-
Notifications
You must be signed in to change notification settings - Fork 1.4k
/
adam.py
581 lines (493 loc) · 22.3 KB
/
adam.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
from __future__ import division
import math
import warnings
import numpy
import chainer
from chainer.backends import cuda
from chainer.backends import intel64
from chainer import optimizer
from chainer import types
if types.TYPE_CHECKING:
import typing_extensions as tpe
class AdamHyperparameter(tpe.Protocol):
"""Protocol class for hyperparameter of Adam.
This is only for PEP 544 compliant static type checkers.
"""
alpha = None # type: float
beta1 = None # type: float
beta2 = None # type: float
eps = None # type: float
eta = None # type: float
weight_decay_rate = None # type: float
amsgrad = None # type: bool
adabound = None # type: bool
final_lr = None # type: float
gamma = None # type: float
_default_hyperparam = optimizer.Hyperparameter() # type: AdamHyperparameter # NOQA
_default_hyperparam.alpha = 0.001
_default_hyperparam.beta1 = 0.9
_default_hyperparam.beta2 = 0.999
_default_hyperparam.eps = 1e-8
_default_hyperparam.eta = 1.0
_default_hyperparam.weight_decay_rate = 0
_default_hyperparam.amsgrad = False
_default_hyperparam.adabound = False
_default_hyperparam.final_lr = 0.1
_default_hyperparam.gamma = 1e-3
def _learning_rate(hp, t):
if t == 0:
raise RuntimeError(
'Can\'t determine the learning rate of Adam optimizer '
'because the update steps have not been started.')
fix1 = 1. - math.pow(hp.beta1, t)
fix2 = 1. - math.pow(hp.beta2, t)
return hp.alpha * math.sqrt(fix2) / fix1
def _get_intermediate_dtype(dtype):
# Returns the dtype for intermediate calculation.
# For float16 input, float32 is used.
# Otherwise the same dtype as the parameter is used.
if dtype == numpy.float16:
return numpy.float32
return dtype
def _inplace_axpby(x, a, b, y):
# in-place axpby: x = a * x + b * y
if isinstance(x, intel64.mdarray):
x.inplace_axpby(a, b, y)
else:
if a == 1:
x += b * y
else:
x[...] = a * x + b * y
class AdamRule(optimizer.UpdateRule):
"""Update rule of Adam optimization algorithm.
See: `Adam: A Method for Stochastic Optimization
<https://arxiv.org/abs/1412.6980v8>`_
Modified for proper weight decay.
See: `Fixing Weight Decay Regularization in Adam
<https://openreview.net/forum?id=rk6qdGgCZ>`_
With option to use AMSGrad variant of Adam.
See: `On the Convergence of Adam and Beyond
<https://openreview.net/forum?id=ryQu7f-RZ>`_
With option to use AdaBound variant of Adam.
See: `Adaptive Gradient Methods with Dynamic Bound of Learning Rate
<https://openreview.net/forum?id=Bkg3g2R9FX>`
See :class:`~chainer.optimizers.Adam` for the default values
of the hyperparameters.
Args:
parent_hyperparam (~chainer.optimizer.Hyperparameter): Hyperparameter
that provides the default values.
alpha (float): Coefficient of learning rate.
beta1 (float): Exponential decay rate of the first order moment.
beta2 (float): Exponential decay rate of the second order moment.
eps (float): Small value for the numerical stability.
eta (float): Schedule multiplier, can be used for warm restarts.
weight_decay_rate (float): Weight decay rate.
amsgrad (bool): Whether to use the AMSGrad variant of Adam.
adabound (bool): Whether to use the AdaBound variant of Adam.
final_lr (float): Final (SGD) learning rate in AdaBound.
gamma (float): Convergence speed of the bound functions in AdaBound.
