/
optimization.html
1140 lines (1076 loc) · 54.2 KB
/
optimization.html
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
<!doctype html>
<html lang="en">
<head>
<meta charset="utf-8">
<meta name="viewport" content="width=device-width, initial-scale=1, minimum-scale=1" />
<meta name="generator" content="pdoc 0.9.2" />
<title>ktrain.lroptimize.optimization API documentation</title>
<meta name="description" content="Functions and classes related to optimization (weight updates)." />
<link rel="preload stylesheet" as="style" href="https://cdnjs.cloudflare.com/ajax/libs/10up-sanitize.css/11.0.1/sanitize.min.css" integrity="sha256-PK9q560IAAa6WVRRh76LtCaI8pjTJ2z11v0miyNNjrs=" crossorigin>
<link rel="preload stylesheet" as="style" href="https://cdnjs.cloudflare.com/ajax/libs/10up-sanitize.css/11.0.1/typography.min.css" integrity="sha256-7l/o7C8jubJiy74VsKTidCy1yBkRtiUGbVkYBylBqUg=" crossorigin>
<link rel="stylesheet preload" as="style" href="https://cdnjs.cloudflare.com/ajax/libs/highlight.js/10.1.1/styles/github.min.css" crossorigin>
<style>:root{--highlight-color:#fe9}.flex{display:flex !important}body{line-height:1.5em}#content{padding:20px}#sidebar{padding:30px;overflow:hidden}#sidebar > *:last-child{margin-bottom:2cm}.http-server-breadcrumbs{font-size:130%;margin:0 0 15px 0}#footer{font-size:.75em;padding:5px 30px;border-top:1px solid #ddd;text-align:right}#footer p{margin:0 0 0 1em;display:inline-block}#footer p:last-child{margin-right:30px}h1,h2,h3,h4,h5{font-weight:300}h1{font-size:2.5em;line-height:1.1em}h2{font-size:1.75em;margin:1em 0 .50em 0}h3{font-size:1.4em;margin:25px 0 10px 0}h4{margin:0;font-size:105%}h1:target,h2:target,h3:target,h4:target,h5:target,h6:target{background:var(--highlight-color);padding:.2em 0}a{color:#058;text-decoration:none;transition:color .3s ease-in-out}a:hover{color:#e82}.title code{font-weight:bold}h2[id^="header-"]{margin-top:2em}.ident{color:#900}pre code{background:#f8f8f8;font-size:.8em;line-height:1.4em}code{background:#f2f2f1;padding:1px 4px;overflow-wrap:break-word}h1 code{background:transparent}pre{background:#f8f8f8;border:0;border-top:1px solid #ccc;border-bottom:1px solid #ccc;margin:1em 0;padding:1ex}#http-server-module-list{display:flex;flex-flow:column}#http-server-module-list div{display:flex}#http-server-module-list dt{min-width:10%}#http-server-module-list p{margin-top:0}.toc ul,#index{list-style-type:none;margin:0;padding:0}#index code{background:transparent}#index h3{border-bottom:1px solid #ddd}#index ul{padding:0}#index h4{margin-top:.6em;font-weight:bold}@media (min-width:200ex){#index .two-column{column-count:2}}@media (min-width:300ex){#index .two-column{column-count:3}}dl{margin-bottom:2em}dl dl:last-child{margin-bottom:4em}dd{margin:0 0 1em 3em}#header-classes + dl > dd{margin-bottom:3em}dd dd{margin-left:2em}dd p{margin:10px 0}.name{background:#eee;font-weight:bold;font-size:.85em;padding:5px 10px;display:inline-block;min-width:40%}.name:hover{background:#e0e0e0}dt:target .name{background:var(--highlight-color)}.name > span:first-child{white-space:nowrap}.name.class > span:nth-child(2){margin-left:.4em}.inherited{color:#999;border-left:5px solid #eee;padding-left:1em}.inheritance em{font-style:normal;font-weight:bold}.desc h2{font-weight:400;font-size:1.25em}.desc h3{font-size:1em}.desc dt code{background:inherit}.source summary,.git-link-div{color:#666;text-align:right;font-weight:400;font-size:.8em;text-transform:uppercase}.source summary > *{white-space:nowrap;cursor:pointer}.git-link{color:inherit;margin-left:1em}.source pre{max-height:500px;overflow:auto;margin:0}.source pre code{font-size:12px;overflow:visible}.hlist{list-style:none}.hlist li{display:inline}.hlist li:after{content:',\2002'}.hlist li:last-child:after{content:none}.hlist .hlist{display:inline;padding-left:1em}img{max-width:100%}td{padding:0 .5em}.admonition{padding:.1em .5em;margin-bottom:1em}.admonition-title{font-weight:bold}.admonition.note,.admonition.info,.admonition.important{background:#aef}.admonition.todo,.admonition.versionadded,.admonition.tip,.admonition.hint{background:#dfd}.admonition.warning,.admonition.versionchanged,.admonition.deprecated{background:#fd4}.admonition.error,.admonition.danger,.admonition.caution{background:lightpink}</style>
<style media="screen and (min-width: 700px)">@media screen and (min-width:700px){#sidebar{width:30%;height:100vh;overflow:auto;position:sticky;top:0}#content{width:70%;max-width:100ch;padding:3em 4em;border-left:1px solid #ddd}pre code{font-size:1em}.item .name{font-size:1em}main{display:flex;flex-direction:row-reverse;justify-content:flex-end}.toc ul ul,#index ul{padding-left:1.5em}.toc > ul > li{margin-top:.5em}}</style>
<style media="print">@media print{#sidebar h1{page-break-before:always}.source{display:none}}@media print{*{background:transparent !important;color:#000 !important;box-shadow:none !important;text-shadow:none !important}a[href]:after{content:" (" attr(href) ")";font-size:90%}a[href][title]:after{content:none}abbr[title]:after{content:" (" attr(title) ")"}.ir a:after,a[href^="javascript:"]:after,a[href^="#"]:after{content:""}pre,blockquote{border:1px solid #999;page-break-inside:avoid}thead{display:table-header-group}tr,img{page-break-inside:avoid}img{max-width:100% !important}@page{margin:0.5cm}p,h2,h3{orphans:3;widows:3}h1,h2,h3,h4,h5,h6{page-break-after:avoid}}</style>
<script defer src="https://cdnjs.cloudflare.com/ajax/libs/highlight.js/10.1.1/highlight.min.js" integrity="sha256-Uv3H6lx7dJmRfRvH8TH6kJD1TSK1aFcwgx+mdg3epi8=" crossorigin></script>
<script>window.addEventListener('DOMContentLoaded', () => hljs.initHighlighting())</script>
</head>
<body>
<main>
<article id="content">
<header>
<h1 class="title">Module <code>ktrain.lroptimize.optimization</code></h1>
</header>
<section id="section-intro">
<p>Functions and classes related to optimization (weight updates).</p>
<details class="source">
<summary>
<span>Expand source code</span>
</summary>
<pre><code class="python"># Copyright 2019 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.
