def create(
cls,
label_size,
intent_size,
# BertEmbedder params
model_name='bert-base-multilingual-cased',
mode="weighted",
is_freeze=True,
# BiLSTM params
embedding_size=768,
hidden_dim=512,
rnn_layers=1,
lstm_dropout=0.3,
# CRFDecoder params
crf_dropout=0.5,
nbest=1,
# Clf params
clf_dropout=0.3,
# Global params
device="cuda"):
embeddings = BERTEmbedder.create(model_name=model_name,
device=device,
mode=mode,
is_freeze=is_freeze)
lstm = BiLSTM.create(embedding_size=embedding_size,
hidden_dim=hidden_dim,
rnn_layers=rnn_layers,
dropout=lstm_dropout)
crf = NCRFDecoder.create(label_size,
hidden_dim,
crf_dropout,
nbest=nbest,
device=device)
clf = ClassDecoder(intent_size, hidden_dim, clf_dropout)
return cls(embeddings, lstm, crf, clf, device)
def create(
cls,
label_size,
# BertEmbedder params
model_name='bert-base-multilingual-cased',
mode="weighted",
is_freeze=True,
# BiLSTM
hidden_dim=512,
rnn_layers=1,
lstm_dropout=0.3,
# Attn params
embedding_size=768,
key_dim=64,
val_dim=64,
num_heads=3,
attn_dropout=0.3,
# CRFDecoder params
crf_dropout=0.5,
# Global params
device="cuda"):
embeddings = BERTEmbedder.create(model_name=model_name,
device=device,
mode=mode,
is_freeze=is_freeze)
lstm = BiLSTM.create(embedding_size=embedding_size,
hidden_dim=hidden_dim,
rnn_layers=rnn_layers,
dropout=lstm_dropout)
attn = MultiHeadAttention(key_dim, val_dim, hidden_dim, num_heads,
attn_dropout)
crf = CRFDecoder.create(label_size, hidden_dim, crf_dropout)
return cls(embeddings, lstm, attn, crf, device)
def create(
cls,
label_size,
# BertEmbedder params
model_name='bert-base-multilingual-cased',
mode="weighted",
is_freeze=True,
# Pos params
num_pos_dict=18,
pos_embedding_size=18,
# BiLSTM params
embedding_size=768,
hidden_dim=512,
rnn_layers=1,
lstm_dropout=0.3,
# CRFDecoder params
crf_dropout=0.5,
# Global params
device="cuda"):
embeddings = BERTEmbedder.create(model_name=model_name,
device=device,
mode=mode,
is_freeze=is_freeze)
posembeddings = torch.nn.Embedding(num_pos_dict, pos_embedding_size)
lstm = BiLSTM.create(embedding_size=embedding_size +
pos_embedding_size,
hidden_dim=hidden_dim,
rnn_layers=rnn_layers,
dropout=lstm_dropout)
crf = CRFDecoder.create(label_size, hidden_dim, crf_dropout)
return cls(embeddings, posembeddings, lstm, crf, device)
def create(
cls,
# Embedding params
num_char_dict=399,
# BiLSTM params
embedding_size=32,
hidden_dim=64,
rnn_layers=1,
lstm_dropout=0.5,
device="cuda",
mode="weighted",
is_freeze=True):
config = {
"embedding_size": embedding_size,
"hidden_dim": hidden_dim,
"rnn_layers": rnn_layers,
"lstm_dropout": lstm_dropout,
"mode": mode,
"is_freeze": is_freeze
}
charembeddings = torch.nn.Embedding(num_char_dict, embedding_size)
lstm = BiLSTM.create(embedding_size=embedding_size,
hidden_dim=hidden_dim,
rnn_layers=rnn_layers,
dropout=lstm_dropout)
self = cls(charembeddings, lstm, config)
return self
def create(
cls,
label_size,
# BertEmbedder params
model_name='bert-base-multilingual-cased',
mode="weighted",
is_freeze=True,
# Char Encoder params
num_char_dict=399,
# BiLSTM params
embedding_size=768,
hidden_dim=512,
rnn_layers=1,
lstm_dropout=0.5,
# BiLSTM params for char encoders
char_lstm_embedding_size=32,
char_lstm_hidden_dim=32,
char_lstm_rnn_layers=1,
char_lstm_dropout=0.5,
# CRFDecoder params
crf_dropout=0.5,
# Global params
device="cuda"):
embeddings = BERTEmbedder.create(model_name=model_name,
device=device,
mode=mode,
is_freeze=is_freeze)
charencodings = CHAREncoder.create(
num_char_dict=num_char_dict,
embedding_size=char_lstm_embedding_size,
hidden_dim=char_lstm_hidden_dim,
rnn_layers=char_lstm_rnn_layers,
lstm_dropout=char_lstm_dropout,
device=device,
mode=mode,
is_freeze=is_freeze)
lstm = BiLSTM.create(embedding_size=embedding_size +
(char_lstm_hidden_dim * 2),
hidden_dim=hidden_dim,
rnn_layers=rnn_layers,
dropout=lstm_dropout)
crf = CRFDecoder.create(label_size, hidden_dim, crf_dropout)
return cls(embeddings, charencodings, lstm, crf, device)
def create(
cls,
label_size,
# BertEmbedder params
model_name='bert-base-multilingual-cased',
mode="weighted",
is_freeze=True,
# Char Encoder params
num_char_dict=189,
max_token_char_len=30,
# BiLSTM params
embedding_size=768,
hidden_dim=512,
rnn_layers=1,
lstm_dropout=0.5,
# BiLSTM params for char encoders
#char_lstm_embedding_size=5670, char_lstm_rnn_layers=1,
char_lstm_hidden_dim=32,
char_lstm_dropout=0.5,
# CRFDecoder params
crf_dropout=0.5,
# Global params
device="cuda"):
embeddings = BERTEmbedder.create(model_name=model_name,
device=device,
mode=mode,
is_freeze=is_freeze)
chart_lstm = torch.nn.LSTM(num_char_dict * max_token_char_len,
char_lstm_hidden_dim,
bidirectional=True,
batch_first=True,
dropout=char_lstm_dropout)
lstm = BiLSTM.create(embedding_size=embedding_size +
(char_lstm_hidden_dim * 2),
hidden_dim=hidden_dim,
rnn_layers=rnn_layers,
dropout=lstm_dropout)
crf = CRFDecoder.create(label_size, hidden_dim, crf_dropout)
return cls(embeddings, chart_lstm, lstm, crf, device)