"""
is_elementwise = True
_kernel = None
_amsgrad_kernel = None
_adabound_kernel = None
_amsbound_kernel = None
# Only used in `update_core_gpu`.
# A dummy ndarray to help ElementwiseKernel deduce generic type T as
# `dtype`.
# It cannot be deduced only by scalar arguments.
_dummy = None
def __init__(self, parent_hyperparam=None,
alpha=None, beta1=None, beta2=None, eps=None,
eta=None, weight_decay_rate=None, amsgrad=None,
adabound=None, final_lr=None, gamma=None):
super(AdamRule, self).__init__(
parent_hyperparam or _default_hyperparam)
if alpha is not None:
self.hyperparam.alpha = alpha
if beta1 is not None:
self.hyperparam.beta1 = beta1
if beta2 is not None:
self.hyperparam.beta2 = beta2
if eps is not None:
self.hyperparam.eps = eps
if eta is not None:
self.hyperparam.eta = eta
if weight_decay_rate is not None:
self.hyperparam.weight_decay_rate = weight_decay_rate
if amsgrad is not None:
self.hyperparam.amsgrad = amsgrad
if adabound is not None:
self.hyperparam.adabound = adabound
if final_lr is not None:
self.hyperparam.final_lr = final_lr
if gamma is not None:
self.hyperparam.gamma = gamma
if self.hyperparam.adabound:
self.initial_alpha = self.hyperparam.alpha
def init_state(self, param):
with chainer.using_device(param.device):
xp = param.device.xp
self.state['m'] = xp.zeros_like(param.data)
self.state['v'] = xp.zeros_like(param.data)
if self.hyperparam.amsgrad:
self.state['vhat'] = xp.zeros_like(param.data)
# For iDeep
if isinstance(param.data, intel64.mdarray):
self.state['m'] = intel64.ideep.array(
self.state['m'], itype=intel64.ideep.wgt_array)
self.state['v'] = intel64.ideep.array(
self.state['v'], itype=intel64.ideep.wgt_array)
if self.hyperparam.amsgrad:
self.state['vhat'] = intel64.ideep.array(
self.state['vhat'], itype=intel64.ideep.wgt_array)
def _check_eps(self, interm_dtype):
# Checks that the eps does not underflow.
hp = self.hyperparam
eps = interm_dtype(hp.eps)
if hp.eps != 0 and eps == 0:
raise ValueError(
'eps of Adam optimizer is too small for {} ({})'.format(
interm_dtype.name, hp.eps))
# Note that the converted `eps` (numpy scalar) is discarded here and
# the original `hp.eps` is used in calculation, because Python
# scalars are faster in cupy elementwise kernels.
def update_core_cpu(self, param):
grad = param.grad
if grad is None:
return
hp = self.hyperparam
dtype = _get_intermediate_dtype(param.dtype.type)
self._check_eps(dtype)
grad = grad.astype(dtype, copy=False)
m, v = self.state['m'], self.state['v']
# m += (1 - beta1) * (grad - m)
_inplace_axpby(m, 1.0, 1.0 - hp.beta1, grad - m)
# v += (1 - beta2) * (grad * grad - v)
_inplace_axpby(v, 1.0, 1.0 - hp.beta2, grad*grad - v)
if hp.amsgrad:
vhat = self.state['vhat']
# For iDeep
if isinstance(vhat, intel64.mdarray):
vhat[...] = numpy.maximum(vhat, v)
else:
numpy.maximum(vhat, v, out=vhat)
else:
vhat = v
vhat = vhat.astype(dtype, copy=False)
step = self.alpha_t / (numpy.sqrt(vhat) + hp.eps)
if hp.adabound:
lower, upper = self.bounds
step = numpy.clip(step, lower, upper)
# param -=
# eta * (step * m - weight_decay_rate * param)
_inplace_axpby(
param.data, 1.0 - hp.eta * hp.weight_decay_rate, -hp.eta, step * m)
def update_core_gpu(self, param):
grad = param.grad
if grad is None:
return
hp = self.hyperparam
dtype = _get_intermediate_dtype(param.dtype.type)
self._check_eps(dtype)
if self._dummy is None:
self._dummy = cuda.cupy.empty((0,), dtype=dtype)
if hp.adabound:
lower, upper = self.bounds
if hp.amsgrad and hp.adabound:
if AdamRule._amsbound_kernel is None:
AdamRule._amsbound_kernel = cuda.elementwise(
'P grad, T alpha_t, T one_minus_beta1, T one_minus_beta2, '
'T lower, T upper, '
'T eps, T eta, T weight_decay_rate, raw T dummy',
'P param, P m, P v, P vhat',
