def __init__(self, n_lstm_hidden=400, n_mlp_sib=100, mlp_dropout=.33, **kwargs):
super().__init__(**kwargs)
self.mlp_sib_s = MLP(n_in=n_lstm_hidden*2, n_out=n_mlp_sib, dropout=mlp_dropout)
self.mlp_sib_d = MLP(n_in=n_lstm_hidden*2, n_out=n_mlp_sib, dropout=mlp_dropout)
self.mlp_sib_h = MLP(n_in=n_lstm_hidden*2, n_out=n_mlp_sib, dropout=mlp_dropout)
self.sib_attn = Triaffine(n_in=n_mlp_sib, bias_x=True, bias_y=True)
self.crf = CRF2oDependency()
def __init__(self,
n_words,
n_rels,
n_tags=None,
n_chars=None,
encoder='lstm',
feat=['char'],
n_embed=100,
n_pretrained=100,
n_feat_embed=100,
n_char_embed=50,
n_char_hidden=100,
char_pad_index=0,
elmo='original_5b',
elmo_bos_eos=(True, False),
bert=None,
n_bert_layers=4,
mix_dropout=.0,
bert_pooling='mean',
bert_pad_index=0,
freeze=True,
embed_dropout=.33,
n_lstm_hidden=400,
n_lstm_layers=3,
encoder_dropout=.33,
n_arc_mlp=500,
n_rel_mlp=100,
mlp_dropout=.33,
scale=0,
pad_index=0,
unk_index=1,
**kwargs):
super().__init__(**Config().update(locals()))
self.arc_mlp_d = MLP(n_in=self.args.n_hidden,
n_out=n_arc_mlp,
dropout=mlp_dropout)
self.arc_mlp_h = MLP(n_in=self.args.n_hidden,
n_out=n_arc_mlp,
dropout=mlp_dropout)
self.rel_mlp_d = MLP(n_in=self.args.n_hidden,
n_out=n_rel_mlp,
dropout=mlp_dropout)
self.rel_mlp_h = MLP(n_in=self.args.n_hidden,
n_out=n_rel_mlp,
dropout=mlp_dropout)
self.arc_attn = Biaffine(n_in=n_arc_mlp,
scale=scale,
bias_x=True,
bias_y=False)
self.rel_attn = Biaffine(n_in=n_rel_mlp,
n_out=n_rels,
bias_x=True,
bias_y=True)
self.criterion = nn.CrossEntropyLoss()
def __init__(self,
n_words,
n_labels,
n_tags=None,
n_chars=None,
encoder='lstm',
feat=['char'],
n_embed=100,
n_pretrained=100,
n_feat_embed=100,
n_char_embed=50,
n_char_hidden=100,
char_pad_index=0,
bert=None,
n_bert_layers=4,
mix_dropout=.0,
bert_pooling='mean',
bert_pad_index=0,
freeze=True,
embed_dropout=.33,
n_lstm_hidden=400,
n_lstm_layers=3,
encoder_dropout=.33,
n_span_mlp=500,
n_label_mlp=100,
mlp_dropout=.33,
pad_index=0,
unk_index=1,
**kwargs):
super().__init__(**Config().update(locals()))
self.span_mlp_l = MLP(n_in=self.args.n_hidden, n_out=n_span_mlp, dropout=mlp_dropout)
self.span_mlp_r = MLP(n_in=self.args.n_hidden, n_out=n_span_mlp, dropout=mlp_dropout)
self.label_mlp_l = MLP(n_in=self.args.n_hidden, n_out=n_label_mlp, dropout=mlp_dropout)
self.label_mlp_r = MLP(n_in=self.args.n_hidden, n_out=n_label_mlp, dropout=mlp_dropout)
self.span_attn = Biaffine(n_in=n_span_mlp, bias_x=True, bias_y=False)
self.label_attn = Biaffine(n_in=n_label_mlp, n_out=n_labels, bias_x=True, bias_y=True)
self.crf = CRFConstituency()
self.criterion = nn.CrossEntropyLoss()
def __init__(self,
n_words,
n_feats,
n_labels,
feat='char',
n_embed=100,
n_feat_embed=100,
n_char_embed=50,
bert=None,
n_bert_layers=4,
max_len=None,
mix_dropout=.0,
embed_dropout=.33,
n_lstm_hidden=400,
n_lstm_layers=3,
lstm_dropout=.33,
n_mlp_span=500,
n_mlp_label=100,
mlp_dropout=.33,
feat_pad_index=0,
pad_index=0,
unk_index=1,
**kwargs):
super().__init__()
self.args = Config().update(locals())
# the embedding layer
self.word_embed = nn.Embedding(num_embeddings=n_words,
embedding_dim=n_embed)
if feat == 'char':
self.feat_embed = CharLSTM(n_chars=n_feats,
n_embed=n_char_embed,
n_out=n_feat_embed,
pad_index=feat_pad_index)
elif feat == 'bert':
self.feat_embed = BertEmbedding(model=bert,
n_layers=n_bert_layers,
n_out=n_feat_embed,
pad_index=feat_pad_index,
max_len=max_len,
dropout=mix_dropout)
self.n_feat_embed = self.feat_embed.n_out
elif feat == 'tag':
self.feat_embed = nn.Embedding(num_embeddings=n_feats,
embedding_dim=n_feat_embed)
else:
raise RuntimeError("The feat type should be in ['char', 'bert', 'tag'].")
