def __init__(self, config, print_info: bool = True):
super(NNCRF_sl, self).__init__()
self.device = config.device
self.encoder = BiLSTMEncoder(config, print_info=print_info)
self.inferencer = LinearCRF(config, print_info=print_info)
self.label2idx = config.label2idx
self.idx2word = config.idx2word
self.idx2labels = config.idx2labels
self.Oid = self.label2idx['O']
self.padid = self.label2idx['<PAD>']
self.startid = self.label2idx['<START>']
self.stopid = self.label2idx['<STOP>']
self.pos_dic, self.type_dic = gen_dic(config.label2idx.keys(),
self.label2idx)
self.tags_num = len(self.idx2labels)
e_type, pos = gen_embedding_table(self.idx2labels, self.type_dic,
self.pos_dic)
self.type_embedding = torch.nn.Embedding(
self.tags_num,
self.tags_num).from_pretrained(e_type,
freeze=True).cuda(self.device)
self.pos_embedding = torch.nn.Embedding(
self.tags_num,
self.tags_num).from_pretrained(pos, freeze=True).cuda(self.device)
def __init__(self, config: Config, print_info: bool = True):
super(NNCRF, self).__init__()
self.device = config.device
self.encoder = BiLSTMEncoder(config, print_info=print_info)
self.inferencer = None
if config.use_crf_layer:
self.inferencer = LinearCRF(config, print_info=print_info)
def __init__(self, config):
super().__init__(config)
if config.text_encoder == 'bilstm':
encoder = BiLSTMEncoder(config)
if config.text_encoder in ['bert', 'bert_freeze']:
encoder = BERTEncoder(config)
if not config.remove_name:
self.name_module = NameModule(config, encoder)
if not config.remove_para or not config.remove_img:
self.cross_modal_module = CrossModalModule(config, encoder)
if config.fusion == 'concatenation':
self.fc = nn.Linear(config.feature_dim, config.label_num)
elif config.fusion == 'attention': # single head (only 3 modules)
self.ent_Q = nn.Parameter(
self.weight_init(config.attention_dim, 'uniform'))
self.ent_K = nn.Linear(config.module_embedding_dim,
config.attention_dim)
self.ent_V = nn.Linear(config.module_embedding_dim,
config.module_embedding_dim)
self.fc = nn.Linear(config.module_embedding_dim, config.label_num)
else:
self.fc = nn.ModuleList([
nn.Linear(config.module_embedding_dim, config.label_num)
for i in range(3)
])
self.dropout = nn.Dropout(config.dropout_rate)
def __init__(self, config):
super().__init__()
if config.encoder == 'textcnn':
self.encoder = TextCNNEncoder(config)
if config.encoder == 'bilstm':
self.encoder = BiLSTMEncoder(config)
if config.encoder in ['bert', 'bert_freeze']:
self.encoder = BERTEncoder(config)