def __init__(self,
readout,
transition,
attention=None,
fork_inputs=None,
add_contexts=True,
**kwargs):
if not fork_inputs:
fork_inputs = [
name for name in transition.apply.sequences if name != 'mask'
]
fork = Fork(fork_inputs)
if attention:
distribute = Distribute(fork_inputs,
attention.take_glimpses.outputs[0])
transition = AttentionRecurrent(transition,
attention,
distribute,
add_contexts=add_contexts,
name="att_trans")
else:
transition = FakeAttentionRecurrent(transition,
name="with_fake_attention")
super(SequenceGenerator, self).__init__(readout, transition, fork,
**kwargs)
def __init__(self,
transition,
context_transition,
attention,
distribute=None,
add_contexts=True,
attended_name=None,
attended_mask_name=None,
**kwargs):
super(AttentionRecurrent, self).__init__(**kwargs)
self._sequence_names = list(transition.apply.sequences)
self._state_names = list(transition.apply.states)
self._context_names = list(transition.apply.contexts)
if add_contexts:
if not attended_name:
attended_name = 'attended_list'
if not attended_mask_name:
attended_mask_name = 'attended_mask_list'
self.posTag_name = 'posTag'
self._context_names += [
attended_name, attended_mask_name, self.posTag_name
]
else:
attended_name = self._context_names[0]
attended_mask_name = self._context_names[1]
if not distribute:
normal_inputs = [
name for name in self._sequence_names if 'mask' not in name
]
distribute = Distribute(normal_inputs,
attention.take_glimpses.outputs[0])
self.transition = transition
self.context_transition = context_transition
self.attention = attention
self.distribute = distribute
self.add_contexts = add_contexts
self.attended_name = attended_name
self.attended_mask_name = attended_mask_name
self.preprocessed_attended_name = "preprocessed_" + self.attended_name
self.preprocessed_posTag_name = 'preprocessed_' + self.posTag_name
self._glimpse_names = self.attention.take_glimpses.outputs #unchanged
# We need to determine which glimpses are fed back.
# Currently we extract it from `take_glimpses` signature.
self.previous_glimpses_needed = [
name for name in self._glimpse_names
if name in self.attention.take_glimpses.inputs
]
self.children = [
self.transition, self.context_transition, self.attention,
self.distribute
]
def __init__(self, recurrent, extra_input_name, extra_input_dim, **kwargs):
self.recurrent = recurrent
self.extra_input_name = extra_input_name
self.extra_input_dim = extra_input_dim
self._normal_inputs = [
name for name in self.recurrent.apply.sequences if name != 'mask'
]
self.distribute = Distribute(self._normal_inputs,
self.extra_input_name)
children = [self.recurrent, self.distribute]
super(RecurrentWithExtraInput, self).__init__(children=children,
**kwargs)
self.apply.sequences = self.recurrent.apply.sequences + [
self.extra_input_name
]
self.apply.outputs = self.recurrent.apply.outputs
self.apply.states = self.recurrent.apply.states
self.apply.contexts = self.recurrent.apply.contexts
self.initial_states.outputs = self.recurrent.initial_states.outputs
class RecurrentWithExtraInput(Initializable):
@lazy(allocation=['extra_input_dim'])
def __init__(self, recurrent, extra_input_name, extra_input_dim, **kwargs):
self.recurrent = recurrent
self.extra_input_name = extra_input_name
self.extra_input_dim = extra_input_dim
self._normal_inputs = [
name for name in self.recurrent.apply.sequences if name != 'mask'
]
self.distribute = Distribute(self._normal_inputs,
self.extra_input_name)
children = [self.recurrent, self.distribute]
super(RecurrentWithExtraInput, self).__init__(children=children,
**kwargs)
self.apply.sequences = self.recurrent.apply.sequences + [
self.extra_input_name
]
self.apply.outputs = self.recurrent.apply.outputs
self.apply.states = self.recurrent.apply.states
self.apply.contexts = self.recurrent.apply.contexts
self.initial_states.outputs = self.recurrent.initial_states.outputs
def _push_allocation_config(self):
self.distribute.source_dim = self.extra_input_dim
self.distribute.target_dims = self.recurrent.get_dims(
self.distribute.target_names)
@application
def apply(self, **kwargs):
# Should handle both "iterate=True" and "iterate=False"
extra_input = kwargs.pop(self.extra_input_name)
mask = kwargs.pop('mask', None)
normal_inputs = dict_subset(kwargs, self._normal_inputs, pop=True)
normal_inputs = self.distribute.apply(
as_dict=True,
**dict_union(normal_inputs, {self.extra_input_name: extra_input}))
return self.recurrent.apply(mask=mask,
**dict_union(normal_inputs, kwargs))
@application
def initial_states(self, *args, **kwargs):
return self.recurrent.initial_states(*args, **kwargs)
def get_dim(self, name):
if name == self.extra_input_name:
return self.extra_input_dim
return self.recurrent.get_dim(name)
def __init__(self,
transition,
attention,
topical_attention,
distribute=None,
topical_distribute=None,
add_contexts=True,
attended_name=None,
attended_mask_name=None,
topical_name=None,
topical_attended_name=None,
topical_attended_mask_name=None,
content_name=None,
**kwargs):
super(AttentionRecurrent, self).__init__(**kwargs)
self._sequence_names = list(transition.apply.sequences)
self._state_names = list(transition.apply.states)
self._context_names = list(transition.apply.contexts)
self._topical_glimpse_names = [
'topical_weighted_averages', 'topical_weights'
]
if add_contexts:
if not attended_name:
attended_name = 'attended'
if not attended_mask_name:
attended_mask_name = 'attended_mask'
self._context_names += [attended_name, attended_mask_name]
else:
attended_name = self._context_names[0]
attended_mask_name = self._context_names[1]
if not distribute:
normal_inputs = [
name for name in self._sequence_names if 'mask' not in name
]
distribute = Distribute(normal_inputs,
attention.take_glimpses.outputs[0],
name='distribute')
if not topical_distribute:
normal_inputs = [
name for name in self._sequence_names if 'mask' not in name
]
topical_distribute = Distribute(normal_inputs,
self._topical_glimpse_names[0],
name='topical_distribute')
self.transition = transition
self.attention = attention
self.topical_attention = topical_attention
self.distribute = distribute
self.topical_distribute = topical_distribute
self.add_contexts = add_contexts
self.attended_name = attended_name
self.attended_mask_name = attended_mask_name
self.topical_attended_name = topical_attended_name
self.topical_attended_mask_name = topical_attended_mask_name
self.content_name = content_name
if not topical_name:
self.topical_name = 'topical_embeddingq'
else:
self.topical_name = topical_name
self.preprocessed_topical_attended_name = "preprocessed_" + self.topical_attended_name
self.preprocessed_attended_name = "preprocessed_" + self.attended_name
self._glimpse_names = self.attention.take_glimpses.outputs
# We need to determine which glimpses are fed back.
# Currently we extract it from `take_glimpses` signature.
self.previous_glimpses_needed = [
name for name in self._glimpse_names
if name in self.attention.take_glimpses.inputs
]
self.children = [
self.transition, self.attention, self.topical_attention,
self.distribute, self.topical_distribute
]