self.embed_dropout = IndependentDropout(p=embed_dropout)
# the lstm layer
self.lstm = BiLSTM(input_size=n_embed+n_feat_embed,
hidden_size=n_lstm_hidden,
num_layers=n_lstm_layers,
dropout=lstm_dropout)
self.lstm_dropout = SharedDropout(p=lstm_dropout)
# the MLP layers
self.mlp_span_l = MLP(n_in=n_lstm_hidden*2,
n_out=n_mlp_span,
dropout=mlp_dropout)
self.mlp_span_r = MLP(n_in=n_lstm_hidden*2,
n_out=n_mlp_span,
dropout=mlp_dropout)
self.mlp_label_l = MLP(n_in=n_lstm_hidden*2,
n_out=n_mlp_label,
dropout=mlp_dropout)
self.mlp_label_r = MLP(n_in=n_lstm_hidden*2,
n_out=n_mlp_label,
dropout=mlp_dropout)
# the Biaffine layers
self.span_attn = Biaffine(n_in=n_mlp_span,
bias_x=True,
bias_y=False)
self.label_attn = Biaffine(n_in=n_mlp_label,
n_out=n_labels,
bias_x=True,
bias_y=True)
self.crf = CRFConstituency()
self.criterion = nn.CrossEntropyLoss()
self.pad_index = pad_index
self.unk_index = unk_index
def __init__(
self,
n_words,
n_rels,
n_tags=None,
n_chars=None,
encoder='lstm',
feat=['char'],
n_embed=100,
n_pretrained=100,
n_feat_embed=100,
n_char_embed=50,
n_char_hidden=100,
char_pad_index=0,
bert=None,
n_bert_layers=4,
mix_dropout=.0,
bert_pooling='mean',
bert_pad_index=0,
freeze=True,
embed_dropout=.33,
n_lstm_hidden=400,
n_lstm_layers=3,
encoder_dropout=.33,
n_arc_mlp=500,
n_rel_mlp=100,
mlp_dropout=.33,
use_attentions=True, # attention
attention_head=0, # attention
attention_layer=6, # attention
scale=0,
pad_index=0,
unk_index=1,
**kwargs):
super().__init__(**Config().update(locals()))
self.arc_mlp_d = MLP(n_in=self.args.n_hidden,
n_out=n_arc_mlp,
dropout=mlp_dropout)
self.arc_mlp_h = MLP(n_in=self.args.n_hidden,
n_out=n_arc_mlp,
dropout=mlp_dropout)
self.rel_mlp_d = MLP(n_in=self.args.n_hidden,
n_out=n_rel_mlp,
dropout=mlp_dropout)
self.rel_mlp_h = MLP(n_in=self.args.n_hidden,
n_out=n_rel_mlp,
dropout=mlp_dropout)
self.arc_attn = Biaffine(n_in=n_arc_mlp,
scale=scale,
bias_x=True,
bias_y=False)
self.rel_attn = Biaffine(n_in=n_rel_mlp,
n_out=n_rels,
bias_x=True,
bias_y=True)
# transformer attention
if use_attentions:
self.attn_mix = nn.Parameter(torch.randn(1))
self.criterion = nn.CrossEntropyLoss()
def __init__(self,
n_words,
n_rels,
n_tags=None,
n_chars=None,
encoder='lstm',
feat=['char'],
n_embed=100,
n_pretrained=100,
n_feat_embed=100,
n_char_embed=50,
n_char_hidden=100,
char_pad_index=0,
bert=None,
n_bert_layers=4,
mix_dropout=.0,
bert_pooling='mean',
bert_pad_index=0,
freeze=True,
embed_dropout=.33,
n_lstm_hidden=400,
n_lstm_layers=3,
encoder_dropout=.33,
n_arc_mlp=500,
n_sib_mlp=100,
n_rel_mlp=100,
mlp_dropout=.33,
scale=0,
inference='mfvi',
max_iter=3,
pad_index=0,
unk_index=1,
**kwargs):
super().__init__(**Config().update(locals()))
self.arc_mlp_d = MLP(n_in=self.args.n_hidden,
n_out=n_arc_mlp,
dropout=mlp_dropout)
self.arc_mlp_h = MLP(n_in=self.args.n_hidden,
n_out=n_arc_mlp,
dropout=mlp_dropout)
self.sib_mlp_s = MLP(n_in=self.args.n_hidden,
n_out=n_sib_mlp,
