
output_shapes = tf_utils.convert_shapes(self._output_shape, to_tuples=False)
return nest.map_structure(_add_batch, output_shapes)
def call(self, inputs, mask=None, training=None):
# We must copy for thread safety, but it only needs to be a shallow copy.
kwargs = {k: v for k, v in self.arguments.items()}
if self._fn_expects_mask_arg:
kwargs['mask'] = mask
if self._fn_expects_training_arg:
kwargs['training'] = training
created_variables = []
def _variable_creator(next_creator, **kwargs):
var = next_creator(**kwargs)
created_variables.append(var)
return var
with backprop.GradientTape(watch_accessed_variables=True) as tape,\
variable_scope.variable_creator_scope(_variable_creator):
result = self.function(inputs, **kwargs)
self._check_variables(created_variables, tape.watched_variables())
return result
def _check_variables(self, created_variables, accessed_variables):
if not created_variables and not accessed_variables:
# In the common case that a Lambda layer does not touch a Variable, we
# don't want to incur the runtime cost of assembling any state used for
# checking only to immediately discard it.
return
tracked_weights = set(v.experimental_ref() for v in self.weights)
untracked_new_vars = [v for v in created_variables
if v.experimental_ref() not in tracked_weights]
if untracked_new_vars:
variable_str = '\n'.join([' {}'.format(i) for i in untracked_new_vars])
error_str = textwrap.dedent(
'''
The following Variables were created within a Lambda layer ({name})
but are not tracked by said layer:
{variable_str}
The layer cannot safely ensure proper Variable reuse across multiple
calls, and consquently this behavior is disallowed for safety. Lambda
layers are not well suited to stateful computation; instead, writing a
subclassed Layer is the recommend way to define layers with
Variables.'''
).format(name=self.name, variable_str=variable_str)
raise ValueError(error_str)
untracked_used_vars = [v for v in accessed_variables
if v.experimental_ref() not in tracked_weights]
if untracked_used_vars and not self._already_warned:
variable_str = '\n'.join([' {}'.format(i) for i in untracked_used_vars])
self._warn(textwrap.dedent(
'''
The following Variables were used a Lambda layer's call ({name}), but
are not present in its tracked objects:
{variable_str}
It is possible that this is intended behavior, but it is more likely
an omission. This is a strong indication that this layer should be
formulated as a subclassed Layer rather than a Lambda layer.'''
).format(name=self.name, variable_str=variable_str))
self._already_warned = True
def _warn(self, msg):
# This method will be overridden in a unit test to raise an error, because
# self.assertWarns is not universally implemented.
return tf_logging.warn(msg)
def compute_mask(self, inputs, mask=None):
if callable(self.mask):
return self.mask(inputs, mask)
return self.mask
def get_config(self):
function_config = self._serialize_function_to_config(self.function)
output_shape_config = self._serialize_function_to_config(self._output_shape,
allow_raw=True)
config = {
'function': function_config[0],
'function_type': function_config[1],
'module': function_config[2],
'output_shape': output_shape_config[0],
'output_shape_type': output_shape_config[1],
'output_shape_module': output_shape_config[2],
}
if self.mask is not None:
mask_config = self._serialize_function_to_config(self.mask)
config.update({
'mask': mask_config[0],
'mask_type': mask_config[1],
'mask_module': mask_config[2]
})
config['arguments'] = self.arguments
base_config = super(Lambda, self).get_config()
return dict(list(base_config.items()) + list(config.items()))
def _serialize_function_to_config(self, inputs, allow_raw=False):
if isinstance(inputs, python_types.LambdaType):
output = generic_utils.func_dump(inputs)
output_type = 'lambda'
module = inputs.__module__
elif callable(inputs):
output = inputs.__name__
output_type = 'function'
module = inputs.__module__
elif allow_raw:
output = inputs
output_type = 'raw'
module = None
else:
raise ValueError(
'Invalid input for serialization, type: %s ' % type(inputs))
return output, output_type, module
@classmethod
def from_config(cls, config, custom_objects=None):
config = config.copy()
function = cls._parse_function_from_config(
config, custom_objects, 'function', 'module', 'function_type')
output_shape = cls._parse_function_from_config(
config, custom_objects, 'output_shape', 'output_shape_module',
'output_shape_type')
if 'mask' in config:
mask = cls._parse_function_from_config(
config, custom_objects, 'mask', 'mask_module', 'mask_type')
else:
mask = None
config['function'] = function
config['output_shape'] = output_shape
config['mask'] = mask
# If arguments were numpy array, they have been saved as
# list. We need to recover the ndarray
if 'arguments' in config:
for key in config['arguments']:
if isinstance(config['arguments'][key], dict):
arg_dict = config['arguments'][key]
if 'type' in arg_dict and arg_dict['type'] == 'ndarray':
# Overwrite the argument with its numpy translation
config['arguments'][key] = np.array(arg_dict['value'])
return cls(**config)
@classmethod
def _parse_function_from_config(
cls, config, custom_objects, func_attr_name, module_attr_name,
func_type_attr_name):
globs = globals()
module = config.pop(module_attr_name, None)
if module in sys.modules:
globs.update(sys.modules[module].__dict__)
elif module is not None:
# Note: we don't know the name of the function if it's a lambda.
warnings.warn('{} is not loaded, but a Lambda layer uses it. '
'It may cause errors.'.format(module)
, UserWarning)
if custom_objects:
globs.update(custom_objects)
function_type = config.pop(func_type_attr_name)
if function_type == 'function':
# Simple lookup in custom objects
function = generic_utils.deserialize_keras_object(
config[func_attr_name],
custom_objects=custom_objects,
printable_module_name='function in Lambda layer')
elif function_type == 'lambda':
# Unsafe deserialization from bytecode
function = generic_utils.func_load(
config[func_attr_name], globs=globs)
elif function_type == 'raw':