# ==============================================================================
"""Functions and classes related to optimization (weight updates)."""
import re
from typing import Callable, List, Optional, Union
import tensorflow as tf
class WarmUp(tf.keras.optimizers.schedules.LearningRateSchedule):
"""
```
Applies a warmup schedule on a given learning rate decay schedule.
Args:
initial_learning_rate (:obj:`float`):
The initial learning rate for the schedule after the warmup (so this will be the learning rate at the end
of the warmup).
decay_schedule_fn (:obj:`Callable`):
The schedule function to apply after the warmup for the rest of training.
warmup_steps (:obj:`int`):
The number of steps for the warmup part of training.
power (:obj:`float`, `optional`, defaults to 1):
The power to use for the polynomial warmup (defaults is a linear warmup).
name (:obj:`str`, `optional`):
Optional name prefix for the returned tensors during the schedule.
```
"""
def __init__(
self,
initial_learning_rate: float,
decay_schedule_fn: Callable,
warmup_steps: int,
power: float = 1.0,
name: str = None,
):
super().__init__()
self.initial_learning_rate = initial_learning_rate
self.warmup_steps = warmup_steps
self.power = power
self.decay_schedule_fn = decay_schedule_fn
self.name = name
def __call__(self, step):
with tf.name_scope(self.name or "WarmUp") as name:
# Implements polynomial warmup. i.e., if global_step < warmup_steps, the
# learning rate will be `global_step/num_warmup_steps * init_lr`.
global_step_float = tf.cast(step, tf.float32)
warmup_steps_float = tf.cast(self.warmup_steps, tf.float32)
warmup_percent_done = global_step_float / warmup_steps_float
warmup_learning_rate = self.initial_learning_rate * tf.math.pow(
warmup_percent_done, self.power
)
return tf.cond(
global_step_float < warmup_steps_float,
lambda: warmup_learning_rate,
lambda: self.decay_schedule_fn(step - self.warmup_steps),
name=name,
)
def get_config(self):
return {
"initial_learning_rate": self.initial_learning_rate,
"decay_schedule_fn": self.decay_schedule_fn,
"warmup_steps": self.warmup_steps,
"power": self.power,
"name": self.name,
}
def create_optimizer(
init_lr: float,
num_train_steps: int,
num_warmup_steps: int,
min_lr_ratio: float = 0.0,
adam_epsilon: float = 1e-8,
weight_decay_rate: float = 0.0,
include_in_weight_decay: Optional[List[str]] = None,
):
"""
```
Creates an optimizer with a learning rate schedule using a warmup phase followed by a linear decay.
Args:
init_lr (:obj:`float`):
The desired learning rate at the end of the warmup phase.
num_train_step (:obj:`int`):
The total number of training steps.
num_warmup_steps (:obj:`int`):
The number of warmup steps.
min_lr_ratio (:obj:`float`, `optional`, defaults to 0):
The final learning rate at the end of the linear decay will be :obj:`init_lr * min_lr_ratio`.
adam_epsilon (:obj:`float`, `optional`, defaults to 1e-8):
The epsilon to use in Adam.
weight_decay_rate (:obj:`float`, `optional`, defaults to 0):
The weight decay to use.
include_in_weight_decay (:obj:`List[str]`, `optional`):
List of the parameter names (or re patterns) to apply weight decay to. If none is passed, weight decay is
applied to all parameters except bias and layer norm parameters.
```
"""
# Implements linear decay of the learning rate.
lr_schedule = tf.keras.optimizers.schedules.PolynomialDecay(
initial_learning_rate=init_lr,
decay_steps=num_train_steps - num_warmup_steps,
end_learning_rate=init_lr * min_lr_ratio,
)
if num_warmup_steps:
lr_schedule = WarmUp(
initial_learning_rate=init_lr,
decay_schedule_fn=lr_schedule,
warmup_steps=num_warmup_steps,
)
if weight_decay_rate > 0.0:
optimizer = AdamWeightDecay(
learning_rate=lr_schedule,
weight_decay_rate=weight_decay_rate,
beta_1=0.9,
beta_2=0.999,
epsilon=adam_epsilon,
exclude_from_weight_decay=["LayerNorm", "layer_norm", "bias"],
include_in_weight_decay=include_in_weight_decay,
)
else:
optimizer = tf.keras.optimizers.Adam(
learning_rate=lr_schedule, epsilon=adam_epsilon
)
# We return the optimizer and the LR scheduler in order to better track the
# evolution of the LR independently of the optimizer.
return optimizer, lr_schedule
class AdamWeightDecay(tf.keras.optimizers.Adam):
"""
```
Adam enables L2 weight decay and clip_by_global_norm on gradients. Just adding the square of the weights to the
loss function is *not* the correct way of using L2 regularization/weight decay with Adam, since that will interact
with the m and v parameters in strange ways as shown in
`Decoupled Weight Decay Regularization <https://arxiv.org/abs/1711.05101>`__.
