-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathNetwork.py
208 lines (176 loc) · 8.41 KB
/
Network.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
from keras.datasets import mnist
from keras.layers import Input, Conv2D, PReLU, MaxPool2D, Dense, Flatten, Embedding, BatchNormalization
from keras.engine.topology import Layer
from keras.layers.merge import concatenate
from keras.models import Model
from keras.optimizers import SGD
from keras import backend as K
from keras.regularizers import l2
from keras import losses
import numpy as np
from keras.utils import to_categorical
from tqdm import tqdm
from keras import initializers
from MyCallback import visualize
def prelu(x, name='default'):
if name == 'default':
return PReLU(alpha_initializer=initializers.Constant(value=0.25))(x)
else:
return PReLU(alpha_initializer=initializers.Constant(value=0.25), name=name)(x)
class Network():
def __init__(self, alpha_center=0.5,
lambda_centerloss=0.1):
self.alpha_center = alpha_center
self.lambda_centerloss = lambda_centerloss
self.pool_name = "pool_"
self.conv_name = "conv_"
self.model = None
def _conv_block(self, input, out_dim, kernel, counter, weight_decay):
x = Conv2D(out_dim, (kernel, kernel), name=self.conv_name + str(counter),
kernel_regularizer=l2(weight_decay), padding='same')(input)
out = prelu(x)
return out
class CenterLayer(Layer):
def __init__(self, num_classes, feature_dim, alpha_center, **kwargs):
super().__init__(**kwargs)
self.alpha_center = alpha_center
self.num_classes = num_classes
self.feature_dim = feature_dim
def build(self, input_shape):
# Create a trainable weight variable for this layer
self.centers = self.add_weight(name='centers',
shape=(self.num_classes, self.feature_dim),
initializer='uniform',
trainable=False)
super().build(input_shape) # Be sure to call this somewhere!
def call(self, x, mask=None):
## x[0] is Nx2, x[1] is Nx10 onehot, self.centers is 10x2
delta_centers = K.dot(K.transpose(x[1]), (K.dot(x[1], self.centers) - x[0])) # 10x2
denominator = K.sum(K.transpose(x[1]), axis=1, keepdims=True) + 1
delta_centers /= denominator
new_centers = self.centers - self.alpha_center * delta_centers
self.add_update((self.centers, new_centers), x)
self.result = (K.dot(x[1], self.centers) - x[0])
self.result = K.sum(self.result ** 2, axis=1, keepdims=True)
return self.result
def prepare_data(self):
(x_train, y_train), (x_test, y_test) = mnist.load_data()
print('x_train shape:', x_train.shape)
print('x_test shape:', x_test.shape)
x_train = x_train.reshape((-1, 28, 28, 1))
x_test = x_test.reshape((-1, 28, 28, 1))
y_train_onehot = to_categorical(y_train, 10)
y_test_onehot = to_categorical(y_test, 10)
return x_train, y_train_onehot, x_test, y_test_onehot, y_train, y_test
def _build_model(self, im_size=28, hidden_dim=128, kernel = 3, weight_decay=0.005,
num_classes=10, feature_dim=2, is_concated = True):
input = Input((im_size, im_size, 1))
labels = Input((num_classes,))
x = BatchNormalization()(input)
x = self._conv_block(input=x, out_dim=hidden_dim, kernel=3, counter=0, weight_decay=weight_decay)
x = self._conv_block(input=x, out_dim=hidden_dim, kernel=3, counter=1, weight_decay=weight_decay)
x = MaxPool2D(pool_size=(2, 2), strides=(2, 2), padding='valid')(x)
x = self._conv_block(input=x, out_dim=hidden_dim, kernel=3, counter=2, weight_decay=weight_decay)
x = MaxPool2D(pool_size=(2, 2), strides=(2, 2), padding='valid')(x)
x = self._conv_block(input=x, out_dim=hidden_dim * 2, kernel=4, counter=3, weight_decay=weight_decay)
x = MaxPool2D(pool_size=(2, 2), strides=(2, 2), padding='valid')(x)
x = self._conv_block(input=x, out_dim=hidden_dim * 2, kernel=2, counter=4, weight_decay=weight_decay)
x1 = MaxPool2D(pool_size=(2, 2), strides=(2, 2), padding='valid')(x)
x2 = self._conv_block(input=x1, out_dim=hidden_dim * 2, kernel=1, counter=5, weight_decay=weight_decay)
if is_concated:
x_concated = concatenate([x1, x2], axis=3)
x_flatten = Flatten()(x_concated)
else:
x_flatten = Flatten()(x2)
