-
Notifications
You must be signed in to change notification settings - Fork 21.3k
/
sgd.py
503 lines (439 loc) · 18.2 KB
/
sgd.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
from typing import List, Optional
import torch
from torch import Tensor
from torch.utils._foreach_utils import _get_fused_kernels_supported_devices
from .optimizer import (
_default_to_fused_or_foreach,
_differentiable_doc,
_foreach_doc,
_fused_doc,
_maximize_doc,
_use_grad_for_differentiable,
DeviceDict,
Optimizer,
)
__all__ = ["SGD", "sgd"]
class SGD(Optimizer):
def __init__(
self,
params,
lr: float = 1e-3,
momentum: float = 0,
dampening: float = 0,
weight_decay: float = 0,
nesterov=False,
*,
maximize: bool = False,
foreach: Optional[bool] = None,
differentiable: bool = False,
fused: Optional[bool] = None,
):
if lr < 0.0:
raise ValueError(f"Invalid learning rate: {lr}")
if momentum < 0.0:
raise ValueError(f"Invalid momentum value: {momentum}")
if weight_decay < 0.0:
raise ValueError(f"Invalid weight_decay value: {weight_decay}")
defaults = dict(
lr=lr,
momentum=momentum,
dampening=dampening,
weight_decay=weight_decay,
nesterov=nesterov,
maximize=maximize,
foreach=foreach,
differentiable=differentiable,
fused=fused,
)
if nesterov and (momentum <= 0 or dampening != 0):
raise ValueError("Nesterov momentum requires a momentum and zero dampening")
super().__init__(params, defaults)
if fused:
self._step_supports_amp_scaling = True
fused_supported_devices = _get_fused_kernels_supported_devices()
if not all(
p.device.type in fused_supported_devices and torch.is_floating_point(p)
for pg in self.param_groups
for p in pg["params"]
):
raise RuntimeError(
"`fused=True` requires all the params to be floating point Tensors of "
f"supported devices: {fused_supported_devices}."
)
if differentiable:
raise RuntimeError("`fused` does not support `differentiable`")
if foreach:
raise RuntimeError("`fused` and `foreach` cannot be `True` together.")
def __setstate__(self, state):
super().__setstate__(state)
for group in self.param_groups:
group.setdefault("nesterov", False)
group.setdefault("maximize", False)
group.setdefault("foreach", None)
group.setdefault("differentiable", False)
group.setdefault("fused", False)
def _init_group(self, group, params, grads, momentum_buffer_list):
has_sparse_grad = False
for p in group["params"]:
if p.grad is not None:
params.append(p)
grads.append(p.grad)
if p.grad.is_sparse:
has_sparse_grad = True
if group["momentum"] != 0:
state = self.state[p]
momentum_buffer_list.append(state.get("momentum_buffer"))
return has_sparse_grad
@_use_grad_for_differentiable
def step(self, closure=None):
"""Performs a single optimization step.
Args:
closure (Callable, optional): A closure that reevaluates the model
and returns the loss.
"""
loss = None
if closure is not None:
with torch.enable_grad():
loss = closure()
for group in self.param_groups:
params: List[Tensor] = []
grads: List[Tensor] = []
momentum_buffer_list: List[Optional[Tensor]] = []
has_sparse_grad = self._init_group(
group, params, grads, momentum_buffer_list
)
sgd(
params,
grads,
momentum_buffer_list,
weight_decay=group["weight_decay"],
momentum=group["momentum"],
lr=group["lr"],
dampening=group["dampening"],
nesterov=group["nesterov"],
maximize=group["maximize"],
has_sparse_grad=has_sparse_grad,
foreach=group["foreach"],
fused=group["fused"],
grad_scale=getattr(self, "grad_scale", None),
found_inf=getattr(self, "found_inf", None),
)
if group["momentum"] != 0:
# update momentum_buffers in state
for p, momentum_buffer in zip(params, momentum_buffer_list):
state = self.state[p]
state["momentum_buffer"] = momentum_buffer
return loss
SGD.__doc__ = (
r"""Implements stochastic gradient descent (optionally with momentum).