dropout=mlp_dropout)
self.sib_mlp_d = MLP(n_in=self.args.n_hidden,
n_out=n_sib_mlp,
dropout=mlp_dropout)
self.sib_mlp_h = MLP(n_in=self.args.n_hidden,
n_out=n_sib_mlp,
dropout=mlp_dropout)
self.rel_mlp_d = MLP(n_in=self.args.n_hidden,
n_out=n_rel_mlp,
dropout=mlp_dropout)
self.rel_mlp_h = MLP(n_in=self.args.n_hidden,
n_out=n_rel_mlp,
dropout=mlp_dropout)
self.arc_attn = Biaffine(n_in=n_arc_mlp,
scale=scale,
bias_x=True,
bias_y=False)
self.sib_attn = Triaffine(n_in=n_sib_mlp,
scale=scale,
bias_x=True,
bias_y=True)
self.rel_attn = Biaffine(n_in=n_rel_mlp,
n_out=n_rels,
bias_x=True,
bias_y=True)
self.inference = (MFVIDependency
if inference == 'mfvi' else LBPDependency)(max_iter)
self.criterion = nn.CrossEntropyLoss()
def __init__(self,
n_words,
n_labels,
n_tags=None,
n_chars=None,
n_lemmas=None,
encoder='lstm',
feat=['tag', 'char', 'lemma'],
n_embed=100,
n_pretrained=125,
n_feat_embed=100,
n_char_embed=50,
n_char_hidden=400,
char_pad_index=0,
char_dropout=0.33,
bert=None,
n_bert_layers=4,
mix_dropout=.0,
bert_pooling='mean',
bert_pad_index=0,
freeze=True,
embed_dropout=.2,
n_lstm_hidden=600,
n_lstm_layers=3,
encoder_dropout=.33,
n_edge_mlp=600,
n_label_mlp=600,
edge_mlp_dropout=.25,
label_mlp_dropout=.33,
interpolation=0.1,
pad_index=0,
unk_index=1,
**kwargs):
super().__init__(**Config().update(locals()))
self.edge_mlp_d = MLP(n_in=self.args.n_hidden,
n_out=n_edge_mlp,
dropout=edge_mlp_dropout,
activation=False)
self.edge_mlp_h = MLP(n_in=self.args.n_hidden,
n_out=n_edge_mlp,
dropout=edge_mlp_dropout,
activation=False)
self.label_mlp_d = MLP(n_in=self.args.n_hidden,
n_out=n_label_mlp,
dropout=label_mlp_dropout,
activation=False)
self.label_mlp_h = MLP(n_in=self.args.n_hidden,
n_out=n_label_mlp,
dropout=label_mlp_dropout,
activation=False)
self.edge_attn = Biaffine(n_in=n_edge_mlp,
n_out=2,
bias_x=True,
bias_y=True)
self.label_attn = Biaffine(n_in=n_label_mlp,
n_out=n_labels,
bias_x=True,
bias_y=True)
self.criterion = nn.CrossEntropyLoss()
def __init__(self,
n_words,
n_labels,
n_tags=None,
n_chars=None,
encoder='lstm',
feat=['char'],
n_embed=100,
n_pretrained=100,
n_feat_embed=100,
n_char_embed=50,
n_char_hidden=100,
char_pad_index=0,
elmo='original_5b',
elmo_bos_eos=(True, True),
bert=None,
n_bert_layers=4,
mix_dropout=.0,
bert_pooling='mean',
bert_pad_index=0,
freeze=True,
embed_dropout=.33,
n_lstm_hidden=400,
n_lstm_layers=3,
encoder_dropout=.33,
n_span_mlp=500,
n_pair_mlp=100,
n_label_mlp=100,
mlp_dropout=.33,
inference='mfvi',
max_iter=3,
interpolation=0.1,
pad_index=0,
unk_index=1,
**kwargs):
super().__init__(**Config().update(locals()))
self.span_mlp_l = MLP(n_in=self.args.n_hidden,
n_out=n_span_mlp,
dropout=mlp_dropout)
self.span_mlp_r = MLP(n_in=self.args.n_hidden,
n_out=n_span_mlp,
dropout=mlp_dropout)
self.pair_mlp_l = MLP(n_in=self.args.n_hidden,
n_out=n_pair_mlp,
dropout=mlp_dropout)
self.pair_mlp_r = MLP(n_in=self.args.n_hidden,
n_out=n_pair_mlp,
dropout=mlp_dropout)