Instead we want ot decay the weights in a manner that doesn't interact with the m/v parameters. This is equivalent
to adding the square of the weights to the loss with plain (non-momentum) SGD.
Args:
learning_rate (:obj:`Union[float, tf.keras.optimizers.schedules.LearningRateSchedule]`, `optional`, defaults to 1e-3):
The learning rate to use or a schedule.
beta_1 (:obj:`float`, `optional`, defaults to 0.9):
The beta1 parameter in Adam, which is the exponential decay rate for the 1st momentum estimates.
beta_2 (:obj:`float`, `optional`, defaults to 0.999):
The beta2 parameter in Adam, which is the exponential decay rate for the 2nd momentum estimates.
epsilon (:obj:`float`, `optional`, defaults to 1e-7):
The epsilon paramenter in Adam, which is a small constant for numerical stability.
amsgrad (:obj:`bool`, `optional`, default to `False`):
Wheter to apply AMSGrad varient of this algorithm or not, see
`On the Convergence of Adam and Beyond <https://arxiv.org/abs/1904.09237>`__.
weight_decay_rate (:obj:`float`, `optional`, defaults to 0):
The weight decay to apply.
include_in_weight_decay (:obj:`List[str]`, `optional`):
List of the parameter names (or re patterns) to apply weight decay to. If none is passed, weight decay is
applied to all parameters by default (unless they are in :obj:`exclude_from_weight_decay`).
exclude_from_weight_decay (:obj:`List[str]`, `optional`):
List of the parameter names (or re patterns) to exclude from applying weight decay to. If a
:obj:`include_in_weight_decay` is passed, the names in it will supersede this list.
name (:obj:`str`, `optional`, defaults to 'AdamWeightDecay'):
Optional name for the operations created when applying gradients.
kwargs:
Keyward arguments. Allowed to be {``clipnorm``, ``clipvalue``, ``lr``, ``decay``}. ``clipnorm`` is clip
gradients by norm; ``clipvalue`` is clip gradients by value, ``decay`` is included for backward
compatibility to allow time inverse decay of learning rate. ``lr`` is included for backward compatibility,
recommended to use ``learning_rate`` instead.
```
"""
def __init__(
self,
learning_rate: Union[
float, tf.keras.optimizers.schedules.LearningRateSchedule
] = 0.001,
beta_1: float = 0.9,
beta_2: float = 0.999,
epsilon: float = 1e-7,
amsgrad: bool = False,
weight_decay_rate: float = 0.0,
include_in_weight_decay: Optional[List[str]] = None,
exclude_from_weight_decay: Optional[List[str]] = None,
name: str = "AdamWeightDecay",
**kwargs
):
super().__init__(
learning_rate, beta_1, beta_2, epsilon, amsgrad, name, **kwargs
)
self.weight_decay_rate = weight_decay_rate
self._include_in_weight_decay = include_in_weight_decay
self._exclude_from_weight_decay = exclude_from_weight_decay
@classmethod
def from_config(cls, config):
"""Creates an optimizer from its config with WarmUp custom object."""
custom_objects = {"WarmUp": WarmUp}
return super(AdamWeightDecay, cls).from_config(
config, custom_objects=custom_objects
)
def _prepare_local(self, var_device, var_dtype, apply_state):
super(AdamWeightDecay, self)._prepare_local(var_device, var_dtype, apply_state)
apply_state[(var_device, var_dtype)]["weight_decay_rate"] = tf.constant(
self.weight_decay_rate, name="adam_weight_decay_rate"
)
def _decay_weights_op(self, var, learning_rate, apply_state):
do_decay = self._do_use_weight_decay(var.name)
if do_decay:
return var.assign_sub(
learning_rate
* var
* apply_state[(var.device, var.dtype.base_dtype)]["weight_decay_rate"],
use_locking=self._use_locking,
)
return tf.no_op()
def apply_gradients(self, grads_and_vars, name=None, **kwargs):
# 2020-07-05: added **kwargs based on https://github.com/tensorflow/addons/issues/1267
grads, tvars = list(zip(*grads_and_vars))
return super(AdamWeightDecay, self).apply_gradients(
zip(grads, tvars),
name=name,
)
def _get_lr(self, var_device, var_dtype, apply_state):
"""Retrieves the learning rate with the given state."""
if apply_state is None:
return self._decayed_lr_t[var_dtype], {}
apply_state = apply_state or {}
coefficients = apply_state.get((var_device, var_dtype))
if coefficients is None:
coefficients = self._fallback_apply_state(var_device, var_dtype)
apply_state[(var_device, var_dtype)] = coefficients
return coefficients["lr_t"], dict(apply_state=apply_state)
def _resource_apply_dense(self, grad, var, apply_state=None):
lr_t, kwargs = self._get_lr(var.device, var.dtype.base_dtype, apply_state)
decay = self._decay_weights_op(var, lr_t, apply_state)
with tf.control_dependencies([decay]):
return super(AdamWeightDecay, self)._resource_apply_dense(
grad, var, **kwargs
)
def _resource_apply_sparse(self, grad, var, indices, apply_state=None):
lr_t, kwargs = self._get_lr(var.device, var.dtype.base_dtype, apply_state)
decay = self._decay_weights_op(var, lr_t, apply_state)
with tf.control_dependencies([decay]):
return super(AdamWeightDecay, self)._resource_apply_sparse(
grad, var, indices, **kwargs
)
def get_config(self):
config = super().get_config()
config.update({"weight_decay_rate": self.weight_decay_rate})
return config
def _do_use_weight_decay(self, param_name):
"""Whether to use L2 weight decay for `param_name`."""
