about word_embedding

Source code use nn.Embedding to load pretrained word embedding 
```
nn.Embedding(self.config.data_word_vec.shape[0], self.config.data_word_vec.shape[1])
self.word_embedding.weight.data.copy_(torch.from_numpy(self.config.data_word_vec))
```
actually optimizer will update both **Embedding layer**.
Is that a bug or just on propose?