def __init__(self, args):
super(Model, self).__init__()
self.args = args
# the embedding layer
self.word_embed = nn.Embedding(num_embeddings=args.n_words,
embedding_dim=args.n_embed)
if args.feat == 'char':
self.feat_embed = CHAR_LSTM(n_chars=args.n_feats,
n_embed=args.n_char_embed,
n_out=args.n_embed)
elif args.feat == 'bert':
self.feat_embed = BertEmbedding(model=args.bert_model,
n_layers=args.n_bert_layers,
n_out=args.n_embed)
else:
self.feat_embed = nn.Embedding(num_embeddings=args.n_feats,
embedding_dim=args.n_embed)
self.embed_dropout = IndependentDropout(p=args.embed_dropout)
# the word-lstm layer
self.lstm = BiLSTM(input_size=args.n_embed * 2,
hidden_size=args.n_lstm_hidden,
num_layers=args.n_lstm_layers,
dropout=args.lstm_dropout)
self.lstm_dropout = SharedDropout(p=args.lstm_dropout)
# the MLP layers
self.mlp_arc_h = MLP(n_in=args.n_lstm_hidden * 2,
n_hidden=args.n_mlp_arc,
dropout=args.mlp_dropout)
self.mlp_arc_d = MLP(n_in=args.n_lstm_hidden * 2,
n_hidden=args.n_mlp_arc,
dropout=args.mlp_dropout)
self.mlp_rel_h = MLP(n_in=args.n_lstm_hidden * 2,
n_hidden=args.n_mlp_rel,
dropout=args.mlp_dropout)
self.mlp_rel_d = MLP(n_in=args.n_lstm_hidden * 2,
n_hidden=args.n_mlp_rel,
dropout=args.mlp_dropout)
# the Biaffine layers
self.arc_attn = Biaffine(n_in=args.n_mlp_arc,
bias_x=True,
bias_y=False)
self.rel_attn = Biaffine(n_in=args.n_mlp_rel,
n_out=args.n_rels,
bias_x=True,
bias_y=True)
self.pad_index = args.pad_index
self.unk_index = args.unk_index
def __init__(self, args):
super(Model, self).__init__()
self.args = args
# the embedding layer
if args.bert is False:
self.word_embed = nn.Embedding(num_embeddings=args.n_words,
embedding_dim=args.word_embed)
if args.freeze_word_emb:
self.word_embed.weight.requires_grad = False
else:
self.word_embed = BertEmbedding(model=args.bert_model,
n_layers=args.n_bert_layers,
n_out=args.word_embed)
self.feat_embed = nn.Embedding(num_embeddings=args.n_feats,
embedding_dim=args.n_embed)
if args.freeze_feat_emb:
self.feat_embed.weight.requires_grad = False
self.embed_dropout = IndependentDropout(p=args.embed_dropout)
# the word-lstm layer
self.lstm = BiLSTM(input_size=args.word_embed + args.n_embed,
hidden_size=args.n_lstm_hidden,
num_layers=args.n_lstm_layers,
dropout=args.lstm_dropout)
self.lstm_dropout = SharedDropout(p=args.lstm_dropout)
# the MLP layers
self.mlp_arc_h = MLP(n_in=args.n_lstm_hidden * 2,
n_hidden=args.n_mlp_arc,
dropout=args.mlp_dropout)
self.mlp_arc_d = MLP(n_in=args.n_lstm_hidden * 2,
n_hidden=args.n_mlp_arc,
dropout=args.mlp_dropout)
# the Biaffine layers
self.arc_attn = Biaffine(n_in=args.n_mlp_arc,
bias_x=True,
bias_y=False)
self.pad_index = args.pad_index
self.unk_index = args.unk_index
self.multinomial = nn.Parameter(torch.ones(args.n_feats, args.n_feats))
def __init__(self, config, embeddings):
super(BiaffineParser, self).__init__()
self.config = config
# the embedding layer
self.pretrained = nn.Embedding.from_pretrained(embeddings)
self.embed = nn.Embedding(num_embeddings=config.n_words,
embedding_dim=config.n_embed)
self.tag_embed = nn.Embedding(num_embeddings=config.n_tags,
embedding_dim=config.n_tag_embed)
self.embed_dropout = IndependentDropout(p=config.embed_dropout)
# the word-lstm layer
self.lstm = BiLSTM(input_size=config.n_embed + config.n_tag_embed,
hidden_size=config.n_lstm_hidden,
num_layers=config.n_lstm_layers,
dropout=config.lstm_dropout)
self.lstm_dropout = SharedDropout(p=config.lstm_dropout)
