def __init__(
self,
vocab: Vocabulary,
model_name: Union[str, AutoModel],
embedding_dropout: float = 0.0,
initializer: InitializerApplicator = InitializerApplicator(),
label_smoothing: float = None,
ignore_span_metric: bool = False,
srl_eval_path: str = DEFAULT_SRL_EVAL_PATH,
restrict_frames: bool = False,
restrict_roles: bool = False,
**kwargs,
) -> None:
# bypass SrlBert constructor
Model.__init__(self, vocab, **kwargs)
self.lemma_frame_dict = load_lemma_frame(LEMMA_FRAME_PATH)
self.frame_role_dict = load_role_frame(FRAME_ROLE_PATH)
self.restrict_frames = restrict_frames
self.restrict_roles = restrict_roles
if isinstance(model_name, str):
self.transformer = AutoModel.from_pretrained(model_name)
else:
self.transformer = model_name
# loss
self.role_criterion = nn.CrossEntropyLoss(ignore_index=0)
self.frame_criterion = nn.CrossEntropyLoss()
# number of classes
self.num_classes = self.vocab.get_vocab_size("labels")
self.frame_num_classes = self.vocab.get_vocab_size("frames_labels")
# metrics
role_set = self.vocab.get_token_to_index_vocabulary("labels")
role_set_filter = [v for k, v in role_set.items() if k != "O"]
self.f1_role_metric = FBetaMeasure(average="micro", labels=role_set_filter)
self.f1_frame_metric = FBetaMeasure(average="micro")
# output layer
self.tag_projection_layer = nn.Linear(self.transformer.config.hidden_size, self.num_classes)
self.frame_projection_layer = nn.Linear(
self.transformer.config.hidden_size, self.frame_num_classes
)
self.embedding_dropout = nn.Dropout(p=embedding_dropout)
self._label_smoothing = label_smoothing
initializer(self)
def __init__(
self,
vocab: Vocabulary,
bert_model: Union[str, AutoModel],
embedding_dropout: float = 0.0,
initializer: InitializerApplicator = InitializerApplicator(),
label_smoothing: float = None,
ignore_span_metric: bool = False,
srl_eval_path: str = DEFAULT_SRL_EVAL_PATH,
restrict_frames: bool = False,
restrict_roles: bool = False,
inventory: str = "verbatlas",
**kwargs,
) -> None:
# bypass SrlBert constructor
Model.__init__(self, vocab, **kwargs)
self.lemma_frame_dict = load_lemma_frame(LEMMA_FRAME_PATH)
self.frame_role_dict = load_role_frame(FRAME_ROLE_PATH)
self.restrict_frames = restrict_frames
self.restrict_roles = restrict_roles
self.transformer = AutoModel.from_pretrained(bert_model)
self.frame_criterion = nn.CrossEntropyLoss()
if inventory == "verbatlas":
# add missing labels
frame_list = load_label_list(FRAME_LIST_PATH)
self.vocab.add_tokens_to_namespace(frame_list, "frames_labels")
self.num_classes = self.vocab.get_vocab_size("labels")
self.frame_num_classes = self.vocab.get_vocab_size("frames_labels")
if srl_eval_path is not None:
# For the span based evaluation, we don't want to consider labels
# for verb, because the verb index is provided to the model.
self.span_metric = SrlEvalScorer(srl_eval_path, ignore_classes=["V"])
else:
self.span_metric = None
self.f1_frame_metric = FBetaMeasure(average="micro")
self.tag_projection_layer = nn.Linear(self.transformer.config.hidden_size, self.num_classes)
self.frame_projection_layer = nn.Linear(
self.transformer.config.hidden_size, self.frame_num_classes
)
self.embedding_dropout = nn.Dropout(p=embedding_dropout)
self._label_smoothing = label_smoothing
self.ignore_span_metric = ignore_span_metric
initializer(self)
def __init__(self,
vocab: Vocabulary,
response_only_predictor: CMVPredictor,
op_response_predictor: CMVPredictor,
output_feedforward: FeedForward,
dropout: float = 0.5,
initializer: InitializerApplicator = InitializerApplicator(),
regularizer: Optional[RegularizerApplicator] = None) -> None:
Model.__init__(self, vocab, regularizer)
self._response_only_predictor = response_only_predictor
self._op_response_predictor = op_response_predictor
if dropout:
self.dropout = torch.nn.Dropout(dropout)
else:
self.dropout = None
self._output_feedforward = output_feedforward
self._num_labels = vocab.get_vocab_size(namespace="labels")
#check_dimensions_match(text_field_embedder.get_output_dim(), encoder.get_input_dim(),
# "text field embedding dim", "encoder input dim")
#check_dimensions_match(encoder.get_output_dim() * 4, projection_feedforward.get_input_dim(),
# "encoder output dim", "projection feedforward input")
#check_dimensions_match(projection_feedforward.get_output_dim(), inference_encoder.get_input_dim(),
# "proj feedforward output dim", "inference lstm input dim")
self._accuracy = BooleanAccuracy()
self._fscore = F1Measure(positive_label=1)
self._fake_accuracy = BooleanAccuracy()
self._fake_fscore = F1Measure(positive_label=1)
self._loss = torch.nn.functional.binary_cross_entropy_with_logits
initializer(self)