NER

# 代码与公式的结合理解
## 字嵌入层
```python
self.char_embedding = nn.Embedding(args.char_alphabet_size, self.char_emb_dim)
```
## 词嵌入层
```python
        if self.use_edge:
            # word embedding
            self.word_embedding = nn.Embedding(args.word_alphabet_size, self.word_emb_dim)
            if data.pretrain_word_embedding is not None:
                scale = np.sqrt(3.0 / self.word_emb_dim)
                data.pretrain_word_embedding[0, :] = np.random.uniform(-scale, scale, [1, self.word_emb_dim])
                self.word_embedding.weight.data.copy_(torch.from_numpy(data.pretrain_word_embedding))
            # bmes embedding
            self.bmes_embedding = nn.Embedding(4, self.bmes_dim)
```
## LSTM层
```python
self.emb_rnn_f = nn.LSTM(self.char_emb_dim, self.hidden_dim, batch_first=True)
        self.emb_rnn_b = nn.LSTM(self.char_emb_dim, self.hidden_dim, batch_first=True)
        # length embedding
        self.length_embedding = nn.Embedding(self.max_word_length, self.length_dim)
        self.dropout = nn.Dropout(self.emb_dropout_rate)
        self.norm = nn.LayerNorm(self.hidden_dim)

```
## 节点聚合
```python
self.edge2node_f = nn.ModuleList(
                [MultiHeadAtt(self.hidden_dim, self.hidden_dim * 2 + self.length_dim,
                              nhead=self.num_head, head_dim=self.head_dim, dropout=self.tf_dropout_rate)
                 for _ in range(self.iters)])
```
## 边聚合
```python
self.node2edge_f = nn.ModuleList(
                [MultiHeadAtt(self.hidden_dim, self.hidden_dim + self.bmes_dim, nhead=self.num_head,
                              head_dim=self.head_dim, dropout=self.tf_dropout_rate)
                 for _ in range(self.iters)])
```
MultiHeadAtt定义与module.py文件中，包含四个卷积层，一个Dropout和一个LayerNorm层
## 全局聚合
```python
self.glo_att_f_node = nn.ModuleList(
                [GloAtt(self.hidden_dim, nhead=self.num_head, head_dim=self.head_dim, dropout=self.tf_dropout_rate)
                 for _ in range(self.iters)])

            if self.use_edge:
                self.glo_att_f_edge = nn.ModuleList(
                    [GloAtt(self.hidden_dim, nhead=self.num_head, head_dim=self.head_dim, dropout=self.tf_dropout_rate)
                     for _ in range(self.iters)])

            # Updating modules
            if self.use_edge:
                self.glo_rnn_f = Global_Cell(self.hidden_dim * 3, self.hidden_dim, dropout=self.cell_dropout_rate)
                self.node_rnn_f = Nodes_Cell(self.hidden_dim * 5, self.hidden_dim, dropout=self.cell_dropout_rate)
                self.edge_rnn_f = Edges_Cell(self.hidden_dim * 4, self.hidden_dim, dropout=self.cell_dropout_rate)
            else:
                self.glo_rnn_f = Global_Cell(self.hidden_dim * 2, self.hidden_dim, dropout=self.cell_dropout_rate)
                self.node_rnn_f = Nodes_Cell(self.hidden_dim * 4, self.hidden_dim, dropout=self.cell_dropout_rate)
```
## 可选的CRF层
```python
        if self.use_crf:
            self.hidden2tag = nn.Linear(output_dim, self.label_size + 2)
            self.crf = CRF(self.label_size, self.gpu)
        else:
            self.hidden2tag = nn.Linear(output_dim, self.label_size)
            self.criterion = nn.CrossEntropyLoss()
```
CRF类定义与crf.py文件中，代码较多，之后画个图吧