-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathmain.py
174 lines (140 loc) · 6.61 KB
/
main.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
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
import torch
import utils.getters_setters as gs
from server import Server
from client import Client
from centralized import Centralized
from fdaCentralized import FdaCentralized
from advCentralized import AdvCentralized
from advServer import AdvServer
from utils.checkpointSaver import CheckpointSaver
from models.discriminator import Discriminator
from utils.args import get_parser
import wandb
wandb.login()
def main():
parser = get_parser()
args = parser.parse_args()
gs.set_seed(args.seed)
args.device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
# in order to include transformations in hyperparameters to save
train_transforms, test_transforms = gs.get_transforms(args)
# if the such args are not set, they are choosen as the best one
if args.hnm is None:
args.hnm = False if args.step == '4' else True
if args.lr is None:
args.lr = 0.025 if args.step in ['1', '2'] else 0.01
config={
"step": args.step,
"seed": args.seed,
"dataset": args.dataset,
"model": args.model,
"num_rounds": args.num_rounds,
"num_epochs": args.num_epochs,
"clients_per_round": args.clients_per_round,
"hnm": args.hnm,
"learning_rate": args.lr,
"schedule": args.schedule,
"batch_size": args.bs,
"weight_decay": args.wd,
"momentum": args.m,
"train_transform": train_transforms,
"test_transform": test_transforms,
"FdaSize": args.L,
"fdaSize_pixelss": args.b,
"T": args.T,
"lambda_adv": args.lam_adv,
"lambda_kd": args.lam_kd,
"path_model": args.path_model,
"path_discriminator": args.path_discriminator,
"L": args.L
}
# to have a config without personalized names
# vars(args) | {"train_transform": train_transforms, "test_transform": test_transforms}
# initialize wandb and save it in args for ease of convenience
run = wandb.init(project="project-2b", config=config)
args.wandb = run
print('Initializing model...')
model = gs.model_init(args)
model.to(args.device)
print('Done.')
if args.path_model is not None:
print('Loading checkpoint...')
gs.loadCheckpoint(path=args.path_model, model=model, device=args.device)
print('Done.')
if args.step in ['5a', '5b', '5c']:
print('Initializing discriminator...')
discriminator = Discriminator(in_channels=gs.get_dataset_num_classes(args.dataset))
discriminator.to(args.device)
print('Done.')
if args.path_discriminator is not None:
print('Loading checkpoint...')
gs.loadCheckpoint(path=args.path_discriminator, model=discriminator, device=args.device)
print('Done.')
print('Generate datasets...')
train_datasets, test_datasets, eval_datasets = gs.get_datasets(args)
print('Done.')
metrics = gs.set_metrics(args)
print('training and testing...')
if args.step == '1':
clf = Centralized(args, model, train_datasets, metrics['train'])
clf.train()
loss_tsd = clf.test(test_datasets[0], metrics['test_same_dom'])
loss_tdd = clf.test(test_datasets[1], metrics['test_diff_dom'])
elif args.step == '2':
train_clients, test_clients = gs.gen_clients(args, train_datasets, test_datasets, model)
clf = Server(args, model, train_clients, metrics['train'])
clf.train()
loss_tsd = clf.test(test_clients[0], metrics['test_same_dom'])
loss_tdd = clf.test(test_clients[1], metrics['test_diff_dom'])
elif args.step == '3a':
clf = Centralized(args, model, train_datasets, metrics['train'])
clf.train(eval_datasets, metrics['eval'])
loss_tsd = clf.test(test_datasets[0], metrics['test_same_dom'])
loss_tdd = clf.test(test_datasets[1], metrics['test_diff_dom'])
elif args.step == '3b':
# generate clients to extract, if needed, the styles from eval_clients
eval_clients, test_clients = gs.gen_clients(args, eval_datasets, test_datasets, model)
clf = FdaCentralized(args, model, train_datasets, metrics['train'], b=args.b, L=args.L, clients=eval_clients)
clf.train(eval_datasets, metrics['eval'])
loss_tsd = clf.test(test_datasets[0], metrics['test_same_dom'])
loss_tdd = clf.test(test_datasets[1], metrics['test_diff_dom'])
elif args.step == '4':
train_clients, test_clients = gs.gen_clients(args, train_datasets, test_datasets, model)
clf = Server(args, model, train_clients, metrics['train'])
clf.train()
loss_tsd = clf.test(test_clients[0], metrics['test_same_dom'])
loss_tdd = clf.test(test_clients[1], metrics['test_diff_dom'])
elif args.step == '5a':
clf = AdvCentralized(args, model, discriminator, train_datasets, metrics['train'], b=args.b)
clf.train(eval_datasets, metrics['eval'])
loss_tsd = clf.test(test_datasets[0], metrics['test_same_dom'])
loss_tdd = clf.test(test_datasets[1], metrics['test_diff_dom'])
elif args.step == '5b':
train_clients, test_clients = gs.gen_clients(args, train_datasets, test_datasets, model, discriminator)
clf = AdvServer(args, model, discriminator, train_clients, metrics['train'])
clf.train()
loss_tsd = clf.test(test_clients[0], metrics['test_same_dom'])
loss_tdd = clf.test(test_clients[1], metrics['test_diff_dom'])
elif args.step == '5c':
train_clients, test_clients = gs.gen_clients(args, train_datasets, test_datasets, model, discriminator)
clf = AdvServer(args, model, discriminator, train_clients, metrics['train'])
clf.train()
loss_tsd = clf.test(test_clients[0], metrics['test_same_dom'])
loss_tdd = clf.test(test_clients[1], metrics['test_diff_dom'])
elif args.step == 'test':
# centralized is used just to exploit its function test
clf = Centralized(args, model, train_datasets, metrics['train'])
loss_tsd = clf.test(test_datasets[0], metrics['test_same_dom'])
loss_tdd = clf.test(test_datasets[1], metrics['test_diff_dom'])
print('Done.')
if args.step in ['1', '2', '4', '5b']: # the ones not saved during training
checkpoint_saver = CheckpointSaver(dirpath='./saved_models', args=args, decreasing=False, top_n=1)
checkpoint_saver(model, metrics['train'].results["Mean IoU"], args.num_epochs)
# log test metrics
wandb.log({
'tsd': metrics['test_same_dom'].results | {'loss': loss_tsd},
'tdd': metrics['test_diff_dom'].results | {'loss': loss_tdd}
})
wandb.finish()
if __name__ == '__main__':
main()