.. math::
\begin{aligned}
&\rule{110mm}{0.4pt} \\
&\textbf{input} : \gamma \text{ (lr)}, \: \theta_0 \text{ (params)}, \: f(\theta)
\text{ (objective)}, \: \lambda \text{ (weight decay)}, \\
&\hspace{13mm} \:\mu \text{ (momentum)}, \:\tau \text{ (dampening)},
\:\textit{ nesterov,}\:\textit{ maximize} \\[-1.ex]
&\rule{110mm}{0.4pt} \\
&\textbf{for} \: t=1 \: \textbf{to} \: \ldots \: \textbf{do} \\
&\hspace{5mm}g_t \leftarrow \nabla_{\theta} f_t (\theta_{t-1}) \\
&\hspace{5mm}\textbf{if} \: \lambda \neq 0 \\
&\hspace{10mm} g_t \leftarrow g_t + \lambda \theta_{t-1} \\
&\hspace{5mm}\textbf{if} \: \mu \neq 0 \\
&\hspace{10mm}\textbf{if} \: t > 1 \\
&\hspace{15mm} \textbf{b}_t \leftarrow \mu \textbf{b}_{t-1} + (1-\tau) g_t \\
&\hspace{10mm}\textbf{else} \\
&\hspace{15mm} \textbf{b}_t \leftarrow g_t \\
&\hspace{10mm}\textbf{if} \: \textit{nesterov} \\
&\hspace{15mm} g_t \leftarrow g_{t} + \mu \textbf{b}_t \\
&\hspace{10mm}\textbf{else} \\[-1.ex]
&\hspace{15mm} g_t \leftarrow \textbf{b}_t \\
&\hspace{5mm}\textbf{if} \: \textit{maximize} \\
&\hspace{10mm}\theta_t \leftarrow \theta_{t-1} + \gamma g_t \\[-1.ex]
&\hspace{5mm}\textbf{else} \\[-1.ex]
&\hspace{10mm}\theta_t \leftarrow \theta_{t-1} - \gamma g_t \\[-1.ex]
&\rule{110mm}{0.4pt} \\[-1.ex]
&\bf{return} \: \theta_t \\[-1.ex]
&\rule{110mm}{0.4pt} \\[-1.ex]
\end{aligned}
Nesterov momentum is based on the formula from
`On the importance of initialization and momentum in deep learning`__.
"""
+ rf"""
Args:
params (iterable): iterable of parameters to optimize or dicts defining
parameter groups
lr (float, optional): learning rate (default: 1e-3)
momentum (float, optional): momentum factor (default: 0)
weight_decay (float, optional): weight decay (L2 penalty) (default: 0)
dampening (float, optional): dampening for momentum (default: 0)
nesterov (bool, optional): enables Nesterov momentum (default: False)
{_maximize_doc}
{_foreach_doc}
{_differentiable_doc}
{_fused_doc}
"""
+ r"""
Example:
>>> # xdoctest: +SKIP
>>> optimizer = torch.optim.SGD(model.parameters(), lr=0.1, momentum=0.9)
>>> optimizer.zero_grad()
>>> loss_fn(model(input), target).backward()
>>> optimizer.step()
__ http://www.cs.toronto.edu/%7Ehinton/absps/momentum.pdf
.. note::
The implementation of SGD with Momentum/Nesterov subtly differs from
Sutskever et. al. and implementations in some other frameworks.
Considering the specific case of Momentum, the update can be written as
.. math::
\begin{aligned}
v_{t+1} & = \mu * v_{t} + g_{t+1}, \\
p_{t+1} & = p_{t} - \text{lr} * v_{t+1},
\end{aligned}
where :math:`p`, :math:`g`, :math:`v` and :math:`\mu` denote the
parameters, gradient, velocity, and momentum respectively.
