def __init__(self, enc_word_vocab_size, dec_word_vocab_size, act_vocab_size, slot_vocab_size,
emb_dim, hid_dim, dropout, num_layers):
super(SLUSystem, self).__init__()
assert emb_dim == hid_dim
# Initialization
self.hid_dim = hid_dim
# embedding
self.dec_word_emb = Embedding(dec_word_vocab_size, emb_dim, Constants.PAD, dropout)
self.act_emb = Embedding(act_vocab_size, emb_dim, Constants.PAD, dropout)
self.slot_emb = Embedding(slot_vocab_size, emb_dim, Constants.PAD, dropout)
# shared encoder
config = BertConfig(
enc_word_vocab_size,
hidden_size=hid_dim,
num_hidden_layers=num_layers,
num_attention_heads=10,
intermediate_size=4*hid_dim, # 4 * hid_dim
hidden_act='gelu',
max_position_embeddings=50
)
self.encoder = BertModel(config)
# act-slot-value predictors
self.act_stc = STC(hid_dim, act_vocab_size, dropout, emb_dim)
self.slot_stc = STC(hid_dim + emb_dim, slot_vocab_size, dropout, emb_dim)
self.enc_to_dec = Encoder2Decoder(hid_dim, hid_dim)
# the hid dim of decoder is 2 * hid_dim
self.value_decoder = LSTMDecoder(self.dec_word_emb, self.act_emb, self.slot_emb,
dec_word_vocab_size, hid_dim, dropout)
def __init__(self, enc_word_vocab_size, dec_word_vocab_size, act_vocab_size, slot_vocab_size,
emb_dim, hid_dim, dropout):
super(SLUSystem, self).__init__()
# Initialization
self.hid_dim = hid_dim
# embedding
self.word_emb = Embedding(enc_word_vocab_size, emb_dim, Constants.PAD, dropout)
self.act_emb = Embedding(act_vocab_size, emb_dim, Constants.PAD, dropout)
self.slot_emb = Embedding(slot_vocab_size, emb_dim, Constants.PAD, dropout)
# shared encoder
self.encoder_bidirectional = True
if self.encoder_bidirectional:
self.enc_hid_all_dim = 2 * hid_dim
else:
self.enc_hid_all_dim = hid_dim
self.encoder = EncoderRNN('LSTM', self.encoder_bidirectional,
1, hid_dim, self.word_emb, 0.)
# act-slot-value predictors
self.act_stc = STC(self.enc_hid_all_dim, act_vocab_size, dropout, emb_dim)
self.slot_stc = STC(self.enc_hid_all_dim + emb_dim, slot_vocab_size, dropout, emb_dim)
self.enc_to_dec = Encoder2Decoder(self.enc_hid_all_dim, hid_dim)
# the hid dim of decoder is 2 * hid_dim
self.value_decoder = LSTMDecoder(self.word_emb, self.act_emb, self.slot_emb,
dec_word_vocab_size, hid_dim, dropout)
def __init__(self, wcn_emb_type, enc_word_vocab_size, dec_word_vocab_size,
act_vocab_size, slot_vocab_size, emb_dim, hid_dim, dropout):
super(SLUSystem, self).__init__()
# Initialization
self.hid_dim = hid_dim
# embedding
self.dec_word_emb = Embedding(dec_word_vocab_size, emb_dim,
Constants.PAD, dropout)
self.act_emb = Embedding(act_vocab_size, emb_dim, Constants.PAD,
dropout)
self.slot_emb = Embedding(slot_vocab_size, emb_dim, Constants.PAD,
dropout)
# shared encoder
if wcn_emb_type == 'swsr':
self.enc_word_emb = SWSREmbedding(enc_word_vocab_size, emb_dim,
Constants.PAD, dropout)
elif wcn_emb_type == 'mwsr':
self.enc_word_emb = MWSREmbedding(enc_word_vocab_size, emb_dim,
Constants.PAD, dropout)
elif wcn_emb_type == 'savg':
self.enc_word_emb = SAVGEmbedding(enc_word_vocab_size, emb_dim,
Constants.PAD, dropout)
elif wcn_emb_type == 'cavg':
self.enc_word_emb = CAVGEmbedding(enc_word_vocab_size, emb_dim,
Constants.PAD, dropout)
else:
raise Exception('Undefined type of embedding.')
self.encoder_bidirectional = True
if self.encoder_bidirectional:
self.enc_hid_all_dim = 2 * hid_dim
else:
self.enc_hid_all_dim = hid_dim
self.encoder = EncoderRNN('LSTM', self.encoder_bidirectional, 1,
hid_dim, self.enc_word_emb, 0.)
# act-slot-value predictors
self.act_stc = STC(self.enc_hid_all_dim, act_vocab_size, dropout,
emb_dim)
self.slot_stc = STC(self.enc_hid_all_dim + emb_dim, slot_vocab_size,
dropout, emb_dim)
self.enc_to_dec = Encoder2Decoder(self.enc_hid_all_dim, hid_dim)
# the hid dim of decoder is 2 * hid_dim
self.value_decoder = LSTMDecoder(self.dec_word_emb, self.act_emb,
self.slot_emb, dec_word_vocab_size,
hid_dim, dropout)
def __init__(self,
vocab_size,
class_size,
emb_dim=100,
rnn_type='LSTM',
bidirectional=True,
hid_dim=128,
num_layers=1,
dropout=0.5):
super(RNN2One, self).__init__()
self.rnn_type = rnn_type
self.num_layers = num_layers
self.bidirectional = bidirectional
if bidirectional:
num_directions = 2
else:
num_directions = 1
self.lin_dim = num_directions * hid_dim
self.emb = Embedding(vocab_size, emb_dim, Constants.PAD, dropout)
self.rnn = EncoderRNN(rnn_type, bidirectional, num_layers, hid_dim,
self.emb, dropout)
self.dropout = nn.Dropout(dropout)
self.lin = nn.Linear(self.lin_dim, class_size)
self.init_params()
def __init__(self, vocab_size, emb_dim, hid_dim, dropout):
super(S2SPointer, self).__init__()
self.word_emb = Embedding(vocab_size, emb_dim, Constants.PAD, dropout)
self.encoder = EncoderRNN('LSTM', True, 1, hid_dim, self.word_emb, 0.)
self.enc_to_dec = Encoder2Decoder(2 * hid_dim, 2 * hid_dim)
self.decoder = LSTMDecoder(self.word_emb, vocab_size, 2 * hid_dim, dropout)
def __init__(self, enc_word_vocab_size, dec_word_vocab_size, emb_dim,
hid_dim, dropout):
super(DAModel, self).__init__()
# Initialization
self.hid_dim = hid_dim
# embedding
self.enc_word_emb = Embedding(enc_word_vocab_size, emb_dim,
Constants.PAD, dropout)
self.dec_word_emb = Embedding(dec_word_vocab_size, emb_dim,
Constants.PAD, dropout)
# encoder-decoder
self.encoder = EncoderRNN('LSTM', True, 1, hid_dim, self.enc_word_emb,
0.)
self.enc_to_dec = Encoder2Decoder(2 * hid_dim, 2 * hid_dim)
self.decoder = LSTMDecoder(self.dec_word_emb, dec_word_vocab_size,
2 * hid_dim, dropout)