Newer
Older
# Copyright (c) OpenMMLab. All rights reserved.
from typing import List
import torch
from mmengine.registry import PARAM_SCHEDULERS
from .param_scheduler import (INF, ConstantParamScheduler,
CosineAnnealingParamScheduler,
ExponentialParamScheduler, LinearParamScheduler,
MultiStepParamScheduler, PolyParamScheduler,
StepParamScheduler)
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
@PARAM_SCHEDULERS.register_module()
class ConstantMomentum(ConstantParamScheduler):
"""Decays the momentum value of each parameter group by a small constant
factor until the number of epoch reaches a pre-defined milestone: ``end``.
Notice that such decay can happen simultaneously with other changes to the
momentum value from outside this scheduler.
Args:
optimizer (Optimizer): Wrapped optimizer.
factor (float): The number we multiply momentum until the milestone.
Defaults to 1./3.
begin (int): Step at which to start updating the momentum.
Defaults to 0.
end (int): Step at which to stop updating the momentum.
Defaults to INF.
last_step (int): The index of last step. Used for resume without state
dict. Defaults to -1.
by_epoch (bool): Whether the scheduled momentum is updated by epochs.
Defaults to True.
verbose (bool): Whether to print the momentum for each update.
Defaults to False.
"""
def __init__(self,
optimizer: torch.optim.Optimizer,
factor: float = 1.0 / 3,
begin: int = 0,
end: int = INF,
last_step: int = -1,
by_epoch: bool = True,
verbose: bool = False):
super().__init__(
optimizer,
param_name='momentum',
factor=factor,
begin=begin,
end=end,
last_step=last_step,
by_epoch=by_epoch,
verbose=verbose)
@PARAM_SCHEDULERS.register_module()
class CosineAnnealingMomentum(CosineAnnealingParamScheduler):
r"""Set the momentum of each parameter group using a cosine annealing
schedule, where :math:`\eta_{max}` is set to the initial value and
:math:`T_{cur}` is the number of epochs since the last restart in SGDR:
.. math::
\begin{aligned}
\eta_t & = \eta_{min} + \frac{1}{2}(\eta_{max} - \eta_{min})\left(1
+ \cos\left(\frac{T_{cur}}{T_{max}}\pi\right)\right),
& T_{cur} \neq (2k+1)T_{max}; \\
\eta_{t+1} & = \eta_{t} + \frac{1}{2}(\eta_{max} - \eta_{min})
\left(1 - \cos\left(\frac{1}{T_{max}}\pi\right)\right),
& T_{cur} = (2k+1)T_{max}.
\end{aligned}
Notice that because the schedule
is defined recursively, the momentum can be simultaneously modified
outside this scheduler by other operators. If the momentum is set
solely by this scheduler, the momentum at each step becomes:
.. math::
\eta_t = \eta_{min} + \frac{1}{2}(\eta_{max} - \eta_{min})\left(1 +
\cos\left(\frac{T_{cur}}{T_{max}}\pi\right)\right)
It has been proposed in
`SGDR: Stochastic Gradient Descent with Warm Restarts`_. Note that this
only implements the cosine annealing part of SGDR, and not the restarts.
Args:
optimizer (Optimizer): Wrapped optimizer.
T_max (int): Maximum number of iterations.
eta_min (float): Minimum momentum value. Defaults to 0.
begin (int): Step at which to start updating the momentum.
Defaults to 0.
end (int): Step at which to stop updating the momentum.
Defaults to INF.
last_step (int): The index of last step. Used for resume without
state dict. Defaults to -1.
by_epoch (bool): Whether the scheduled momentum is updated by
epochs. Defaults to True.
verbose (bool): Whether to print the momentum for each update.
Defaults to False.
.. _SGDR\: Stochastic Gradient Descent with Warm Restarts:
https://arxiv.org/abs/1608.03983
"""
def __init__(self,
optimizer: torch.optim.Optimizer,
T_max: int,
eta_min: int = 0,
begin: int = 0,
end: int = INF,
last_step: int = -1,
by_epoch: bool = True,
verbose: bool = False):
super().__init__(
optimizer,
param_name='momentum',
T_max=T_max,
eta_min=eta_min,
begin=begin,
end=end,
last_step=last_step,
by_epoch=by_epoch,
verbose=verbose)
@PARAM_SCHEDULERS.register_module()
class ExponentialMomentum(ExponentialParamScheduler):
"""Decays the momentum of each parameter group by gamma every epoch.
Args:
optimizer (Optimizer): Wrapped optimizer.
gamma (float): Multiplicative factor of momentum value decay.
begin (int): Step at which to start updating the momentum.
Defaults to 0.
end (int): Step at which to stop updating the momentum.
Defaults to INF.
last_step (int): The index of last step. Used for resume without
state dict. Defaults to -1.
by_epoch (bool): Whether the scheduled momentum is updated by
epochs. Defaults to True.
verbose (bool): Whether to print the momentum for each update.
Defaults to False.
"""
def __init__(self,
optimizer: torch.optim.Optimizer,
gamma: float,
begin: int = 0,
end: int = INF,
last_step: int = -1,
by_epoch: bool = True,
verbose: bool = False):
super().__init__(
optimizer,
param_name='momentum',
gamma=gamma,
begin=begin,
end=end,
last_step=last_step,
by_epoch=by_epoch,
verbose=verbose)
@PARAM_SCHEDULERS.register_module()
class LinearMomentum(LinearParamScheduler):
"""Decays the momentum of each parameter group by linearly changing
small multiplicative factor until the number of epoch reaches a pre-defined
milestone: ``end``.