This is in contrast to Sutskever et. al. and
other frameworks which employ an update of the form
.. math::
\begin{aligned}
v_{t+1} & = \mu * v_{t} + \text{lr} * g_{t+1}, \\
p_{t+1} & = p_{t} - v_{t+1}.
\end{aligned}
The Nesterov version is analogously modified.
Moreover, the initial value of the momentum buffer is set to the
gradient value at the first step. This is in contrast to some other
frameworks that initialize it to all zeros.
"""
)
def sgd(
params: List[Tensor],
d_p_list: List[Tensor],
momentum_buffer_list: List[Optional[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
has_sparse_grad: bool = False,
foreach: Optional[bool] = None,
fused: Optional[bool] = None,
grad_scale: Optional[Tensor] = None,
found_inf: Optional[Tensor] = None,
*,
weight_decay: float,
momentum: float,
lr: float,
dampening: float,
nesterov: bool,
maximize: bool,
):
r"""Functional API that performs SGD algorithm computation.
See :class:`~torch.optim.SGD` for details.
"""
# Respect when the user inputs False/True for foreach or fused. We only want to change
# the default when neither have been user-specified. Note that we default to foreach
# and pass False to use_fused. This is not a mistake--we want to give the fused impl
# bake-in time before making it the default, even if it is typically faster.
if foreach is None and fused is None:
# why must we be explicit about an if statement for torch.jit.is_scripting here?
# because JIT can't handle Optionals nor fancy conditionals when scripting
if not torch.jit.is_scripting():
fused, foreach = _default_to_fused_or_foreach(
params, differentiable=False, use_fused=False
)
else:
foreach = False
fused = False
if foreach is None:
foreach = False
if fused is None:
fused = False
if foreach and torch.jit.is_scripting():
raise RuntimeError("torch.jit.script not supported with foreach optimizers")
if fused and torch.jit.is_scripting():
raise RuntimeError("torch.jit.script not supported with fused optimizers")
if foreach and not torch.jit.is_scripting():
func = _multi_tensor_sgd
elif fused and not torch.jit.is_scripting():
func = _fused_sgd
else:
func = _single_tensor_sgd
func(
params,
d_p_list,
momentum_buffer_list,
weight_decay=weight_decay,
momentum=momentum,
lr=lr,
dampening=dampening,
nesterov=nesterov,
has_sparse_grad=has_sparse_grad,
maximize=maximize,
grad_scale=grad_scale,
found_inf=found_inf,
)
def _single_tensor_sgd(
params: List[Tensor],
grads: List[Tensor],
momentum_buffer_list: List[Optional[Tensor]],
grad_scale: Optional[Tensor],
found_inf: Optional[Tensor],
*,
weight_decay: float,
momentum: float,
lr: float,
dampening: float,
nesterov: bool,
maximize: bool,
has_sparse_grad: bool,
):
assert grad_scale is None and found_inf is None
for i, param in enumerate(params):
grad = grads[i] if not maximize else -grads[i]
if weight_decay != 0:
grad = grad.add(param, alpha=weight_decay)
if momentum != 0:
buf = momentum_buffer_list[i]
if buf is None:
buf = torch.clone(grad).detach()
momentum_buffer_list[i] = buf
else:
buf.mul_(momentum).add_(grad, alpha=1 - dampening)
if nesterov:
grad = grad.add(buf, alpha=momentum)
else:
grad = buf
param.add_(grad, alpha=-lr)
def _multi_tensor_sgd(
params: List[Tensor],
grads: List[Tensor],
momentum_buffer_list: List[Optional[Tensor]],
grad_scale: Optional[Tensor],
found_inf: Optional[Tensor],
*,
weight_decay: float,
momentum: float,
lr: float,
dampening: float,
nesterov: bool,
maximize: bool,
has_sparse_grad: bool,
):
assert grad_scale is None and found_inf is None
if len(params) == 0:
return
grouped_tensors = Optimizer._group_tensors_by_device_and_dtype(
