-
Notifications
You must be signed in to change notification settings - Fork 10
/
main_fl_mrcm.py
184 lines (173 loc) · 8.03 KB
/
main_fl_mrcm.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
175
176
177
178
179
180
181
182
#!/usr/bin/env python
# -*- coding: utf-8 -*-
# Python version: 3.6
# import matplotlib
# matplotlib.use('Agg')
# import matplotlib.pyplot as plt
import copy
import numpy as np
import torch
import os
from utils.options import args_parser
from models.recon_Update import LocalUpdate_ad_da
from models.Fed import FedAvg
from models.test import evaluator
from data.mri_data import SliceData, DataTransform
from data.subsample import create_mask_for_mask_type
from models.unet_model import UnetModel_ad_da, Feature_discriminator
from tensorboardX import SummaryWriter
import pathlib
if __name__ == '__main__':
os.environ["HDF5_USE_FILE_LOCKING"] = 'FALSE'
# parse args
args = args_parser()
path_dict = {'B': pathlib.Path('Dataset dir B'),
'F': pathlib.Path('Dataset dir F'),
'H': pathlib.Path('Dataset dir H'),
'I': pathlib.Path('Dataset dir I')}
rate_dict = {'B': 1.0, 'F': 1.0, 'H': 1.0, 'I': 1.0} # control the sample rate for each dataset
print(rate_dict)
args.device = torch.device('cuda:{}'.format(args.gpu[0]) if torch.cuda.is_available() and args.gpu != -1 else 'cpu')
if not os.path.exists(args.save_dir):
os.mkdir(args.save_dir)
writer = SummaryWriter(log_dir=args.save_dir/ 'summary')
def save_networks(net, epoch, local=False, local_no = None):
"""Save all the networks to the disk.
Parameters:
epoch (int) -- current epoch; used in the file name '%s_net_%s.pth' % (epoch, name)
"""
if local:
save_filename = '%s_net_D_%s.pth' % (epoch,local_no)
else:
save_filename = '%s_net.pth' % (epoch)
save_path = os.path.join(args.save_dir, save_filename)
if len(args.gpu) > 1 and torch.cuda.is_available():
torch.save(net.module.cpu().state_dict(), save_path)
net.to(args.device)
else:
torch.save(net.cpu().state_dict(), save_path)
net.to(args.device)
# data loader
def _create_dataset(data_path,data_transform, data_partition, sequence, sample_rate=None, seed=42):
dataset = SliceData(
root=data_path / data_partition,
transform=data_transform,
sample_rate=sample_rate,
challenge=args.challenge,
sequence=sequence,
seed=seed
)
return dataset
# load dataset and split users
if args.dataset == 'mri':
mask = create_mask_for_mask_type(args.mask_type, args.center_fractions,
args.accelerations)
train_data_transform = DataTransform(args.resolution, args.challenge, mask, use_seed=False)
val_data_transform = DataTransform(args.resolution, args.challenge, mask)
datasets_list = []
if args.phase == 'train':
for data in args.train_datasets:
dataset_train = _create_dataset(path_dict[data]/args.sequence,train_data_transform, 'train', args.sequence,rate_dict[data], args.seed)
datasets_list.append(dataset_train)
dataset_val = _create_dataset(path_dict[args.test_dataset]/args.sequence,val_data_transform, 'val', args.sequence, args.val_sample_rate)
else:
exit('Error: unrecognized dataset')
#make target domain dataset has the same number of sample as max train dataset
target_datasets_list=[]
for dataset in datasets_list:
tmp_list = []
dataset_target_domain = _create_dataset(path_dict[args.test_dataset] / args.sequence, train_data_transform,'train', args.sequence,rate_dict[args.test_dataset])
while len(tmp_list)<len(dataset.examples):
for sample in dataset_target_domain.examples:
