def __init__(self,
hidden_size,
num_labels,
crf=False) -> None:
super().__init__()
self.classifier = torch.nn.Linear(hidden_size, num_labels)
self.crf = CRF(num_labels) if crf else None
def __init__(self, hidden_size, num_labels, crf=False) -> None:
"""A linear layer with an optional CRF (:cite:`lafferty2001conditional`) layer on top of it.
Args:
hidden_size: Size of hidden states.
num_labels: Size of tag set.
crf: ``True`` to enable CRF (:cite:`lafferty2001conditional`).
"""
super().__init__()
self.classifier = torch.nn.Linear(hidden_size, num_labels)
self.crf = CRF(num_labels) if crf else None
def __init__(self,
encoder: TransformerEncoder,
num_labels,
crf=False,
secondary_encoder=None) -> None:
"""
A shallow tagging model use transformer as decoder.
Args:
encoder: A pretrained transformer.
num_labels: Size of tagset.
crf: True to enable CRF.
crf_constraints: The allowed transitions (from_label_id, to_label_id).
"""
super().__init__()
self.encoder = encoder
self.secondary_encoder = secondary_encoder
# noinspection PyUnresolvedReferences
self.classifier = nn.Linear(encoder.transformer.config.hidden_size, num_labels)
self.crf = CRF(num_labels) if crf else None
def __init__(self,
n_rels,
hidden_size,
n_mlp_rel=300,
mlp_dropout=0.2,
crf=False) -> None:
super().__init__()
self.mlp_rel_h = MLP(n_in=hidden_size,
n_out=n_mlp_rel,
dropout=mlp_dropout)
self.mlp_rel_d = MLP(n_in=hidden_size,
n_out=n_mlp_rel,
dropout=mlp_dropout)
self.rel_attn = Biaffine(n_in=n_mlp_rel,
n_out=n_rels,
bias_x=True,
bias_y=True)
bias = 1 / math.sqrt(self.rel_attn.weight.size(1))
nn.init.uniform_(self.rel_attn.weight, -bias, bias)
self.crf = CRF(n_rels) if crf else None
def __init__(self,
embed: Union[nn.Embedding, int],
rnn_input,
rnn_hidden,
n_out,
drop=0.5,
crf=True,
crf_constraints=None):
super(RNNTaggingModel, self).__init__()
# the embedding layer
if isinstance(embed, nn.Module):
self.embed = embed
n_embed = embed.embedding_dim
else:
self.embed = None
n_embed = embed
if rnn_input:
self.embed_to_rnn = nn.Linear(n_embed, rnn_input)
else:
self.embed_to_rnn = None
rnn_input = n_embed
# the word-lstm layer
self.word_lstm = nn.LSTM(input_size=rnn_input,
hidden_size=rnn_hidden,
batch_first=True,
bidirectional=True)
# the output layer
self.out = nn.Linear(rnn_hidden * 2, n_out)
# the CRF layer
self.crf = CRF(n_out, crf_constraints) if crf else None
self.drop = nn.Dropout(drop)
# self.drop = SharedDropout(drop)
# self.drop = LockedDropout(drop)
self.reset_parameters()
def __init__(self,
encoder: TransformerEncoder,
num_labels,
crf=False,
secondary_encoder=None,
extra_embeddings: EmbeddingDim = None) -> None:
"""
A shallow tagging model use transformer as decoder.
Args:
encoder: A pretrained transformer.
num_labels: Size of tagset.
crf: True to enable CRF.
extra_embeddings: Extra embeddings which will be concatenated to the encoder outputs.
"""
super().__init__()
self.encoder = encoder
self.secondary_encoder = secondary_encoder
self.extra_embeddings = extra_embeddings
# noinspection PyUnresolvedReferences
feature_size = encoder.transformer.config.hidden_size
if extra_embeddings:
feature_size += extra_embeddings.get_output_dim()
self.classifier = nn.Linear(feature_size, num_labels)
self.crf = CRF(num_labels) if crf else None