if self.weight_decay_rate == 0:
return False
if self._include_in_weight_decay:
for r in self._include_in_weight_decay:
if re.search(r, param_name) is not None:
return True
if self._exclude_from_weight_decay:
for r in self._exclude_from_weight_decay:
if re.search(r, param_name) is not None:
return False
return True
# Extracted from https://github.com/OpenNMT/OpenNMT-tf/blob/master/opennmt/optimizers/utils.py
class GradientAccumulator(object):
"""Gradient accumulation utility.
When used with a distribution strategy, the accumulator should be called in a
replica context. Gradients will be accumulated locally on each replica and
without synchronization. Users should then call ``.gradients``, scale the
gradients if required, and pass the result to ``apply_gradients``.
"""
# We use the ON_READ synchronization policy so that no synchronization is
# performed on assignment. To get the value, we call .value() which returns the
# value on the current replica without synchronization.
def __init__(self):
"""Initializes the accumulator."""
self._gradients = []
self._accum_steps = None
@property
def step(self):
"""Number of accumulated steps."""
if self._accum_steps is None:
self._accum_steps = tf.Variable(
tf.constant(0, dtype=tf.int64),
trainable=False,
synchronization=tf.VariableSynchronization.ON_READ,
aggregation=tf.VariableAggregation.ONLY_FIRST_REPLICA,
)
return self._accum_steps.value()
@property
def gradients(self):
"""The accumulated gradients on the current replica."""
if not self._gradients:
raise ValueError(
"The accumulator should be called first to initialize the gradients"
)
return list(
gradient.value() if gradient is not None else gradient
for gradient in self._gradients
)
def __call__(self, gradients):
"""Accumulates :obj:`gradients` on the current replica."""
if not self._gradients:
_ = self.step # Create the step variable.
self._gradients.extend(
[
tf.Variable(
tf.zeros_like(gradient),
trainable=False,
synchronization=tf.VariableSynchronization.ON_READ,
aggregation=tf.VariableAggregation.ONLY_FIRST_REPLICA,
)
if gradient is not None
else gradient
for gradient in gradients
]
)
if len(gradients) != len(self._gradients):
raise ValueError(
"Expected %s gradients, but got %d"
% (len(self._gradients), len(gradients))
)
for accum_gradient, gradient in zip(self._gradients, gradients):
if accum_gradient is not None and gradient is not None:
accum_gradient.assign_add(gradient)
self._accum_steps.assign_add(1)
def reset(self):
"""Resets the accumulated gradients on the current replica."""
if not self._gradients:
return
self._accum_steps.assign(0)
for gradient in self._gradients:
if gradient is not None:
gradient.assign(tf.zeros_like(gradient))</code></pre>
</details>
</section>
<section>
</section>
<section>
</section>
<section>
<h2 class="section-title" id="header-functions">Functions</h2>
<dl>
<dt id="ktrain.lroptimize.optimization.create_optimizer"><code class="name flex">
<span>def <span class="ident">create_optimizer</span></span>(<span>init_lr: float, num_train_steps: int, num_warmup_steps: int, min_lr_ratio: float = 0.0, adam_epsilon: float = 1e-08, weight_decay_rate: float = 0.0, include_in_weight_decay: Union[List[str], NoneType] = None)</span>
</code></dt>
<dd>
<div class="desc"><pre><code>Creates an optimizer with a learning rate schedule using a warmup phase followed by a linear decay.
Args:
init_lr (:obj:`float`):
The desired learning rate at the end of the warmup phase.
num_train_step (:obj:`int`):
The total number of training steps.
num_warmup_steps (:obj:`int`):
The number of warmup steps.
min_lr_ratio (:obj:`float`, `optional`, defaults to 0):
The final learning rate at the end of the linear decay will be :obj:`init_lr * min_lr_ratio`.
adam_epsilon (:obj:`float`, `optional`, defaults to 1e-8):
The epsilon to use in Adam.
weight_decay_rate (:obj:`float`, `optional`, defaults to 0):
The weight decay to use.
include_in_weight_decay (:obj:`List[str]`, `optional`):
List of the parameter names (or re patterns) to apply weight decay to. If none is passed, weight decay is
applied to all parameters except bias and layer norm parameters.
</code></pre></div>
<details class="source">
<summary>
<span>Expand source code</span>
</summary>
<pre><code class="python">def create_optimizer(
init_lr: float,
num_train_steps: int,
num_warmup_steps: int,
min_lr_ratio: float = 0.0,
adam_epsilon: float = 1e-8,
weight_decay_rate: float = 0.0,
include_in_weight_decay: Optional[List[str]] = None,
):
"""
```
Creates an optimizer with a learning rate schedule using a warmup phase followed by a linear decay.
Args:
init_lr (:obj:`float`):
The desired learning rate at the end of the warmup phase.
num_train_step (:obj:`int`):
The total number of training steps.
num_warmup_steps (:obj:`int`):
The number of warmup steps.
min_lr_ratio (:obj:`float`, `optional`, defaults to 0):
The final learning rate at the end of the linear decay will be :obj:`init_lr * min_lr_ratio`.
adam_epsilon (:obj:`float`, `optional`, defaults to 1e-8):
The epsilon to use in Adam.
weight_decay_rate (:obj:`float`, `optional`, defaults to 0):
The weight decay to use.
include_in_weight_decay (:obj:`List[str]`, `optional`):
List of the parameter names (or re patterns) to apply weight decay to. If none is passed, weight decay is
applied to all parameters except bias and layer norm parameters.
