def _build_cell(self,
mode,
batch_size,
initial_state=None,
memory=None,
memory_sequence_length=None,
dtype=None,
alignment_history=False):
attention_mechanisms = [
_build_attention_mechanism(
attention_mechanism,
self.num_units,
memory,
memory_sequence_length=memory_sequence_length)
for attention_mechanism in self.attention_mechanism_class
]
cell = build_cell(self.num_layers,
self.num_units,
mode,
dropout=self.dropout,
residual_connections=self.residual_connections,
cell_class=self.cell_class,
attention_layers=self.attention_layers,
attention_mechanisms=attention_mechanisms)
initial_state = cell.zero_state(batch_size, memory.dtype)
return cell, initial_state
def _build_cell(self,
mode,
batch_size,
initial_state=None,
memory=None,
memory_sequence_length=None,
dtype=None,
alignment_history=False):
_ = memory_sequence_length
_ = alignment_history
if memory is None and dtype is None:
raise ValueError(
"dtype argument is required when memory is not set")
cell = build_cell(self.num_layers,
self.num_units,
mode,
dropout=self.dropout,
residual_connections=self.residual_connections,
cell_class=self.cell_class)
initial_state = self._init_state(cell.zero_state(
batch_size, dtype or memory.dtype),
initial_state=initial_state)
return cell, initial_state
def make_inputs(self, features, training=None):
inputs = features["char_ids"]
flat_inputs = tf.reshape(inputs, [-1, tf.shape(inputs)[-1]])
embeddings = self._embed(flat_inputs, training)
sequence_length = tf.count_nonzero(flat_inputs, axis=1)
cell = build_cell(1,
self.num_units,
tf.estimator.ModeKeys.TRAIN if training else None,
dropout=self.dropout,
cell_class=self.cell_class)
rnn_outputs, rnn_state = tf.nn.dynamic_rnn(
cell,
embeddings,
sequence_length=sequence_length,
dtype=embeddings.dtype)
if self.encoding == "average":
encoding = tf.reduce_mean(rnn_outputs, axis=1)
elif self.encoding == "last":
encoding = last_encoding_from_state(rnn_state)
outputs = tf.reshape(encoding,
[-1, tf.shape(inputs)[1], self.num_units])
return outputs
def _build_cell(self, mode):
return build_cell(self.num_layers,
self.num_units,
mode,
dropout=self.dropout,
residual_connections=self.residual_connections,
cell_class=self.cell_class)
def _build_cell(self,
mode,
batch_size,
initial_state=None,
memory=None,
memory_sequence_length=None,
dtype=None):
attention_mechanisms = [
_build_attention_mechanism(
attention_mechanism,
self.num_units,
memory,
memory_sequence_length=memory_sequence_length)
for attention_mechanism in self.attention_mechanism_class]
cell = build_cell(
self.num_layers,
self.num_units,
mode,
dropout=self.dropout,
residual_connections=self.residual_connections,
cell_class=self.cell_class,
attention_layers=self.attention_layers,
attention_mechanisms=attention_mechanisms)
initial_state = cell.zero_state(batch_size, memory.dtype)
return cell, initial_state
def _build_cell(self, mode):
return build_cell(
self.num_layers,
self.num_units,
mode,
dropout=self.dropout,
residual_connections=self.residual_connections,
cell_class=self.cell_class)
def _build_cell(self,
mode,
batch_size,
initial_state=None,
memory=None,
memory_sequence_length=None):
_ = memory
_ = memory_sequence_length
cell = build_cell(self.num_layers,
self.num_units,
mode,
dropout=self.dropout,
residual_connections=self.residual_connections,
cell_class=self.cell_class)
initial_state = self._init_state(cell.zero_state(
batch_size, tf.float32),
initial_state=initial_state)
return cell, initial_state
def transform(self, inputs, mode):
flat_inputs = tf.reshape(inputs, [-1, tf.shape(inputs)[-1]])
embeddings = self._embed(flat_inputs, mode)
sequence_length = tf.count_nonzero(flat_inputs, axis=1)
cell = build_cell(1,
self.num_units,
mode,
dropout=self.dropout,
cell_class=self.cell_class)
rnn_outputs, rnn_state = tf.nn.dynamic_rnn(
cell,
embeddings,
sequence_length=sequence_length,
dtype=embeddings.dtype)
if self.encoding == "average":
encoding = tf.reduce_mean(rnn_outputs, axis=1)
elif self.encoding == "last":
encoding = last_encoding_from_state(rnn_state)
outputs = tf.reshape(encoding,
[-1, tf.shape(inputs)[1], self.num_units])
return outputs
def _build_cell(self,
mode,
batch_size,
initial_state=None,
memory=None,
memory_sequence_length=None,
dtype=None):
_ = memory_sequence_length
if memory is None and dtype is None:
raise ValueError("dtype argument is required when memory is not set")
cell = build_cell(
self.num_layers,
self.num_units,
mode,
dropout=self.dropout,
residual_connections=self.residual_connections,
cell_class=self.cell_class)
initial_state = self._init_state(
cell.zero_state(batch_size, dtype or memory.dtype), initial_state=initial_state)
return cell, initial_state