self.pair_mlp_b = MLP(n_in=self.args.n_hidden,
n_out=n_pair_mlp,
dropout=mlp_dropout)
self.label_mlp_l = MLP(n_in=self.args.n_hidden,
n_out=n_label_mlp,
dropout=mlp_dropout)
self.label_mlp_r = MLP(n_in=self.args.n_hidden,
n_out=n_label_mlp,
dropout=mlp_dropout)
self.span_attn = Biaffine(n_in=n_span_mlp, bias_x=True, bias_y=False)
self.pair_attn = Triaffine(n_in=n_pair_mlp, bias_x=True, bias_y=False)
self.label_attn = Biaffine(n_in=n_label_mlp,
n_out=n_labels,
bias_x=True,
bias_y=True)
self.inference = (ConstituencyMFVI if inference == 'mfvi' else
ConstituencyLBP)(max_iter)
self.criterion = nn.CrossEntropyLoss()
def __init__(self,
n_words,
n_feats,
n_rels,
encoder='lstm',
feat='char',
n_embed=100,
n_feat_embed=100,
n_char_embed=50,
bert=None,
n_bert_layers=4,
mix_dropout=.0,
embed_dropout=.33,
n_lstm_hidden=400,
n_lstm_layers=3,
n_att_layers=6,
lstm_dropout=.33,
n_mlp_arc=500,
n_mlp_rel=100,
mlp_dropout=.33,
feat_pad_index=0,
pad_index=0,
unk_index=1,
**kwargs):
super().__init__()
self.args = Config().update(locals())
self.word_embed = nn.Embedding(num_embeddings=n_words,
embedding_dim=n_embed)
# can trained embed:word in train
self.feat_embed = CharLSTM(n_chars=n_feats,
n_embed=n_char_embed,
n_out=n_feat_embed,
pad_index=feat_pad_index)
self.embed_dropout = IndependentDropout(p=embed_dropout)
# 输入层的dropout，采用独立dropout
self.encoder_type=encoder
if(encoder=='lstm'):
self.encoder = BiLSTM(input_size=n_embed + n_feat_embed,
hidden_size=n_lstm_hidden,
num_layers=n_lstm_layers,
dropout=lstm_dropout)
self.lstm_dropout = SharedDropout(p=lstm_dropout)
# 编码层lstm以及shared dropout
elif(encoder=='att'):
d_input=n_embed + n_feat_embed
self.linear1=nn.Linear(d_input,n_lstm_hidden * 2) # 前加
self.encoder=Attention_encoder(d_model=n_lstm_hidden * 2,n_layers=n_att_layers)
# self.linear2=nn.Linear(512,n_lstm_hidden * 2,bias=False) # 后加
self.mlp_arc_d = MLP(n_in=n_lstm_hidden * 2,
n_out=n_mlp_arc,
dropout=mlp_dropout)
self.mlp_arc_h = MLP(n_in=n_lstm_hidden * 2,
n_out=n_mlp_arc,
dropout=mlp_dropout)
self.mlp_rel_d = MLP(n_in=n_lstm_hidden * 2,
n_out=n_mlp_rel,
dropout=mlp_dropout)
self.mlp_rel_h = MLP(n_in=n_lstm_hidden * 2,
n_out=n_mlp_rel,
dropout=mlp_dropout)
# 四个不同的全连接层，映射到对应的维度
self.arc_attn = Biaffine(n_in=n_mlp_arc,
bias_x=True,
bias_y=False)
self.rel_attn = Biaffine(n_in=n_mlp_rel,
n_out=n_rels,
bias_x=True,
bias_y=True)
self.criterion = nn.CrossEntropyLoss()
self.pad_index = pad_index
self.unk_index = unk_index
def __init__(self,
n_words,
n_labels,
n_tags=None,
n_chars=None,
n_lemmas=None,
encoder='lstm',
feat=['tag', 'char', 'lemma'],
n_embed=100,
n_pretrained=125,
n_feat_embed=100,
n_char_embed=50,
n_char_hidden=100,
char_pad_index=0,
char_dropout=0,
elmo='original_5b',
elmo_bos_eos=(True, False),
bert=None,
n_bert_layers=4,
mix_dropout=.0,
bert_pooling='mean',
bert_pad_index=0,
freeze=True,
embed_dropout=.2,
n_lstm_hidden=600,
n_lstm_layers=3,
encoder_dropout=.33,
n_edge_mlp=600,