# the MLP layers
self.mlp_arc_h = MLP(n_in=config.n_lstm_hidden * 2,
n_hidden=config.n_mlp_arc,
dropout=config.mlp_dropout)
self.mlp_arc_d = MLP(n_in=config.n_lstm_hidden * 2,
n_hidden=config.n_mlp_arc,
dropout=config.mlp_dropout)
self.mlp_rel_h = MLP(n_in=config.n_lstm_hidden * 2,
n_hidden=config.n_mlp_rel,
dropout=config.mlp_dropout)
self.mlp_rel_d = MLP(n_in=config.n_lstm_hidden * 2,
n_hidden=config.n_mlp_rel,
dropout=config.mlp_dropout)
# the Biaffine layers
self.arc_attn = Biaffine(n_in=config.n_mlp_arc,
bias_x=True,
bias_y=False)
self.rel_attn = Biaffine(n_in=config.n_mlp_rel,
n_out=config.n_rels,
bias_x=True,
bias_y=True)
self.pad_index = config.pad_index
self.unk_index = config.unk_index
self.reset_parameters()
def __init__(self, args, mask_token_id=0):
super().__init__()
self.args = args
if args.n_embed:
# the embedding layer
self.word_embed = nn.Embedding(num_embeddings=args.n_words,
embedding_dim=args.n_embed)
self.unk_index = args.unk_index
else:
self.word_embed = None
if args.feat == 'char':
self.feat_embed = CharLSTM(n_chars=args.n_feats,
n_embed=args.n_char_embed,
n_out=args.n_feat_embed,
pad_index=args.feat_pad_index)
self.pad_index = args.pad_index
elif args.feat == 'bert':
self.feat_embed = BertEmbedding(model=args.bert_model,
n_layers=args.n_bert_layers,
n_out=args.n_feat_embed,
requires_grad=args.bert_fine_tune,
mask_token_id=mask_token_id,
token_dropout=args.token_dropout,
mix_dropout=args.mix_dropout,
use_hidden_states=args.use_hidden_states,
use_attentions=args.use_attentions,
attention_layer=args.attention_layer)
#self.args.n_mlp_arc = self.feat_embed.bert.config.max_position_embeddings
self.args.n_feat_embed = self.feat_embed.n_out # taken from the model
self.args.n_bert_layers = self.feat_embed.n_layers # taken from the model
self.pad_index = self.feat_embed.pad_index # taken from the model
self.args.pad_index = self.pad_index # update
else:
self.feat_embed = nn.Embedding(num_embeddings=args.n_feats,
embedding_dim=args.n_feat_embed)
self.pad_index = args.pad_index
self.embed_dropout = IndependentDropout(p=args.embed_dropout)
if args.n_lstm_layers:
# the lstm layer
self.lstm = BiLSTM(input_size=args.n_embed+args.n_feat_embed,
hidden_size=args.n_lstm_hidden,
num_layers=args.n_lstm_layers,
dropout=args.lstm_dropout)
self.lstm_dropout = SharedDropout(p=args.lstm_dropout)
mlp_input_size = args.n_lstm_hidden*2
else:
self.lstm = None
mlp_input_size = args.n_embed + args.n_feat_embed
# the MLP layers
self.mlp_arc_d = MLP(n_in=mlp_input_size,
n_out=args.n_mlp_arc,
dropout=args.mlp_dropout)
self.mlp_arc_h = MLP(n_in=mlp_input_size,
n_out=args.n_mlp_arc,
dropout=args.mlp_dropout)
self.mlp_rel_d = MLP(n_in=mlp_input_size,
n_out=args.n_mlp_rel,
dropout=args.mlp_dropout)
self.mlp_rel_h = MLP(n_in=mlp_input_size,
n_out=args.n_mlp_rel,
dropout=args.mlp_dropout)
# the Biaffine layers
self.arc_attn = Biaffine(n_in=args.n_mlp_arc,
bias_x=True,
bias_y=False)
self.rel_attn = Biaffine(n_in=args.n_mlp_rel,
n_out=args.n_rels,
bias_x=True,
bias_y=True)
# transformer attention
if args.use_attentions:
self.attn_mix = nn.Parameter(torch.randn(1)) #2)) # 1))
# # distance
# self.args.distance = False # DEBUG
# if self.args.distance:
# self.distance = DeepBiaffine(mlp_input_size, mlp_input_size, self.args.deep_biaff_hidden_dim, 1, dropout=args.mlp_dropout)
self.criterion = nn.CrossEntropyLoss()
def __init__(self, config, embed):
