-
Notifications
You must be signed in to change notification settings - Fork 0
/
utils.py
159 lines (128 loc) · 5.08 KB
/
utils.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
import os
import random
import sys
import numpy as np
import torch
def str_to_bool(val):
"""Convert a string representation of truth to true (1) or false (0).
Copied from the python implementation distutils.utils.strtobool
True values are 'y', 'yes', 't', 'true', 'on', and '1'; false values
are 'n', 'no', 'f', 'false', 'off', and '0'. Raises ValueError if
'val' is anything else.
>>> str_to_bool('YES')
1
>>> str_to_bool('FALSE')
0
"""
val = val.lower()
if val in ('y', 'yes', 't', 'true', 'on', '1'):
return True
if val in ('n', 'no', 'f', 'false', 'off', '0'):
return False
raise ValueError('invalid truth value {}'.format(val))
def cosine_annealing(step, total_steps, lr_max, lr_min):
"""Cosine Annealing for learning rate decay scheduler"""
return lr_min + (lr_max -
lr_min) * 0.5 * (1 + np.cos(step / total_steps * np.pi))
def keras_decay(step, decay=0.0001):
"""Learning rate decay in Keras-style"""
return 1. / (1. + decay * step)
class SGDRScheduler(torch.optim.lr_scheduler._LRScheduler):
"""SGD with restarts scheduler"""
def __init__(self, optimizer, T0, T_mul, eta_min, last_epoch=-1):
self.Ti = T0
self.T_mul = T_mul
self.eta_min = eta_min
self.last_restart = 0
super().__init__(optimizer, last_epoch)
def get_lr(self):
T_cur = self.last_epoch - self.last_restart
if T_cur >= self.Ti:
self.last_restart = self.last_epoch
self.Ti = self.Ti * self.T_mul
T_cur = 0
return [
self.eta_min + (base_lr - self.eta_min) *
(1 + np.cos(np.pi * T_cur / self.Ti)) / 2
for base_lr in self.base_lrs
]
def _get_optimizer(optim_config):
"""Defines optimizer according to the given config"""
optimizer_name = optim_config['optimizer']
if optimizer_name == 'sgd':
optimizer = torch.optim.SGD
"""
optimizer = torch.optim.SGD(model_parameters,
lr=optim_config['base_lr'],
momentum=optim_config['momentum'],
weight_decay=optim_config['weight_decay'],
nesterov=optim_config['nesterov'])
"""
elif optimizer_name == 'adam':
optimizer = torch.optim.Adam
"""
optimizer = torch.optim.Adam(model_parameters,
lr=optim_config['base_lr'],
betas=optim_config['betas'],
weight_decay=optim_config['weight_decay'],
amsgrad=str_to_bool(
optim_config['amsgrad']))
"""
else:
print('Un-known optimizer', optimizer_name)
sys.exit()
return optimizer
def _get_scheduler(optimizer, optim_config):
"""
Defines learning rate scheduler according to the given config
"""
if optim_config['scheduler'] == 'multistep':
scheduler = torch.optim.lr_scheduler.MultiStepLR(
optimizer,
milestones=optim_config['milestones'],
gamma=optim_config['lr_decay'])
elif optim_config['scheduler'] == 'sgdr':
scheduler = SGDRScheduler(optimizer, optim_config['T0'],
optim_config['Tmult'],
optim_config['lr_min'])
elif optim_config['scheduler'] == 'cosine':
total_steps = optim_config['epochs'] * \
optim_config['steps_per_epoch']
scheduler = torch.optim.lr_scheduler.LambdaLR(
optimizer,
lr_lambda=lambda step: cosine_annealing(
step,
total_steps,
1, # since lr_lambda computes multiplicative factor
optim_config['lr_min'] / optim_config['base_lr']))
elif optim_config['scheduler'] == 'keras_decay':
scheduler = torch.optim.lr_scheduler.LambdaLR(
optimizer, lr_lambda=lambda step: keras_decay(step))
else:
scheduler = None
return scheduler
def create_optimizer(model_parameters, optim_config):
"""Defines an optimizer and a scheduler"""
optimizer = _get_optimizer(model_parameters, optim_config)
scheduler = _get_scheduler(optimizer, optim_config)
return optimizer, scheduler
def seed_worker(worker_id):
"""
Used in generating seed for the worker of torch.utils.data.Dataloader
"""
worker_seed = torch.initial_seed() % 2**32
np.random.seed(worker_seed)
random.seed(worker_seed)
def set_seed(seed, config = None):
"""
set initial seed for reproduction
"""
if config is None:
raise ValueError("config should not be None")
random.seed(seed)
np.random.seed(seed)
torch.manual_seed(seed)
if torch.cuda.is_available():
torch.cuda.manual_seed_all(seed)
torch.backends.cudnn.deterministic = str_to_bool(config["cudnn_deterministic_toggle"])
torch.backends.cudnn.benchmark = str_to_bool(config["cudnn_benchmark_toggle"])