-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
311 lines (270 loc) · 14.6 KB
/
train.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
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
import torch
import torch.nn as nn
from torch.nn import functional as F
import numpy as np
import copy
from lib.utils.federated_utils import *
from lib.utils.avgmeter import AverageMeter
from train.utils import *
from train.loss import *
from train.context import disable_tracking_bn_stats
from train.ramps import exp_rampup
def train_irc(train_dloader_list, test_dloader_list, model_list, classifier_list, optimizer_list, classifier_optimizer_list, epoch, writer,
num_classes, domain_weight, source_domains, batchnorm_mmd, batch_per_epoch, confidence_gate_begin,
confidence_gate_end, communication_rounds, total_epochs, malicious_domain, attack_level, args=None, pre_models=None, pre_classifiers=None, mean=None):
task_criterion = nn.CrossEntropyLoss().cuda()
logsoftmax = nn.LogSoftmax(dim=1).cuda()
cos = nn.CosineSimilarity(dim=1).cuda()
tau = 0.05
source_domain_num = len(train_dloader_list[1:])
for model in model_list:
model.train()
for classifier in classifier_list:
classifier.train()
# If communication rounds <1,
# then we perform parameter aggregation after (1/communication_rounds) epochs
# If communication rounds >=1:
# then we extend the training epochs and use fewer samples in each epoch.
if communication_rounds in [0.2, 0.5]:
model_aggregation_frequency = round(1 / communication_rounds)
else:
model_aggregation_frequency = 1
# # train local source domain models
for f in range(model_aggregation_frequency):
current_domain_index = 0
# Train model locally on source domains
for train_dloader, model, classifier, optimizer, classifier_optimizer in zip(train_dloader_list[1:],
model_list[1:],
classifier_list[1:],
optimizer_list[1:],
classifier_optimizer_list[1:]):
# check if the source domain is the malicious domain with poisoning attack
source_domain = source_domains[current_domain_index]
current_domain_index += 1
if source_domain == malicious_domain and attack_level > 0:
poisoning_attack = True
else:
poisoning_attack = False
for i, (image_s, label_s) in enumerate(train_dloader):
if i >= batch_per_epoch:
break
image_s_w = image_s[0].cuda()
image_s_s = image_s[1].cuda()
label_s = label_s.long().cuda()
true_label = label_s
if poisoning_attack:
# perform poison attack on source domain
corrupted_num = round(label_s.size(0) * attack_level)
# provide fake labels for those corrupted data
label_s[:corrupted_num, ...] = (label_s[:corrupted_num, ...] + 1) % num_classes
# reset grad
optimizer.zero_grad()
classifier_optimizer.zero_grad()
# each source domain do optimize
feature_s, _ = model(image_s_w)
if args.pj == 1:
_, feature_cur = model(image_s_s)
else:
feature_cur, _ = model(image_s_s)
output_s = classifier(feature_s)
# label smooth
log_probs = torch.log_softmax(output_s, dim=1)
label_s = torch.zeros(log_probs.size()).scatter_(1, label_s.unsqueeze(1).cpu(), 1)
label_s = label_s.cuda()
alpha = 0.1
label_s = (1-alpha) * label_s + alpha / num_classes
task_loss_s = (- label_s * log_probs).mean(0).sum()
# Instance-Instance similarity graph
feature_sims = []
sort_idxs = []
target_sim = []
target_gcc = []
with torch.no_grad():
# calculate the instance2instance similarity, from offline local target domain model (weak augmentation)
temp = F.normalize(pre_models[0](image_s_w)[0])
sim = torch.matmul(temp, torch.t(temp))
idx = torch.sort(-sim)[1]
feature_sims.append(sim)
sort_idxs.append(idx)
k=12
target_tmp = torch.zeros_like(sim)
for j in range(image_s_w.size(0)):
target_tmp[j][idx[j][:k]] = 1.0
target_sim.append(target_tmp)
