-
Notifications
You must be signed in to change notification settings - Fork 17
/
repoptimizer_adamw.py
455 lines (378 loc) · 20.1 KB
/
repoptimizer_adamw.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
# --------------------------------------------------------
# Re-parameterizing Your Optimizers rather than Architectures (https://arxiv.org/abs/2205.15242)
# Github source: https://github.com/DingXiaoH/RepOptimizers
# Licensed under The MIT License [see LICENSE for details]
# --------------------------------------------------------
import math
import torch
from torch import Tensor
from torch.optim.optimizer import Optimizer
from typing import List, Optional
class RepOptimizerAdamW(Optimizer):
r"""Implements AdamW algorithm.
.. math::
\begin{aligned}
&\rule{110mm}{0.4pt} \\
&\textbf{input} : \gamma \text{(lr)}, \: \beta_1, \beta_2
\text{(betas)}, \: \theta_0 \text{(params)}, \: f(\theta) \text{(objective)},
\: \epsilon \text{ (epsilon)} \\
&\hspace{13mm} \lambda \text{(weight decay)}, \: \textit{amsgrad},
\: \textit{maximize} \\
&\textbf{initialize} : m_0 \leftarrow 0 \text{ (first moment)}, v_0 \leftarrow 0
\text{ ( second moment)}, \: \widehat{v_0}^{max}\leftarrow 0 \\[-1.ex]
&\rule{110mm}{0.4pt} \\
&\textbf{for} \: t=1 \: \textbf{to} \: \ldots \: \textbf{do} \\
&\hspace{5mm}\textbf{if} \: \textit{maximize}: \\
&\hspace{10mm}g_t \leftarrow -\nabla_{\theta} f_t (\theta_{t-1}) \\
&\hspace{5mm}\textbf{else} \\
&\hspace{10mm}g_t \leftarrow \nabla_{\theta} f_t (\theta_{t-1}) \\
&\hspace{5mm} \theta_t \leftarrow \theta_{t-1} - \gamma \lambda \theta_{t-1} \\
&\hspace{5mm}m_t \leftarrow \beta_1 m_{t-1} + (1 - \beta_1) g_t \\
&\hspace{5mm}v_t \leftarrow \beta_2 v_{t-1} + (1-\beta_2) g^2_t \\
&\hspace{5mm}\widehat{m_t} \leftarrow m_t/\big(1-\beta_1^t \big) \\
&\hspace{5mm}\widehat{v_t} \leftarrow v_t/\big(1-\beta_2^t \big) \\
&\hspace{5mm}\textbf{if} \: amsgrad \\
&\hspace{10mm}\widehat{v_t}^{max} \leftarrow \mathrm{max}(\widehat{v_t}^{max},
\widehat{v_t}) \\
&\hspace{10mm}\theta_t \leftarrow \theta_t - \gamma \widehat{m_t}/
\big(\sqrt{\widehat{v_t}^{max}} + \epsilon \big) \\
&\hspace{5mm}\textbf{else} \\
&\hspace{10mm}\theta_t \leftarrow \theta_t - \gamma \widehat{m_t}/
\big(\sqrt{\widehat{v_t}} + \epsilon \big) \\
&\rule{110mm}{0.4pt} \\[-1.ex]
&\bf{return} \: \theta_t \\[-1.ex]
&\rule{110mm}{0.4pt} \\[-1.ex]
\end{aligned}
For further details regarding the algorithm we refer to `Decoupled Weight Decay Regularization`_.
Args:
grad_mult_map (dict): output from a RepOptimizerHandler in the form of dict{param : mult}
params (iterable): iterable of parameters to optimize or dicts defining
parameter groups
lr (float, optional): learning rate (default: 1e-3)
betas (Tuple[float, float], optional): coefficients used for computing
running averages of gradient and its square (default: (0.9, 0.999))
eps (float, optional): term added to the denominator to improve
numerical stability (default: 1e-8)
weight_decay (float, optional): weight decay coefficient (default: 1e-2)
amsgrad (bool, optional): whether to use the AMSGrad variant of this
algorithm from the paper `On the Convergence of Adam and Beyond`_
(default: False)
maximize (bool, optional): maximize the params based on the objective, instead of
minimizing (default: False)
foreach (bool, optional): whether foreach implementation of optimizer
is used (default: None)
capturable (bool, optional): whether this instance is safe to capture in a CUDA graph.
