class FSTTransition(BaseRecurrent, Initializable):
def __init__(self, fst, remap_table, no_transition_cost, **kwargs):
"""Wrap FST in a recurrent brick.
Parameters
----------
fst : FST instance
remap_table : dict
Maps neutral network characters to FST characters.
no_transition_cost : float
Cost of going to the start state when no arc for an input
symbol is available.
"""
super(FSTTransition, self).__init__(**kwargs)
self.fst = fst
self.transition = FSTTransitionOp(fst, remap_table)
self.probability_computer = FSTCostsOp(
fst, remap_table, no_transition_cost)
self.out_dim = len(remap_table)
@recurrent(sequences=['inputs', 'mask'],
states=['states', 'weights', 'add'],
outputs=['states', 'weights', 'add'], contexts=[])
def apply(self, inputs, states, weights, add,
mask=None):
new_states, new_weights = self.transition(states, weights, inputs)
if mask:
# In fact I don't really understand why we do this:
# anyway states not covered by masks should have no effect
# on the cost...
new_states = tensor.cast(mask * new_states +
(1. - mask) * states, 'int64')
new_weights = mask * new_weights + (1. - mask) * weights
new_add = self.probability_computer(new_states, new_weights)
return new_states, new_weights, new_add
@application(outputs=['states', 'weights', 'add'])
def initial_states(self, batch_size, *args, **kwargs):
states_dict = self.fst.expand({self.fst.fst.start: 0.0})
states = tensor.as_tensor_variable(
self.transition.pad(states_dict.keys(), NOT_STATE))
states = tensor.tile(states[None, :], (batch_size, 1))
weights = tensor.as_tensor_variable(
self.transition.pad(states_dict.values(), 0))
weights = tensor.tile(weights[None, :], (batch_size, 1))
add = self.probability_computer(states, weights)
return states, weights, add
def get_dim(self, name):
if name == 'states' or name == 'weights':
return MAX_STATES
if name == 'add':
return self.out_dim
if name == 'inputs':
return 0
return super(FSTTransition, self).get_dim(name)
class FSTTransition(BaseRecurrent, Initializable):
def __init__(self, fst, remap_table, no_transition_cost, **kwargs):
"""Wrap FST in a recurrent brick.
Parameters
----------
fst : FST instance
remap_table : dict
Maps neutral network characters to FST characters.
no_transition_cost : float
Cost of going to the start state when no arc for an input
symbol is available.
"""
super(FSTTransition, self).__init__(**kwargs)
self.fst = fst
self.transition = FSTTransitionOp(fst, remap_table)
self.probability_computer = FSTCostsOp(fst, remap_table,
no_transition_cost)
self.out_dim = len(remap_table)
@recurrent(sequences=['inputs', 'mask'],
states=['states', 'weights', 'add'],
outputs=['states', 'weights', 'add'],
contexts=[])
def apply(self, inputs, states, weights, add, mask=None):
new_states, new_weights = self.transition(states, weights, inputs)
if mask:
# In fact I don't really understand why we do this:
# anyway states not covered by masks should have no effect
# on the cost...
new_states = tensor.cast(mask * new_states + (1. - mask) * states,
'int64')
new_weights = mask * new_weights + (1. - mask) * weights
new_add = self.probability_computer(new_states, new_weights)
return new_states, new_weights, new_add
@application(outputs=['states', 'weights', 'add'])
def initial_states(self, batch_size, *args, **kwargs):
states_dict = self.fst.expand({self.fst.fst.start: 0.0})
states = tensor.as_tensor_variable(
self.transition.pad(states_dict.keys(), NOT_STATE))
states = tensor.tile(states[None, :], (batch_size, 1))
weights = tensor.as_tensor_variable(
self.transition.pad(states_dict.values(), 0))
weights = tensor.tile(weights[None, :], (batch_size, 1))
add = self.probability_computer(states, weights)
return states, weights, add
def get_dim(self, name):
if name == 'states' or name == 'weights':
return MAX_STATES
if name == 'add':
return self.out_dim
if name == 'inputs':
return 0
return super(FSTTransition, self).get_dim(name)
def __init__(self, fst, remap_table, no_transition_cost, **kwargs):
"""Wrap FST in a recurrent brick.
Parameters
----------
fst : FST instance
remap_table : dict
Maps neutral network characters to FST characters.
no_transition_cost : float
Cost of going to the start state when no arc for an input
symbol is available.
"""
super(FSTTransition, self).__init__(**kwargs)
self.fst = fst
self.transition = FSTTransitionOp(fst, remap_table)
self.probability_computer = FSTCostsOp(fst, remap_table,
no_transition_cost)
self.out_dim = len(remap_table)
def __init__(self, fst, remap_table, no_transition_cost, **kwargs):
"""Wrap FST in a recurrent brick.
Parameters
----------
fst : FST instance
remap_table : dict
Maps neutral network characters to FST characters.
no_transition_cost : float
Cost of going to the start state when no arc for an input
symbol is available.
"""
super(FSTTransition, self).__init__(**kwargs)
self.fst = fst
self.transition = FSTTransitionOp(fst, remap_table)
self.probability_computer = FSTCostsOp(
fst, remap_table, no_transition_cost)
self.out_dim = len(remap_table)