模型的整体结构

相关代码

import torch
import torch.nn as nn
from transformers import BertModel, BertPreTrainedModel


class FCLayer(nn.Module):
    def __init__(self, input_dim, output_dim, dropout_rate=0.0, use_activation=True):
        super(FCLayer, self).__init__()
        self.use_activation = use_activation
        self.dropout = nn.Dropout(dropout_rate)
        self.linear = nn.Linear(input_dim, output_dim)
        self.tanh = nn.Tanh()

    def forward(self, x):
        x = self.dropout(x)
        if self.use_activation:
            x = self.tanh(x)
        return self.linear(x)


class RBERT(BertPreTrainedModel):
    def __init__(self, config, args):
        super(RBERT, self).__init__(config)
        self.bert = BertModel(config=config)  # Load pretrained bert

        self.num_labels = config.num_labels

        self.cls_fc_layer = FCLayer(config.hidden_size, config.hidden_size, args.dropout_rate)
        self.entity_fc_layer = FCLayer(config.hidden_size, config.hidden_size, args.dropout_rate)
        self.label_classifier = FCLayer(
            config.hidden_size * 3,
            config.num_labels,
            args.dropout_rate,
            use_activation=False,
        )

    @staticmethod
    def entity_average(hidden_output, e_mask):
        """
        Average the entity hidden state vectors (H_i ~ H_j)
        :param hidden_output: [batch_size, j-i+1, dim]
        :param e_mask: [batch_size, max_seq_len]
                e.g. e_mask[0] == [0, 0, 0, 1, 1, 1, 0, 0, ... 0]
        :return: [batch_size, dim]
        """
        e_mask_unsqueeze = e_mask.unsqueeze(1)  # [b, 1, j-i+1]
        length_tensor = (e_mask != 0).sum(dim=1).unsqueeze(1)  # [batch_size, 1]

        # [b, 1, j-i+1] * [b, j-i+1, dim] = [b, 1, dim] -> [b, dim]
        sum_vector = torch.bmm(e_mask_unsqueeze.float(), hidden_output).squeeze(1)
        avg_vector = sum_vector.float() / length_tensor.float()  # broadcasting
        return avg_vector

    def forward(self, input_ids, attention_mask, token_type_ids, labels, e1_mask, e2_mask):
        outputs = self.bert(
            input_ids, attention_mask=attention_mask, token_type_ids=token_type_ids
        )  # sequence_output, pooled_output, (hidden_states), (attentions)
        sequence_output = outputs[0]
        pooled_output = outputs[1]  # [CLS]

        # Average
        e1_h = self.entity_average(sequence_output, e1_mask)
        e2_h = self.entity_average(sequence_output, e2_mask)

        # Dropout -> tanh -> fc_layer (Share FC layer for e1 and e2)
        pooled_output = self.cls_fc_layer(pooled_output)
        e1_h = self.entity_fc_layer(e1_h)
        e2_h = self.entity_fc_layer(e2_h)

        # Concat -> fc_layer
        concat_h = torch.cat([pooled_output, e1_h, e2_h], dim=-1)
        logits = self.label_classifier(concat_h)

        outputs = (logits,) + outputs[2:]  # add hidden states and attention if they are here

        # Softmax
        if labels is not None:
            if self.num_labels == 1:
                loss_fct = nn.MSELoss()
                loss = loss_fct(logits.view(-1), labels.view(-1))
            else:
                loss_fct = nn.CrossEntropyLoss()
                loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))

            outputs = (loss,) + outputs

        return outputs  # (loss), logits, (hidden_states), (attentions)

代码解析

• 首先我们来看RBERT类，它继承了BertPreTrainedModel类，在类初始化的时候要传入两个参数：config和args，config是模型相关的，args是其它的一些配置。
• 假设输入的input_ids, attention_mask, token_type_ids, labels, e1_mask, e2_mask的维度分别是：(16表示的是batchsize的大小，384表示的是设置的句子的最大长度)
  input_ids.shape= torch.Size([16, 384])
  attention_mask.shape= torch.Size([16, 384])
  token_type_ids.shape= torch.Size([16, 384])
  labels.shape= torch.Size([16])
  e1_mask.shape= torch.Size([16, 384])
  e2_mask.shape= torch.Size([16, 384])
  经过原始的bert之后得到output，其中outputs[0]的维度是[16,384,768]，也就是每一个句子的表示，outputs[1]表示的是经过池化之后的句子表示，维度是[16,768]，意思是将384个字的每个维度的特征通过池化将信息聚合在一起。
• 对于sequence_output, e1_mask或者sequence_output, e2_mask，我们将他们分别传入到entity_averag函数中，针对于e1_mask或者e2_mask，他们的维度都是[16,384],然后进行变换为[16,1,384]，通过将[16,1,384]和[16,384,768]进行矩阵相乘，就得到了实体的特征表示，维度是[16,1,768]，去除掉第1维再除以实体的长度进行归一化，最终得到一个[16,768]的表示。
• 我们将cls，也就是outputs[1]，和实体1以及实体2的特征表示进行拼接，得到一个维度为[16,2304]的张量，再经过一个全连接层映射成[16,19]，这里的19是类别的数目，最后使用相关的损失函数计算损失即可。

使用

最后是这么使用的：

定义相关参数以及设置

        self.args = args
        self.train_dataset = train_dataset
        self.dev_dataset = dev_dataset
        self.test_dataset = test_dataset

        self.label_lst = get_label(args)
        self.num_labels = len(self.label_lst)

        self.config = BertConfig.from_pretrained(
            args.model_name_or_path,
            num_labels=self.num_labels,
            finetuning_task=args.task,
            id2label={str(i): label for i, label in enumerate(self.label_lst)},
            label2id={label: i for i, label in enumerate(self.label_lst)},
        )
        self.model = RBERT.from_pretrained(args.model_name_or_path, config=self.config, args=args)
        # GPU or CPU
        self.device = "cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu"
        self.model.to(self.device)

代码来源：https://github.com/monologg/R-BERT/