forked from clovaai/rebias
/
main_imagenet.py
157 lines (140 loc) · 5.84 KB
/
main_imagenet.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
"""ReBias
Copyright (c) 2020-present NAVER Corp.
MIT license
Entry point of 9-Class ImageNet experiments.
This script provides full implementations including
- Various methods (ReBias, Vanilla, Biased, LearnedMixIn, RUBi)
- Target network: ResNet-18
- Biased network: BagNet-18
- We do not provide Stylised ImageNet implementation here. See README.md for details.
- Sub-sampled 9-Class ImageNet / ImageNet-A from the full ImageNet / ImageNet-A folder.
- Please see datasets/imagenet.py for details.
- Cluster-based unbiased accuracies.
- For curious readers, `make_clusters.py` shows how to make texture clusters.
Usage:
python main_imagenet.py --train_root /path/to/your/imagenet/train
--val_root /path/to/your/imagenet/val
--imageneta_root /path/to/your/imagenet_a
"""
import fire
from datasets.imagenet import get_imagenet_dataloader
from evaluator import ImageNetEvaluator
from logger import PythonLogger
from trainer import Trainer
from models import resnet18, bagnet18, ReBiasModels
class ImageNetTrainer(Trainer):
def _set_models(self):
f_net = resnet18(**self.options.f_config)
g_nets = [bagnet18(**self.options.g_config)
for _ in range(self.options.n_g_nets)]
self.model = ReBiasModels(f_net, g_nets)
self.evaluator = ImageNetEvaluator(device=self.device)
def main(train_root,
val_root,
imageneta_root,
batch_size=128,
num_classes=9,
# optimizer config
lr=0.001,
optim='Adam',
n_epochs=120,
lr_step_size=30,
scheduler='CosineAnnealingLR',
n_f_pretrain_epochs=0,
n_g_pretrain_epochs=0,
f_lambda_outer=1,
g_lambda_inner=1,
n_g_update=1,
update_g_cls=True,
# criterion config
outer_criterion='RbfHSIC',
inner_criterion='MinusRbfHSIC',
rbf_sigma_scale_x=1,
rbf_sigma_scale_y=1,
rbf_sigma_x='median',
rbf_sigma_y='median',
update_sigma_per_epoch=True,
hsic_alg='unbiased',
feature_pos='post',
# model configs
n_g_nets=1,
final_bottleneck_dim=0,
# logging
log_step=10,
# others
save_dir='./checkpoints',
):
logger = PythonLogger()
logger.log('preparing train loader...')
tr_loader = get_imagenet_dataloader(train_root,
batch_size=batch_size,
train=True)
logger.log('preparing val loader...')
val_loaders = {}
val_loaders['biased'] = get_imagenet_dataloader(val_root,
batch_size=batch_size,
train=False)
val_loaders['unbiased'] = get_imagenet_dataloader(val_root,
batch_size=batch_size,
train=False)
val_loaders['imagenet-a'] = get_imagenet_dataloader(imageneta_root,
batch_size=batch_size,
train=False,
val_data='ImageNet-A')
logger.log('preparing trainer...')
if scheduler == 'StepLR':
f_scheduler_config = {'step_size': lr_step_size}
g_scheduler_config = {'step_size': lr_step_size}
elif scheduler == 'CosineAnnealingLR':
f_scheduler_config = {'T_max': n_epochs}
g_scheduler_config = {'T_max': n_epochs}
else:
raise NotImplementedError
if outer_criterion == 'LearnedMixin':
outer_criterion_config = {'feat_dim': 512, 'num_classes': 9}
elif outer_criterion == 'RUBi':
outer_criterion_config = {'feat_dim': 512}
else:
outer_criterion_config = {'sigma_x': rbf_sigma_x, 'sigma_y': rbf_sigma_y,
'algorithm': hsic_alg},
engine = ImageNetTrainer(
outer_criterion=outer_criterion,
inner_criterion=inner_criterion,
outer_criterion_config=outer_criterion_config,
outer_criterion_detail={'sigma_x_type': rbf_sigma_x,
'sigma_y_type': rbf_sigma_y,
'sigma_x_scale': rbf_sigma_scale_x,
'sigma_y_scale': rbf_sigma_scale_y},
inner_criterion_config={'sigma_x': rbf_sigma_x, 'sigma_y': rbf_sigma_y,
'algorithm': hsic_alg},
inner_criterion_detail={'sigma_x_type': rbf_sigma_x,
'sigma_y_type': rbf_sigma_y,
'sigma_x_scale': rbf_sigma_scale_x,
'sigma_y_scale': rbf_sigma_scale_y},
n_epochs=n_epochs,
n_f_pretrain_epochs=n_f_pretrain_epochs,
n_g_pretrain_epochs=n_g_pretrain_epochs,
f_config={'feature_pos': feature_pos,
'num_classes': num_classes},
g_config={'feature_pos': feature_pos,
'num_classes': num_classes},
optimizer=optim,
f_optim_config={'lr': lr, 'weight_decay': 1e-4},
g_optim_config={'lr': lr, 'weight_decay': 1e-4},
f_scheduler_config=f_scheduler_config,
g_scheduler_config=g_scheduler_config,
scheduler=scheduler,
f_lambda_outer=f_lambda_outer,
g_lambda_inner=g_lambda_inner,
n_g_update=n_g_update,
update_g_cls=update_g_cls,
n_g_nets=n_g_nets,
train_loader=tr_loader,
logger=logger,
log_step=log_step)
engine.train(tr_loader, val_loaders=val_loaders,
val_epoch_step=1,
update_sigma_per_epoch=update_sigma_per_epoch,
save_dir=save_dir)
if __name__ == '__main__':
fire.Fire(main)