def transition_block(x,
filter_num,
training,
name,
keep_prob,
compression=1.0):
"""
:param x:
:param filter_num:
:param training:
:param name:
:param compression:
:param keep_prob:
:return:
"""
with tf.variable_scope(name):
x = tf.layers.batch_normalization(x, training=training)
x = tf.nn.relu(x)
x = conv2d(x,
kernel_size=1,
stride=1,
filter_num=int(filter_num * compression),
padding="SAME",
name="conv")
x = tf.nn.dropout(x, keep_prob=keep_prob)
x = avg_pool(x,
kernel_size=2,
stride=2,
name="avg_pool",
padding="SAME")
return x
def __build_graph(self):
"""
:return:
"""
x = conv2d(self.__X, kernel_size=3, stride=1, filter_num=self.filter_num, padding="SAME", name="conv1")
x = tf.layers.batch_normalization(x, training=self.__is_training)
x = tf.nn.relu(x)
for block_idx in range(self.dense_block_num - 1):
dense_block_name = "dense_block" + str(block_idx + 1)
transition_block_name = "transition_block" + str(block_idx + 1)
x, filter_num = dense_block(x, self.layers_num[block_idx], self.filter_num,
self.growth_rate, training=self.__is_training,
name=dense_block_name, keep_prob=self.__keep_prob)
x = transition_block(x, self.filter_num, training=self.__is_training,
name=transition_block_name, keep_prob=self.__keep_prob,
compression=self.compression)
self.filter_num = int(self.filter_num * self.compression)
dense_block_name = "dense_block" + str(self.dense_block_num)
x = conv2d(x, kernel_size=3, stride=1, filter_num=self.filter_num, padding="VALID", name="conv2")
x = tf.layers.batch_normalization(x, training=self.__is_training)
x = tf.nn.relu(x)
x, filter_num = dense_block(x, self.layers_num[-1], self.filter_num, self.growth_rate,
training=self.__is_training, name=dense_block_name,
keep_prob=self.__keep_prob)
x = tf.layers.batch_normalization(x, training=self.__is_training)
x = tf.nn.relu(x)
x = conv2d(x, kernel_size=3, stride=1, filter_num=self.filter_num, padding="VALID", name="conv3")
x = tf.layers.batch_normalization(x, training=self.__is_training)
x = tf.nn.relu(x)
self.deep_features = x
pool_size = x.get_shape()[1]
x = avg_pool(x, kernel_size=pool_size, stride=1, padding="VALID", name="avg_pool")
input_dim = x.get_shape()[-1]
x = tf.reshape(x, [-1, input_dim])
x = fc_connect(x, self.class_num, name="fc_connect")
x = tf.nn.softmax(x)
return x