'''T grad_ = static_cast<T>(grad);
T m_ = static_cast<T>(m);
T v_ = static_cast<T>(v);
T vhat_ = static_cast<T>(vhat);
m_ += one_minus_beta1 * (grad_ - m_);
v_ += one_minus_beta2 * (grad_ * grad_ - v_);
vhat_ = max(vhat_, v_);
vhat = static_cast<T>(vhat_);
m = static_cast<P>(m_);
v = static_cast<P>(v_);
param -= eta *
(max(min(alpha_t / (sqrt(vhat_) + eps), upper),
lower) * m_ + weight_decay_rate * param);''',
'amsbound')
AdamRule._amsbound_kernel(
grad, self.alpha_t, 1 - hp.beta1,
1 - hp.beta2, lower, upper, hp.eps,
hp.eta, hp.weight_decay_rate, self._dummy,
param.data, self.state['m'], self.state['v'],
self.state['vhat'])
elif hp.adabound:
if AdamRule._adabound_kernel is None:
AdamRule._adabound_kernel = cuda.elementwise(
'P grad, T alpha_t, T one_minus_beta1, T one_minus_beta2, '
'T lower, T upper, '
'T eps, T eta, T weight_decay_rate, raw T dummy',
'P param, P m, P v',
'''T grad_ = static_cast<T>(grad);
T m_ = static_cast<T>(m);
T v_ = static_cast<T>(v);
m_ += one_minus_beta1 * (grad_ - m_);
v_ += one_minus_beta2 * (grad_ * grad_ - v_);
m = static_cast<P>(m_);
v = static_cast<P>(v_);
param -= eta *
(max(min(alpha_t / (sqrt(v_) + eps), upper),
lower) * m_ + weight_decay_rate * param);''',
'adabound')
AdamRule._adabound_kernel(
grad, self.alpha_t, 1 - hp.beta1,
1 - hp.beta2, lower, upper, hp.eps,
hp.eta, hp.weight_decay_rate, self._dummy,
param.data, self.state['m'], self.state['v'])
elif hp.amsgrad:
if AdamRule._amsgrad_kernel is None:
AdamRule._amsgrad_kernel = cuda.elementwise(
'P grad, T alpha_t, T one_minus_beta1, T one_minus_beta2, '
'T eps, T eta, T weight_decay_rate, raw T dummy',
'P param, P m, P v, P vhat',
'''T grad_ = static_cast<T>(grad);
T m_ = static_cast<T>(m);
T v_ = static_cast<T>(v);
T vhat_ = static_cast<T>(vhat);
m_ += one_minus_beta1 * (grad_ - m_);
v_ += one_minus_beta2 * (grad_ * grad_ - v_);
vhat_ = max(vhat_, v_);
vhat = static_cast<T>(vhat_);
m = static_cast<P>(m_);
v = static_cast<P>(v_);
param -= eta * (alpha_t * m_ / (sqrt(vhat_) + eps) +
weight_decay_rate * param);''',
'adam')
AdamRule._amsgrad_kernel(
grad, self.alpha_t, 1 - hp.beta1,
1 - hp.beta2, hp.eps,
hp.eta, hp.weight_decay_rate, self._dummy,
param.data, self.state['m'], self.state['v'],
self.state['vhat'])
else:
if AdamRule._kernel is None:
AdamRule._kernel = cuda.elementwise(
'P grad, T alpha_t, T one_minus_beta1, T one_minus_beta2, '
'T eps, T eta, T weight_decay_rate, raw T dummy',
'P param, P m, P v',
'''T grad_ = static_cast<T>(grad);
T m_ = static_cast<T>(m);
T v_ = static_cast<T>(v);
m_ += one_minus_beta1 * (grad_ - m_);
v_ += one_minus_beta2 * (grad_ * grad_ - v_);
m = static_cast<P>(m_);
v = static_cast<P>(v_);
param -= eta * (alpha_t * m_ / (sqrt(v_) + eps) +
weight_decay_rate * param);''',
'adam')
AdamRule._kernel(
grad, self.alpha_t, 1 - hp.beta1,
1 - hp.beta2, hp.eps,
hp.eta, hp.weight_decay_rate, self._dummy,
param.data, self.state['m'], self.state['v'])
@property
def alpha_t(self):
return _learning_rate(self.hyperparam, self.t)
@property
def lr(self):
warnings.warn(
'AdamRule.lr has been renamed to AdamRule.alpha_t. '
'Use of AdamRule.lr is deprecated in Chainer v6.',
DeprecationWarning)
return self.alpha_t
@property
def bounds(self):
if self.t == 0:
raise RuntimeError(
'Can\'t determine the bounds of AdaBound optimizer '
'because the update steps have not been started.')
hp = self.hyperparam
# Workaround to reflect changing `alpha` in `final_lr`.
# (by some of `chainer.training.extensions`)
final_lr = hp.final_lr * hp.alpha / self.initial_alpha
lower = final_lr * (1.0 - 1.0 / (hp.gamma * self.t + 1))
upper = final_lr * (1.0 + 1.0 / (hp.gamma * self.t))
return lower, upper
class Adam(optimizer.GradientMethod):
"""Adam optimizer.