[params, grads, momentum_buffer_list], with_indices=True # type: ignore[list-item]
)
for (
device_params,
device_grads,
device_momentum_buffer_list,
), indices in grouped_tensors.values():
device_has_sparse_grad = has_sparse_grad and any(
grad.is_sparse for grad in device_grads
)
if maximize:
device_grads = torch._foreach_neg(device_grads) # type: ignore[assignment]
if weight_decay != 0:
# Re-use the intermediate memory (device_grads) already allocated for maximize
if maximize:
torch._foreach_add_(device_grads, device_params, alpha=weight_decay)
else:
device_grads = torch._foreach_add( # type: ignore[assignment]
device_grads, device_params, alpha=weight_decay
)
if momentum != 0:
bufs = []
all_states_with_momentum_buffer = True
for i in range(len(device_momentum_buffer_list)):
if device_momentum_buffer_list[i] is None:
all_states_with_momentum_buffer = False
break
else:
bufs.append(device_momentum_buffer_list[i])
if all_states_with_momentum_buffer:
torch._foreach_mul_(bufs, momentum)
torch._foreach_add_(bufs, device_grads, alpha=1 - dampening)
else:
bufs = []
for i in range(len(device_momentum_buffer_list)):
if device_momentum_buffer_list[i] is None:
buf = device_momentum_buffer_list[i] = momentum_buffer_list[
indices[i]
] = torch.clone(device_grads[i]).detach()
else:
buf = device_momentum_buffer_list[i]
buf.mul_(momentum).add_(device_grads[i], alpha=1 - dampening)
bufs.append(buf)
if nesterov:
torch._foreach_add_(device_grads, bufs, alpha=momentum)
else:
device_grads = bufs
if not device_has_sparse_grad:
# handle internal item() call if lr is a tensor
if isinstance(lr, torch.Tensor) and torch._utils.is_compiling():
grads_x_lr = torch._foreach_mul(device_grads, -lr)
torch._foreach_add_(device_params, grads_x_lr)
else:
torch._foreach_add_(device_params, device_grads, alpha=-lr)
else:
# foreach APIs don't support sparse
for i in range(len(device_params)):
device_params[i].add_(device_grads[i], alpha=-lr)
def _fused_sgd(
params: List[Tensor],
grads: List[Tensor],
momentum_buffer_list: List[Optional[Tensor]],
grad_scale: Optional[Tensor],
found_inf: Optional[Tensor],
*,
weight_decay: float,
momentum: float,
lr: float,
dampening: float,
nesterov: bool,
maximize: bool,
has_sparse_grad: bool,
) -> None:
if not params:
return
if has_sparse_grad:
raise RuntimeError("`_fused_sgd` does not support sparse gradients")
grad_scale_dict: DeviceDict = (
{grad_scale.device: grad_scale} if grad_scale is not None else {}
)
found_inf_dict: DeviceDict = (
{found_inf.device: found_inf} if found_inf is not None else {}
)
no_momentum_buffer = momentum == 0
is_first_step = (
all(t is None for t in momentum_buffer_list) and not no_momentum_buffer
)
if is_first_step:
for i, g in enumerate(grads):
momentum_buffer_list[i] = torch.empty_like(g)
grouped_tensors = Optimizer._group_tensors_by_device_and_dtype(
[params, grads, momentum_buffer_list], with_indices=False # type: ignore[list-item]
)
for (device, _), (
(device_params, device_grads, device_momentum_buffer_list),
_,
) in grouped_tensors.items():
device_grad_scale, device_found_inf = None, None
if grad_scale is not None:
device_grad_scale = grad_scale_dict.setdefault(
device, grad_scale.to(device)
)
if found_inf_dict is not None and found_inf is not None:
device_found_inf = found_inf_dict.setdefault(device, found_inf.to(device))
torch._fused_sgd_(
device_params,
device_grads,
[] if no_momentum_buffer else device_momentum_buffer_list,
weight_decay=weight_decay,
momentum=momentum,
lr=lr,
dampening=dampening,
nesterov=nesterov,
maximize=maximize,
is_first_step=is_first_step,
grad_scale=device_grad_scale,
found_inf=device_found_inf,
)