-
Notifications
You must be signed in to change notification settings - Fork 2
/
pfedme.py
146 lines (114 loc) · 5.61 KB
/
pfedme.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
import random
import copy
import numpy as np
import torch
from pfedme.config import get_args
from model import simplecnn, textcnn
from test import compute_local_test_accuracy, compute_acc
from pfedme.utils import pFedMeOptimizer
from prepare_data import get_dataloader
from attack import *
def local_train_fedavg(args, nets_this_round, train_local_dls, val_local_dls, test_dl, data_distributions, best_val_acc_list, best_test_acc_list):
for net_id, net in nets_this_round.items():
train_local_dl = train_local_dls[net_id]
data_distribution = data_distributions[net_id]
# Pre-Trainging Test Accuracy
optimizer = pFedMeOptimizer(filter(lambda p: p.requires_grad, net.parameters()), lr=args.personalized_learning_rate, lamda=args.lamda)
criterion = torch.nn.CrossEntropyLoss().cuda()
net.cuda()
net.train()
local_params = copy.deepcopy(list(net.parameters()))
iterator = iter(train_local_dl)
for iteration in range(args.num_local_iterations):
try:
x, target = next(iterator)
except StopIteration:
iterator = iter(train_local_dl)
x, target = next(iterator)
x, target = x.cuda(), target.cuda()
target = target.long()
for i in range(args.K):
optimizer.zero_grad()
out = net(x)
loss = criterion(out, target)
loss.backward()
personalized_params = optimizer.step(local_params)
for new_param, localweight in zip(personalized_params, local_params):
localweight.data = localweight.data - args.lamda * args.lr * (localweight.data - new_param.data)
for param, new_param in zip(net.parameters(), local_params):
param.data = new_param.data
if net_id in benign_client_list:
val_acc = compute_acc(net, val_local_dls[net_id])
personalized_test_acc, generalized_test_acc = compute_local_test_accuracy(net, test_dl, data_distribution)
if val_acc > best_val_acc_list[net_id]:
best_val_acc_list[net_id] = val_acc
best_test_acc_list[net_id] = personalized_test_acc
print('>> Client {} | Personalized Test Acc: {:.5f} | Generalized Test Acc: {:.5f}'.format(net_id, personalized_test_acc, generalized_test_acc))
net.to('cpu')
return np.array(best_test_acc_list)[np.array(benign_client_list)].mean()
args, cfg = get_args()
print(args)
seed = args.init_seed
np.random.seed(seed)
torch.manual_seed(seed)
if torch.cuda.is_available():
torch.cuda.manual_seed(seed)
random.seed(seed)
n_party_per_round = int(args.n_parties * args.sample_fraction)
party_list = [i for i in range(args.n_parties)]
party_list_rounds = []
if n_party_per_round != args.n_parties:
for i in range(args.comm_round):
party_list_rounds.append(random.sample(party_list, n_party_per_round))
else:
for i in range(args.comm_round):
party_list_rounds.append(party_list)
benign_client_list = random.sample(party_list, int(args.n_parties * (1-args.attack_ratio)))
benign_client_list.sort()
print(f'>> -------- Benign clients: {benign_client_list} --------')
train_local_dls, val_local_dls, test_dl, net_dataidx_map, traindata_cls_counts, data_distributions = get_dataloader(args)
if args.dataset == 'cifar10':
model = simplecnn
elif args.dataset == 'cifar100':
model = simplecnn
elif args.dataset == 'yahoo_answers':
model = textcnn
global_model = model(cfg['classes_size'])
global_parameters = global_model.state_dict()
local_models = []
best_val_acc_list, best_test_acc_list = [],[]
for i in range(args.n_parties):
local_models.append(model(cfg['classes_size']))
best_val_acc_list.append(0)
best_test_acc_list.append(0)
for round in range(args.comm_round): # Federated round loop
party_list_this_round = party_list_rounds[round]
if args.sample_fraction<1.0:
print(f'>> Clients in this round : {party_list_this_round}')
global_w = global_model.state_dict() # Global Model Initialization
nets_this_round = {k: local_models[k] for k in party_list_this_round}
for net in nets_this_round.values():
net.load_state_dict(global_w)
# Local Model Training
mean_personalized_acc = local_train_fedavg(args, nets_this_round, train_local_dls, val_local_dls, test_dl, data_distributions, best_val_acc_list, best_test_acc_list)
# Aggregation Weight Calculation
total_data_points = sum([len(net_dataidx_map[r]) for r in party_list_this_round])
fed_avg_freqs = [len(net_dataidx_map[r]) / total_data_points for r in party_list_this_round]
if round==0 or args.sample_fraction<1.0:
print(f'Dataset size weight : {fed_avg_freqs}')
manipulate_gradient(args, global_model, nets_this_round, benign_client_list)
previous_global_model = copy.deepcopy(global_w)
# Model Aggregation
for net_id, net in enumerate(nets_this_round.values()):
net_para = net.state_dict()
if net_id == 0:
for key in net_para:
global_w[key] = net_para[key] * fed_avg_freqs[net_id]
else:
for key in net_para:
global_w[key] += net_para[key] * fed_avg_freqs[net_id]
for key in previous_global_model:
global_w[key] = (1 - args.alpha) * previous_global_model[key] + args.alpha * global_w[key]
global_model.load_state_dict(global_w) # Update the global model
print('>> (Current) Round {} | Local Per: {:.5f}'.format(round, mean_personalized_acc))
print('-'*80)