x = Dense(units=feature_dim, kernel_regularizer=l2(weight_decay))(x_flatten)
x = PReLU(name="deep_deatures")(x)
y_out = Dense(num_classes, activation="softmax", kernel_regularizer=l2(weight_decay))(x)
y_side = self.CenterLayer(num_classes=num_classes,
alpha_center=self.alpha_center,
feature_dim=feature_dim, name="centerlosslayer")([x, labels])
model = Model(inputs=[input, labels], outputs=[y_out, y_side])
model.summary()
self.model = model
def my_model(self, im_size = 28, num_classes = 10, weight_decay = 0.005):
input = Input((im_size, im_size, 1))
labels = Input((num_classes,))
x = BatchNormalization()(input)
#
x = Conv2D(filters=32, kernel_size=(5, 5), strides=(1, 1), padding='same',
kernel_regularizer=l2(weight_decay))(x)
x = prelu(x)
x = Conv2D(filters=32, kernel_size=(5, 5), strides=(1, 1), padding='same',
kernel_regularizer=l2(weight_decay))(x)
x = prelu(x)
x = MaxPool2D(pool_size=(2, 2), strides=(2, 2), padding='valid')(x)
#
x = Conv2D(filters=64, kernel_size=(5, 5), strides=(1, 1), padding='same',
kernel_regularizer=l2(weight_decay))(x)
x = prelu(x)
x = Conv2D(filters=64, kernel_size=(5, 5), strides=(1, 1), padding='same',
kernel_regularizer=l2(weight_decay))(x)
x = prelu(x)
x = MaxPool2D(pool_size=(2, 2), strides=(2, 2), padding='valid')(x)
#
x = Conv2D(filters=128, kernel_size=(5, 5), strides=(1, 1), padding='same',
kernel_regularizer=l2(weight_decay))(x)
x = prelu(x)
x = Conv2D(filters=128, kernel_size=(5, 5), strides=(1, 1), padding='same',
kernel_regularizer=l2(weight_decay))(x)
x = prelu(x)
x = MaxPool2D(pool_size=(2, 2), strides=(2, 2), padding='valid')(x)
#
x = Flatten()(x)
x = Dense(2, kernel_regularizer=l2(weight_decay))(x)
x = prelu(x, name='deep_deatures')
#
main = Dense(10, activation='softmax', name='main_out', kernel_regularizer=l2(weight_decay))(x)
side = self.CenterLayer(num_classes=num_classes, feature_dim=2, alpha_center=0.5,
name='centerlosslayer')([x, labels])
model = Model(inputs=[input, labels], outputs=[main, side])
model.summary()
self.model = model
def center_loss(self, y_true, y_pred):
return 0.5 * K.sum(y_pred, axis=0)
def _fit(self, x_train, y_train, x_test, y_test,
learning_rate=0.01, momentum=0.9, epochs=50,
batch_size=64, train_percentage=0.1):
if self.model is None:
print("first you need to init model")
self._build_model()
dummy1 = np.zeros((x_train.shape[0], 1))
dummy2 = np.zeros((x_test.shape[0], 1))
optimizer = SGD(lr=learning_rate, momentum=momentum)
self.model.compile(optimizer=optimizer,
loss=[losses.categorical_crossentropy, self.center_loss],
loss_weights=[1, self.lambda_centerloss])
N = x_train.shape[0]
n = int(train_percentage * N)
self.model.fit([x_train[:n], y_train[:n]], [y_train[:n], dummy1[:n]],
batch_size=batch_size,
epochs=epochs,
verbose=2,
validation_data=([x_test[:n], y_test[:n]], [y_test[:n], dummy2[:n]]))
def _plot_results(self, x_train, y_train, x_test, y_test, epochs=50, train_percentage=0.1):
N = x_train.shape[0]
n = int(train_percentage * N)
reduced_model = Model(inputs=self.model.input[0], outputs=self.model.get_layer('deep_deatures').output)
feats = reduced_model.predict(x_train[:n])
visualize(feats[:n], y_train[:n], epoch=epochs - 1,
centers=self.model.get_layer('centerlosslayer').get_weights()[0],
lambda_cl=self.lambda_centerloss, is_train=True)
feats = reduced_model.predict(x_test[:n])
visualize(feats[:n], y_test[:n], epoch=epochs - 1,
centers=self.model.get_layer('centerlosslayer').get_weights()[0],
lambda_cl=self.lambda_centerloss, is_train=False)
def test():
network = Network(alpha_center=0.5, lambda_centerloss=0.1)
network._build_model(is_concated=True)
#network.my_model()
x_train, y_train_onehot, x_test, y_test_onehot, y_train, y_test = network.prepare_data()
for i in tqdm(range(1, 11)):
network._fit(x_train, y_train_onehot, x_test, y_test_onehot, train_percentage=0.3,
epochs=10, learning_rate=0.001, momentum=0.9, batch_size=64)
network._plot_results(x_train, y_train, x_test, y_test, train_percentage=0.3, epochs=10 * i + 1)
if __name__ == "__main__":
test()