def encode(self, inputs, sequence_lengths, is_training=False):
'''
get the high level feature representation
Args:
inputs: the input to the layer as a
[batch_size, max_length, dim] tensor
sequence_length: the length of the input sequences
is_training: whether or not the network is in training mode
Returns:
the output of the layer as a [bath_size, max_length, output_dim]
tensor
'''
outputs = inputs
output_seq_lengths = sequence_lengths
with tf.variable_scope('inlayer'):
#apply the linear layer
outputs = self.hidden_layer(outputs)
#apply the nonlinearity
outputs = tf.nn.relu(outputs)
if float(self.conf['listener_dropout']) < 1 and is_training:
outputs = tf.nn.dropout(
outputs, float(self.conf['listener_dropout']))
for l in range(int(self.conf['listener_numlayers'])):
with tf.variable_scope('layer%d' % l):
#apply the linear layer
hidden = self.hidden_layer(outputs)
#apply the nonlinearity
outputs = (tf.nn.relu(hidden) + outputs)/2
if float(self.conf['listener_dropout']) < 1 and is_training:
outputs = tf.nn.dropout(
outputs, float(self.conf['listener_dropout']))
#apply the pyramid stack
outputs, output_seq_lengths = pyramid_stack(outputs,
output_seq_lengths)
outputs = self.outlayer(outputs, output_seq_lengths)
if float(self.conf['listener_dropout']) < 1 and is_training:
outputs = tf.nn.dropout(outputs,
float(self.conf['listener_dropout']))
return outputs
def __call__(self, inputs, sequence_lengths, scope=None):
"""
Create the variables and do the forward computation
Args:
inputs: A time minor tensor of shape [batch_size, time,
input_size],
sequence_lengths: the length of the input sequences
scope: The variable scope sets the namespace under which
the variables created during this call will be stored.
Returns:
the output of the layer, the concatenated outputs of the
forward and backward pass shape [batch_size, time/2, input_size*2].
"""
with tf.variable_scope(scope or type(self).__name__):
#apply blstm layer
outputs = self.blstm(inputs, sequence_lengths)
stacked_outputs, output_seq_lengths = ops.pyramid_stack(
outputs, sequence_lengths)
return stacked_outputs, output_seq_lengths