-
Notifications
You must be signed in to change notification settings - Fork 3
/
optim_factory.py
129 lines (118 loc) · 4.73 KB
/
optim_factory.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
""" Optimizer Factory w/ Custom Weight Decay
Hacked together by / Copyright 2020 Ross Wightman
"""
import torch
from torch import optim as optim
from .adafactor import Adafactor
from .adahessian import Adahessian
from .adamp import AdamP
from .lookahead import Lookahead
from .nadam import Nadam
from .novograd import NovoGrad
from .nvnovograd import NvNovoGrad
from .radam import RAdam
from .rmsprop_tf import RMSpropTF
from .sgdp import SGDP
try:
from apex.optimizers import FusedNovoGrad, FusedAdam, FusedLAMB, FusedSGD
has_apex = True
except ImportError:
has_apex = False
def add_weight_decay(model, weight_decay=1e-5, skip_list=()):
decay = []
no_decay = []
for name, param in model.named_parameters():
if not param.requires_grad:
continue # frozen weights
if len(param.shape) == 1 or name.endswith(".bias") or name in skip_list:
no_decay.append(param)
else:
decay.append(param)
return [
{'params': no_decay, 'weight_decay': 0.},
{'params': decay, 'weight_decay': weight_decay}]
def create_optimizer(args, model, filter_bias_and_bn=True):
"""
args : dict, arguments of optimizer
model: nn.Module, initialized model
"""
print('[setup] optimizer:', args)
opt_lower = args.opt.lower()
weight_decay = args.weight_decay
if weight_decay and filter_bias_and_bn:
skip = {}
if hasattr(model, 'no_weight_decay'):
skip = model.no_weight_decay()
parameters = add_weight_decay(model, weight_decay, skip)
weight_decay = 0.
else:
parameters = model.parameters()
if 'fused' in opt_lower:
assert has_apex and torch.cuda.is_available(), 'APEX and CUDA required for fused optimizers'
opt_args = dict(lr=args.lr, weight_decay=weight_decay)
if hasattr(args, 'opt_eps') and args.opt_eps is not None:
opt_args['eps'] = args.opt_eps
if hasattr(args, 'opt_betas') and args.opt_betas is not None:
opt_args['betas'] = args.opt_betas
if hasattr(args, 'momentum') and args.momentum is not None:
opt_args['momentum'] = args.momentum
# if hasattr(args, 'opt_args') and args.opt_args is not None:
# opt_args.update(args.opt_args)
opt_split = opt_lower.split('_')
opt_lower = opt_split[-1]
if opt_lower == 'sgd' or opt_lower == 'nesterov':
opt_args.pop('eps', None)
optimizer = optim.SGD(parameters, nesterov=True, **opt_args)
elif opt_lower == 'momentum':
opt_args.pop('eps', None)
optimizer = optim.SGD(parameters, nesterov=False, **opt_args)
elif opt_lower == 'adam':
optimizer = optim.Adam(parameters, **opt_args)
elif opt_lower == 'adamw':
optimizer = optim.AdamW(parameters, **opt_args)
elif opt_lower == 'nadam':
optimizer = Nadam(parameters, **opt_args)
elif opt_lower == 'radam':
optimizer = RAdam(parameters, **opt_args)
elif opt_lower == 'adamp':
optimizer = AdamP(parameters, wd_ratio=0.01, nesterov=True, **opt_args)
elif opt_lower == 'sgdp':
optimizer = SGDP(parameters, nesterov=True, **opt_args)
elif opt_lower == 'adadelta':
optimizer = optim.Adadelta(parameters, **opt_args)
elif opt_lower == 'adafactor':
if not args.lr:
opt_args['lr'] = None
optimizer = Adafactor(parameters, **opt_args)
elif opt_lower == 'adahessian':
optimizer = Adahessian(parameters, **opt_args)
elif opt_lower == 'rmsprop':
optimizer = optim.RMSprop(parameters, alpha=0.9, **opt_args)
elif opt_lower == 'rmsproptf':
optimizer = RMSpropTF(parameters, alpha=0.9, **opt_args)
elif opt_lower == 'novograd':
optimizer = NovoGrad(parameters, **opt_args)
elif opt_lower == 'nvnovograd':
optimizer = NvNovoGrad(parameters, **opt_args)
elif opt_lower == 'fusedsgd':
opt_args.pop('eps', None)
optimizer = FusedSGD(parameters, nesterov=True, **opt_args)
elif opt_lower == 'fusedmomentum':
opt_args.pop('eps', None)
optimizer = FusedSGD(parameters, nesterov=False, **opt_args)
elif opt_lower == 'fusedadam':
optimizer = FusedAdam(parameters, adam_w_mode=False, **opt_args)
elif opt_lower == 'fusedadamw':
optimizer = FusedAdam(parameters, adam_w_mode=True, **opt_args)
elif opt_lower == 'fusedlamb':
optimizer = FusedLAMB(parameters, **opt_args)
elif opt_lower == 'fusednovograd':
opt_args.setdefault('betas', (0.95, 0.98))
optimizer = FusedNovoGrad(parameters, **opt_args)
else:
assert False and "Invalid optimizer"
raise ValueError
if len(opt_split) > 1:
if opt_split[0] == 'lookahead':
optimizer = Lookahead(optimizer)
return optimizer