def set_network(self):
##########################################
# Recurrent layer
##########################################
with tf.variable_scope('LSTMcell'):
# 여러개의 셀을 조합한 RNN 셀을 생성합니다.
multi_cells = tf.contrib.rnn.MultiRNNCell(
[utils.LSTM_cell(FLAGS.RNN_HIDDEN_DIMENSION, self.Dropout_Rate1, self.Dropout_Rate2) for _ in range(FLAGS.N_LAYERS)])
# RNN 신경망을 생성 (SEQ를 통해 매 sentence의 길이까지만 계산을 해 효율성 증대)
self.outputs, _states = tf.nn.dynamic_rnn(cell=multi_cells, inputs=self.X, dtype=tf.float32) # sequence_length=self.SEQ,
# 마지막에 해당하는 결과물을 가져옴 (utils의 last_relevant 함수를 이용)
self.rnn_outputs = utils.last_relevant(self.outputs, self.SEQ)
##########################################
# Fully connected network
##########################################
with tf.variable_scope('FC-layer'):
FC1 = tf.contrib.layers.fully_connected(self.rnn_outputs, FLAGS.FC_HIDDEN_DIMENSION, activation_fn=None)
FC_act1 = tf.nn.relu(tf.layers.batch_normalization(FC1, momentum=0.9, training=self.TRAIN_PH))
FC2 = tf.contrib.layers.fully_connected(FC_act1, FLAGS.FC_HIDDEN_DIMENSION, activation_fn=None)
FC_act2 = tf.nn.relu(tf.layers.batch_normalization(FC2, momentum=0.9, training=self.TRAIN_PH))
self.y_logits = tf.contrib.layers.fully_connected(FC_act2, FLAGS.NUM_OF_CLASS, activation_fn=None)
def set_network(self):
##########################################
# Embedding
##########################################
with tf.variable_scope('Embedding'):
self.W = tf.get_variable(
name='embedding_matrix',
shape=[FLAGS.VOCAB_SIZE, FLAGS.EMBEDDING_SIZE],
initializer=tf.random_normal_initializer(stddev=0.1))
self.X = tf.nn.embedding_lookup(self.W, self.X_idx)
##########################################
# Convolutional layer
##########################################
net = self.X
for i, width in enumerate(FLAGS.CONV_KERNEL_WIDTH):
with tf.variable_scope('conv_{}'.format(i + 1)):
# 1d-conv
net = utils.conv_1d(net, width, FLAGS.HIDDEN_DIMENSION,
self.TRAIN_PH)
# 1d-max pool
# if i % 2 == 1:
# net = tf.layers.max_pooling1d(net, pool_size=2, strides=2, padding='SAME')
##########################################
# Recurrent layer
##########################################
with tf.variable_scope('LSTMcell'):
# 여러개의 셀을 조합한 RNN 셀을 생성합니다.
multi_cells = tf.contrib.rnn.MultiRNNCell([
utils.LSTM_cell(FLAGS.RNN_HIDDEN_DIMENSION, self.Dropout_Rate1,
self.Dropout_Rate2)
for _ in range(FLAGS.N_LAYERS)
])
# RNN 신경망을 생성 (SEQ를 통해 매 sentence의 길이까지만 계산을 해 효율성 증대)
outputs, _states = tf.nn.dynamic_rnn(
cell=multi_cells, inputs=net,
dtype=tf.float32) #sequence_length=self.SEQ,
# Attention
with tf.variable_scope('Attention'):
self.att_matrix, self.att_alpha = utils.attention(
INPUTS=outputs,
ATTENTION_SIZE=FLAGS.ATTENTION_SIZE,
SEQ=self.SEQ,
time_major=False,
return_alphas=True)
##########################################
# Fully connected network
##########################################
with tf.variable_scope('FC-layer'):
FC1 = tf.contrib.layers.fully_connected(
self.att_matrix,
FLAGS.FC_HIDDEN_DIMENSION,
weights_initializer=utils.he_init,
activation_fn=None)
FC_act1 = tf.nn.relu(
tf.layers.batch_normalization(FC1,
momentum=0.9,
training=self.TRAIN_PH))
FC2 = tf.contrib.layers.fully_connected(
FC_act1,
FLAGS.FC_HIDDEN_DIMENSION,
weights_initializer=utils.he_init,
activation_fn=None)
FC_act2 = tf.nn.relu(
tf.layers.batch_normalization(FC2,
momentum=0.9,
training=self.TRAIN_PH))
self.y_logits = tf.contrib.layers.fully_connected(
FC_act2, FLAGS.NUM_OF_CLASS, activation_fn=None)