def net_discriminator(input_tensor, output_dim=1, keep_prob=0.5):
"""
To build the discriminative network
:param input_tensor:
:param keep_prob:
:return:
"""
conv_1 = layers.conv_relu_layer('conv_1',
input_tensor,
kernel_size=3,
stride=1,
output_dim=32)
pool_1 = layers.pooling_layer('pool_1', conv_1, kernel_size=2, stride=1)
conv_2 = layers.conv_relu_layer('conv_2',
pool_1,
kernel_size=3,
stride=1,
output_dim=32)
pool_2 = layers.pooling_layer('pool_2', conv_2, kernel_size=2, stride=1)
conv_3 = layers.conv_relu_layer('conv_3',
pool_2,
kernel_size=3,
stride=1,
output_dim=32)
pool_3 = layers.pooling_layer('pool_3', conv_3, kernel_size=4, stride=1)
fc_1 = layers.fc_relu_layer('fc_1', pool_3, output_dim=512)
fc_1 = tf.nn.dropout(fc_1, keep_prob=keep_prob)
fc_2 = layers.fc_relu_layer('fc_2', fc_1, output_dim=512)
fc_2 = tf.nn.dropout(fc_2, keep_prob=keep_prob)
fc_3 = layers.fc_layer('fc_3', fc_2, output_dim=output_dim)
return fc_3
def recognition_net(input_tensor, label_size):
"""
A recognition network as an auxiliary to the learning process.
:param input_tensor:
:param label_size: the number of all classes involved
:return prob:
"""
with tf.variable_scope(NAME_SCOPE_RECOGNITION):
fc_1 = layers.fc_relu_layer('Fc1', input_tensor, 1024)
fc_2 = layers.fc_relu_layer('Fc2', fc_1, 1024)
prob = layers.fc_layer('Prob', fc_2, label_size)
return prob
def generative_net(input_tensor, code_length):
"""
A network to rebuild the data from latent representation.
:param input_tensor:
:param code_length: hashing length
:return fc_3:
"""
with tf.variable_scope(NAME_SCOPE_GENERATION):
fc_1 = layers.fc_relu_layer('Fc1', input_tensor, 1024)
fc_2 = layers.fc_relu_layer('Fc2', fc_1, 1024)
fc_3 = tf.sigmoid(layers.fc_layer('Fc3', fc_2, code_length)) * 2 - 1
return fc_3
def variational_net(input_tensor, latent_size, sub_scope):
"""
A variational approximation of p(z) with neural network.
:param input_tensor:
:param latent_size: the size of the latent space
:param sub_scope:
:return variational_mean:
:return variational_log_sigma:
"""
with tf.variable_scope(NAME_SCOPE_VARIATION + '/' + sub_scope):
fc_1 = layers.fc_relu_layer('Fc1', input_tensor, 1024)
fc_2 = layers.fc_relu_layer('Fc2', fc_1, 1024)
variational_mean = tf.sigmoid(layers.fc_layer('VMean', fc_2, latent_size))
# variational_log_sigma = tf.sigmoid(layers.fc_layer('LogSigma', fc_2, latent_size))
return variational_mean # , variational_log_sigma