See: `Adam: A Method for Stochastic Optimization
<https://arxiv.org/abs/1412.6980v8>`_
Modified for proper weight decay (also called
:class:`~chainer.optimizers.AdamW`).
AdamW introduces the additional parameters ``eta``
and ``weight_decay_rate``, which can be used to properly scale the
learning rate, and decouple the weight decay rate from ``alpha``,
as shown in the below paper.
Note that with the default values ``eta = 1`` and
``weight_decay_rate = 0``, this implementation is identical to
the standard Adam method.
See: `Fixing Weight Decay Regularization in Adam
<https://openreview.net/forum?id=rk6qdGgCZ>`_
A flag ``amsgrad`` to use the :class:`~chainer.optimizers.AMSGrad`
variant of Adam from the paper:
`On the Convergence of Adam and Beyond
<https://openreview.net/forum?id=ryQu7f-RZ>`_
A flag ``adabound`` to use the :class:`~chainer.optimizers.AdaBound`
variant of Adam from the paper:
`Adaptive Gradient Methods with Dynamic Bound of Learning Rate
<https://openreview.net/forum?id=Bkg3g2R9FX>`_
If both ``amsgrad`` and ``adabound`` are ``True``, the optimizer is
equivalent to :class:`~chainer.optimizers.AMSBound` proposed in the
AdaBound paper.
Args:
alpha (float): Coefficient of learning rate.
beta1 (float): Exponential decay rate of the first order moment.
beta2 (float): Exponential decay rate of the second order moment.
eps (float): Small value for the numerical stability.
eta (float): Schedule multiplier, can be used for warm restarts.
weight_decay_rate (float): Weight decay rate.
amsgrad (bool): Whether to use AMSGrad variant of Adam.
adabound (bool): Whether to use the AdaBound variant of Adam.
final_lr (float): Final (SGD) learning rate in AdaBound.
gamma (float): Convergence speed of the bound functions in AdaBound.
"""
def __init__(self,
alpha=_default_hyperparam.alpha,
beta1=_default_hyperparam.beta1,
beta2=_default_hyperparam.beta2,
eps=_default_hyperparam.eps,
eta=_default_hyperparam.eta,
weight_decay_rate=_default_hyperparam.weight_decay_rate,
amsgrad=_default_hyperparam.amsgrad,
adabound=_default_hyperparam.adabound,
final_lr=_default_hyperparam.final_lr,
gamma=_default_hyperparam.gamma):
super(Adam, self).__init__()
self.hyperparam.alpha = alpha
self.hyperparam.beta1 = beta1
self.hyperparam.beta2 = beta2
self.hyperparam.eps = eps
self.hyperparam.eta = eta
self.hyperparam.weight_decay_rate = weight_decay_rate
self.hyperparam.amsgrad = amsgrad
self.hyperparam.adabound = adabound
self.hyperparam.final_lr = final_lr
self.hyperparam.gamma = gamma
alpha = optimizer.HyperparameterProxy('alpha')
beta1 = optimizer.HyperparameterProxy('beta1')
beta2 = optimizer.HyperparameterProxy('beta2')
eps = optimizer.HyperparameterProxy('eps')
eta = optimizer.HyperparameterProxy('eta')
weight_decay_rate = optimizer.HyperparameterProxy('weight_decay_rate')
amsgrad = optimizer.HyperparameterProxy('amsgrad')
adabound = optimizer.HyperparameterProxy('adabound')
final_lr = optimizer.HyperparameterProxy('final_lr')
gamma = optimizer.HyperparameterProxy('gamma')
def create_update_rule(self):
return AdamRule(self.hyperparam)
@property
def alpha_t(self):
return _learning_rate(self.hyperparam, self.t)
@property
def lr(self):
warnings.warn(
'Adam.lr has been renamed to AdamRule.alpha_t. '
'Use of Adam.lr is deprecated in Chainer v6.',
DeprecationWarning)
return self.alpha_t
class AdamW(Adam):
"""AdamW optimizer.
This class is a special case of :class:`~chainer.optimizers.Adam`.
