def _construct_siamese_architecture(self, learning_rate_multipliers,
l2_regularization_penalization):
""" Constructs the siamese architecture and stores it in the class
Arguments:
learning_rate_multipliers
l2_regularization_penalization
"""
# Let's define the cnn architecture
convolutional_net = Sequential()
convolutional_net.add(
Dense(units=128, activation='sigmoid',
kernel_regularizer=l2(
l2_regularization_penalization['Dense1']),
name='Dense1'))
# Now the pairs of images
input_image_1 = Input(self.input_shape)
input_image_2 = Input(self.input_shape)
encoded_image_1 = convolutional_net(input_image_1)
encoded_image_2 = convolutional_net(input_image_2)
# L1 distance layer between the two encoded outputs
# One could use Subtract from Keras, but we want the absolute value
l1_distance_layer = Lambda(
lambda tensors: K.abs(tensors[0] - tensors[1]))
l1_distance = l1_distance_layer([encoded_image_1, encoded_image_2])
# Same class or not prediction
prediction = Dense(units=1, activation='sigmoid')(l1_distance)
self.model = Model(
inputs=[input_image_1, input_image_2], outputs=prediction)
# Define the optimizer and compile the model
optimizer = Modified_SGD(
lr=self.learning_rate,
lr_multipliers=learning_rate_multipliers,
momentum=0.5)
self.model.compile(loss='binary_crossentropy', metrics=['binary_accuracy'],
optimizer=optimizer)
def _construct_siamese_architecture(self, learning_rate_multipliers,
l2_regularization_penalization):
""" Constructs the siamese architecture and stores it in the class
Arguments:
learning_rate_multipliers
l2_regularization_penalization
"""
# Let's define the cnn architecture
convolutional_net = Sequential()
'''
#struct 1：
#简单定义孪生网络的主体结构为MLP，属于一个全连接神经网络
convolutional_net.add(Dense(128, activation=tf.nn.relu, input_shape=self.input_shape))
convolutional_net.add(Dense(64, activation=tf.nn.relu, input_shape=self.input_shape))
'''
K.set_image_dim_ordering('tf')
#struct 2：
#孪生网络的主体结构为CNN，属于一个卷积神经网络
convolutional_net.add(
Conv2D(filters=64,
kernel_size=(3, 3),
activation='relu',
input_shape=self.input_shape,
kernel_regularizer=l2(
l2_regularization_penalization['Conv1']),
name='Conv1'))
convolutional_net.add(MaxPool2D(pool_size=1))
# convolutional_net.add(Conv2D(filters=128, kernel_size=(7, 3),
# activation='relu',
# kernel_regularizer=l2(
# l2_regularization_penalization['Conv2']),
# name='Conv2'))
# convolutional_net.add(MaxPool2D())
# convolutional_net.add(Conv1D(filters=128, kernel_size=(4, 4),
# activation='relu',
# kernel_regularizer=l2(
# l2_regularization_penalization['Conv3']),
# name='Conv3'))
# convolutional_net.add(MaxPool2D())
#
# convolutional_net.add(Conv1D(filters=256, kernel_size=(4, 4),
# activation='relu',
# kernel_regularizer=l2(
# l2_regularization_penalization['Conv4']),
# name='Conv4'))
convolutional_net.add(Flatten())
convolutional_net.add(
Dense(64,
activation='sigmoid',
kernel_regularizer=l2(
l2_regularization_penalization['Dense1']),
name='Dense1'))
# Now the pairs of images
print(self.input_shape)
input_image_1 = Input(self.input_shape)
input_image_2 = Input(self.input_shape)
use_CNN = True
if use_CNN:
input_image_1 = tf.reshape(input_image_1, (32, 24, 1))
input_image_2 = tf.reshape(input_image_1, (32, 24, 1))
print(input_image_1.shape)
encoded_image_1 = convolutional_net(input_image_1)
encoded_image_2 = convolutional_net(input_image_2)
print(encoded_image_1.shape)
# L1 distance layer between the two encoded outputs
# One could use Subtract from Keras, but we want the absolute value
l1_distance_layer = Lambda(
lambda tensors: K.abs(tensors[0] - tensors[1]))
l1_distance = l1_distance_layer([encoded_image_1, encoded_image_2])
print(l1_distance.shape)
###
#todo 这个地方可以接着修改，中间加入一层dense作为mlp的隐层，再进最后的分类sigmod层
# Same class or not prediction
prediction = Dense(1, activation='sigmoid')(l1_distance)
self.model = Model(inputs=[input_image_1, input_image_2],
outputs=prediction)
self.model.summary()
# Define the optimizer and compile the model
optimizer = Modified_SGD(lr=self.learning_rate,
lr_multipliers=learning_rate_multipliers,
momentum=0.5)
self.model.compile(loss='binary_crossentropy',
metrics=['binary_accuracy'],
optimizer=optimizer)