n_pair_mlp=150,
n_label_mlp=600,
edge_mlp_dropout=.25,
pair_mlp_dropout=.25,
label_mlp_dropout=.33,
inference='mfvi',
max_iter=3,
interpolation=0.1,
pad_index=0,
unk_index=1,
**kwargs):
super().__init__(**Config().update(locals()))
self.edge_mlp_d = MLP(n_in=self.args.n_hidden,
n_out=n_edge_mlp,
dropout=edge_mlp_dropout,
activation=False)
self.edge_mlp_h = MLP(n_in=self.args.n_hidden,
n_out=n_edge_mlp,
dropout=edge_mlp_dropout,
activation=False)
self.pair_mlp_d = MLP(n_in=self.args.n_hidden,
n_out=n_pair_mlp,
dropout=pair_mlp_dropout,
activation=False)
self.pair_mlp_h = MLP(n_in=self.args.n_hidden,
n_out=n_pair_mlp,
dropout=pair_mlp_dropout,
activation=False)
self.pair_mlp_g = MLP(n_in=self.args.n_hidden,
n_out=n_pair_mlp,
dropout=pair_mlp_dropout,
activation=False)
self.label_mlp_d = MLP(n_in=self.args.n_hidden,
n_out=n_label_mlp,
dropout=label_mlp_dropout,
activation=False)
self.label_mlp_h = MLP(n_in=self.args.n_hidden,
n_out=n_label_mlp,
dropout=label_mlp_dropout,
activation=False)
self.edge_attn = Biaffine(n_in=n_edge_mlp, bias_x=True, bias_y=True)
self.sib_attn = Triaffine(n_in=n_pair_mlp, bias_x=True, bias_y=True)
self.cop_attn = Triaffine(n_in=n_pair_mlp, bias_x=True, bias_y=True)
self.grd_attn = Triaffine(n_in=n_pair_mlp, bias_x=True, bias_y=True)
self.label_attn = Biaffine(n_in=n_label_mlp,
n_out=n_labels,
bias_x=True,
bias_y=True)
self.inference = (SemanticDependencyMFVI if inference == 'mfvi' else
SemanticDependencyLBP)(max_iter)
self.criterion = nn.CrossEntropyLoss()
def __init__(self,
n_words,
n_labels,
n_tags=None,
n_chars=None,
n_lemmas=None,
feat='tag,char,lemma',
n_embed=100,
n_embed_proj=125,
n_feat_embed=100,
n_char_embed=50,
char_pad_index=0,
bert=None,
n_bert_layers=4,
mix_dropout=.0,
bert_pad_index=0,
embed_dropout=.2,
n_lstm_hidden=600,
n_lstm_layers=3,
lstm_dropout=.33,
n_mlp_edge=600,
n_mlp_label=600,
edge_mlp_dropout=.25,
label_mlp_dropout=.33,
interpolation=0.1,
pad_index=0,
unk_index=1,
**kwargs):
super().__init__()
self.args = Config().update(locals())
# the embedding layer
self.word_embed = nn.Embedding(num_embeddings=n_words,
embedding_dim=n_embed)
self.embed_proj = nn.Linear(n_embed, n_embed_proj)
self.n_input = n_embed + n_embed_proj
if 'tag' in feat:
self.tag_embed = nn.Embedding(num_embeddings=n_tags,
embedding_dim=n_feat_embed)
self.n_input += n_feat_embed
if 'char' in feat:
self.char_embed = CharLSTM(n_chars=n_chars,
n_embed=n_char_embed,
n_out=n_feat_embed,
pad_index=char_pad_index)
self.n_input += n_feat_embed
if 'lemma' in feat:
self.lemma_embed = nn.Embedding(num_embeddings=n_lemmas,
embedding_dim=n_feat_embed)
self.n_input += n_feat_embed
if 'bert' in feat:
self.bert_embed = BertEmbedding(model=bert,
n_layers=n_bert_layers,
pad_index=bert_pad_index,
dropout=mix_dropout)
self.n_input += self.bert_embed.n_out
self.embed_dropout = IndependentDropout(p=embed_dropout)
# the lstm layer
self.lstm = LSTM(input_size=self.n_input,
