def _get_state_cost(self, state: WikiTablesDecoderState) -> torch.Tensor:
if not state.is_finished():
raise RuntimeError(
"_get_state_cost() is not defined for unfinished states!")
# Our checklist cost is a sum of squared error from where we want to be, making sure we
# take into account the mask. We clamp the lower limit of the balance at 0 to avoid
# penalizing agenda actions produced multiple times.
checklist_balance = torch.clamp(state.checklist_state[0].get_balance(),
min=0.0)
checklist_cost = torch.sum((checklist_balance)**2)
# This is the number of items on the agenda that we want to see in the decoded sequence.
# We use this as the denotation cost if the path is incorrect.
denotation_cost = torch.sum(
state.checklist_state[0].checklist_target.float())
checklist_cost = self._checklist_cost_weight * checklist_cost
action_history = state.action_history[0]
batch_index = state.batch_indices[0]
action_strings = [
state.possible_actions[batch_index][i][0] for i in action_history
]
logical_form = state.world[batch_index].get_logical_form(
action_strings)
lisp_string = state.example_lisp_string[batch_index]
if self._denotation_accuracy.evaluate_logical_form(
logical_form, lisp_string):
cost = checklist_cost
else:
cost = checklist_cost + (
1 - self._checklist_cost_weight) * denotation_cost
return cost
def _get_state_cost(self, state: WikiTablesDecoderState) -> torch.Tensor:
if not state.is_finished():
raise RuntimeError("_get_state_cost() is not defined for unfinished states!")
action_history = state.action_history[0]
batch_index = state.batch_indices[0]
action_strings = [state.possible_actions[batch_index][i][0] for i in action_history]
logical_form = state.world[batch_index].get_logical_form(action_strings)
lisp_string = state.example_lisp_string[batch_index]
if self._denotation_accuracy.evaluate_logical_form(logical_form, lisp_string):
cost = torch.FloatTensor([0.0])
else:
cost = torch.FloatTensor([10])
return Variable(state.flattened_linking_scores.data.new(cost)).float()
def _get_state_cost(self, state: WikiTablesDecoderState) -> torch.Tensor:
if not state.is_finished():
raise RuntimeError("_get_state_cost() is not defined for unfinished states!")
# Our checklist cost is a sum of squared error from where we want to be, making sure we
# take into account the mask. We clamp the lower limit of the balance at 0 to avoid
# penalizing agenda actions produced multiple times.
checklist_balance = torch.clamp(state.checklist_state[0].get_balance(), min=0.0)
checklist_cost = torch.sum((checklist_balance) ** 2)
# This is the number of items on the agenda that we want to see in the decoded sequence.
# We use this as the denotation cost if the path is incorrect.
denotation_cost = torch.sum(state.checklist_state[0].checklist_target.float())
checklist_cost = self._checklist_cost_weight * checklist_cost
action_history = state.action_history[0]
batch_index = state.batch_indices[0]
action_strings = [state.possible_actions[batch_index][i][0] for i in action_history]
logical_form = state.world[batch_index].get_logical_form(action_strings)
lisp_string = state.example_lisp_string[batch_index]
if self._denotation_accuracy.evaluate_logical_form(logical_form, lisp_string):
cost = checklist_cost
else:
cost = checklist_cost + (1 - self._checklist_cost_weight) * denotation_cost
return cost