-
Notifications
You must be signed in to change notification settings - Fork 14
/
Copy pathlocal_unsupervised.py
152 lines (120 loc) · 6.23 KB
/
local_unsupervised.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
from torch.utils.data import Dataset
import copy
import torch
import torch.optim
import torch.nn.functional as F
from options import args_parser
from networks.models import ModelFedCon
from utils import losses, ramps
import torch.nn as nn
from utils_SimPLE import label_guessing, sharpen
from loss.loss import UnsupervisedLoss # , build_pair_loss
import logging
from torchvision import transforms
from ramp import LinearRampUp
args = args_parser()
def get_current_consistency_weight(epoch):
return args.consistency * ramps.sigmoid_rampup(epoch, args.consistency_rampup)
# alpha=0.999
def update_ema_variables(model, ema_model, alpha, global_step):
alpha = min(1 - 1 / (global_step + 1), alpha)
for ema_param, param in zip(ema_model.parameters(), model.parameters()):
ema_param.data.mul_(alpha).add_(param.data, alpha=1 - alpha)
class DatasetSplit(Dataset):
def __init__(self, dataset, idxs):
self.dataset = dataset
self.idxs = list(idxs)
def __len__(self):
return len(self.idxs)
def __getitem__(self, item):
items, index, weak_aug, strong_aug, label = self.dataset[self.idxs[item]]
return items, index, weak_aug, strong_aug, label
class UnsupervisedLocalUpdate(object):
def __init__(self, args, idxs, n_classes):
net = ModelFedCon(args.model, args.out_dim, n_classes=n_classes)
net_ema = ModelFedCon(args.model, args.out_dim, n_classes=n_classes)
if len(args.gpu.split(',')) > 1:
net = torch.nn.DataParallel(net, device_ids=[i for i in range(round(len(args.gpu) / 2))])
net_ema = torch.nn.DataParallel(net_ema, device_ids=[i for i in range(round(len(args.gpu) / 2))])
self.ema_model = net_ema.cuda()
self.model = net.cuda()
for param in self.ema_model.parameters():
param.detach_()
self.data_idxs = idxs
self.epoch = 0
self.iter_num = 0
self.flag = True
self.unsup_lr = args.unsup_lr
self.softmax = nn.Softmax()
self.max_grad_norm = args.max_grad_norm
self.norm = transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
self.max_step = args.rounds * round(len(self.data_idxs) / args.batch_size)
if args.opt == 'adam':
self.optimizer = torch.optim.Adam(self.model.parameters(), lr=args.unsup_lr,
betas=(0.9, 0.999), weight_decay=5e-4)
elif args.opt == 'sgd':
self.optimizer = torch.optim.SGD(self.model.parameters(), lr=args.unsup_lr, momentum=0.9,
weight_decay=5e-4)
elif args.opt == 'adamw':
self.optimizer = torch.optim.AdamW(self.model.parameters(), lr=args.unsup_lr,
weight_decay=0.02)
self.max_warmup_step = round(len(self.data_idxs) / args.batch_size) * args.num_warmup_epochs
self.ramp_up = LinearRampUp(length=self.max_warmup_step)
def train(self, args, net_w, op_dict, epoch, unlabeled_idx, train_dl_local, n_classes):
self.model.load_state_dict(copy.deepcopy(net_w))
self.model.train()
self.ema_model.eval()
self.model.cuda()
self.ema_model.cuda()
self.optimizer.load_state_dict(op_dict)
for param_group in self.optimizer.param_groups:
param_group['lr'] = self.unsup_lr
self.epoch = epoch
if self.flag:
self.ema_model.load_state_dict(copy.deepcopy(net_w))
self.flag = False
logging.info('EMA model initialized')
epoch_loss = []
logging.info('Unlabeled client %d begin unsupervised training' % unlabeled_idx)
correct_pseu = 0
all_pseu = 0
test_right = 0
test_right_ema = 0
train_right = 0
same_total = 0
for epoch in range(args.local_ep):
batch_loss = []
for i, (_, weak_aug_batch, label_batch) in enumerate(train_dl_local):
weak_aug_batch = [weak_aug_batch[version].cuda() for version in range(len(weak_aug_batch))]
with torch.no_grad():
guessed = label_guessing(self.ema_model, [weak_aug_batch[0]], args.model)
sharpened = sharpen(guessed)
pseu = torch.argmax(sharpened, dim=1)
label = label_batch.squeeze()
if len(label.shape) == 0:
label = label.unsqueeze(dim=0)
correct_pseu += torch.sum(label[torch.max(sharpened, dim=1)[0] > args.confidence_threshold] == pseu[
torch.max(sharpened, dim=1)[0] > args.confidence_threshold].cpu()).item()
all_pseu += len(pseu[torch.max(sharpened, dim=1)[0] > args.confidence_threshold])
train_right += sum([pseu[i].cpu() == label_batch[i].int() for i in range(label_batch.shape[0])])
logits_str = self.model(weak_aug_batch[1], model=args.model)[2]
probs_str = F.softmax(logits_str, dim=1)
pred_label = torch.argmax(logits_str, dim=1)
same_total += sum([pred_label[sam] == pseu[sam] for sam in range(logits_str.shape[0])])
loss_u = torch.sum(losses.softmax_mse_loss(probs_str, sharpened)) / args.batch_size
ramp_up_value = self.ramp_up(current=self.epoch)
loss = ramp_up_value * args.lambda_u * loss_u
self.optimizer.zero_grad()
loss.backward()
torch.nn.utils.clip_grad_norm_(self.model.parameters(),
max_norm=self.max_grad_norm)
self.optimizer.step()
update_ema_variables(self.model, self.ema_model, args.ema_decay, self.iter_num)
batch_loss.append(loss.item())
self.iter_num = self.iter_num + 1
epoch_loss.append(sum(batch_loss) / len(batch_loss))
self.epoch = self.epoch + 1
self.model.cpu()
self.ema_model.cpu()
return self.model.state_dict(), self.ema_model.state_dict(), sum(epoch_loss) / len(epoch_loss), copy.deepcopy(
self.optimizer.state_dict()), ramp_up_value, correct_pseu, all_pseu, test_right, train_right.cpu().item(), test_right_ema, same_total.cpu().item()