super(BiaffineParser, self).__init__()
self.config = config
# the embedding layer
self.pretrained = nn.Embedding.from_pretrained(embed)
self.word_embed = nn.Embedding(num_embeddings=config.n_words,
embedding_dim=config.n_embed)
# the char-lstm layer
self.char_lstm = CHAR_LSTM(n_chars=config.n_chars,
n_embed=config.n_char_embed,
n_out=config.n_embed)
self.embed_dropout = IndependentDropout(p=config.embed_dropout)
self.tag_lstm = BiLSTM(input_size=config.n_embed * 2,
hidden_size=config.n_lstm_hidden,
num_layers=config.n_lstm_layers,
dropout=config.lstm_dropout)
self.dep_lstm = BiLSTM(input_size=config.n_embed * 2 +
config.n_mlp_arc,
hidden_size=config.n_lstm_hidden,
num_layers=config.n_lstm_layers,
dropout=config.lstm_dropout)
if config.weight:
self.tag_mix = ScalarMix(n_layers=config.n_lstm_layers)
self.dep_mix = ScalarMix(n_layers=config.n_lstm_layers)
self.lstm_dropout = SharedDropout(p=config.lstm_dropout)
# the MLP layers
self.mlp_tag = MLP(n_in=config.n_lstm_hidden * 2,
n_hidden=config.n_mlp_arc,
dropout=0.5)
self.mlp_dep = MLP(n_in=config.n_lstm_hidden * 2,
n_hidden=config.n_mlp_arc,
dropout=config.mlp_dropout)
self.mlp_arc_h = MLP(n_in=config.n_lstm_hidden * 2,
n_hidden=config.n_mlp_arc,
dropout=config.mlp_dropout)
self.mlp_arc_d = MLP(n_in=config.n_lstm_hidden * 2,
n_hidden=config.n_mlp_arc,
dropout=config.mlp_dropout)
self.mlp_rel_h = MLP(n_in=config.n_lstm_hidden * 2,
n_hidden=config.n_mlp_rel,
dropout=config.mlp_dropout)
self.mlp_rel_d = MLP(n_in=config.n_lstm_hidden * 2,
n_hidden=config.n_mlp_rel,
dropout=config.mlp_dropout)
self.ffn_pos_tag = nn.Linear(config.n_mlp_arc, config.n_pos_tags)
self.ffn_dep_tag = nn.Linear(config.n_mlp_arc, config.n_dep_tags)
# the Biaffine layers
self.arc_attn = Biaffine(n_in=config.n_mlp_arc,
bias_x=True,
bias_y=False)
self.rel_attn = Biaffine(n_in=config.n_mlp_rel,
n_out=config.n_rels,
bias_x=True,
bias_y=True)
self.weight = config.weight
self.pad_index = config.pad_index
self.unk_index = config.unk_index
self.criterion = nn.CrossEntropyLoss()
self.reset_parameters()
def __init__(self, args):
super(Model, self).__init__()
self.args = args
self.pretrained = False
# the embedding layer
self.char_embed = nn.Embedding(num_embeddings=args.n_chars,
embedding_dim=args.n_embed)
n_lstm_input = args.n_embed
if args.feat == 'bert':
self.feat_embed = BertEmbedding(model=args.bert_model,
n_layers=args.n_bert_layers,
n_out=args.n_feat_embed)
n_lstm_input += args.n_feat_embed
if self.args.feat in {'bigram', 'trigram'}:
self.bigram_embed = nn.Embedding(num_embeddings=args.n_bigrams,
embedding_dim=args.n_embed)
n_lstm_input += args.n_embed
if self.args.feat == 'trigram':
self.trigram_embed = nn.Embedding(num_embeddings=args.n_trigrams,
embedding_dim=args.n_embed)
n_lstm_input += args.n_embed
self.embed_dropout = IndependentDropout(p=args.embed_dropout)
# the lstm layer
self.lstm = BiLSTM(input_size=n_lstm_input,
hidden_size=args.n_lstm_hidden,
num_layers=args.n_lstm_layers,
dropout=args.lstm_dropout)
self.lstm_dropout = SharedDropout(p=args.lstm_dropout)
# the MLP layers
self.mlp_span_l = MLP(n_in=args.n_lstm_hidden * 2,
n_out=args.n_mlp_span,
dropout=args.mlp_dropout)
self.mlp_span_r = MLP(n_in=args.n_lstm_hidden * 2,
n_out=args.n_mlp_span,
dropout=args.mlp_dropout)
# the Biaffine layers
self.span_attn = Biaffine(n_in=args.n_mlp_span,
bias_x=True,
bias_y=False)
if args.link == 'mlp':
# a representation that a fencepost is a split point
self.mlp_span_s = MLP(n_in=args.n_lstm_hidden * 2,