Passing True can impair ungraphed performance, so if you don't intend to
graph capture this instance, leave it False (default: False)
.. _Decoupled Weight Decay Regularization:
https://arxiv.org/abs/1711.05101
.. _On the Convergence of Adam and Beyond:
https://openreview.net/forum?id=ryQu7f-RZ
"""
def __init__(self,
grad_mult_map,
params, lr=1e-3, betas=(0.9, 0.999), eps=1e-8,
weight_decay=1e-2, amsgrad=False, *, maximize: bool = False,
foreach: Optional[bool] = None,
capturable: bool = False):
if not 0.0 <= lr:
raise ValueError("Invalid learning rate: {}".format(lr))
if not 0.0 <= eps:
raise ValueError("Invalid epsilon value: {}".format(eps))
if not 0.0 <= betas[0] < 1.0:
raise ValueError("Invalid beta parameter at index 0: {}".format(betas[0]))
if not 0.0 <= betas[1] < 1.0:
raise ValueError("Invalid beta parameter at index 1: {}".format(betas[1]))
if not 0.0 <= weight_decay:
raise ValueError("Invalid weight_decay value: {}".format(weight_decay))
defaults = dict(lr=lr, betas=betas, eps=eps,
weight_decay=weight_decay, amsgrad=amsgrad,
foreach=foreach, maximize=maximize, capturable=capturable)
super(RepOptimizerAdamW, self).__init__(params, defaults)
self.grad_mult_map = grad_mult_map
def __setstate__(self, state):
super().__setstate__(state)
for group in self.param_groups:
group.setdefault('amsgrad', False)
group.setdefault('maximize', False)
group.setdefault('foreach', None)
group.setdefault('capturable', False)
state_values = list(self.state.values())
step_is_tensor = (len(state_values) != 0) and torch.is_tensor(state_values[0]['step'])
if not step_is_tensor:
for s in state_values:
s['step'] = torch.tensor(float(s['step']))
@torch.no_grad()
def step(self, closure=None):
"""Performs a single optimization step.
Args:
closure (Callable, optional): A closure that reevaluates the model
and returns the loss.
"""
if hasattr(self, '_cuda_graph_capture_health_check'):
self._cuda_graph_capture_health_check()
loss = None
if closure is not None:
with torch.enable_grad():
loss = closure()
for group in self.param_groups:
params_with_grad = []
grads = []
exp_avgs = []
exp_avg_sqs = []
max_exp_avg_sqs = []
state_steps = []
amsgrad = group['amsgrad']
beta1, beta2 = group['betas']
for p in group['params']:
if p.grad is None:
continue
if p in self.grad_mult_map:
p.grad *= self.grad_mult_map[p]
params_with_grad.append(p)
if p.grad.is_sparse:
raise RuntimeError('AdamW does not support sparse gradients')
grads.append(p.grad)
state = self.state[p]
# State initialization
if len(state) == 0:
state['step'] = torch.zeros((1,), dtype=torch.float, device=p.device) \
if self.defaults['capturable'] else torch.tensor(0.)
# Exponential moving average of gradient values
state['exp_avg'] = torch.zeros_like(p, memory_format=torch.preserve_format)
# Exponential moving average of squared gradient values
state['exp_avg_sq'] = torch.zeros_like(p, memory_format=torch.preserve_format)
if amsgrad:
# Maintains max of all exp. moving avg. of sq. grad. values
state['max_exp_avg_sq'] = torch.zeros_like(p, memory_format=torch.preserve_format)
exp_avgs.append(state['exp_avg'])
exp_avg_sqs.append(state['exp_avg_sq'])
if amsgrad:
max_exp_avg_sqs.append(state['max_exp_avg_sq'])
state_steps.append(state['step'])
adamw(params_with_grad,
grads,
exp_avgs,
exp_avg_sqs,
max_exp_avg_sqs,
state_steps,
amsgrad=amsgrad,
beta1=beta1,
beta2=beta2,
lr=group['lr'],
weight_decay=group['weight_decay'],
eps=group['eps'],
maximize=group['maximize'],
foreach=group['foreach'],
capturable=group['capturable'],
grad_mult_map=self.grad_mult_map)
return loss
def adamw(params: List[Tensor],
grads: List[Tensor],
exp_avgs: List[Tensor],
exp_avg_sqs: List[Tensor],
max_exp_avg_sqs: List[Tensor],
state_steps: List[Tensor],
# kwonly args with defaults are not supported by functions compiled with torchscript issue #70627
# setting this as kwarg for now as functional API is compiled by torch/distributed/optim
foreach: bool = None,
capturable: bool = False,
*,
amsgrad: bool,
beta1: float,
beta2: float,
lr: float,
weight_decay: float,
eps: float,
maximize: bool,
grad_mult_map: dict):
r"""Functional API that performs AdamW algorithm computation.
See :class:`~torch.optim.AdamW` for details.
