def __init__(self,
vocab: Vocabulary,
sentence_embedder: TextFieldEmbedder,
action_embedding_dim: int,
encoder: Seq2SeqEncoder,
attention_function: SimilarityFunction,
beam_size: int,
max_num_finished_states: int,
max_decoding_steps: int,
dropout: float = 0.0,
normalize_beam_score_by_length: bool = False,
checklist_cost_weight: float = 0.6,
dynamic_cost_weight: Dict[str, Union[int, float]] = None,
penalize_non_agenda_actions: bool = False,
initial_mml_model_file: str = None) -> None:
super(NlvrCoverageSemanticParser, self).__init__(vocab=vocab,
sentence_embedder=sentence_embedder,
action_embedding_dim=action_embedding_dim,
encoder=encoder,
dropout=dropout)
self._agenda_coverage = Average()
self._decoder_trainer: DecoderTrainer[Callable[[NlvrDecoderState], torch.Tensor]] = \
ExpectedRiskMinimization(beam_size=beam_size,
normalize_by_length=normalize_beam_score_by_length,
max_decoding_steps=max_decoding_steps,
max_num_finished_states=max_num_finished_states)
# Instantiating an empty NlvrWorld just to get the number of terminals.
self._terminal_productions = set(NlvrWorld([]).terminal_productions.values())
self._decoder_step = NlvrDecoderStep(encoder_output_dim=self._encoder.get_output_dim(),
action_embedding_dim=action_embedding_dim,
attention_function=attention_function,
dropout=dropout,
use_coverage=True)
self._checklist_cost_weight = checklist_cost_weight
self._dynamic_cost_wait_epochs = None
self._dynamic_cost_rate = None
if dynamic_cost_weight:
self._dynamic_cost_wait_epochs = dynamic_cost_weight["wait_num_epochs"]
self._dynamic_cost_rate = dynamic_cost_weight["rate"]
self._penalize_non_agenda_actions = penalize_non_agenda_actions
self._last_epoch_in_forward: int = None
# TODO (pradeep): Checking whether file exists here to avoid raising an error when we've
# copied a trained ERM model from a different machine and the original MML model that was
# used to initialize it does not exist on the current machine. This may not be the best
# solution for the problem.
if initial_mml_model_file is not None:
if os.path.isfile(initial_mml_model_file):
archive = load_archive(initial_mml_model_file)
self._initialize_weights_from_archive(archive)
else:
# A model file is passed, but it does not exist. This is expected to happen when
# you're using a trained ERM model to decode. But it may also happen if the path to
# the file is really just incorrect. So throwing a warning.
logger.warning("MML model file for initializing weights is passed, but does not exist."
" This is fine if you're just decoding.")
def __init__(self, vocab: Vocabulary, sentence_embedder: TextFieldEmbedder,
action_embedding_dim: int, encoder: Seq2SeqEncoder,
attention_function: SimilarityFunction,
decoder_beam_search: BeamSearch,
max_decoding_steps: int) -> None:
super(NlvrDirectSemanticParser,
self).__init__(vocab=vocab,
sentence_embedder=sentence_embedder,
action_embedding_dim=action_embedding_dim,
encoder=encoder)
self._decoder_trainer = MaximumMarginalLikelihood()
self._decoder_step = NlvrDecoderStep(
encoder_output_dim=self._encoder.get_output_dim(),
action_embedding_dim=action_embedding_dim,
attention_function=attention_function)
self._decoder_beam_search = decoder_beam_search
self._max_decoding_steps = max_decoding_steps
self._action_padding_index = -1
def __init__(self,
vocab,
sentence_embedder,
action_embedding_dim,
encoder,
attention,
decoder_beam_search,
max_decoding_steps,
dropout=0.0):
super(NlvrDirectSemanticParser,
self).__init__(vocab=vocab,
sentence_embedder=sentence_embedder,
action_embedding_dim=action_embedding_dim,
encoder=encoder,
dropout=dropout)
self._decoder_trainer = MaximumMarginalLikelihood()
self._decoder_step = NlvrDecoderStep(
encoder_output_dim=self._encoder.get_output_dim(),
action_embedding_dim=action_embedding_dim,
input_attention=attention,
dropout=dropout)
self._decoder_beam_search = decoder_beam_search
self._max_decoding_steps = max_decoding_steps
self._action_padding_index = -1