def _build_body(self):
# input projection
_units = tf.layers.dense(self._features, self.dense_size,
kernel_regularizer=tf.nn.l2_loss,
kernel_initializer=xav())
if self.attn:
attn_scope = "attention_mechanism/{}".format(self.attn.type)
with tf.variable_scope(attn_scope):
if self.attn.type == 'general':
_attn_output = am.general_attention(
self._key,
self._emb_context,
hidden_size=self.attn.hidden_size,
projected_align=self.attn.projected_align)
elif self.attn.type == 'bahdanau':
_attn_output = am.bahdanau_attention(
self._key,
self._emb_context,
hidden_size=self.attn.hidden_size,
projected_align=self.attn.projected_align)
elif self.attn.type == 'cs_general':
_attn_output = am.cs_general_attention(
self._key,
self._emb_context,
hidden_size=self.attn.hidden_size,
depth=self.attn.depth,
projected_align=self.attn.projected_align)
elif self.attn.type == 'cs_bahdanau':
_attn_output = am.cs_bahdanau_attention(
self._key,
self._emb_context,
hidden_size=self.attn.hidden_size,
depth=self.attn.depth,
projected_align=self.attn.projected_align)
elif self.attn.type == 'light_general':
_attn_output = am.light_general_attention(
self._key,
self._emb_context,
hidden_size=self.attn.hidden_size,
projected_align=self.attn.projected_align)
elif self.attn.type == 'light_bahdanau':
_attn_output = am.light_bahdanau_attention(
self._key,
self._emb_context,
hidden_size=self.attn.hidden_size,
projected_align=self.attn.projected_align)
else:
raise ValueError("wrong value for attention mechanism type")
_units = tf.concat([_units, _attn_output], -1)
_units = tf_layers.variational_dropout(_units,
keep_prob=self._dropout_keep_prob)
# recurrent network unit
_lstm_cell = tf.nn.rnn_cell.LSTMCell(self.hidden_size)
_utter_lengths = tf.to_int32(tf.reduce_sum(self._utterance_mask, axis=-1))
_output, _state = tf.nn.dynamic_rnn(_lstm_cell,
_units,
time_major=False,
initial_state=self._initial_state,
sequence_length=_utter_lengths)
_output = tf.reshape(_output, (self._batch_size, -1, self.hidden_size))
_output = tf_layers.variational_dropout(_output,
keep_prob=self._dropout_keep_prob)
# output projection
_logits = tf.layers.dense(_output, self.action_size,
kernel_regularizer=tf.nn.l2_loss,
kernel_initializer=xav(), name='logits')
return _logits, _state
def _build_body(self):
# input projection
_units = tf.layers.dense(self._features, self.dense_size,
kernel_regularizer=tf.nn.l2_loss,
kernel_initializer=xav())
if self.attn:
attn_scope = "attention_mechanism/{}".format(self.attn.type)
with tf.variable_scope(attn_scope):
if self.attn.type == 'general':
_attn_output = am.general_attention(
self._key,
self._emb_context,
hidden_size=self.attn.hidden_size,
projected_align=self.attn.projected_align)
elif self.attn.type == 'bahdanau':
_attn_output = am.bahdanau_attention(
self._key,
self._emb_context,
hidden_size=self.attn.hidden_size,
projected_align=self.attn.projected_align)
elif self.attn.type == 'cs_general':
_attn_output = am.cs_general_attention(
self._key,
self._emb_context,
hidden_size=self.attn.hidden_size,
depth=self.attn.depth,
projected_align=self.attn.projected_align)
elif self.attn.type == 'cs_bahdanau':
_attn_output = am.cs_bahdanau_attention(
self._key,
self._emb_context,
hidden_size=self.attn.hidden_size,
depth=self.attn.depth,
projected_align=self.attn.projected_align)
elif self.attn.type == 'light_general':
_attn_output = am.light_general_attention(
self._key,
self._emb_context,
hidden_size=self.attn.hidden_size,
projected_align=self.attn.projected_align)
elif self.attn.type == 'light_bahdanau':
_attn_output = am.light_bahdanau_attention(
self._key,
self._emb_context,
hidden_size=self.attn.hidden_size,
projected_align=self.attn.projected_align)
else:
raise ValueError("wrong value for attention mechanism type")
_units = tf.concat([_units, _attn_output], -1)
_units = tf_layers.variational_dropout(_units,
keep_prob=self._dropout_keep_prob)
# recurrent network unit
_lstm_cell = tf.nn.rnn_cell.LSTMCell(self.hidden_size)
_utter_lengths = tf.to_int32(tf.reduce_sum(self._utterance_mask, axis=-1))
_output, _state = tf.nn.dynamic_rnn(_lstm_cell,
_units,
initial_state=self._initial_state,
sequence_length=_utter_lengths)
# output projection
_logits = tf.layers.dense(_output, self.action_size,
kernel_regularizer=tf.nn.l2_loss,
kernel_initializer=xav(), name='logits')
return _logits, _state