-
Notifications
You must be signed in to change notification settings - Fork 0
/
AiFedIE_noniid_cifar10.py
277 lines (256 loc) · 12.4 KB
/
AiFedIE_noniid_cifar10.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
# -*- coding: utf-8 -*-
"""
Created on Tue May 25 19:15:35 2021
@author: liush
"""
import numpy as np
import tensorflow.compat.v1.keras
from tensorflow.compat.v1.keras.datasets import cifar10
from tensorflow.compat.v1.keras import backend as K
from tensorflow.keras.models import load_model
from information_function import compute_label_number
import random
import os
#import matplotlib.pyplot as plt
from function import plot_result
import tensorflow
import pandas as pd
import time
import math
#from scipy.spatial.distance import pdist
#os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
#os.environ["CUDA_VISIBLE_DEVICES"] = "0" # 使用第一, 三块GPU
#tensorflow.compat.v1.GPUOptions(per_process_gpu_memory_fraction=0.99)
#gpu_options = tensorflow.compat.v1.GPUOptions(allow_growth=True)
#sess = tensorflow.compat.v1.Session(config=tensorflow.compat.v1.ConfigProto(gpu_options=gpu_options))
#tensorflow.compat.v1.disable_eager_execution()
img_rows, img_cols = 32, 32
batch_size = 48
num_classes = 10
epochs = 5
rounds = 400 #communication round
time_range_1, time_range_2 = [3,30]
a = 1.36
k, c = 40, 0.2 #total number of clients, fraction
m = int(k*c)
(X_train, Y_train), (x_test, y_test) = cifar10.load_data()
x_train_all, y_train_all, x_test_all, y_test_all = [], [], [], []
type_dateset = 'noniid'
name_dataset = 'cifar10'
save_path = 'results_'+type_dateset+'_'+name_dataset+'_save/AiFedLN_0602_result1.xlsx'
path = 'data_'+name_dataset+'_'+type_dateset+'/'
model = load_model('w0_'+name_dataset+'.h5')#initialized model w0
#save_path = 'results_fmnist_save/ALU_0308_result1_direct.xlsx'
#model = load_model('w0_fmnist.h5')#initialized model w0
#path = "data_fmnist_noniid/"
#------------------------------------------------------------------------------local test
#name_x_test_local = 'data_test_local/'+name_dataset+'_x_test_local.npy'
#name_y_test_local = 'data_test_local/'+name_dataset+'_y_test_local.npy'
#x_test_local = np.load(name_x_test_local)
#y_test_local = np.load(name_y_test_local)
#y_test_local = tensorflow.keras.utils.to_categorical(y_test_local, num_classes)
#local_model_acc_all = np.zeros(shape = (k,rounds))
#delta_local_model_acc_all = np.zeros(shape = (k,rounds))
#local_model_acc_m = np.zeros(shape = (m,rounds))
#delta_local_model_acc_m = np.zeros(shape = (m,rounds))
#------------------------------------------------------------------------------consistency
from compute_RDM import compute_rda
from compute_RDM import compute_rc_simp
name_x_stimulus = 'data_stimulus/' + name_dataset +'_x_stimulus.npy'
stimulus = np.load(name_x_stimulus)
#distance_measure = 'correlation', 'cosine', 'euclidean'
distance_measure = 'cosine'
layer_idx_all = [6]
#[1,2,6,7]#related to model structure, model.layers in compute_rdm function
#1,2-->cnn layers; 3-->max pooling layer; 4--> dropout layer; 5--> flatten; 6,7--> FC layers
global_rdm_allayer =[]
for layer_idx in layer_idx_all:
global_rdm_onelayer = compute_rda(model, stimulus, distance_measure, layer_idx-1)#initialize global_rdm
global_rdm_allayer.append(global_rdm_onelayer)
#global_rdm_allayer = np.array(global_rdm_allayer)
flag_upload_deeplayer = np.zeros(shape = (k,rounds))
update_prob = np.ones(shape = (m,1))/m
for i in range(m):
update_prob[i] *= (i+1)
#representational_consistency_all = np.zeros(shape = (k, len(layer_idx_all), rounds))
#representational_consistency_all = np.zeros(shape = (k, rounds))
#delta_representational_consistency_all = np.zeros(shape = (k, rounds))
representational_consistency_m = np.zeros(shape = (m, rounds))
#delta_representational_consistency_m = np.zeros(shape = (m, rounds))
#save_path_rc_all = 'results_'+type_dateset+'_'+name_dataset+'_save/ALMU_0601_result1_rc_all'#-----------