"""
if not all(isinstance(t, torch.Tensor) for t in state_steps):
raise RuntimeError("API has changed, `state_steps` argument must contain a list of singleton tensors")
if foreach is None:
# Placeholder for more complex foreach logic to be added when value is not set
foreach = False
if foreach and torch.jit.is_scripting():
raise RuntimeError('torch.jit.script not supported with foreach optimizers')
if foreach and not torch.jit.is_scripting():
func = _multi_tensor_adamw
else:
func = _single_tensor_adamw
func(params,
grads,
exp_avgs,
exp_avg_sqs,
max_exp_avg_sqs,
state_steps,
amsgrad=amsgrad,
beta1=beta1,
beta2=beta2,
lr=lr,
weight_decay=weight_decay,
eps=eps,
maximize=maximize,
capturable=capturable,
grad_mult_map=grad_mult_map)
def _single_tensor_adamw(params: List[Tensor],
grads: List[Tensor],
exp_avgs: List[Tensor],
exp_avg_sqs: List[Tensor],
max_exp_avg_sqs: List[Tensor],
state_steps: List[Tensor],
*,
amsgrad: bool,
beta1: float,
beta2: float,
lr: float,
weight_decay: float,
eps: float,
maximize: bool,
capturable: bool,
grad_mult_map: dict):
for i, param in enumerate(params):
grad = grads[i] if not maximize else -grads[i]
exp_avg = exp_avgs[i]
exp_avg_sq = exp_avg_sqs[i]
step_t = state_steps[i]
if param in grad_mult_map:
update_mult = grad_mult_map[param]
else:
update_mult = 1
if capturable:
assert param.is_cuda and step_t.is_cuda, "If capturable=True, params and state_steps must be CUDA tensors."
if torch.is_complex(param):
grad = torch.view_as_real(grad)
exp_avg = torch.view_as_real(exp_avg)
exp_avg_sq = torch.view_as_real(exp_avg_sq)
param = torch.view_as_real(param)
# update step
step_t += 1
# Perform stepweight decay
param.mul_(1 - lr * weight_decay)
# Decay the first and second moment running average coefficient
exp_avg.mul_(beta1).add_(grad, alpha=1 - beta1)
exp_avg_sq.mul_(beta2).addcmul_(grad, grad, value=1 - beta2)
if capturable:
step = step_t
# 1 - beta1 ** step can't be captured in a CUDA graph, even if step is a CUDA tensor
# (incurs "RuntimeError: CUDA error: operation not permitted when stream is capturing")
bias_correction1 = 1 - torch.pow(beta1, step)
bias_correction2 = 1 - torch.pow(beta2, step)
step_size = lr / bias_correction1
step_size_neg = step_size.neg()
bias_correction2_sqrt = bias_correction2.sqrt()
if amsgrad:
# Maintains the maximum of all 2nd moment running avg. till now
torch.maximum(max_exp_avg_sqs[i], exp_avg_sq, out=max_exp_avg_sqs[i])
# Uses the max. for normalizing running avg. of gradient
# Folds in (admittedly ugly) 1-elem step_size math here to avoid extra param-set-sized read+write
# (can't fold it into addcdiv_ below because addcdiv_ requires value is a Number, not a Tensor)
denom = (max_exp_avg_sqs[i].sqrt() / (bias_correction2_sqrt * step_size_neg)).add_(eps / step_size_neg)
else:
denom = (exp_avg_sq.sqrt() / (bias_correction2_sqrt * step_size_neg)).add_(eps / step_size_neg)
param.addcdiv_(exp_avg * update_mult, denom)
else:
step = step_t.item()
bias_correction1 = 1 - beta1 ** step
bias_correction2 = 1 - beta2 ** step
step_size = lr / bias_correction1
bias_correction2_sqrt = math.sqrt(bias_correction2)
if amsgrad:
# Maintains the maximum of all 2nd moment running avg. till now
torch.maximum(max_exp_avg_sqs[i], exp_avg_sq, out=max_exp_avg_sqs[i])
# Use the max. for normalizing running avg. of gradient
denom = (max_exp_avg_sqs[i].sqrt() / bias_correction2_sqrt).add_(eps)
else:
denom = (exp_avg_sq.sqrt() / bias_correction2_sqrt).add_(eps)
param.addcdiv_(exp_avg * update_mult, denom, value=-step_size)
# Note: I did not test this part since I never used "foreach". Only added three lines of code related to "mults"
def _multi_tensor_adamw(params: List[Tensor],
grads: List[Tensor],
exp_avgs: List[Tensor],
exp_avg_sqs: List[Tensor],
max_exp_avg_sqs: List[Tensor],
state_steps: List[Tensor],
*,
amsgrad: bool,
beta1: float,
beta2: float,
lr: float,
weight_decay: float,
eps: float,
maximize: bool,
capturable: bool,
grad_mult_map: dict):
if len(params) == 0:
return
if capturable:
assert all(p.is_cuda and step.is_cuda for p, step in zip(params, state_steps)), \
"If capturable=True, params and state_steps must be CUDA tensors."
