-
Notifications
You must be signed in to change notification settings - Fork 1
/
IID_LoAdaBoost_evaluation.py
219 lines (165 loc) · 8.99 KB
/
IID_LoAdaBoost_evaluation.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
import pandas as pd
import numpy as np
from keras.layers import Input, Dense, Dropout
from keras.models import Model
from sklearn.metrics import auc
from sklearn.metrics import roc_curve
from model_average import *
import matplotlib.pyplot as plt
from keras import initializers
def ann(X_train, Y_train, random_seed, batch_size_specified=100, dropout_rate=0.5):
#model
input_shape = X_train.shape[1]
# input layer
input_layer = Input(shape=(input_shape,))
# hidden layers
hidden_layer1 = Dense(20, activation='relu', kernel_initializer=initializers.glorot_uniform(seed=random_seed))(input_layer)
hidden_layer1 = Dropout(dropout_rate)(hidden_layer1)
hidden_layer2 = Dense(10, activation='relu')(hidden_layer1)
hidden_layer2 = Dropout(dropout_rate)(hidden_layer2)
hidden_layer3 = Dense(5, activation='relu')(hidden_layer2)
hidden_layer3 = Dropout(dropout_rate)(hidden_layer3)
# output layer
output_layer = Dense(1, activation='sigmoid')(hidden_layer3)
ann_model = Model(inputs=input_layer, outputs=output_layer)
ann_model.compile(optimizer='adam', loss='binary_crossentropy')
#ann_model.summary()
history = ann_model.fit(X_train, Y_train,
epochs=3,
batch_size=batch_size_specified,
shuffle=True,
verbose=False)
loss = history.history["loss"][-1]
return ann_model, loss
def ann2(X_train, Y_train, initializers, batch_size_specified=100, dropout_rate=0.5):
kernel_indices = [0, 2, 4, 6]
bias_indices = [1, 3, 5, 7]
kernel_initializers = np.array(initializers)[kernel_indices]
bias_initializers = np.array(initializers)[bias_indices]
#model
input_shape = X_train.shape[1]
# input layer
input_layer = Input(shape=(input_shape,))
# hidden layers
hidden_layer1 = Dense(20, activation='relu', weights=[kernel_initializers[0], bias_initializers[0]])(input_layer)
hidden_layer1 = Dropout(dropout_rate)(hidden_layer1)
hidden_layer2 = Dense(10, activation='relu', weights=[kernel_initializers[1], bias_initializers[1]])(hidden_layer1)
hidden_layer2 = Dropout(dropout_rate)(hidden_layer2)
hidden_layer3 = Dense(5, activation='relu', weights=[kernel_initializers[2], bias_initializers[2]])(hidden_layer2)
hidden_layer3 = Dropout(dropout_rate)(hidden_layer3)
# output layer
output_layer = Dense(1, activation='sigmoid', weights=[kernel_initializers[3], bias_initializers[3]])(hidden_layer3)
ann_model = Model(inputs=input_layer, outputs=output_layer)
ann_model.compile(optimizer='adam', loss='binary_crossentropy')
#ann_model.summary()
history = ann_model.fit(X_train, Y_train,
epochs=3,
batch_size=batch_size_specified,
shuffle=True,
verbose=False)
loss = history.history["loss"][-1]
return ann_model, loss
def calculate_auc(model, X_test, Y_test):
Y_pred = model.predict(X_test)
fpr, tpr, thresholds = roc_curve(Y_test, Y_pred, pos_label=1)
roc_auc = auc(fpr, tpr)
return roc_auc
def federated_learning(num_of_clients):
average_epoch_counts = []
average_training_aucs = []
end_of_loop_test_aucs = []
previous_median_loss = 1
# global loops
for t in range(30):
np.random.seed(t+1)
indices = np.random.choice(100, num_of_clients, replace=False)
#print indices
print "round " + str(t+1) +" start, random seed=" + str(t+1)
X_train_clients = X_train_100_shares[indices]
Y_train_clients = Y_train_100_shares[indices]
anns = []
roc_aucs = []
test_aucs = []
losses = []
for i in range(num_of_clients):
if t == 0:
ann_model, loss = ann(np.array(X_train_clients[i]), np.array(Y_train_clients[i]), random_seed=t+1,
batch_size_specified=30,
dropout_rate=0.0)
else:
ann_model, loss = ann2(np.array(X_train_clients[i]), np.array(Y_train_clients[i]),
initializers=weights, batch_size_specified=30, dropout_rate=0.0)
anns.append(ann_model)
# calculate auc for model trained with each client
roc_auc = calculate_auc(ann_model, np.array(X_train_clients[i]), np.array(Y_train_clients[i]))