hidden_size=n_lstm_hidden,
num_layers=n_lstm_layers,
bidirectional=True,
dropout=lstm_dropout)
self.lstm_dropout = SharedDropout(p=lstm_dropout)
# the MLP layers
self.mlp_edge_d = MLP(n_in=n_lstm_hidden*2, n_out=n_mlp_edge, dropout=edge_mlp_dropout, activation=False)
self.mlp_edge_h = MLP(n_in=n_lstm_hidden*2, n_out=n_mlp_edge, dropout=edge_mlp_dropout, activation=False)
self.mlp_label_d = MLP(n_in=n_lstm_hidden*2, n_out=n_mlp_label, dropout=label_mlp_dropout, activation=False)
self.mlp_label_h = MLP(n_in=n_lstm_hidden*2, n_out=n_mlp_label, dropout=label_mlp_dropout, activation=False)
# the Biaffine layers
self.edge_attn = Biaffine(n_in=n_mlp_edge, n_out=2, bias_x=True, bias_y=True)
self.label_attn = Biaffine(n_in=n_mlp_label, n_out=n_labels, bias_x=True, bias_y=True)
self.criterion = nn.CrossEntropyLoss()
self.interpolation = interpolation
self.pad_index = pad_index
self.unk_index = unk_index
def __init__(self,
n_words,
n_feats,
n_labels,
feat='char',
n_embed=100,
n_feat_embed=100,
n_char_embed=50,
bert=None,
n_bert_layers=4,
mix_dropout=.0,
embed_dropout=.33,
n_lstm_hidden=600,
n_lstm_layers=3,
lstm_dropout=.33,
n_mlp_edge=600,
n_mlp_label=600,
mlp_dropout=.33,
feat_pad_index=0,
pad_index=0,
unk_index=1,
**kwargs):
super().__init__()
self.args = Config().update(locals())
# the embedding layer
self.word_embed = nn.Embedding(num_embeddings=n_words,
embedding_dim=n_embed)
# the linear to transform 100d glove to 125d
self.glove_linear = nn.Linear(100, n_embed) # 用的glove_100d
if feat == 'char':
self.feat_embed = CharLSTM(n_chars=n_feats,
n_embed=n_char_embed,
n_out=n_feat_embed,
pad_index=feat_pad_index)
elif feat == 'bert':
self.feat_embed = BertEmbedding(model=bert,
n_layers=n_bert_layers,
n_out=n_feat_embed,
pad_index=feat_pad_index,
dropout=mix_dropout)
self.n_feat_embed = self.feat_embed.n_out
elif feat == 'tag':
self.feat_embed = nn.Embedding(num_embeddings=n_feats,
embedding_dim=n_feat_embed)
else:
raise RuntimeError(
"The feat type should be in ['char', 'bert', 'tag'].")
self.embed_dropout = IndependentDropout(p=embed_dropout)
# the lstm layer
self.lstm = LSTM(input_size=n_embed + n_feat_embed,
hidden_size=n_lstm_hidden,
num_layers=n_lstm_layers,
bidirectional=True,
dropout=lstm_dropout)
self.lstm_dropout = SharedDropout(p=lstm_dropout)
# the MLP layers
self.mlp_edge_d = MLP(n_in=n_lstm_hidden * 2,
n_out=n_mlp_edge,
dropout=mlp_dropout)
self.mlp_edge_h = MLP(n_in=n_lstm_hidden * 2,
n_out=n_mlp_edge,
dropout=mlp_dropout)
self.mlp_label_d = MLP(n_in=n_lstm_hidden * 2,
n_out=n_mlp_label,
dropout=mlp_dropout)
self.mlp_label_h = MLP(n_in=n_lstm_hidden * 2,
n_out=n_mlp_label,
dropout=mlp_dropout)
# the Biaffine layers
self.edge_attn = Biaffine(n_in=n_mlp_edge,
n_out=2,
bias_x=True,
bias_y=True)
self.label_attn = Biaffine(n_in=n_mlp_label,
n_out=n_labels,
bias_x=True,
bias_y=True)
self.criterion = nn.CrossEntropyLoss()
self.pad_index = pad_index
self.unk_index = unk_index