for i in range(len(feature_sims)):
if i == 0:
target_gcc = feature_sims[i]
else:
target_gcc += feature_sims[i]
tau = args.s_tau
target_gcc = target_gcc/tau
target_gcc = F.softmax(target_gcc, dim=1)
# calculate the instance2instance similarity, from online local source domain model (weak augmentation)
intra_temp = F.normalize(model(image_s_w)[0])
intra_sim = torch.matmul(intra_temp, torch.t(intra_temp))
intra_gcc = intra_sim/tau
intra_gcc = F.softmax(intra_gcc, dim=1)
# calculate the instance2instance similarity, from online local source domain model (weak augmentation)
online_temp = F.normalize(model(image_s_w)[0])
online_sim = torch.matmul(online_temp, torch.t(online_temp))
online_sim_tmp = logsoftmax(online_sim)
# inter-irc loss, online local source model (weak augmentation) -> offline local target model (weak augmentation), cross domains
inter_gcc_loss = online_sim_tmp * target_gcc
inter_gcc_loss = - inter_gcc_loss.mean(0).sum()
# calculate the instance2instance similarity, from online local source domain model (strong augmentation)
online_intra_temp = F.normalize(model(image_s_s)[0])
online_intra_sim = torch.matmul(online_intra_temp, torch.t(online_intra_temp))
online_intra_sim_tmp = logsoftmax(online_intra_sim)
# intra-irc loss, online local source model (weak augmentation) -> online local source model (strong augmentation), cross data views
intra_gcc_loss = online_intra_sim_tmp * intra_gcc
intra_gcc_loss = - intra_gcc_loss.mean(0).sum()
loss = task_loss_s + args.s_inter * inter_gcc_loss + args.s_intra * intra_gcc_loss
loss.backward()
optimizer.step()
classifier_optimizer.step()
# the epochs of local target domain
epochs_target = 1
# 模型聚合权重
target_weight = [0, 0]
consensus_focus_dict = {}
# train local target domain model
for f in range(epochs_target):
# train local target domain model by pseudo-labeling strategy, such as knowledge vote strategy (KD3A)
confidence_gate = (confidence_gate_end - confidence_gate_begin) * (epoch / total_epochs) + confidence_gate_begin
for i in range(1, len(train_dloader_list)):
consensus_focus_dict[i] = 0
for i, (image_t, label_t) in enumerate(train_dloader_list[0]):
if i >= batch_per_epoch:
break
optimizer_list[0].zero_grad()
classifier_optimizer_list[0].zero_grad()
image_w = image_t[0].cuda()
image_s = image_t[1].cuda()
# knowledge vote
with torch.no_grad():
knowledge_list = [torch.softmax(classifier_list[i](model_list[i](image_w)[0]), dim=1).unsqueeze(1) for
i in range(0, len(classifier_list))]
knowledge_list = torch.cat(knowledge_list, 1)
_, kv_pred, kv_mask = knowledge_vote(knowledge_list, confidence_gate,
num_classes=num_classes)
target_weight[0] += torch.sum(kv_mask).item()
target_weight[1] += kv_mask.size(0)
consensus_focus_dict = calculate_consensus_focus(consensus_focus_dict, knowledge_list, confidence_gate,
source_domain_num, num_classes)
graph_weights = []
sum_weights = 0.0
for k,v in consensus_focus_dict.items():
graph_weights.append(v)
sum_weights += v
for i in range(len(graph_weights)):
graph_weights[i]/=(sum_weights + 1e-5)
# choose pseudo-labeling strategies
pseudo_label = []
label_mask = []
# if args.pl == 1:
# # Max predictioin
# pseudo_label = max_pred
# label_mask = max_pred_mask
# elif args.pl == 2:
# # mean prediction
# pseudo_label = mean_pred
# label_mask = mean_pred_mask
# elif args.pl == 3:
# # knowledge vote
# pseudo_label = kv_pred
# label_mask = kv_mask
# else:
# pseudo_label = weighted_mean_pred
# label_mask = weighted_mean_pred_mask
# knowledge vote, from KD3A
pseudo_label = kv_pred
label_mask = kv_mask