save_path_rc_m = 'results_'+type_dateset+'_'+name_dataset+'_save/AiFedLN_0602_result1_rc_m'#-----------
#####################################################################
file_list = []
file_list = os.listdir(path)
for s in file_list:
#print(s)
b = s[10] #client_001_x_train.npy, client_001_x_test.npy, etc
path_s = path + s
if ((b == 'x') & (len(s)==21)):
x_train_all.append(np.load(path_s))
elif ((b=='y') & (len(s)==21)):
y_train_all.append(np.load(path_s))
elif ((b=='x') & (len(s)==20)):
x_test_all.append(np.load(path_s))
elif ((b=='y') & (len(s)==20)):
y_test_all.append(np.load(path_s))
start=time.time()
iw_all = np.zeros(shape = k)#informative weights
s0_all = []
twf_all = np.zeros(shape = k)
timestamp_all = np.zeros(shape = k)
clients_index = []
for i in range(0,k):
clients_index.append(i)
img_rows, img_cols = x_train_all[0].shape[1], x_train_all[0].shape[2]
num_channel = x_train_all[0].shape[3]
if K.image_data_format() == 'channels_first':
X_train = X_train.reshape(X_train.shape[0], num_channel, img_rows, img_cols)
x_test = x_test.reshape(x_test.shape[0], num_channel, img_rows, img_cols)
input_shape = (num_channel, img_rows, img_cols)
else:
X_train =X_train.reshape(X_train.shape[0], img_rows, img_cols, num_channel)
x_test = x_test.reshape(x_test.shape[0], img_rows, img_cols, num_channel)
input_shape = (img_rows, img_cols, num_channel)
y_test = tensorflow.keras.utils.to_categorical(y_test, num_classes)
global_model_weights = model.get_weights()
weights = []#weights of each client
weights_glo = []
length_all = np.zeros(shape = k)#total size of data
time_cost_all=np.zeros(shape=k)
for i in range(0,k):
x_train = x_train_all[i]
y_train = y_train_all[i]
length_all[i] = len(x_train)
iw_all[i] = compute_label_number(y_train)#informative weights
weights.append(np.array(model.get_weights())) #initialize
weights_glo.append(np.array(model.get_weights()))
time_cost_all[i] = random.uniform(time_range_1, time_range_2)#--------------------------------
size_all = length_all/sum(length_all)
info_all = iw_all/sum(iw_all)
set_es = [0]
global_model_weights = np.array(model.get_weights())
global_model_test_loss = []
global_model_test_acc = []
#layer_dist_euclidean_all = np.zeros(shape = (8,rounds))
#layer_dist_cosine_all = np.zeros(shape = (8,rounds))
cost_unit_model = 39.69#-------------------------------------------------------MB
cost_unit_shallow = 3.68
accum_time = []#accumulated time of each round
accum_cost = []#---------------------------------------------------------------accum_cost
sum_wait_time = 0
sum_cost = 0
for r in range(0,rounds):
s0 = random.sample(clients_index, int(m)) #clients of rounds r
s0_all.append(s0)
time_thisround = []
#length_sum = 0
for i in range(0,int(m)):
time_thisround.append(time_cost_all[s0[i]])
time_cost_all[s0[i]] = random.uniform(time_range_1, time_range_2)#new cost time for finished clients
timestamp_all[s0[i]] = r+1#update timestamp
x_train = x_train_all[s0[i]]
y_train = y_train_all[s0[i]]
y_train = tensorflow.keras.utils.to_categorical(y_train, num_classes)
model.set_weights(global_model_weights) #current local model
history = model.fit(x_train, y_train,
batch_size=batch_size,
epochs=epochs,
verbose=0)
weights[s0[i]] = np.array(model.get_weights()) #local model weights update
#score_local = model.evaluate(x_test_local,y_test_local, verbose=0)
#local_model_acc_all[s0[i]][r] = score_local[1]
#local_model_acc_m[i][r] = score_local[1]
#if r>=1:
#delta_local_model_acc_all[s0[i]][r] = local_model_acc_all[s0[i]][r]-global_model_test_acc[r-1]
#delta_local_model_acc_m[i][r] = local_model_acc_m[s0[i]][r]-global_model_test_acc[r-1]
for i in range(k):
twf_all[i] = math.pow(a,-r+timestamp_all[i]-1)
alpha_all = list(twf_all)/sum(twf_all)
coe_all = alpha_all*size_all*info_all
coe_norm = coe_all/sum(coe_all)
wait_time = max(time_thisround)