```
"""
# Implements linear decay of the learning rate.
lr_schedule = tf.keras.optimizers.schedules.PolynomialDecay(
initial_learning_rate=init_lr,
decay_steps=num_train_steps - num_warmup_steps,
end_learning_rate=init_lr * min_lr_ratio,
)
if num_warmup_steps:
lr_schedule = WarmUp(
initial_learning_rate=init_lr,
decay_schedule_fn=lr_schedule,
warmup_steps=num_warmup_steps,
)
if weight_decay_rate > 0.0:
optimizer = AdamWeightDecay(
learning_rate=lr_schedule,
weight_decay_rate=weight_decay_rate,
beta_1=0.9,
beta_2=0.999,
epsilon=adam_epsilon,
exclude_from_weight_decay=["LayerNorm", "layer_norm", "bias"],
include_in_weight_decay=include_in_weight_decay,
)
else:
optimizer = tf.keras.optimizers.Adam(
learning_rate=lr_schedule, epsilon=adam_epsilon
)
# We return the optimizer and the LR scheduler in order to better track the
# evolution of the LR independently of the optimizer.
return optimizer, lr_schedule</code></pre>
</details>
</dd>
</dl>
</section>
<section>
<h2 class="section-title" id="header-classes">Classes</h2>
<dl>
<dt id="ktrain.lroptimize.optimization.AdamWeightDecay"><code class="flex name class">
<span>class <span class="ident">AdamWeightDecay</span></span>
<span>(</span><span>learning_rate: Union[float, tensorflow.python.keras.optimizer_v2.learning_rate_schedule.LearningRateSchedule] = 0.001, beta_1: float = 0.9, beta_2: float = 0.999, epsilon: float = 1e-07, amsgrad: bool = False, weight_decay_rate: float = 0.0, include_in_weight_decay: Union[List[str], NoneType] = None, exclude_from_weight_decay: Union[List[str], NoneType] = None, name: str = 'AdamWeightDecay', **kwargs)</span>
</code></dt>
<dd>
<div class="desc"><pre><code>Adam enables L2 weight decay and clip_by_global_norm on gradients. Just adding the square of the weights to the
loss function is *not* the correct way of using L2 regularization/weight decay with Adam, since that will interact
with the m and v parameters in strange ways as shown in
`Decoupled Weight Decay Regularization <https://arxiv.org/abs/1711.05101>`__.
Instead we want ot decay the weights in a manner that doesn't interact with the m/v parameters. This is equivalent
to adding the square of the weights to the loss with plain (non-momentum) SGD.
Args:
learning_rate (:obj:`Union[float, tf.keras.optimizers.schedules.LearningRateSchedule]`, `optional`, defaults to 1e-3):
The learning rate to use or a schedule.
beta_1 (:obj:`float`, `optional`, defaults to 0.9):
The beta1 parameter in Adam, which is the exponential decay rate for the 1st momentum estimates.
beta_2 (:obj:`float`, `optional`, defaults to 0.999):
The beta2 parameter in Adam, which is the exponential decay rate for the 2nd momentum estimates.
epsilon (:obj:`float`, `optional`, defaults to 1e-7):
The epsilon paramenter in Adam, which is a small constant for numerical stability.
amsgrad (:obj:`bool`, `optional`, default to `False`):
Wheter to apply AMSGrad varient of this algorithm or not, see
`On the Convergence of Adam and Beyond <https://arxiv.org/abs/1904.09237>`__.
weight_decay_rate (:obj:`float`, `optional`, defaults to 0):
The weight decay to apply.
include_in_weight_decay (:obj:`List[str]`, `optional`):
List of the parameter names (or re patterns) to apply weight decay to. If none is passed, weight decay is
applied to all parameters by default (unless they are in :obj:`exclude_from_weight_decay`).
exclude_from_weight_decay (:obj:`List[str]`, `optional`):
List of the parameter names (or re patterns) to exclude from applying weight decay to. If a
:obj:`include_in_weight_decay` is passed, the names in it will supersede this list.
name (:obj:`str`, `optional`, defaults to 'AdamWeightDecay'):
Optional name for the operations created when applying gradients.
kwargs:
Keyward arguments. Allowed to be {``clipnorm``, ``clipvalue``, ``lr``, ``decay``}. ``clipnorm`` is clip
gradients by norm; ``clipvalue`` is clip gradients by value, ``decay`` is included for backward
compatibility to allow time inverse decay of learning rate. ``lr`` is included for backward compatibility,
recommended to use ``learning_rate`` instead.
</code></pre>
<p>Create a new Optimizer.</p>
<p>This must be called by the constructors of subclasses.
Note that Optimizer instances should not bind to a single graph,
and so shouldn't keep Tensors as member variables. Generally
you should be able to use the _set_hyper()/state.get_hyper()
facility instead.</p>
<p>This class in stateful and thread-compatible.</p>
<h2 id="args">Args</h2>
<dl>
<dt><strong><code>name</code></strong></dt>
<dd>A non-empty string.