Notice that such decay can happen simultaneously with other changes to the
momentum from outside this scheduler.
Args:
optimizer (Optimizer): Wrapped optimizer.
start_factor (float): The number we multiply momentum in the
first epoch. The multiplication factor changes towards end_factor
in the following epochs. Defaults to 1./3.
end_factor (float): The number we multiply momentum at the end
of linear changing process. Defaults to 1.0.
begin (int): Step at which to start updating the momentum.
Defaults to 0.
end (int): Step at which to stop updating the momentum.
Defaults to INF.
last_step (int): The index of last step. Used for resume without
state dict. Defaults to -1.
by_epoch (bool): Whether the scheduled momentum is updated by
epochs. Defaults to True.
verbose (bool): Whether to print the momentum for each update.
Defaults to False.
"""
def __init__(self,
optimizer: torch.optim.Optimizer,
start_factor: float = 1.0 / 3,
end_factor: float = 1.0,
begin: int = 0,
end: int = INF,
last_step: int = -1,
by_epoch: bool = True,
verbose: bool = False):
super().__init__(
optimizer,
param_name='momentum',
start_factor=start_factor,
end_factor=end_factor,
begin=begin,
end=end,
last_step=last_step,
by_epoch=by_epoch,
verbose=verbose)
@PARAM_SCHEDULERS.register_module()
class MultiStepMomentum(MultiStepParamScheduler):
"""Decays the specified momentum in each parameter group by gamma once the
number of epoch reaches one of the milestones. Notice that such decay can
happen simultaneously with other changes to the momentum from outside this
scheduler.
Args:
optimizer (Optimizer): Wrapped optimizer.
milestones (list): List of epoch indices. Must be increasing.
gamma (float): Multiplicative factor of momentum value decay.
Defaults to 0.1.
begin (int): Step at which to start updating the momentum.
Defaults to 0.
end (int): Step at which to stop updating the momentum.
Defaults to INF.
last_step (int): The index of last step. Used for resume without
state dict. Defaults to -1.
by_epoch (bool): Whether the scheduled momentum is updated by
epochs. Defaults to True.
verbose (bool): Whether to print the momentum for each update.
Defaults to False.
"""
def __init__(self,
optimizer: torch.optim.Optimizer,
milestones: List[int],
gamma: float = 0.1,
last_step: int = -1,
begin: int = 0,
end: int = INF,
by_epoch: bool = True,
verbose: bool = False):
super().__init__(
optimizer,
param_name='momentum',
milestones=milestones,
gamma=gamma,
last_step=last_step,
begin=begin,
end=end,
by_epoch=by_epoch,
verbose=verbose)
@PARAM_SCHEDULERS.register_module()
class StepMomentum(StepParamScheduler):
"""Decays the momentum of each parameter group by gamma every step_size
epochs. Notice that such decay can happen simultaneously with other changes
to the momentum from outside this scheduler.
Args:
optimizer (Optimizer): Wrapped optimizer.
step_size (int): Period of momentum value decay.
gamma (float): Multiplicative factor of momentum value decay.
Defaults to 0.1.
begin (int): Step at which to start updating the momentum.
Defaults to 0.
end (int): Step at which to stop updating the momentum.
Defaults to INF.
last_step (int): The index of last step. Used for resume without
state dict. Defaults to -1.
by_epoch (bool): Whether the scheduled momentum is updated by
epochs. Defaults to True.
verbose (bool): Whether to print the momentum for each update.
Defaults to False.
"""
def __init__(self,
optimizer: torch.optim.Optimizer,
step_size: int,
gamma: float = 0.1,
begin: int = 0,
end: int = INF,
last_step: int = -1,
by_epoch: bool = True,
verbose: bool = False):
super().__init__(
optimizer,
param_name='momentum',
step_size=step_size,
gamma=gamma,
begin=begin,
end=end,
last_step=last_step,
by_epoch=by_epoch,
verbose=verbose)
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
@PARAM_SCHEDULERS.register_module()
class PolyMomentum(PolyParamScheduler):
"""Decays the momentum of each parameter group in a polynomial decay
scheme.
Notice that such decay can happen simultaneously with other changes to the
parameter value from outside this scheduler.
Args:
optimizer (Optimizer): Wrapped optimizer.
eta_min (float): Minimum momentum at the end of scheduling.
Defaults to 0.
power (float): The power of the polynomial. Defaults to 1.0.
begin (int): Step at which to start updating the parameters.
Defaults to 0.
end (int): Step at which to stop updating the parameters.
Defaults to INF.
last_step (int): The index of last step. Used for resume without
state dict. Defaults to -1.
by_epoch (bool): Whether the scheduled parameters are updated by
epochs. Defaults to True.
verbose (bool): Whether to print the value for each update.
Defaults to False.
"""
def __init__(self,
optimizer: torch.optim.Optimizer,
eta_min: float = 0,
power: float = 1,
begin: int = 0,
end: int = INF,
last_step: int = -1,
by_epoch: bool = True,
verbose: bool = False):
super().__init__(
optimizer,
param_name='momentum',
eta_min=eta_min,
power=power,
begin=begin,
end=end,
last_step=last_step,
by_epoch=by_epoch,
verbose=verbose)