def __init__(self, config):
"""
Args:
config(Config):
"""
super(ModelBaseLineSQC, self).__init__(config)
# self.config = config
# embedding
self.bert = load_bert(config.bert_config)
self.embed_context = Embedding(self.bert,
config.bert_config.max_position_embeddings,
config.bert_config.hidden_size)
self.embed_question = Embedding(self.bert, config.max_query_length,
config.bert_config.hidden_size)
# encoder
self.attention_direction = config.direction
self.encoder = Encoder(config.encoder_hidden_layer_number,
config.bert_config.hidden_size,
config.bert_config.max_position_embeddings,
config.encoder_intermediate_dim,
config.attention_head_num,
config.attention_droup_out,
config.attention_use_bias,
bi_direction_attention=config.bi_direction_attention,
max_query_position=config.max_query_length,
attention_direction=config.direction
)
# pointer
self.pointer_linear = torch.nn.Linear(config.bert_config.hidden_size, 2)
self.query_pointor_linear = torch.nn.Linear(config.bert_config.hidden_size, int(512 * 2 / config.max_query_length))
def __init__(self, config):
super(ModelCLS, self).__init__(config)
self.bert = load_bert(config.bert_config)
self.embed_context = Embedding(self.bert,
config.bert_config.max_position_embeddings,
config.bert_config.hidden_size)
self.embed_question = Embedding(self.bert, config.max_query_length,
config.bert_config.hidden_size)
self.start_pointer = torch.nn.Linear(self.config.bert_config.hidden_size * 2,
self.config.bert_config.max_position_embeddings)
self.end_pointer = torch.nn.Linear(self.config.bert_config.hidden_size * 2,
self.config.bert_config.max_position_embeddings)
def __init__(self, config):
"""
Args:
config(Config):
"""
super(ModelBaseLineNER, self).__init__(config)
# self.config = config
# embedding
self.bert = load_bert(config.bert_config)
self.embed_word = Embedding(self.bert,
config.bert_config.max_position_embeddings,
config.bert_config.hidden_size)
# encoder
self.encoder = Encoder(config.encoder_hidden_layer_number,
config.bert_config.hidden_size,
config.bert_config.max_position_embeddings,
config.encoder_intermediate_dim,
config.attention_head_num,
config.attention_droup_out,
config.attention_use_bias)
self.rnn = torch.nn.LSTM(config.bert_config.hidden_size,
config.lstm_hidden_size,
num_layers=config.lstm_layer_num,
batch_first=config.lstm_batch_first,
bidirectional=config.lstm_bi_direction)
self.rnn_linear = torch.nn.Linear(self.config.lstm_hidden_size * (2 if self.config.lstm_bi_direction else 1),
self.config.crf_target_size + 2)
self.rnn_noraml = torch.nn.LayerNorm((self.config.bert_config.max_position_embeddings,
self.config.crf_target_size + 2))
self.crf = CRF(self.config.crf_target_size, self.config.device=="cuda", self.config.crf_average_batch)
# pointer
self.pointer_linear = torch.nn.Linear(config.bert_config.hidden_size, self.config.crf_target_size + 2)