sum_wait_time += wait_time
accum_time.append(sum_wait_time)
#model.summary() #model structure
#weights = np.array(model.get_weights())
#layer_weights_new, layer_weights_old = [], []
#layer_dist_euclidean = np.zeros(shape = 8)
#layer_dist_cosine = np.zeros(shape = 8)
#----------------------------------------------------------------------compute rdms
rc_this_round = []
for i in range(m):
num_layer = 0
model.set_weights(weights[s0[i]])
for layer_idx in layer_idx_all:
local_rdm_onelayer = compute_rda(model, stimulus, distance_measure, layer_idx-1)
Pearson_coefficient_onelayer = compute_rc_simp(
local_rdm_onelayer, global_rdm_allayer[num_layer])
representational_consistency_onelayer = Pearson_coefficient_onelayer[0]*\
Pearson_coefficient_onelayer[0]
#representational_consistency_all[i][num_layer][r]= representational_consistency_onelayer
representational_consistency_m[i][r]= representational_consistency_onelayer
rc_this_round.append(representational_consistency_onelayer)
#if r>=1:
#delta_representational_consistency_all[i][r]=representational_consistency_all[i][r]\
#-representational_consistency_all[i][r-1]
#num_layer+=1
#Pearson_coefficient_layer = compute_representational_consistency(local_rdm[j],global_rdm[j])
#representational_consistency_layer = Pearson_coefficient_layer[0]*Pearson_coefficient_layer[0]
#representational_consistency_all[i][j][r]= representational_consistency_layer
#layer_weights_new.append(model.get_weights()[j])#layers of the trained model
#layer_weights_old.append(global_model_weights[r][j])#layers of the current global model
#layer_weights_new, layer_weights_old = np.array(layer_weights_new), np.array(layer_weights_old)
rc_this_round = np.array(rc_this_round)
idx_rc_m = rc_this_round.argsort()#sorted idx, min first
#idx_rc_m = rc_this_round.argsort()[::-1]#sorted idx, max first
#sum_upload_deeplayer = 0
for idx in idx_rc_m:
AT = random.random()#[0,1]
if AT >= update_prob[idx]:#upload
flag_upload_deeplayer[s0[idx]][r] = 1
#sum_upload_deeplayer += 1
sum_cost += cost_unit_model
else:#not upload
weights[s0[idx]][4] = global_model_weights[4]
sum_cost += cost_unit_shallow
accum_cost.append(sum_cost)
weights_new = coe_norm[0]*weights[0]
for i in range(1,k):
weights_new += coe_norm[i]*weights[i] # aggregate
model.set_weights(weights_new) # global model update
global_model_weights = model.get_weights()
for i in range(0,m):
weights_glo[s0[i]] = model.get_weights()
num_layer = 0
for layer_idx in layer_idx_all:
global_rdm_onelayer = compute_rda(model, stimulus, distance_measure, layer_idx-1)#initialize global_rdm
global_rdm_allayer[num_layer] = (global_rdm_onelayer)
num_layer+=1
#global_rdm_allayer = np.array(global_rdm_allayer)
score = model.evaluate(x_test, y_test, verbose=0)
global_model_test_loss.append(score[0])
global_model_test_acc.append(score[1])
if (r+1)<=40 or (r+1)%50 ==0:
print ("round %d:"%(r+1),end = '\n')
print('Global Model Test loss:', score[0])
print('Global Model Test accuracy:', score[1])
print('\n')
plot_result(global_model_test_acc,'AiFedLN','accuracy')
save_name = list(zip(global_model_test_acc, global_model_test_loss,accum_time,accum_cost))
dataframe = pd.DataFrame(save_name, columns=['accuracy', 'loss','accum_time','accum_cost'])
dataframe.to_excel(save_path, index=False)
#------------------------------------------------------------------------------aver rc
#for i in range()
#------------------------------------------------------------------------------save rc
#np.save(save_path_rc_all, representational_consistency_all)
np.save(save_path_rc_m, representational_consistency_m)
'''
save_name_flag = list(flag_update_deeplayer)
dataframe_flag = pd.DataFrame(save_name_flag)
dataframe_flag.to_excel(save_path_flag, index=False)
'''
##############################################################################
end=time.time()
run_time = end-start
print('Running time: %d Seconds'%(run_time))