def siamese_layer(
self,
sequence_len,
model_cfg,
):
hidden_size = model_cfg['PARAMS'].getint('hidden_size')
cell_type = model_cfg['PARAMS'].get('cell_type')
outputs_sen1 = rnn_layer(
embedded_x=self.embedded_x1,
hidden_size=hidden_size,
bidirectional=False,
cell_type=cell_type,
)
outputs_sen2 = rnn_layer(
embedded_x=self.embedded_x2,
hidden_size=hidden_size,
bidirectional=False,
cell_type=cell_type,
reuse=True,
)
out1 = tf.reduce_mean(outputs_sen1, axis=1)
out2 = tf.reduce_mean(outputs_sen2, axis=1)
return manhattan_similarity(out1, out2)
def testRNNBidirectionalNetwork(self):
with self.test_session():
embedded_x = tf.random_normal(
[1, 2, 5]) # batch x sentence length x embedding size
hidden_size = 10
rnn_output = rnn_layer(embedded_x, hidden_size, bidirectional=True)
actual_output = rnn_output.get_shape().as_list()[-1]
self.assertEqual(actual_output, 2 * hidden_size)
def siamese_layer(self, sequence_len, model_cfg):
hidden_size = model_cfg['PARAMS'].getint('hidden_size')
cell_type = model_cfg['PARAMS'].get('cell_type')
outputs_sen = rnn_layer(self.embedded_x, hidden_size, cell_type)
with tf.name_scope('classifier'):
L1 = tf.layers.dropout(
tf.layers.dense(outputs_sen, 100, activation=tf.nn.relu, name='L1'),
rate=self.dropout, training=self.is_training)
y = tf.layers.dense(L1, 1, activation=tf.nn.softmax, name='y')
return y
def testRNNUnidirectionalNetwork(self):
embedded_x = tf.random_normal(
[1, 2, 5]) # batch x sentence length x embedding size
hidden_size = 10
rnn_output = rnn_layer(
embedded_x=embedded_x,
hidden_size=hidden_size,
bidirectional=False,
cell_type='GRU',
)
actual_output = rnn_output.get_shape().as_list()[-1]
self.assertEqual(actual_output, hidden_size)