-
Notifications
You must be signed in to change notification settings - Fork 1
/
cnn_cifar10.py
222 lines (201 loc) · 9.55 KB
/
cnn_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
# -*- coding: utf-8 -*-
"""
==============================================================================
Created on Dec 11 2019
Author: Yaxiong Yuan
==============================================================================
"""
import os
import pickle
import numpy as np
import matplotlib.pyplot as plt
from callbacks_custom import callbacks_custom_CNN
from binary_tanh_custom.binary_ops import binary_tanh
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Dense, Flatten, Input, Activation, BatchNormalization, Conv2D, MaxPooling2D, Dropout
from tensorflow.keras.utils import to_categorical
from tensorflow.keras.datasets import cifar10
from tensorflow.keras import regularizers, optimizers
from tensorflow.keras.preprocessing.image import ImageDataGenerator
from optimizers_custom.ProxSGD_optimizer import ProxSGD
from optimizers_custom.ProxSGD_BNN_optimizer import ProxSGD_BNN
from optimizers_custom.ADABound_optimizer import ADABound
from optimizers_custom.AMSGrad_optimizer import AMSGrad
def load_data():
(x_train, y_train), (x_test, y_test) = cifar10.load_data()
x_train = x_train.astype('float32')
x_test = x_test.astype('float32')
mean = np.mean(x_train, axis=(0, 1, 2, 3))
std = np.std(x_train, axis=(0, 1, 2, 3))
x_train = (x_train - mean) / (std + 1e-7)
x_test = (x_test - mean) / (std + 1e-7)
num_classes = 10
y_train = to_categorical(y_train, num_classes)
y_test = to_categorical(y_test, num_classes)
return x_train, y_train, x_test, y_test
def create_model(active_bnn=0):
num_classes = 10
model = Sequential()
if active_bnn == 1:
model.add(Conv2D(32, (3, 3), padding='same', input_shape=(32, 32, 3), name='conv1'))
model.add(BatchNormalization())
model.add(Activation(binary_tanh))
model.add(Conv2D(32, (3, 3), padding='same', name='conv2'))
model.add(BatchNormalization())
model.add(Activation(binary_tanh))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Dropout(0.2))
model.add(Conv2D(64, (3, 3), padding='same', name='conv3'))
model.add(BatchNormalization())
model.add(Activation(binary_tanh))
model.add(Conv2D(64, (3, 3), padding='same', name='conv4'))
model.add(BatchNormalization())
model.add(Activation(binary_tanh))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Dropout(0.2))
model.add(Conv2D(128, (3, 3), padding='same', name='conv5'))
model.add(BatchNormalization())
model.add(Activation(binary_tanh))
model.add(Conv2D(128, (3, 3), padding='same', name='conv6'))
model.add(BatchNormalization())
model.add(Activation(binary_tanh))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Dropout(0.2))
model.add(Flatten())
model.add(Dense(num_classes, activation='softmax', name='fc1'))
else:
model.add(Conv2D(32, (3, 3), padding='same', input_shape=(32, 32, 3),
kernel_regularizer=regularizers.l1(1e-4),
name='conv1'))
model.add(BatchNormalization())
model.add(Activation('elu'))
model.add(Conv2D(32, (3, 3), padding='same',
kernel_regularizer=regularizers.l1(1e-4),
name='conv2'))
model.add(BatchNormalization())
model.add(Activation('elu'))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Dropout(0.2))
model.add(Conv2D(64, (3, 3), padding='same',
kernel_regularizer=regularizers.l1(1e-4),
name='conv3'))
model.add(BatchNormalization())
model.add(Activation('elu'))
model.add(Conv2D(64, (3, 3), padding='same',
kernel_regularizer=regularizers.l1(1e-4),
name='conv4'))
model.add(BatchNormalization())
model.add(Activation('elu'))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Dropout(0.2))
model.add(Conv2D(128, (3, 3), padding='same',
kernel_regularizer=regularizers.l1(1e-4),
name='conv5'))
model.add(BatchNormalization())
model.add(Activation('elu'))
model.add(Conv2D(128, (3, 3), padding='same',
kernel_regularizer=regularizers.l1(1e-4),
name='conv6'))
model.add(BatchNormalization())
model.add(Activation('elu'))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Dropout(0.2))
model.add(Flatten())
model.add(Dense(num_classes, activation='softmax',
kernel_regularizer=regularizers.l1(1e-4),
name='fc1'))
return model
def train_model(active_bnn=0, optimtype='proxsgd', verbose=True, Epoch=5, T=1):