The name to use for accumulators created
for the optimizer.</dd>
<dt><strong><code>**kwargs</code></strong></dt>
<dd>keyword arguments. Allowed to be {<code>clipnorm</code>, <code>clipvalue</code>, <code>lr</code>,
<code>decay</code>}. <code>clipnorm</code> is clip gradients by norm; <code>clipvalue</code> is clip
gradients by value, <code>decay</code> is included for backward compatibility to
allow time inverse decay of learning rate. <code>lr</code> is included for backward
compatibility, recommended to use <code>learning_rate</code> instead.</dd>
</dl>
<h2 id="raises">Raises</h2>
<dl>
<dt><code>ValueError</code></dt>
<dd>If name is malformed.</dd>
</dl></div>
<details class="source">
<summary>
<span>Expand source code</span>
</summary>
<pre><code class="python">class AdamWeightDecay(tf.keras.optimizers.Adam):
"""
```
Adam enables L2 weight decay and clip_by_global_norm on gradients. Just adding the square of the weights to the
loss function is *not* the correct way of using L2 regularization/weight decay with Adam, since that will interact
with the m and v parameters in strange ways as shown in
`Decoupled Weight Decay Regularization <https://arxiv.org/abs/1711.05101>`__.
Instead we want ot decay the weights in a manner that doesn't interact with the m/v parameters. This is equivalent
to adding the square of the weights to the loss with plain (non-momentum) SGD.
Args:
learning_rate (:obj:`Union[float, tf.keras.optimizers.schedules.LearningRateSchedule]`, `optional`, defaults to 1e-3):
The learning rate to use or a schedule.
beta_1 (:obj:`float`, `optional`, defaults to 0.9):
The beta1 parameter in Adam, which is the exponential decay rate for the 1st momentum estimates.
beta_2 (:obj:`float`, `optional`, defaults to 0.999):
The beta2 parameter in Adam, which is the exponential decay rate for the 2nd momentum estimates.
epsilon (:obj:`float`, `optional`, defaults to 1e-7):
The epsilon paramenter in Adam, which is a small constant for numerical stability.
amsgrad (:obj:`bool`, `optional`, default to `False`):
Wheter to apply AMSGrad varient of this algorithm or not, see
`On the Convergence of Adam and Beyond <https://arxiv.org/abs/1904.09237>`__.
weight_decay_rate (:obj:`float`, `optional`, defaults to 0):
The weight decay to apply.
include_in_weight_decay (:obj:`List[str]`, `optional`):
List of the parameter names (or re patterns) to apply weight decay to. If none is passed, weight decay is
applied to all parameters by default (unless they are in :obj:`exclude_from_weight_decay`).
exclude_from_weight_decay (:obj:`List[str]`, `optional`):
List of the parameter names (or re patterns) to exclude from applying weight decay to. If a
:obj:`include_in_weight_decay` is passed, the names in it will supersede this list.
name (:obj:`str`, `optional`, defaults to 'AdamWeightDecay'):
Optional name for the operations created when applying gradients.
kwargs:
Keyward arguments. Allowed to be {``clipnorm``, ``clipvalue``, ``lr``, ``decay``}. ``clipnorm`` is clip
gradients by norm; ``clipvalue`` is clip gradients by value, ``decay`` is included for backward
compatibility to allow time inverse decay of learning rate. ``lr`` is included for backward compatibility,
recommended to use ``learning_rate`` instead.
```
"""
def __init__(
self,
learning_rate: Union[
float, tf.keras.optimizers.schedules.LearningRateSchedule
] = 0.001,
beta_1: float = 0.9,
beta_2: float = 0.999,
epsilon: float = 1e-7,
amsgrad: bool = False,
weight_decay_rate: float = 0.0,
include_in_weight_decay: Optional[List[str]] = None,
exclude_from_weight_decay: Optional[List[str]] = None,
name: str = "AdamWeightDecay",
**kwargs
):
super().__init__(
learning_rate, beta_1, beta_2, epsilon, amsgrad, name, **kwargs
)
self.weight_decay_rate = weight_decay_rate
self._include_in_weight_decay = include_in_weight_decay
self._exclude_from_weight_decay = exclude_from_weight_decay
@classmethod
def from_config(cls, config):
"""Creates an optimizer from its config with WarmUp custom object."""
custom_objects = {"WarmUp": WarmUp}
return super(AdamWeightDecay, cls).from_config(
config, custom_objects=custom_objects
)
def _prepare_local(self, var_device, var_dtype, apply_state):
super(AdamWeightDecay, self)._prepare_local(var_device, var_dtype, apply_state)
apply_state[(var_device, var_dtype)]["weight_decay_rate"] = tf.constant(
self.weight_decay_rate, name="adam_weight_decay_rate"
)
def _decay_weights_op(self, var, learning_rate, apply_state):
do_decay = self._do_use_weight_decay(var.name)
if do_decay:
return var.assign_sub(
learning_rate
* var
* apply_state[(var.device, var.dtype.base_dtype)]["weight_decay_rate"],
use_locking=self._use_locking,
)
return tf.no_op()
def apply_gradients(self, grads_and_vars, name=None, **kwargs):
# 2020-07-05: added **kwargs based on https://github.com/tensorflow/addons/issues/1267
grads, tvars = list(zip(*grads_and_vars))
return super(AdamWeightDecay, self).apply_gradients(
zip(grads, tvars),
name=name,
)
def _get_lr(self, var_device, var_dtype, apply_state):
"""Retrieves the learning rate with the given state."""
if apply_state is None:
return self._decayed_lr_t[var_dtype], {}
apply_state = apply_state or {}
coefficients = apply_state.get((var_device, var_dtype))
if coefficients is None:
coefficients = self._fallback_apply_state(var_device, var_dtype)
apply_state[(var_device, var_dtype)] = coefficients
return coefficients["lr_t"], dict(apply_state=apply_state)
def _resource_apply_dense(self, grad, var, apply_state=None):
lr_t, kwargs = self._get_lr(var.device, var.dtype.base_dtype, apply_state)
decay = self._decay_weights_op(var, lr_t, apply_state)
with tf.control_dependencies([decay]):
return super(AdamWeightDecay, self)._resource_apply_dense(
grad, var, **kwargs
)
def _resource_apply_sparse(self, grad, var, indices, apply_state=None):
lr_t, kwargs = self._get_lr(var.device, var.dtype.base_dtype, apply_state)
decay = self._decay_weights_op(var, lr_t, apply_state)
with tf.control_dependencies([decay]):
return super(AdamWeightDecay, self)._resource_apply_sparse(
grad, var, indices, **kwargs
)
def get_config(self):
config = super().get_config()
config.update({"weight_decay_rate": self.weight_decay_rate})
return config
def _do_use_weight_decay(self, param_name):
"""Whether to use L2 weight decay for `param_name`."""
