-
Notifications
You must be signed in to change notification settings - Fork 0
/
lr_scheduler.py
183 lines (163 loc) · 7.84 KB
/
lr_scheduler.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
"""
Authors : inzapp
Github url : https://github.com/inzapp/lr-scheduler
Copyright (c) 2022 Inzapp
Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be included in all
copies or substantial portions of the Software.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
SOFTWARE.
"""
import numpy as np
class LRScheduler:
def __init__(self,
iterations,
lr,
policy,
lrf=0.05,
warm_up=0.0,
min_momentum=0.85,
max_momentum=0.95,
initial_cycle_length=2500,
cycle_weight=2):
assert 0.0 <= lr <= 1.0
assert 0.0 <= lrf <= 1.0
assert 0.0 <= warm_up
assert 0.0 <= min_momentum <= 1.0
assert 0.0 <= max_momentum <= 1.0
assert policy in ['constant', 'step', 'step2', 'cosine', 'onecycle']
self.lr = lr
self.policy = policy
self.max_lr = self.lr
self.min_momentum = min_momentum
self.max_momentum = max_momentum
self.iterations = iterations
self.warm_up_iterations = (self.iterations * warm_up) if ((type(warm_up) is float) and (warm_up <= 1.0)) else int(warm_up)
if self.warm_up_iterations > self.iterations:
self.warm_up_iterations = self.iterations
self.cycle_length = initial_cycle_length
self.cycle_weight = cycle_weight
self.min_lr = self.lr * lrf
self.step_weight = np.sqrt(lrf)
self.step2_weight = np.power(lrf, 1.0 / 4.0)
self.cycle_step = 0
def update(self, optimizer, iteration_count):
if self.policy == 'step':
lr = self.__schedule_step_decay(optimizer, iteration_count)
elif self.policy == 'step2':
lr = self.__schedule_step_decay_2(optimizer, iteration_count)
elif self.policy == 'cosine':
lr = self.__schedule_cosine_warm_restart(optimizer, iteration_count)
elif self.policy == 'onecycle':
lr = self.__schedule_one_cycle(optimizer, iteration_count)
elif self.policy == 'constant':
lr = self.lr
else:
print(f'{self.policy} is invalid lr policy.')
lr = None
return lr
def __set_lr(self, optimizer, lr):
optimizer.__setattr__('learning_rate', lr)
def __set_momentum(self, optimizer, momentum):
optimizer_str = optimizer.__str__().lower()
if optimizer_str.find('sgd') > -1:
optimizer.__setattr__('momentum', momentum)
elif optimizer_str.find('adam') > -1:
optimizer.__setattr__('beta_1', momentum)
def __warm_up_lr(self, iteration_count, warm_up_iterations):
return ((np.cos(((iteration_count * np.pi) / warm_up_iterations) + np.pi) + 1.0) * 0.5) * self.lr # cosine warm up
def __schedule_step_decay(self, optimizer, iteration_count):
if self.warm_up_iterations > 0 and iteration_count <= self.warm_up_iterations:
lr = self.__warm_up_lr(iteration_count, self.warm_up_iterations)
elif iteration_count >= int(self.iterations * 0.9):
lr = self.lr * self.step_weight ** 2.0
elif iteration_count >= int(self.iterations * 0.8):
lr = self.lr * self.step_weight
else:
lr = self.lr
self.__set_lr(optimizer, lr)
return lr
def __schedule_step_decay_2(self, optimizer, iteration_count):
if self.warm_up_iterations > 0 and iteration_count <= self.warm_up_iterations:
lr = self.__warm_up_lr(iteration_count, self.warm_up_iterations)
else:
decay_interval = (self.iterations - self.warm_up_iterations) // 5
if iteration_count > self.warm_up_iterations + (decay_interval * 4.0):
lr = self.lr * self.step2_weight ** 4.0
elif iteration_count > self.warm_up_iterations + (decay_interval * 3.0):
lr = self.lr * self.step2_weight ** 3.0
elif iteration_count > self.warm_up_iterations + (decay_interval * 2.0):
lr = self.lr * self.step2_weight ** 2.0
elif iteration_count > self.warm_up_iterations + (decay_interval * 1.0):
lr = self.lr * self.step2_weight
else:
lr = self.lr
self.__set_lr(optimizer, lr)
return lr
def __schedule_one_cycle(self, optimizer, iteration_count):
min_lr = 0.0
max_lr = self.max_lr
min_mm = self.min_momentum
max_mm = self.max_momentum
if self.warm_up_iterations > 0 and iteration_count <= self.warm_up_iterations:
iterations = self.warm_up_iterations
lr = ((np.cos(((iteration_count * np.pi) / iterations) + np.pi) + 1.0) * 0.5) * (max_lr - min_lr) + min_lr # increase only until target iterations
mm = ((np.cos(((iteration_count * np.pi) / iterations) + 0.0) + 1.0) * 0.5) * (max_mm - min_mm) + min_mm # decrease only until target iterations
self.__set_lr(optimizer, lr)
self.__set_momentum(optimizer, mm)
else:
min_lr = self.min_lr
iteration_count -= self.warm_up_iterations + 1
iterations = self.iterations - self.warm_up_iterations
lr = ((np.cos(((iteration_count * np.pi) / iterations) + 0.0) + 1.0) * 0.5) * (max_lr - min_lr) + min_lr # decrease only until target iterations
mm = ((np.cos(((iteration_count * np.pi) / iterations) + np.pi) + 1.0) * 0.5) * (max_mm - min_mm) + min_mm # increase only until target iterations
self.__set_lr(optimizer, lr)
self.__set_momentum(optimizer, mm)
return lr
def __schedule_cosine_warm_restart(self, optimizer, iteration_count):
if self.warm_up_iterations > 0 and iteration_count <= self.warm_up_iterations:
lr = self.__warm_up_lr(iteration_count, self.warm_up_iterations)
else:
if self.cycle_step % self.cycle_length == 0 and self.cycle_step != 0:
self.cycle_step = 0
self.cycle_length = int(self.cycle_length * self.cycle_weight)
lr = self.min_lr + 0.5 * (self.max_lr - self.min_lr) * (1.0 + np.cos(((1.0 / self.cycle_length) * np.pi * (self.cycle_step % self.cycle_length)))) # down and down
self.cycle_step += 1
self.__set_lr(optimizer, lr)
return lr
def plot_lr(policy):
import tensorflow as tf
from matplotlib import pyplot as plt
lr = 0.001
warm_up = 0.3
decay_step = 0.2
iterations = 37500
iterations = int(iterations / (1.0 - warm_up))
optimizer = tf.keras.optimizers.SGD()
lr_scheduler = LRScheduler(iterations=iterations, lr=lr, warm_up=warm_up, policy=policy)
lrs = []
for i in range(iterations):
lr = lr_scheduler.update(optimizer=optimizer, iteration_count=i)
lrs.append(lr)
plt.figure(figsize=(10, 6))
plt.plot(lrs)
plt.legend(['lr'])
plt.xlabel('iterations')
plt.tight_layout(pad=0.5)
plt.show()
if __name__ == '__main__':
plot_lr('constant')
plot_lr('step')
plot_lr('step2')
plot_lr('onecycle')
plot_lr('cosine')