def __init__(self,
vector_size,
num_label,
rnn_layers=3,
arc_mlp_size=500,
label_mlp_size=100,
dropout=0.3,
use_greedy_infer=False):
super(BiaffineParser, self).__init__()
rnn_out_size = vector_size
self.timestep_drop = TimestepDropout(dropout)
self.mlp = nn.Sequential(nn.Linear(rnn_out_size, arc_mlp_size * 2 + label_mlp_size * 2),
nn.LeakyReLU(0.1),
TimestepDropout(p=dropout),)
self.arc_mlp_size = arc_mlp_size
self.label_mlp_size = label_mlp_size
self.arc_predictor = ArcBiaffine(arc_mlp_size, bias=True)
self.label_predictor = LabelBilinear(label_mlp_size, label_mlp_size, num_label, bias=True)
self.use_greedy_infer = use_greedy_infer
self.reset_parameters()
self.dropout = dropout
self.num_label = num_label
def __init__(self, embed, num_label, arc_mlp_size=500, label_mlp_size=100, dropout=0.5, use_greedy_infer=False, app_index=0):
super(BiaffineParser, self).__init__()
self.embed = embed
self.mlp = nn.Sequential(nn.Linear(self.embed.embed_size, arc_mlp_size * 2 + label_mlp_size * 2),
nn.LeakyReLU(0.1),
TimestepDropout(p=dropout),)
self.arc_mlp_size = arc_mlp_size
self.label_mlp_size = label_mlp_size
self.arc_predictor = ArcBiaffine(arc_mlp_size, bias=True)
self.label_predictor = LabelBilinear(label_mlp_size, label_mlp_size, num_label, bias=True)
self.use_greedy_infer = use_greedy_infer
self.reset_parameters()
self.app_index = app_index
self.num_label = num_label
if self.app_index != 0:
raise ValueError("现在app_index必须等于0")
self.dropout = nn.Dropout(dropout)
def __init__(self, char_vocab_size,
emb_dim,
bigram_vocab_size,
trigram_vocab_size,
num_label,
rnn_layers=3,
rnn_hidden_size=800, #单向的数量
arc_mlp_size=500,
label_mlp_size=100,
dropout=0.3,
encoder='lstm',
use_greedy_infer=False,
app_index = 0,
pre_chars_embed=None,
pre_bigrams_embed=None,
pre_trigrams_embed=None):
super(BiaffineParser, self).__init__()
rnn_out_size = 2 * rnn_hidden_size
self.char_embed = Embedding((char_vocab_size, emb_dim))
self.bigram_embed = Embedding((bigram_vocab_size, emb_dim))
self.trigram_embed = Embedding((trigram_vocab_size, emb_dim))
if pre_chars_embed:
self.pre_char_embed = Embedding(pre_chars_embed)
self.pre_char_embed.requires_grad = False
if pre_bigrams_embed:
self.pre_bigram_embed = Embedding(pre_bigrams_embed)
self.pre_bigram_embed.requires_grad = False
if pre_trigrams_embed:
self.pre_trigram_embed = Embedding(pre_trigrams_embed)
self.pre_trigram_embed.requires_grad = False
self.timestep_drop = TimestepDropout(dropout)
self.encoder_name = encoder
if encoder == 'var-lstm':
self.encoder = VarLSTM(input_size=emb_dim*3,
hidden_size=rnn_hidden_size,
num_layers=rnn_layers,
bias=True,
batch_first=True,
input_dropout=dropout,
hidden_dropout=dropout,
bidirectional=True)
elif encoder == 'lstm':
self.encoder = nn.LSTM(input_size=emb_dim*3,
hidden_size=rnn_hidden_size,
num_layers=rnn_layers,
bias=True,
batch_first=True,
dropout=dropout,
bidirectional=True)
else:
raise ValueError('unsupported encoder type: {}'.format(encoder))
self.mlp = nn.Sequential(nn.Linear(rnn_out_size, arc_mlp_size * 2 + label_mlp_size * 2),
nn.LeakyReLU(0.1),
TimestepDropout(p=dropout),)
self.arc_mlp_size = arc_mlp_size
self.label_mlp_size = label_mlp_size
self.arc_predictor = ArcBiaffine(arc_mlp_size, bias=True)
self.label_predictor = LabelBilinear(label_mlp_size, label_mlp_size, num_label, bias=True)
self.use_greedy_infer = use_greedy_infer
self.reset_parameters()
self.dropout = dropout
self.app_index = app_index
self.num_label = num_label
if self.app_index != 0:
raise ValueError("现在app_index必须等于0")