forked from apple/learning-subspaces
-
Notifications
You must be signed in to change notification settings - Fork 1
/
schedulers.py
260 lines (200 loc) · 7.24 KB
/
schedulers.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
#
# For licensing see accompanying LICENSE file.
# Copyright (C) 2020 Apple Inc. All Rights Reserved.
#
# modified from https://github.com/allenai/hidden-networks/blob/master/utils/schedulers.py
import numpy as np
__all__ = ["multistep_lr", "cosine_lr", "constant_lr", "get_policy"]
def get_policy(name):
if name is None:
return constant_lr
out_dict = {
"constant_lr": constant_lr,
"cosine_lr": cosine_lr,
"multistep_lr": multistep_lr,
"linear": linear_lr,
"swa_constant_lr": swa_constant_lr,
"swa_cyc_lr": swa_cyc_lr,
"cyc_lr": cyc_lr,
"lth_lr": lth_lr,
"snapshot_cyc_lr": snapshot_cyc_lr,
}
return out_dict[name]
def assign_learning_rate(optimizer, new_lr):
for param_group in optimizer.param_groups:
param_group["lr"] = new_lr
def constant_lr(optimizer, args, **kwargs):
def _lr_adjuster(epoch, iteration):
if epoch < args.warmup_length:
lr = _warmup_lr(args.lr, args.warmup_length, epoch)
else:
lr = args.lr
assign_learning_rate(optimizer, lr)
return lr
return _lr_adjuster
def cosine_lr(optimizer, args, **kwargs):
def _lr_adjuster(epoch, iteration):
if epoch < args.warmup_length:
lr = _warmup_lr(args.lr, args.warmup_length, epoch)
else:
e = epoch - args.warmup_length
es = args.epochs - args.warmup_length
lr = 0.5 * (1 + np.cos(np.pi * e / es)) * args.lr
assign_learning_rate(optimizer, lr)
print(lr)
return lr
return _lr_adjuster
def linear_lr(optimizer, args, **kwargs):
def _lr_adjuster(epoch, iteration):
lr = (1.0 - (epoch / args.epochs)) * args.lr
assign_learning_rate(optimizer, lr)
print("linear", lr)
return lr
return _lr_adjuster
def exp_increase(optimizer, args, **kwargs):
def _lr_adjuster(epoch, iteration):
lr = (1.0 + (epoch / args.epochs)) * args.lr
assign_learning_rate(optimizer, lr)
print("linear", lr)
return lr
return _lr_adjuster
def multistep_lr(optimizer, args, **kwargs):
"""Sets the learning rate to the initial LR decayed by 10 every 30 epochs"""
def _lr_adjuster(epoch, iteration):
lr = args.lr * (args.lr_gamma ** (epoch // args.lr_adjust))
assign_learning_rate(optimizer, lr)
return lr
return _lr_adjuster
def _warmup_lr(base_lr, warmup_length, epoch):
return base_lr * (epoch + 1) / warmup_length
def swa_constant_lr(optimizer, args, **kwargs):
def _lr_adjuster(epoch, iteration):
if epoch < args.warmup_length:
lr = _warmup_lr(args.lr, args.warmup_length, epoch)
else:
t = float(epoch / args.epochs)
lr_ratio = args.swa_lr / args.lr
if t <= 0.5:
factor = 1.0
elif t <= 0.9:
factor = 1.0 - (1.0 - lr_ratio) * (t - 0.5) / 0.4
else:
factor = lr_ratio
lr = args.lr * factor
assign_learning_rate(optimizer, lr)
print(lr)
return lr
return _lr_adjuster
# modified from https://github.com/timgaripov/swa
def swa_cyc_lr(optimizer, args, **kwargs):
def _lr_adjuster(epoch, iteration):
if epoch < args.warmup_length:
lr = _warmup_lr(args.lr, args.warmup_length, epoch)
else:
b = args.epochs
# TODO: this only works for 160 epochs as of now.
a = int(args.swa_start * 160)
ts = [b]
if args.num_models >= 2:
ts = [
x
for x in np.arange(
a, b + 1, (b - a) // (args.num_models - 1)
)
]
if len(ts) > args.num_models:
ts = [
x
for x in np.arange(
a, b, (b - a) // (args.num_models - 1)
)
]
if ts[-1] != b:
ts[-1] = b
if epoch < ts[0]:
lrbase = args.lr
e = epoch - args.warmup_length
es = ts[0] - args.warmup_length
lr = 0.5 * (1 + np.cos(np.pi * e / es)) * lrbase
else:
for i in range(len(ts)):
if epoch >= ts[i]:
lrbase = args.swa_lr
e = epoch - ts[i]
es = ts[i + 1] - ts[i]
lr = 0.5 * (1 + np.cos(np.pi * e / es)) * lrbase
assign_learning_rate(optimizer, lr)
print(lr)
return lr
return _lr_adjuster
def snapshot_cyc_lr(optimizer, args, **kwargs):
def _lr_adjuster(epoch, iteration):
if epoch < args.warmup_length:
lr = _warmup_lr(args.lr, args.warmup_length, epoch)
else:
b = args.epochs
a = args.warmup_length
ts = [b]
if args.num_models >= 2:
ts = [
x for x in np.arange(a, b + 1, (b - a) // args.num_models)
]
if ts[-1] != b:
ts[-1] = b
if epoch < ts[0]:
lrbase = args.lr
e = epoch - args.warmup_length
es = ts[0] - args.warmup_length
lr = 0.5 * (1 + np.cos(np.pi * e / es)) * lrbase
else:
for i in range(len(ts)):
if epoch >= ts[i]:
lrbase = args.lr
e = epoch - ts[i]
es = ts[i + 1] - ts[i]
lr = 0.5 * (1 + np.cos(np.pi * e / es)) * lrbase
assign_learning_rate(optimizer, lr)
print(lr)
return lr
return _lr_adjuster
def cyc_lr(optimizer, args, **kwargs):
def _lr_adjuster(epoch, iteration):
if epoch < args.warmup_length:
lr = _warmup_lr(args.lr, args.warmup_length, epoch)
else:
b = args.epochs
a = int(args.swa_start * 160)
ts = [b]
nm = 3
if nm >= 2:
ts = [x for x in np.arange(a, b + 1, (b - a) // (nm - 1))]
if ts[-1] != b:
ts[-1] = b
if epoch < ts[0]:
lrbase = args.lr
e = epoch - args.warmup_length
es = ts[0] - args.warmup_length
lr = 0.5 * (1 + np.cos(np.pi * e / es)) * lrbase
else:
for i in range(len(ts)):
if epoch >= ts[i]:
lrbase = args.swa_lr
e = epoch - ts[i]
es = ts[i + 1] - ts[i]
lr = 0.5 * (1 + np.cos(np.pi * e / es)) * lrbase
assign_learning_rate(optimizer, lr)
print(lr)
return lr
return _lr_adjuster
def lth_lr(optimizer, args, **kwargs):
"""Sets the learning rate to the initial LR decayed by 10 every 30 epochs"""
def _lr_adjuster(epoch, iteration):
if epoch < 80:
lr = args.lr
elif epoch < 120:
lr = args.lr * 0.1
else:
lr = args.lr * 0.01
assign_learning_rate(optimizer, lr)
return lr
return _lr_adjuster