if self.weight_decay_rate == 0:
return False
if self._include_in_weight_decay:
for r in self._include_in_weight_decay:
if re.search(r, param_name) is not None:
return True
if self._exclude_from_weight_decay:
for r in self._exclude_from_weight_decay:
if re.search(r, param_name) is not None:
return False
return True</code></pre>
</details>
<h3>Ancestors</h3>
<ul class="hlist">
<li>tensorflow.python.keras.optimizer_v2.adam.Adam</li>
<li>tensorflow.python.keras.optimizer_v2.optimizer_v2.OptimizerV2</li>
<li>tensorflow.python.training.tracking.base.Trackable</li>
</ul>
<h3>Static methods</h3>
<dl>
<dt id="ktrain.lroptimize.optimization.AdamWeightDecay.from_config"><code class="name flex">
<span>def <span class="ident">from_config</span></span>(<span>config)</span>
</code></dt>
<dd>
<div class="desc"><p>Creates an optimizer from its config with WarmUp custom object.</p></div>
<details class="source">
<summary>
<span>Expand source code</span>
</summary>
<pre><code class="python">@classmethod
def from_config(cls, config):
"""Creates an optimizer from its config with WarmUp custom object."""
custom_objects = {"WarmUp": WarmUp}
return super(AdamWeightDecay, cls).from_config(
config, custom_objects=custom_objects
)</code></pre>
</details>
</dd>
</dl>
<h3>Methods</h3>
<dl>
<dt id="ktrain.lroptimize.optimization.AdamWeightDecay.apply_gradients"><code class="name flex">
<span>def <span class="ident">apply_gradients</span></span>(<span>self, grads_and_vars, name=None, **kwargs)</span>
</code></dt>
<dd>
<div class="desc"><p>Apply gradients to variables.</p>
<p>This is the second part of <code>minimize()</code>. It returns an <code>Operation</code> that
applies gradients.</p>
<p>The method sums gradients from all replicas in the presence of
<code>tf.distribute.Strategy</code> by default. You can aggregate gradients yourself by
passing <code>experimental_aggregate_gradients=False</code>.</p>
<p>Example:</p>
<pre><code class="language-python">grads = tape.gradient(loss, vars)
grads = tf.distribute.get_replica_context().all_reduce('sum', grads)
# Processing aggregated gradients.
optimizer.apply_gradients(zip(grads, vars),
experimental_aggregate_gradients=False)
</code></pre>
<h2 id="args">Args</h2>
<dl>
<dt><strong><code>grads_and_vars</code></strong></dt>
<dd>List of (gradient, variable) pairs.</dd>
<dt><strong><code>name</code></strong></dt>
<dd>Optional name for the returned operation. Default to the name passed
to the <code>Optimizer</code> constructor.</dd>
<dt><strong><code>experimental_aggregate_gradients</code></strong></dt>
<dd>Whether to sum gradients from different
replicas in the presense of <code>tf.distribute.Strategy</code>. If False, it's
user responsibility to aggregate the gradients. Default to True.</dd>
</dl>
<h2 id="returns">Returns</h2>
<p>An <code>Operation</code> that applies the specified gradients. The <code>iterations</code>
will be automatically increased by 1.</p>
<h2 id="raises">Raises</h2>
<dl>
<dt><code>TypeError</code></dt>
<dd>If <code>grads_and_vars</code> is malformed.</dd>
<dt><code>ValueError</code></dt>
<dd>If none of the variables have gradients.</dd>
</dl></div>
<details class="source">
<summary>
<span>Expand source code</span>
</summary>
<pre><code class="python">def apply_gradients(self, grads_and_vars, name=None, **kwargs):
# 2020-07-05: added **kwargs based on https://github.com/tensorflow/addons/issues/1267
grads, tvars = list(zip(*grads_and_vars))
return super(AdamWeightDecay, self).apply_gradients(
zip(grads, tvars),
name=name,
)</code></pre>
</details>
</dd>
<dt id="ktrain.lroptimize.optimization.AdamWeightDecay.get_config"><code class="name flex">
<span>def <span class="ident">get_config</span></span>(<span>self)</span>
</code></dt>
<dd>
<div class="desc"><p>Returns the config of the optimizer.</p>
<p>An optimizer config is a Python dictionary (serializable)
containing the configuration of an optimizer.
The same optimizer can be reinstantiated later
(without any saved state) from this configuration.</p>
<h2 id="returns">Returns</h2>
<p>Python dictionary.</p></div>
<details class="source">
<summary>
<span>Expand source code</span>
</summary>
<pre><code class="python">def get_config(self):
config = super().get_config()
config.update({"weight_decay_rate": self.weight_decay_rate})
return config</code></pre>
</details>
</dd>
</dl>
</dd>
<dt id="ktrain.lroptimize.optimization.GradientAccumulator"><code class="flex name class">
<span>class <span class="ident">GradientAccumulator</span></span>
</code></dt>
<dd>
<div class="desc"><p>Gradient accumulation utility.
