-
Notifications
You must be signed in to change notification settings - Fork 8
/
Copy pathpdro_args.py
executable file
·138 lines (130 loc) · 7.5 KB
/
pdro_args.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
#!/usr/bin/env python3
from src.optim import optimizers, lr_schedulers
from src.models import architecture_list
from src.configuration import ArgumentGroup
def add_model_args(experiment):
model_args = ArgumentGroup("Model")
model_args.add_argument('--architecture', type=str,
default="bert-base-uncased",
include_in_name=True,
choices=list(architecture_list.keys()))
model_args.add_argument('--input-format', type=str, default=None,
choices=[None, "bert-base-uncased", "gpt2"],
help="Format (tok+vocabulary) for text input. "
"If None: decided based on the architecture.")
model_args.add_argument("--hidden-dropout-prob", type=float, default=.5)
model_args.add_argument("--input-dropout-prob", type=float, default=.2)
model_args.add_argument("--load-model-from-file", default=None, type=str)
experiment.add_configuration(model_args)
def add_adversary_args(experiment):
adv_args = ArgumentGroup("Adversary")
adv_args.add_argument("--pdro", action="store_true",
help="Train with Adv-DRO", include_in_name=True)
adv_args.add_argument("--adv-architecture",
type=str, default="ff_lm",
choices=list(architecture_list.keys()),
include_in_name="pdro")
adv_args.add_argument("--adv-filename",
type=str, default=None,
include_in_name="pdro")
adv_args.add_argument("--adv-optimizer", type=str, default=None,
choices=list(optimizers.keys()),
include_in_name="pdro")
adv_args.add_argument("--ewc-penalty", type=float, default=0)
adv_args.add_argument("--renorm-ratios", action="store_true")
adv_args.add_argument("--joint", action="store_true",
include_in_name="pdro")
adv_args.add_argument("--class-conditional", action="store_true",
include_in_name="pdro")
adv_args.add_argument("--alpha", type=float, default=1.0,
include_in_name="pdro")
adv_args.add_argument("--beta", type=float, default=1.0,
include_in_name="pdro")
adv_args.add_argument("--ratio-model", action="store_true",
help="Model the ratio directly",
include_in_name="pdro")
adv_args.add_argument("--self-norm-lambda", type=float, default=0,
help="self normalization penalty",
include_in_name="ratio_model")
adv_args.add_argument("--adv-obj", type=str, default="exp_kl",
include_in_name="pdro")
adv_args.add_argument("--tau", type=float, default=1.0,
include_in_name="pdro", help="Temperature")
adv_args.add_argument("--cvar-constraint", action="store_true",
include_in_name="pdro",
help="Use a cvar constraint rather than KL")
adv_args.add_argument("--adv-on-acc", action="store_true",
help="Train adversary to maximize error "
"rate (not loss)",
include_in_name="pdro",)
adv_args.add_argument("--norm-k-model", type=float, default=None,
include_in_name="pdro")
adv_args.add_argument("--norm-k-adv", type=float, default=None,
include_in_name="pdro")
adv_args.add_argument("--norm-model-only", action="store_true",
include_in_name="pdro")
adv_args.add_argument("--norm-adv-only", action="store_true",
include_in_name="pdro")
adv_args.add_argument("--adv-lr", type=float, default=None,
include_in_name="pdro")
adv_args.add_argument("--adv-valid-on-acc", action="store_true",
help="Adversarial validation using accuracy "
"rather than loss",)
adv_args.add_argument("--adv-mom", type=float, default=0,
include_in_name="pdro")
adv_args.add_argument("--non-param", action="store_true",
include_in_name="pdro")
adv_args.add_argument("--kappa", type=float, default=None,
include_in_name="non_param")
adv_args.add_argument("--chi2-eta", type=float, default=None,
include_in_name="non_param")
adv_args.add_argument("--cvar-alpha", type=float, default=None,
include_in_name="non_param")
adv_args.add_argument("--clip-grad-adv", type=float, default=None,
include_in_name="pdro")
adv_args.add_argument("--adv-update-every", type=int, default=1)
adv_args.add_argument("--filter-advs-by", type=str, default="none",
choices=["none", "reverse_kl", "alpha_coverage"])
adv_args.add_argument("--adv-threshold", type=float, default=10000000)
experiment.add_configuration(adv_args)
def add_optimization_args(experiment):
optim_args = ArgumentGroup("Optimization")
optim_args.add_argument("--optimizer", type=str, default="sgd",
choices=list(optimizers.keys()))
optim_args.add_argument("--lr-scheduler", type=str, default="constant",
choices=list(lr_schedulers.keys()))
optim_args.add_argument("--n-steps", type=int, default=500)
optim_args.add_argument("--update-every", type=int, default=1)
optim_args.add_argument("--lm-update-every", type=int, default=1)
optim_args.add_argument("--n-epochs", type=int, default=None)
optim_args.add_argument("--lr", type=float, default=1e-1)
optim_args.add_argument("--weight-decay", type=float, default=0)
optim_args.add_argument("--batch-size", type=int, default=32)
optim_args.add_argument("--num-workers", type=int, default=1)
optim_args.add_argument("--max-tokens-per-batch", type=int, default=None)
optim_args.add_argument("--test-batch-size", type=int, default=32)
optim_args.add_argument("--clip-grad", type=float, default=10.0)
optim_args.add_argument("--valid-interval", default=250)
optim_args.add_argument("--l2-reg", type=float, default=0,
include_in_name=True)
experiment.add_configuration(optim_args)
def add_group_dro_args(experiment):
dro_args = ArgumentGroup("Group-DRO")
dro_args.add_argument("--group-dro", action="store_true",
help="Train with Group DRO",
include_in_name=True)
dro_args.add_argument("--gold-groups", action="store_true",
help="Train with Group DRO on gold domains",
include_in_name="group_dro")
dro_args.add_argument("--group-specifications",
type=str, default=None, nargs="*",
include_in_name="group_dro")
dro_args.add_argument("--train-group-file", type=str, default=None,
include_in_name="group_dro")
dro_args.add_argument("--dev-group-file", type=str, default=None,
include_in_name="group_dro")
dro_args.add_argument("--eta-q", type=float, default=0.1,
include_in_name="group_dro")
dro_args.add_argument("--baseline", type=float, default=None,
nargs="*", include_in_name="group_dro")
experiment.add_configuration(dro_args)