See: `Fixing Weight Decay Regularization in Adam
<https://openreview.net/forum?id=rk6qdGgCZ>`_
Args:
alpha (float): Coefficient of learning rate.
beta1 (float): Exponential decay rate of the first order moment.
beta2 (float): Exponential decay rate of the second order moment.
eps (float): Small value for the numerical stability.
eta (float): Schedule multiplier, can be used for warm restarts.
The default value is 1.0.
weight_decay_rate (float): Weight decay rate.
The default value is 0.
"""
def __init__(self,
alpha=_default_hyperparam.alpha,
beta1=_default_hyperparam.beta1,
beta2=_default_hyperparam.beta2,
eps=_default_hyperparam.eps,
eta=_default_hyperparam.eta,
weight_decay_rate=_default_hyperparam.weight_decay_rate):
super(AdamW, self).__init__(
alpha=alpha, beta1=beta1, beta2=beta2, eps=eps, eta=eta,
weight_decay_rate=weight_decay_rate)
class AMSGrad(Adam):
"""AMSGrad optimizer.
This class is a special case of :class:`~chainer.optimizers.Adam`.
See: `On the Convergence of Adam and Beyond
<https://openreview.net/forum?id=ryQu7f-RZ>`_
Args:
alpha (float): Coefficient of learning rate.
beta1 (float): Exponential decay rate of the first order moment.
beta2 (float): Exponential decay rate of the second order moment.
eps (float): Small value for the numerical stability.
eta (float): Schedule multiplier, can be used for warm restarts.
"""
def __init__(self,
alpha=_default_hyperparam.alpha,
beta1=_default_hyperparam.beta1,
beta2=_default_hyperparam.beta2,
eps=_default_hyperparam.eps,
eta=_default_hyperparam.eta):
super(AMSGrad, self).__init__(
alpha=alpha, beta1=beta1, beta2=beta2, eps=eps, eta=eta,
amsgrad=True)
class AdaBound(Adam):
"""AdaBound optimizer.
This class is a special case of :class:`~chainer.optimizers.Adam`.
See: `Adaptive Gradient Methods with Dynamic Bound of Learning Rate
<https://openreview.net/forum?id=Bkg3g2R9FX>`_
Args:
alpha (float): Coefficient of learning rate.
beta1 (float): Exponential decay rate of the first order moment.
beta2 (float): Exponential decay rate of the second order moment.
final_lr (float): Final (SGD) learning rate in AdaBound.
gamma (float): Convergence speed of the bound functions in AdaBound.
eps (float): Small value for the numerical stability.
eta (float): Schedule multiplier, can be used for warm restarts.
"""
def __init__(self,
alpha=_default_hyperparam.alpha,
beta1=_default_hyperparam.beta1,
beta2=_default_hyperparam.beta2,
final_lr=_default_hyperparam.final_lr,
gamma=_default_hyperparam.gamma,
eps=_default_hyperparam.eps,
eta=_default_hyperparam.eta):
super(AdaBound, self).__init__(
alpha=alpha, beta1=beta1, beta2=beta2, eps=eps, eta=eta,
amsgrad=False, adabound=True, final_lr=final_lr, gamma=gamma)
class AMSBound(Adam):
"""AMSBound optimizer.
This class is a special case of :class:`~chainer.optimizers.Adam`.
See: `Adaptive Gradient Methods with Dynamic Bound of Learning Rate
<https://openreview.net/forum?id=Bkg3g2R9FX>`_
Args:
alpha (float): Coefficient of learning rate.
beta1 (float): Exponential decay rate of the first order moment.
beta2 (float): Exponential decay rate of the second order moment.
final_lr (float): Final (SGD) learning rate in AdaBound.
gamma (float): Convergence speed of the bound functions in AdaBound.
eps (float): Small value for the numerical stability.
eta (float): Schedule multiplier, can be used for warm restarts.
"""
def __init__(self,
alpha=_default_hyperparam.alpha,
beta1=_default_hyperparam.beta1,
beta2=_default_hyperparam.beta2,
final_lr=_default_hyperparam.final_lr,
gamma=_default_hyperparam.gamma,
eps=_default_hyperparam.eps,
eta=_default_hyperparam.eta):
super(AMSBound, self).__init__(
alpha=alpha, beta1=beta1, beta2=beta2, eps=eps, eta=eta,
amsgrad=True, adabound=True, final_lr=final_lr, gamma=gamma)