-
Notifications
You must be signed in to change notification settings - Fork 4
/
get_adam_optimizer.py
54 lines (40 loc) · 1.32 KB
/
get_adam_optimizer.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
from typing import Tuple
from torch.optim import AdamW, Adam
# optimizer
def separate_weight_decayable_params(params):
wd_params, no_wd_params = [], []
for param in params:
param_list = no_wd_params if param.ndim < 2 else wd_params
param_list.append(param)
return wd_params, no_wd_params
def get_adam_optimizer(
params,
lr: float = 1e-4,
wd: float = 1e-2,
betas: Tuple[int, int] = (0.9, 0.99),
eps: float = 1e-8,
filter_by_requires_grad = False,
omit_gammas_and_betas_from_wd = True,
**kwargs
):
has_weight_decay = wd > 0.
if filter_by_requires_grad:
params = [t for t in params if t.requires_grad]
opt_kwargs = dict(
lr = lr,
betas = betas,
eps = eps
)
if not has_weight_decay:
return Adam(params, **opt_kwargs)
opt_kwargs = {'weight_decay': wd, **opt_kwargs}
if not omit_gammas_and_betas_from_wd:
return AdamW(params, **opt_kwargs)
# there is an early practice where betas and gammas are omitted from weight decay in transformers
# unsure whether it is really needed or not
wd_params, no_wd_params = separate_weight_decayable_params(params)
params = [
{'params': wd_params},
{'params': no_wd_params, 'weight_decay': 0},
]
return AdamW(params, **opt_kwargs)