# # Mixup strategy
# lam = np.random.beta(2, 2)
# batch_size = image_w.size(0)
# index = torch.randperm(batch_size).cuda()
# mixed_image = lam * image_w + (1 - lam) * image_w[index, :]
# mixed_label = lam * pseudo_label + (1 - lam) * pseudo_label[index, :]
# feature_t, _ = model_list[0](mixed_image)
# output_t_cls = classifier_list[0](feature_t)
# output_t = torch.log_softmax(output_t_cls, dim=1)
# l_u = (-mixed_label * output_t).sum(1)
# task_loss_t = (l_u * label_mask).mean()
# task loss
feature_t, _ = model_list[0](image_w)
output_t_cls = classifier_list[0](feature_t)
output_t = torch.log_softmax(output_t_cls, dim=1)
l_u = (-pseudo_label * output_t).sum(1)
task_loss_t = (l_u * label_mask).mean()
# Instance-Instance similarity
feature_sims = []
sort_idxs = []
target_sim = []
target_gcc = []
with torch.no_grad():
# calculate the instance2instance similarity, from multiple offline local source domain models (weak augmentation)
for i in range(1, len(model_list)):
# if i == current_domain_index:
# continue
temp = F.normalize(model_list[i](image_w)[0])
sim = torch.matmul(temp, torch.t(temp))
idx = torch.sort(-sim)[1]
feature_sims.append(sim)
sort_idxs.append(idx)
k=12
target_tmp = torch.zeros_like(sim)
for j in range(image_w.size(0)):
target_tmp[j][idx[j][:k]] = 1.0
target_sim.append(target_tmp)
for i in range(len(feature_sims)):
if i == 0:
target_gcc = feature_sims[i] * graph_weights[i]
else:
target_gcc += feature_sims[i] * graph_weights[i]
tau = args.t_tau
target_gcc = target_gcc/tau
target_gcc = F.softmax(target_gcc, dim=1)
# calculate the instance2instance similarity, from online local target domain model (weak augmentation)
local_temp = F.normalize(model_list[0](image_w)[0])
local_sim = torch.matmul(local_temp, torch.t(local_temp))
target_local = local_sim/tau
target_local = F.softmax(target_local, dim=1)
# calculate the instance2instance similarity, from online local target domain model (weak augmentation)
online_temp = F.normalize(model_list[0](image_w)[0])
online_sim = torch.matmul(online_temp, torch.t(online_temp))
online_sim_tmp = logsoftmax(online_sim)
# inter_irc loss, online local target model (weak augmentation) -> offline local source models (weak augmentation), cross domains
gcc_loss = online_sim_tmp * target_gcc
gcc_loss = - gcc_loss.mean(0).sum()
# calculate the instance2instance similarity, from online local target domain model (strong augmentation)
online_temp = F.normalize(model_list[0](image_s)[0])
online_sim = torch.matmul(online_temp, torch.t(online_temp))
online_sim_tmp_strong = logsoftmax(online_sim)
# intra_irc loss, online local target model (strong augmentation) -> online local target model (weak augmentation), cross data views
gcc_local = online_sim_tmp_strong * target_local
gcc_local = - gcc_local.mean(0).sum()
# overall losses
loss = task_loss_t + args.t_inter * gcc_loss + args.t_intra * gcc_local
loss.backward()
optimizer_list[0].step()
classifier_optimizer_list[0].step()
# save the local models, before model aggregation
pre_models = []
pre_classifiers = []
for i in range(0, len(model_list)):
pre_models.append(copy.deepcopy(model_list[i]))
pre_classifiers.append(copy.deepcopy(classifier_list[i]))
# test the accuracy of local target domain model, before model aggregation
target_domain = '******target******'
if args is not None:
target_domain = args.target_domain
acc = test(target_domain, source_domains, test_dloader_list, model_list, classifier_list, epoch, writer, num_classes, states='local')
# aggregating weights of local models: average strategy
domain_weight = []
num_domains = len(model_list)
for i in range(num_domains):
domain_weight.append(1.0/num_domains)
# model aggregation
federated_avg(model_list, domain_weight, mode='fedavg')
federated_avg(classifier_list, domain_weight, mode='fedavg')
return acc, pre_models, pre_classifiers