When used with a distribution strategy, the accumulator should be called in a
replica context. Gradients will be accumulated locally on each replica and
without synchronization. Users should then call <code>.gradients</code>, scale the
gradients if required, and pass the result to <code>apply_gradients</code>.</p>
<p>Initializes the accumulator.</p></div>
<details class="source">
<summary>
<span>Expand source code</span>
</summary>
<pre><code class="python">class GradientAccumulator(object):
"""Gradient accumulation utility.
When used with a distribution strategy, the accumulator should be called in a
replica context. Gradients will be accumulated locally on each replica and
without synchronization. Users should then call ``.gradients``, scale the
gradients if required, and pass the result to ``apply_gradients``.
"""
# We use the ON_READ synchronization policy so that no synchronization is
# performed on assignment. To get the value, we call .value() which returns the
# value on the current replica without synchronization.
def __init__(self):
"""Initializes the accumulator."""
self._gradients = []
self._accum_steps = None
@property
def step(self):
"""Number of accumulated steps."""
if self._accum_steps is None:
self._accum_steps = tf.Variable(
tf.constant(0, dtype=tf.int64),
trainable=False,
synchronization=tf.VariableSynchronization.ON_READ,
aggregation=tf.VariableAggregation.ONLY_FIRST_REPLICA,
)
return self._accum_steps.value()
@property
def gradients(self):
"""The accumulated gradients on the current replica."""
if not self._gradients:
raise ValueError(
"The accumulator should be called first to initialize the gradients"
)
return list(
gradient.value() if gradient is not None else gradient
for gradient in self._gradients
)
def __call__(self, gradients):
"""Accumulates :obj:`gradients` on the current replica."""
if not self._gradients:
_ = self.step # Create the step variable.
self._gradients.extend(
[
tf.Variable(
tf.zeros_like(gradient),
trainable=False,
synchronization=tf.VariableSynchronization.ON_READ,
aggregation=tf.VariableAggregation.ONLY_FIRST_REPLICA,
)
if gradient is not None
else gradient
for gradient in gradients
]
)
if len(gradients) != len(self._gradients):
raise ValueError(
"Expected %s gradients, but got %d"
% (len(self._gradients), len(gradients))
)
for accum_gradient, gradient in zip(self._gradients, gradients):
if accum_gradient is not None and gradient is not None:
accum_gradient.assign_add(gradient)
self._accum_steps.assign_add(1)
def reset(self):
"""Resets the accumulated gradients on the current replica."""
if not self._gradients:
return
self._accum_steps.assign(0)
for gradient in self._gradients:
if gradient is not None:
gradient.assign(tf.zeros_like(gradient))</code></pre>
</details>
<h3>Instance variables</h3>
<dl>
<dt id="ktrain.lroptimize.optimization.GradientAccumulator.gradients"><code class="name">var <span class="ident">gradients</span></code></dt>
<dd>
<div class="desc"><p>The accumulated gradients on the current replica.</p></div>
<details class="source">
<summary>
<span>Expand source code</span>
</summary>
<pre><code class="python">@property
def gradients(self):
"""The accumulated gradients on the current replica."""
if not self._gradients:
raise ValueError(
"The accumulator should be called first to initialize the gradients"
)
return list(
gradient.value() if gradient is not None else gradient
for gradient in self._gradients
)</code></pre>
</details>
</dd>
<dt id="ktrain.lroptimize.optimization.GradientAccumulator.step"><code class="name">var <span class="ident">step</span></code></dt>
<dd>
<div class="desc"><p>Number of accumulated steps.</p></div>
<details class="source">
<summary>
<span>Expand source code</span>
</summary>
<pre><code class="python">@property
def step(self):
"""Number of accumulated steps."""
if self._accum_steps is None:
self._accum_steps = tf.Variable(
tf.constant(0, dtype=tf.int64),
trainable=False,
synchronization=tf.VariableSynchronization.ON_READ,
aggregation=tf.VariableAggregation.ONLY_FIRST_REPLICA,
)
return self._accum_steps.value()</code></pre>
</details>
</dd>
</dl>
<h3>Methods</h3>
<dl>
<dt id="ktrain.lroptimize.optimization.GradientAccumulator.reset"><code class="name flex">
<span>def <span class="ident">reset</span></span>(<span>self)</span>
</code></dt>
<dd>
<div class="desc"><p>Resets the accumulated gradients on the current replica.</p></div>
<details class="source">
<summary>
<span>Expand source code</span>
</summary>
<pre><code class="python">def reset(self):
"""Resets the accumulated gradients on the current replica."""
if not self._gradients:
return
self._accum_steps.assign(0)
for gradient in self._gradients:
if gradient is not None:
gradient.assign(tf.zeros_like(gradient))</code></pre>
</details>
</dd>
</dl>
</dd>
<dt id="ktrain.lroptimize.optimization.WarmUp"><code class="flex name class">
<span>class <span class="ident">WarmUp</span></span>
<span>(</span><span>initial_learning_rate: float, decay_schedule_fn: Callable, warmup_steps: int, power: float = 1.0, name: str = None)</span>
</code></dt>
<dd>
<div class="desc"><pre><code>Applies a warmup schedule on a given learning rate decay schedule.
Args:
initial_learning_rate (:obj:`float`):
The initial learning rate for the schedule after the warmup (so this will be the learning rate at the end
of the warmup).
decay_schedule_fn (:obj:`Callable`):
The schedule function to apply after the warmup for the rest of training.
warmup_steps (:obj:`int`):
The number of steps for the warmup part of training.
power (:obj:`float`, `optional`, defaults to 1):
The power to use for the polynomial warmup (defaults is a linear warmup).
name (:obj:`str`, `optional`):
Optional name prefix for the returned tensors during the schedule.
</code></pre></div>
<details class="source">
<summary>
<span>Expand source code</span>
</summary>