def _build_encoder(self):
with tf.variable_scope('encoder'):
if self.hparams.encoder_type == 'pbilstm':
cells_fw = [model_utils.single_cell("lstm", self.hparams.encoder_num_units // 2, self.mode) for _ in
range(self.hparams.num_encoder_layers)]
cells_bw = [model_utils.single_cell("lstm", self.hparams.encoder_num_units // 2, self.mode) for _ in
range(self.hparams.num_encoder_layers)]
prev_layer = self.inputs
prev_seq_len = self.input_seq_len
with tf.variable_scope("stack_p_bidirectional_rnn"):
state_fw = state_bw = None
for i, (cell_fw, cell_bw) in enumerate(zip(cells_fw, cells_bw)):
initial_state_fw = None
initial_state_bw = None
size = tf.cast(tf.floor(tf.shape(prev_layer)[1] / 2), tf.int32)
prev_layer = prev_layer[:, :size * 2, :]
prev_layer = tf.reshape(prev_layer,
[tf.shape(prev_layer)[0], size, prev_layer.shape[2] * 2])
prev_seq_len = tf.cast(tf.floor(prev_seq_len / 2), tf.int32)
with tf.variable_scope("cell_%d" % i):
outputs, (state_fw, state_bw) = tf.nn.bidirectional_dynamic_rnn(
cell_fw,
cell_bw,
prev_layer,
initial_state_fw=initial_state_fw,
initial_state_bw=initial_state_bw,
sequence_length=prev_seq_len,
dtype=tf.float32
)
# Concat the outputs to create the new input.
prev_layer = tf.concat(outputs, axis=2)
# states_fw.append(state_fw)
# states_bw.append(state_bw)
return prev_layer, tf.concat([state_fw, state_bw], -1)
if self.hparams.encoder_type == 'bilstm':
cells_fw = [model_utils.single_cell("lstm", self.hparams.encoder_num_units // 2, self.mode) for _ in
range(self.hparams.num_encoder_layers)]
cells_bw = [model_utils.single_cell("lstm", self.hparams.encoder_num_units // 2, self.mode) for _ in
range(self.hparams.num_encoder_layers)]
outputs, output_states_fw, output_states_bw = tf.contrib.rnn.stack_bidirectional_dynamic_rnn(
cells_fw, cells_bw,
self.inputs, sequence_length=self.input_seq_len,
dtype=tf.float32)
return outputs, tf.concat([output_states_fw, output_states_bw], -1)
elif self.hparams.encoder_type == 'lstm':
cells = [model_utils.single_cell("lstm", self.hparams.encoder_num_units, self.mode) for _ in
range(self.hparams.num_encoder_layers)]
cell = tf.nn.rnn_cell.MultiRNNCell(cells)
outputs, state = tf.nn.dynamic_rnn(cell, self.inputs, sequence_length=self.input_seq_len,
dtype=tf.float32)
return outputs, state
def _get_decoder_cell(self):
if self.hparams.num_decoder_layers == 1:
decoder_cell = model_utils.single_cell(
"lstm", self.hparams.decoder_num_units, self.mode)
else:
cells = [model_utils.single_cell("lstm", self.hparams.decoder_num_units, self.mode) for _ in
range(self.hparams.num_decoder_layers)]
decoder_cell = tf.nn.rnn_cell.MultiRNNCell(cells)
return decoder_cell
def _build_da_word_encoder(self, inputs, input_seq_len):
#if True:
with tf.variable_scope('encoder') as encoder_scope:
if self.hparams.da_word_encoder_type == 'bilstm':
cells_fw = [
model_utils.single_cell(
"lstm",
self.hparams.da_word_encoder_num_units // 2,
self.mode,
dropout=self.hparams.dropout)
for _ in range(self.hparams.num_da_word_encoder_layers)
]
cells_bw = [
model_utils.single_cell(
"lstm",
self.hparams.da_word_encoder_num_units // 2,
self.mode,
dropout=self.hparams.dropout)
for _ in range(self.hparams.num_da_word_encoder_layers)
]
outputs, output_states_fw, output_states_bw = tf.contrib.rnn.stack_bidirectional_dynamic_rnn(
cells_fw,
cells_bw,
inputs,
sequence_length=input_seq_len,
dtype=tf.float32)
return outputs, tf.concat([output_states_fw, output_states_bw],
-1)
elif self.hparams.da_word_encoder_type == 'lstm':
cells = [
model_utils.single_cell(
"lstm", self.hparams.da_word_encoder_num_units,
self.mode)
for _ in range(self.hparams.num_da_word_encoder_layers)
]
cell = tf.nn.rnn_cell.MultiRNNCell(cells)
outputs, state = tf.nn.dynamic_rnn(
cell,
inputs,
sequence_length=input_seq_len,
dtype=tf.float32)
return outputs, state
def _build_utt_encoder(self, history_inputs, history_input_seq_len):
# encoder at utterance-level
history_encoder_outputs, history_encoder_state = tf.nn.dynamic_rnn(
model_utils.single_cell("lstm",
self.hparams.utt_encoder_num_units,
self.mode,
dropout=self.hparams.dropout),
history_inputs,
sequence_length=history_input_seq_len,
dtype=tf.float32)
return history_encoder_state[0]