def __init__(self, config):
super(LSRPIPELINE, self).__init__()
self.config = config
self.finetune_emb = config.finetune_emb
self.word_emb = nn.Embedding(config.data_word_vec.shape[0], config.data_word_vec.shape[1])
self.word_emb.weight.data.copy_(torch.from_numpy(config.data_word_vec))
if not self.finetune_emb:
self.word_emb.weight.requires_grad = False
self.ner_emb = nn.Embedding(13, config.entity_type_size, padding_idx=0)
self.coref_embed = nn.Embedding(config.max_length, config.coref_size, padding_idx=0)
hidden_size = config.rnn_hidden
input_size = config.data_word_vec.shape[1] + config.coref_size + config.entity_type_size #+ char_hidden
self.linear_re = nn.Linear(hidden_size * 2, hidden_size)
self.linear_sent = nn.Linear(hidden_size * 2, hidden_size)
self.bili = torch.nn.Bilinear(hidden_size, hidden_size, hidden_size)
self.self_att = SelfAttention(hidden_size)
self.bili = torch.nn.Bilinear(hidden_size+config.dis_size, hidden_size+config.dis_size, hidden_size)
self.dis_embed = nn.Embedding(20, config.dis_size, padding_idx=10)
self.linear_output = nn.Linear(2 * hidden_size, config.relation_num)
self.relu = nn.ReLU()
self.dropout_rate = nn.Dropout(config.dropout_rate)
self.rnn_sent = Encoder(input_size, hidden_size, config.dropout_emb, config.dropout_rate)
self.hidden_size = hidden_size
self.use_struct_att = config.use_struct_att
if self.use_struct_att == True:
self.structInduction = StructInduction(hidden_size // 2, hidden_size, True)
self.dropout_gcn = nn.Dropout(config.dropout_gcn)
self.reasoner_layer_first = config.reasoner_layer_first
self.reasoner_layer_second = config.reasoner_layer_second
self.use_reasoning_block = config.use_reasoning_block
if self.use_reasoning_block:
self.reasoner = nn.ModuleList()
self.reasoner.append(DynamicReasoner(hidden_size, self.reasoner_layer_first, self.dropout_gcn))
self.reasoner.append(DynamicReasoner(hidden_size, self.reasoner_layer_second, self.dropout_gcn))
def __init__(self, config):
super(LSR, self).__init__()
self.config = config
self.bert = BertModel.from_pretrained('bert-base-uncased')
print("loaded bert-base-uncased")
hidden_size = config.rnn_hidden
bert_hidden_size = 768
self.linear_re = nn.Linear(bert_hidden_size, hidden_size)
self.linear_sent = nn.Linear(hidden_size * 2, hidden_size)
self.bili = torch.nn.Bilinear(hidden_size, hidden_size, hidden_size)
self.self_att = SelfAttention(hidden_size)
self.bili = torch.nn.Bilinear(hidden_size+config.dis_size, hidden_size+config.dis_size, hidden_size)
self.dis_embed = nn.Embedding(20, config.dis_size, padding_idx=10)
self.linear_output = nn.Linear(2 * hidden_size, config.relation_num)
self.relu = nn.ReLU()
self.dropout_rate = nn.Dropout(config.dropout_rate)
#self.rnn_sent = Encoder(input_size, hidden_size, config.dropout_emb, config.dropout_rate)
self.hidden_size = hidden_size
self.use_struct_att = config.use_struct_att
if self.use_struct_att == True:
self.structInduction = StructInduction(hidden_size // 2, hidden_size, True)
self.dropout_gcn = nn.Dropout(config.dropout_gcn)
self.reasoner_layer_first = config.reasoner_layer_first
self.reasoner_layer_second = config.reasoner_layer_second
self.use_reasoning_block = config.use_reasoning_block
if self.use_reasoning_block:
self.reasoner = nn.ModuleList()
self.reasoner.append(DynamicReasoner(hidden_size, self.reasoner_layer_first, self.dropout_gcn))
self.reasoner.append(DynamicReasoner(hidden_size, self.reasoner_layer_second, self.dropout_gcn))