roc_aucs.append(roc_auc)
#loss
losses.append(loss)
#test auc
test_auc = calculate_auc(ann_model, X_test, Y_test)
test_aucs.append(test_auc)
#print "round " + str(t+2) + " client " + str(i+1) + " loss=" + str(loss) +\
# " training auc=" + str(roc_auc) + " test auc=" + str(test_auc)
average_training_auc = np.average(roc_aucs)
print "round " + str(t+1) + " average training auc=" + str(average_training_auc)
average_training_aucs.append(average_training_auc)
print "round " + str(t+1) + " average training loss=" + str(np.average(losses))
# retrain
anns, epoch_counts = loss_based_retrain(anns, losses, num_of_clients, X_train_clients, Y_train_clients, previous_median_loss, 3)
average_epoch_count = np.average(epoch_counts)
average_epoch_counts.append(average_epoch_count)
print "round " + str(t+1) + " average epoch count=" + str(average_epoch_count)
anns[0].set_weights(average(anns))
end_of_loop_test_auc = calculate_auc(anns[0], X_test, Y_test)
print "round " + str(t+1) + " test auc=" + str(end_of_loop_test_auc)
end_of_loop_test_aucs.append(end_of_loop_test_auc)
weights = anns[0].get_weights()
previous_median_loss = np.percentile(losses, 50)
federated_ann = anns[0]
federated_ann.summary()
average_epoch_count = np.average(average_epoch_counts)
print "average epoch count=" + str(average_epoch_count)
return federated_ann, average_training_aucs, end_of_loop_test_aucs, average_epoch_counts
def prepare_data():
X_train = pd.read_csv("./IID_data/X_train.csv", dtype="int", header=None).values
Y_train = pd.read_csv("./IID_data/Y_train.csv", dtype="int", header=None).values
X_train_100_shares = np.array(np.array_split(X_train, 100))
Y_train_100_shares = np.array(np.array_split(Y_train, 100))
test_data_path = "./IID_data/X_test.csv"
X_test = pd.read_csv(test_data_path, dtype="int", header=None).values
Y_test = pd.read_csv("./IID_data/Y_test.csv", dtype="int", header=None).values
return X_train_100_shares, Y_train_100_shares, X_test, Y_test
def loss_based_retrain(anns, losses, num_of_clients, X_train_clients, Y_train_clients, previous_median_loss, starting_epochs):
epoch_counts = starting_epochs * np.ones(num_of_clients)
for i in range(num_of_clients):
if losses[i] > previous_median_loss:
retrain_count = 0
epoch_count = starting_epochs
original_loss = losses[i]
new_loss = original_loss
while new_loss > previous_median_loss:
retrain_count += 1
retrain_epochs = (starting_epochs - retrain_count+1) if starting_epochs > retrain_count else 1
if epoch_count >= 3*starting_epochs:
break
history = anns[i].fit(X_train_clients[i], Y_train_clients[i], epochs=retrain_epochs, batch_size=30, shuffle=True, verbose=False)
new_loss = history.history["loss"][-1]
epoch_count += retrain_epochs
epoch_counts[i] = epoch_count
print "model " + str(i+1) + " retrained, original loss=" + str(original_loss) + ", retrained loss=" + str(new_loss) + ", epoch count=" + str(epoch_counts[i])
return anns, epoch_counts
def federated_learning_evaluation(num_of_clients):
print "CLIENTS " + str(num_of_clients) + "% START"
federated_ann, average_training_aucs, end_of_loop_test_aucs, average_epoch_counts = federated_learning(num_of_clients)
#pd.DataFrame(average_training_aucs).to_csv("./IID_evaluation/"+ str(num_of_clients) + "adaboost_training_aucs.csv",
# header=False, index=False)
pd.DataFrame(end_of_loop_test_aucs).to_csv("./IID_evaluation/"+ str(num_of_clients) + "LoAdaBoost_test_aucs.csv",
header=False, index=False)
pd.DataFrame(average_epoch_counts).to_csv("./IID_evaluation/"+ str(num_of_clients) + "LoAdaBoost_epochs_per_client_per_communication_round.csv",
header=False, index=False)
print "CLIENTS " + str(num_of_clients) + "% END"
X_train_100_shares, Y_train_100_shares, X_test, Y_test = prepare_data()
federated_learning_evaluation(10)
#federated_learning_evaluation(20)
#federated_learning_evaluation(50)
#federated_learning_evaluation(90)