if maximize:
grads = torch._foreach_neg(tuple(grads)) # type: ignore[assignment]
grads = [torch.view_as_real(x) if torch.is_complex(x) else x for x in grads]
exp_avgs = [torch.view_as_real(x) if torch.is_complex(x) else x for x in exp_avgs]
exp_avg_sqs = [torch.view_as_real(x) if torch.is_complex(x) else x for x in exp_avg_sqs]
params = [torch.view_as_real(x) if torch.is_complex(x) else x for x in params]
mults = [grad_mult_map[x] if x in grad_mult_map else torch.ones_like(x).to(x.device) for x in params] # not tested
# update steps
torch._foreach_add_(state_steps, 1)
# Perform stepweight decay
torch._foreach_mul_(params, 1 - lr * weight_decay)
# Decay the first and second moment running average coefficient
torch._foreach_mul_(exp_avgs, beta1)
torch._foreach_add_(exp_avgs, grads, alpha=1 - beta1)
torch._foreach_mul_(exp_avg_sqs, beta2)
torch._foreach_addcmul_(exp_avg_sqs, grads, grads, 1 - beta2)
if capturable:
# TODO: use foreach_pow if/when foreach_pow is added
bias_correction1 = [torch.pow(beta1, step) for step in state_steps]
bias_correction2 = [torch.pow(beta2, step) for step in state_steps]
# foreach_sub doesn't allow a scalar as the first arg
torch._foreach_sub_(bias_correction1, 1)
torch._foreach_sub_(bias_correction2, 1)
torch._foreach_neg_(bias_correction1)
torch._foreach_neg_(bias_correction2)
# foreach_div doesn't allow a scalar as the first arg
step_size = torch._foreach_div(bias_correction1, lr)
torch._foreach_reciprocal_(step_size)
torch._foreach_neg_(step_size)
bias_correction2_sqrt = torch._foreach_sqrt(bias_correction2)
if amsgrad:
# Maintains the maximum of all 2nd moment running avg. till now
torch._foreach_maximum_(max_exp_avg_sqs, exp_avg_sqs)
# Use the max. for normalizing running avg. of gradient
max_exp_avg_sq_sqrt = torch._foreach_sqrt(max_exp_avg_sqs)
# Folds in (admittedly ugly) 1-elem step_size math here to avoid extra param-set-sized read+write
# (can't fold it into addcdiv_ below because addcdiv_ requires value is a Number, not a Tensor)
torch._foreach_div_(max_exp_avg_sq_sqrt, torch._foreach_mul(bias_correction2_sqrt, step_size))
eps_over_step_size = torch._foreach_div(step_size, eps)
torch._foreach_reciprocal_(eps_over_step_size)
denom = torch._foreach_add(max_exp_avg_sq_sqrt, eps_over_step_size)
else:
exp_avg_sq_sqrt = torch._foreach_sqrt(exp_avg_sqs)
torch._foreach_div_(exp_avg_sq_sqrt, torch._foreach_mul(bias_correction2_sqrt, step_size))
eps_over_step_size = torch._foreach_div(step_size, eps)
torch._foreach_reciprocal_(eps_over_step_size)
denom = torch._foreach_add(exp_avg_sq_sqrt, eps_over_step_size)
torch._foreach_mul_(exp_avgs, mults) # not tested
torch._foreach_addcdiv_(params, exp_avgs, denom)
else:
bias_correction1 = [1 - beta1 ** step.item() for step in state_steps]
bias_correction2 = [1 - beta2 ** step.item() for step in state_steps]
step_size = [(lr / bc) * -1 for bc in bias_correction1]
bias_correction2_sqrt = [math.sqrt(bc) for bc in bias_correction2]
if amsgrad:
# Maintains the maximum of all 2nd moment running avg. till now
torch._foreach_maximum_(max_exp_avg_sqs, exp_avg_sqs)
# Use the max. for normalizing running avg. of gradient
max_exp_avg_sq_sqrt = torch._foreach_sqrt(max_exp_avg_sqs)
torch._foreach_div_(max_exp_avg_sq_sqrt, bias_correction2_sqrt)
denom = torch._foreach_add(max_exp_avg_sq_sqrt, eps)
else:
exp_avg_sq_sqrt = torch._foreach_sqrt(exp_avg_sqs)
torch._foreach_div_(exp_avg_sq_sqrt, bias_correction2_sqrt)
denom = torch._foreach_add(exp_avg_sq_sqrt, eps)
torch._foreach_mul_(exp_avgs, mults) # not tested
torch._foreach_addcdiv_(params, exp_avgs, denom, step_size)