def initial_loop_state(self) -> AttentionLoopState:
length = sum(tf.shape(s)[1] for s in self._encoders_tensors)
if self._use_sentinels:
length += 1
return empty_attention_loop_state(self.batch_size, length,
self.context_vector_size)
def initial_loop_state(self) -> HierarchicalLoopState:
length = len(self.attentions)
if self._use_sentinels:
length += 1
return HierarchicalLoopState(
child_loop_states=[a.initial_loop_state()
for a in self.attentions],
loop_state=empty_attention_loop_state(
self.batch_size, length, self.context_vector_size))
def initial_loop_state(self) -> HierarchicalLoopState:
length = len(self.attentions)
if self._use_sentinels:
length += 1
return HierarchicalLoopState(
child_loop_states=[a.initial_loop_state()
for a in self.attentions],
loop_state=empty_attention_loop_state(
self.batch_size, length, self.context_vector_size))
def initial_loop_state(self) -> AttentionLoopState:
# Here we need to make sure that the hidden_features and attention_mask
# are pre-computed. If this is used in combination with a decoder which
# has train and runtime while loops, these tensors need to be created
# outside of any of those loops in order to be available to both.
# Note that we are not breaking lazy loading here because this method
# is called from a lazy tensor.
debug("Pre-computing attention tensors", "bless")
debug("Hidden features: {}".format(self.hidden_features), "bless")
debug("Hidden mask: {}".format(self.attention_mask), "bless")
return empty_attention_loop_state(
self.batch_size,
tf.shape(self.attention_states)[1],
self.context_vector_size)
def initial_loop_state(self) -> AttentionLoopState:
# Similarly to the feed_forward attention, we need to build the encoder
# projections and masks before the while loop is entered so they are
# not created as a part of the loop
# pylint: disable=not-an-iterable
for val in self.encoder_projections_for_logits:
debug(val, "bless")
debug(self.masks_concat, "bless")
length = sum(tf.shape(s)[1] for s in self._encoders_tensors)
# pylint: enable=not-an-iterable
if self._use_sentinels:
length += 1
return empty_attention_loop_state(self.batch_size, length,
self.context_vector_size)
def initial_loop_state(self) -> AttentionLoopState:
# Similarly to the feed_forward attention, we need to build the encoder
# projections and masks before the while loop is entered so they are
# not created as a part of the loop
# pylint: disable=not-an-iterable
for val in self.encoder_projections_for_logits:
debug(val, "bless")
debug(self.masks_concat, "bless")
length = sum(tf.shape(s)[1] for s in self._encoders_tensors)
# pylint: enable=not-an-iterable
if self._use_sentinels:
length += 1
return empty_attention_loop_state(self.batch_size, length,
self.context_vector_size)
def initial_loop_state(self) -> AttentionLoopState:
return empty_attention_loop_state(self.batch_size,
tf.shape(self.attention_states)[1],
self.context_vector_size)
def initial_loop_state(self) -> AttentionLoopState:
return empty_attention_loop_state()
def initial_loop_state(self) -> HierarchicalLoopState:
return HierarchicalLoopState(
child_loop_states=[a.initial_loop_state()
for a in self.attentions],
loop_state=empty_attention_loop_state())
def initial_loop_state(self) -> AttentionLoopState:
return empty_attention_loop_state(
self.batch_size, 1,
self.context_vector_size)
