-
Notifications
You must be signed in to change notification settings - Fork 3
/
server.py
132 lines (116 loc) · 4.88 KB
/
server.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
import torch
from torch import nn, optim
import torch.nn.functional as F
import time
import copy
import numpy as np
import torch
from torch import nn, optim
import torch.nn.functional as F
import time
import copy
import numpy as np
from utils import init_dict, save_dict, curve_save, time_mark, print_cz, update_lr
def avg_freq(
weights,
L=0.1,
is_conv=True
):
client_num = len(weights)
if is_conv:
N, C, D1, D2 = weights[0].size()
else:
N = 1
C = 1
D1, D2 = weights[0].size()
#print(N, C, D1, D2)
temp_low = np.zeros((C*D1, D2*N), dtype=float)
for i in range(client_num):
# N, C, D1, D2 = weights[i].size()
#weights[i] = weights[i].cpu().numpy()
if is_conv:
weights[i] = weights[i].permute(1, 2, 3, 0).reshape((C*D1, D2*N))
weights[i] = weights[i].cpu().numpy()
client_fft = np.fft.fft2(weights[i], axes=(-2, -1))
amp_fft, pha_fft = np.abs(client_fft), np.angle(client_fft) # FFT
low_part = np.fft.fftshift(amp_fft, axes=(-2, -1))
temp_low += low_part
temp_low = temp_low / 4 # avg the low-frequency
for i in range(client_num):
client_fft = np.fft.fft2(weights[i], axes=(-2, -1))
amp_fft, pha_fft = np.abs(client_fft), np.angle(client_fft)
low_part = np.fft.fftshift(amp_fft, axes=(-2, -1))
h, w = low_part.shape
b_h = (np.floor(h *L / 2)).astype(int)
b_w = (np.floor(w *L / 2)).astype(int)
c_h = np.floor(h/2.0).astype(int)
c_w = np.floor(w/2.0).astype(int)
h1 = c_h-b_h
h2 = c_h+b_h
w1 = c_w-b_w
w2 = c_w+b_w
low_part[h1:h2,w1:w2] = temp_low[h1:h2,w1:w2] # averaged low-freq + individual high-freq
low_part = np.fft.ifftshift(low_part, axes=(-2, -1))
fft_back_ = low_part * np.exp(1j * pha_fft) #
# get the mutated image
fft_back_ = np.fft.ifft2(fft_back_, axes=(-2, -1))
weights[i] = torch.FloatTensor(np.real(fft_back_))
if is_conv:
weights[i] = weights[i].reshape(C, D1, D2, N).permute(3, 0, 1, 2)
return weights
def PFA(
weights,
L,
is_conv
):
return avg_freq(weights=weights, L=L, is_conv=is_conv)
################# Key Function ########################
def communication(
args,
server_model,
models,
original_models,
client_weights,
a_iter
):
pfa_rate = args.l_rate + (a_iter / args.iters) * (0.95 - args.l_rate)
client_num = len(client_weights) #
with torch.no_grad():
# aggregate params
for key in server_model.state_dict().keys():
if 'bn' not in key: #not bn
if 'conv' in key and 'weight' in key:
temp_weights = PFA(
[
models[0].state_dict()[key].data,
models[1].state_dict()[key].data,
models[2].state_dict()[key].data,
models[3].state_dict()[key].data
],
L=pfa_rate,
is_conv=True
)
for client_idx in range(client_num): # copy from server to each client
models[client_idx].state_dict()[key].data.copy_(temp_weights[client_idx])
elif 'linear' in key and 'weight' in key:
temp_weights = PFA(
[
models[0].state_dict()[key].data,
models[1].state_dict()[key].data,
models[2].state_dict()[key].data,
models[3].state_dict()[key].data
],
L=pfa_rate,
is_conv=False
)
for client_idx in range(client_num): #
models[client_idx].state_dict()[key].data.copy_(temp_weights[client_idx])
else:
print(key, '\t not bn, conv, fc layer, with param!')
temp = torch.zeros_like(server_model.state_dict()[key], dtype=torch.float32)
for client_idx in range(client_num):
temp += client_weights[client_idx] * models[client_idx].state_dict()[key]
server_model.state_dict()[key].data.copy_(temp) # non-bn layer，update the server model
for client_idx in range(client_num): # non-bn layer, from server to each client
models[client_idx].state_dict()[key].data.copy_(server_model.state_dict()[key])
return server_model, models