tmp_list.append(sample)
if len(tmp_list) ==len(dataset.examples):
break
dataset_target_domain.examples = tmp_list
target_datasets_list.append(dataset_target_domain)
assert (len(datasets_list)==args.num_users)
# build model
if args.model == 'unet':
net_glob = UnetModel_ad_da(
in_chans=1,
out_chans=1,
chans=32,
num_pool_layers=4,
drop_prob=0.0
).to(args.device)
else:
exit('Error: unrecognized model')
print(net_glob)
net_glob.train()
G_s = []
FD = []
for i in range(args.num_users):
if len(args.gpu) > 1:
G_s.append(torch.nn.DataParallel(UnetModel_ad_da(in_chans=1, out_chans=1, chans=32, num_pool_layers=4,
drop_prob=0.0).to(args.device),args.gpu))
FD.append(torch.nn.DataParallel(Feature_discriminator().to(args.device),args.gpu))
else:
G_s.append(UnetModel_ad_da(in_chans=1,out_chans=1,chans=32,num_pool_layers=4,drop_prob=0.0).to(args.device))
FD.append(Feature_discriminator().to(args.device))
# setting optimizer
opt_g_s = []
opt_FD= []
for i in range(args.num_users):
opt_g_s.append(torch.optim.RMSprop(G_s[i].parameters(), lr=args.lr))
opt_FD.append(torch.optim.RMSprop(FD[i].parameters(), lr=args.lr*10))
# copy weights
if len(args.gpu) > 1:
net_glob = torch.nn.DataParallel(net_glob, args.gpu)
w_glob = net_glob.state_dict()
else:
w_glob = net_glob.state_dict()
# initilize parameters
for G in G_s:
for net, net_cardinal in zip(G.named_parameters(), net_glob.named_parameters()):
net[1].data = net_cardinal[1].data.clone()
# training
if args.phase == 'train':
start_epoch = -1
if args.continues:
if len(args.gpu) > 1:
net_glob.module.load_state_dict(torch.load(args.checkpoint))
print('Load checkpoint :', args.checkpoint)
for i, net_d in enumerate(FD):
path = args.checkpoint.split('.')[0]+'_D_%s.pth'%(i)
print('Load checkpoint :', path)
net_d.module.load_state_dict(torch.load(path))
start_epoch = int(args.checkpoint.split('/')[-1].split('_')[0])
else:
net_glob.load_state_dict(torch.load(args.checkpoint))
print('Load checkpoint :', args.checkpoint)
for i, net_d in enumerate(FD):
path = args.checkpoint.split('.')[0] + '_D_%s.pth' % (i)
print('Load checkpoint :', path)
net_d.load_state_dict(torch.load(path))
start_epoch = int(args.checkpoint.split('/')[-1].split('_')[0])
for iter in range(start_epoch+1,args.epochs):
w_locals, loss_locals = [], []
for idx, dataset_train in enumerate(datasets_list):
flag = args.train_datasets[idx] == args.test_dataset # for disable adv loss for target dataset
local = LocalUpdate_ad_da(args=args, device=args.device, dataset=dataset_train,
dataset_target = target_datasets_list[idx], optimizer=opt_g_s[idx],optimizer_fd=opt_FD[idx],flag=flag)
# models communication
G_s[idx].load_state_dict(net_glob.state_dict())
# global update
w, loss = local.train(net=G_s[idx],net_fd=FD[idx] ,epoch=iter, idx=idx, writer=writer)
w_locals.append(copy.deepcopy(w))
loss_locals.append(copy.deepcopy(loss))
# update global weights
w_glob = FedAvg(w_locals)
# copy weight to net_glob
net_glob.load_state_dict(w_glob)
# print loss
loss_avg = np.sum(loss_locals) / len(loss_locals)
print('Round {:3d}, Average loss {:.3f}'.format(iter, loss_avg))
print('saving the model at the end of epoch %d' % (iter))
save_networks(net_glob, iter)
for i, net_d in enumerate(FD):
save_networks(net_d, iter, local=True, local_no=i)
print('Evaluation ...')
validation = evaluator(dataset_val, args, writer,args.device)
validation.evaluate_recon(net_glob,iter)
writer.close()