n_out=args.n_mlp_span,
dropout=args.mlp_dropout)
# scores for split points
self.score_split = nn.Linear(args.n_mlp_span, 1)
elif args.link == 'att':
self.split_attn = ElementWiseBiaffine(n_in=args.n_lstm_hidden,
bias_x=True,
bias_y=False)
self.pad_index = args.pad_index
self.unk_index = args.unk_index
def __init__(self, args):
super(Model, self).__init__()
self.args = args
# the embedding layer
self.word_embed = nn.Embedding(num_embeddings=args.n_words,
embedding_dim=args.n_embed)
if args.use_char:
self.char_embed = CHAR_LSTM(n_chars=args.n_char_feats,
n_embed=args.n_char_embed,
n_out=args.n_embed)
if args.use_bert:
self.bert_embed = BertEmbedding(model=args.bert_model,
n_layers=args.n_bert_layers,
n_out=args.n_embed)
if args.use_pos:
self.pos_embed = nn.Embedding(num_embeddings=args.n_pos_feats,
embedding_dim=args.n_embed)
self.embed_dropout = IndependentDropout(p=args.embed_dropout)
# the word-lstm layer
self.lstm = BiLSTM(input_size=args.n_embed *
(args.use_char + args.use_bert + args.use_pos + 1),
hidden_size=args.n_lstm_hidden,
num_layers=args.n_lstm_layers,
dropout=args.lstm_dropout)
self.lstm_dropout = SharedDropout(p=args.lstm_dropout)
# the MLP layers
self.mlp_arc_h = MLP(n_in=args.n_lstm_hidden * 2,
n_hidden=args.n_mlp_arc,
dropout=args.mlp_dropout)
self.mlp_arc_d = MLP(n_in=args.n_lstm_hidden * 2,
n_hidden=args.n_mlp_arc,
dropout=args.mlp_dropout)
self.mlp_rel_h = MLP(n_in=args.n_lstm_hidden * 2,
n_hidden=args.n_mlp_rel,
dropout=args.mlp_dropout)
self.mlp_rel_d = MLP(n_in=args.n_lstm_hidden * 2,
n_hidden=args.n_mlp_rel,
dropout=args.mlp_dropout)
# the Biaffine layers
self.arc_attn = Biaffine(n_in=args.n_mlp_arc,
bias_x=True,
bias_y=False)
self.rel_attn = Biaffine(n_in=args.n_mlp_rel,
n_out=args.n_rels,
bias_x=True,
bias_y=True)
self.binary = args.binary
# the Second Order Parts
if self.args.use_second_order:
self.use_sib = args.use_sib
self.use_cop = args.use_cop
self.use_gp = args.use_gp
if args.use_sib:
self.mlp_sib_h = MLP(n_in=args.n_lstm_hidden * 2,
n_hidden=args.n_mlp_sec,
dropout=args.mlp_dropout,
identity=self.binary)
self.mlp_sib_d = MLP(n_in=args.n_lstm_hidden * 2,
n_hidden=args.n_mlp_sec,
dropout=args.mlp_dropout,
identity=self.binary)
self.trilinear_sib = TrilinearScorer(args.n_mlp_sec,
args.n_mlp_sec,
args.n_mlp_sec,
init_std=args.init_std,
rank=args.n_mlp_sec,
factorize=args.factorize)
if args.use_cop:
self.mlp_cop_h = MLP(n_in=args.n_lstm_hidden * 2,
n_hidden=args.n_mlp_sec,
dropout=args.mlp_dropout,
identity=self.binary)
self.mlp_cop_d = MLP(n_in=args.n_lstm_hidden * 2,
n_hidden=args.n_mlp_sec,
dropout=args.mlp_dropout,
identity=self.binary)
self.trilinear_cop = TrilinearScorer(args.n_mlp_sec,
args.n_mlp_sec,
args.n_mlp_sec,
init_std=args.init_std,
rank=args.n_mlp_sec,
factorize=args.factorize)
if args.use_gp:
self.mlp_gp_h = MLP(n_in=args.n_lstm_hidden * 2,
n_hidden=args.n_mlp_sec,
dropout=args.mlp_dropout,
identity=self.binary)
self.mlp_gp_d = MLP(n_in=args.n_lstm_hidden * 2,
n_hidden=args.n_mlp_sec,
dropout=args.mlp_dropout,
identity=self.binary)
self.mlp_gp_hd = MLP(n_in=args.n_lstm_hidden * 2,
n_hidden=args.n_mlp_sec,
dropout=args.mlp_dropout,
identity=self.binary)
self.trilinear_gp = TrilinearScorer(args.n_mlp_sec,
args.n_mlp_sec,
args.n_mlp_sec,
init_std=args.init_std,
rank=args.n_mlp_sec,
factorize=args.factorize)
self.pad_index = args.pad_index
self.unk_index = args.unk_index