x_train, y_train, x_test, y_test = load_data()
# data augmentation
datagen = ImageDataGenerator(rotation_range=15, width_shift_range=0.1, height_shift_range=0.1, horizontal_flip=True)
datagen.fit(x_train)
if optimtype == 'proxsgd':
optim = ProxSGD(epsilon_initial=0.06, epsilon_decay=0.5, rho_initial=0.9, rho_decay=0.5, beta=0.999, mu=5e-5,
clip_bounds=None)
model = create_model(active_bnn)
elif optimtype == 'proxsgd_bnn':
optim = ProxSGD_BNN(epsilon_initial=0.06, epsilon_decay=0.5, rho_initial=0.9, rho_decay=0.5, beta=0.999, mu_x=None,
clip_bounds_x=[-1.0, 1.0], mu_a=None, clip_bounds_a=[0.0, 1.0])
model = create_model(active_bnn)
elif optimtype == 'ams':
optim = AMSGrad(lr=0.001, beta_1=0.9, beta_2=0.999, epsilon=None, decay=0.0, amsgrad=True)
model = create_model(active_bnn)
elif optimtype == 'adabound':
optim = ADABound(lr=0.001, beta_1=0.9, beta_2=0.999, epsilon=None, decay=0.0)
model = create_model(active_bnn)
else:
optim = optimizers.Adam(lr=0.001, beta_1=0.9, beta_2=0.999, epsilon=None, decay=0.0, amsgrad=False,)
model = create_model(active_bnn)
model.summary()
model.compile(loss='categorical_crossentropy', optimizer=optim, metrics=['accuracy'])
if verbose:
cb = [callbacks_custom_CNN.metrics(validation_data=(x_test, y_test), training_data=(x_train, y_train), Times=T)]
else:
cb = []
batch_size = 64
model.fit_generator(datagen.flow(x_train, y_train, batch_size=batch_size), \
steps_per_epoch=x_train.shape[0] // batch_size, epochs=Epoch, \
callbacks=cb, verbose=verbose, validation_data=(x_test, y_test))
model.evaluate(x_test, y_test, verbose=0)
if __name__ == "__main__":
Round = 1
Epoch = 125
Loss_matrix = np.zeros([Round, Epoch])
Accuracy_matrix = np.zeros([Round, Epoch])
NormL1_matrix = np.zeros([Round, Epoch])
Accuracy_train_matrix = np.zeros([Round, Epoch])
Loss_train_matrix = np.zeros([Round, Epoch])
for t in range(Round):
train_model(active_bnn=0, optimtype='proxsgd', verbose=1, Epoch=Epoch, T=t)
Loss_proxsgd = pickle.load(open('./variables_cnn/Loss_proxsgd', 'rb'))
Loss_matrix[t, :] = Loss_proxsgd
Accuracy_proxsgd = pickle.load(open('./variables_cnn/Accuracy_proxsgd', 'rb'))
Accuracy_matrix[t, :] = Accuracy_proxsgd
Loss_proxsgd_train = pickle.load(open('./variables_cnn/Loss_proxsgd_train', 'rb'))
Loss_train_matrix[t, :] = Loss_proxsgd_train
Accuracy_proxsgd_train = pickle.load(open('./variables_cnn/Accuracy_proxsgd_train', 'rb'))
Accuracy_train_matrix[t, :] = Accuracy_proxsgd_train
NormL1_proxsgd = pickle.load(open('./variables_cnn/NormL1_proxsgd', 'rb'))
NormL1_matrix[t, :] = NormL1_proxsgd
pickle.dump(Loss_matrix, open('./variables_cnn/Loss_table', 'wb'))
pickle.dump(Accuracy_matrix, open('./variables_cnn/Accuracy_table', 'wb'))
pickle.dump(Loss_train_matrix, open('./variables_cnn/Loss_train_table', 'wb'))
pickle.dump(Accuracy_train_matrix, open('./variables_cnn/Accuracy_train_table', 'wb'))
pickle.dump(NormL1_matrix, open('./variables_cnn/NormL1_table', 'wb'))
Weights_values = pickle.load(open('./variables_cnn/Weights_proxsgd', 'rb'))
data = Weights_values
data_flattened = np.concatenate((data[0].flatten(), data[1].flatten(), data[2].flatten(),
data[3].flatten(), data[4].flatten(), data[5].flatten()), axis=None)
if os.path.isdir('./figures'):
pass
else:
os.mkdir('./figures')
Loss_values = pickle.load(open('./variables_cnn/Loss_train_table', 'rb'))
Mean_loss_values = np.mean(Loss_values, axis=0)
ep = np.arange(Epoch)
plt.figure(1)
plt.plot(ep, Mean_loss_values)
plt.grid(True)
plt.savefig('./figures/cifar10_loss.png')
plt.show()
Accuracy_values = pickle.load(open('./variables_cnn/Accuracy_table', 'rb'))
Mean_accuracy_values = np.mean(Accuracy_values, axis=0)
ep = np.arange(Epoch)
plt.figure(2)
plt.plot(ep, Mean_accuracy_values)
plt.grid(True)
plt.savefig('./figures/cifar10_acc.png')
plt.show()
Weights_values = pickle.load(open('./variables_cnn/Weights_proxsgd', 'rb'))
data = Weights_values
data_flattened = np.concatenate((data[0].flatten(), data[1].flatten(), data[2].flatten(),
data[3].flatten(), data[4].flatten(), data[5].flatten(), data[6].flatten()), axis=None)
x = np.sort(data_flattened)
y = np.arange(1, len(x)+1)/len(x)
plt.figure(3)
plt.plot(x, y)
plt.xlim((-3, 3))
plt.grid(True)
plt.savefig('./figures/cifar10_cdf.png')
plt.show()