bert系列第一篇： bert进行embedding

bert理解

一句话概括， bert就是一个抽取器。输入一句话（词序列），输出抽取后的embedding序列。
再简单理解就是，它就是一个 encoder。

简单机理

我们可以用transformer在语言模型上做预训练。因为transformer是encoder-decoder结构，语言模型就只需要encoder部分就够了。BERT，利用transformer的encoder来进行预训练。

那么什么是transformer？
这是一个新的训练结构，发展历程而言就是CNN，RNN，transformer； transformer是基于attention机理发展而来。

transformer由编码器和解码器组成。编码器和解码器都是基于attention机制。如下图

什么是注意力机制，一图简单领会，后面我们单独开一篇动手实践一下

注意力机制就是，当前词的含义，必须结合结合上下文才能更好的理解。

encoder输入输出

• 输入会加入特殊的[CLS]代表整句话的含义，可以用于分类。

• input的词help，prince，mayuko等，一共512，这是截取的最大长度。

• 然后经过12层的encoder

• 最后输出的是每个token对应的embedding序列，每个token对应一个768维的向量。这个应该很好理解。

输出的结果

out = bert(xx)

 Return:
        :obj:`tuple(torch.FloatTensor)` comprising various elements depending on the configuration (:class:`~transformers.BertConfig`) and inputs:
        **last_hidden_state** (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`):
            Sequence of hidden-states at the output of the last layer of the model.
        **pooler_output** (:obj:`torch.FloatTensor`: of shape :obj:`(batch_size, hidden_size)`):
            Last layer hidden-state of the first token of the sequence (classification token)
            further processed by a Linear layer and a Tanh activation function. The Linear
            layer weights are trained from the next sentence prediction (classification)
            objective during pre-training.

            This output is usually *not* a good summary
            of the semantic content of the input, you're often better with averaging or pooling
            the sequence of hidden-states for the whole input sequence.
        **hidden_states** (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``config.output_hidden_states=True``):
            Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
            of shape :obj:`(batch_size, sequence_length, hidden_size)`.

            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
        **attentions** (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``config.output_attentions=True``):
            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.

            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
            heads.

作用

有了词序列对应的embedding向量，就可以对词分类、句子向量构建，句子分类、句子相似度比较等。

code(notebook)

#%% md

# bert

#%%

!pip install transformers

#%%

import torch
from transformers import BertModel, BertTokenizer

#%%

tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')

#%%

input_ids = tokenizer.encode('hello world bert!')
input_ids

#%%

type(input_ids)

#%%

ids = torch.LongTensor(input_ids)
ids

#%%

text = tokenizer.convert_ids_to_tokens(input_ids)
text

#%%

model = BertModel.from_pretrained('bert-base-uncased', output_hidden_states=True)
# Set the device to GPU (cuda) if available, otherwise stick with CPU
device = 'cuda' if torch.cuda.is_available() else 'cpu'

model = model.to(device)
ids = ids.to(device)

model.eval()

#%%

print(ids.size())
# unsqueeze IDs to get batch size of 1 as added dimension
granola_ids = ids.unsqueeze(0)
print(granola_ids.size())


#%% md

In the example below, an additional argument has been given to the model initialisation. output_hidden_states will give us more output information. By default, a BertModel will return a tuple but the contents of that tuple differ depending on the configuration of the model. When passing output_hidden_states=True, the tuple will contain (in order; shape in brackets):

1. the last hidden state (batch_size, sequence_length, hidden_size)
1. the pooler_output of the classification token (batch_size, hidden_size)
1. the hidden_states of the outputs of the model at each layer and the initial embedding outputs (batch_size, sequence_length, hidden_size)

#%%

out = model(input_ids=granola_ids) # tuple

hidden_states = out[2]
print("last hidden state:",out[0].shape) #torch.Size([1, 6, 768])
print("pooler_output of classification token:",out[1].shape)#[1,768] cls
print("all hidden_states:", len(out[2]))

#%%

for i, each_layer in enumerate(hidden_states):
    print('layer=',i, each_layer)

#%%

sentence_embedding = torch.mean(hidden_states[-1], dim=1).squeeze()
print(sentence_embedding)
print(sentence_embedding.size())

#%%


# get last four layers
last_four_layers = [hidden_states[i] for i in (-1, -2, -3, -4)]
# cast layers to a tuple and concatenate over the last dimension
cat_hidden_states = torch.cat(tuple(last_four_layers), dim=-1)
print(cat_hidden_states.size())

# take the mean of the concatenated vector over the token dimension
cat_sentence_embedding = torch.mean(cat_hidden_states, dim=1).squeeze()
print(cat_sentence_embedding)
print(cat_sentence_embedding.size())

不同的emebdding组合会带来不一样的结果，参考。

利用concat的向量，最优结果。

总结

1. 不同的层代表不同的特征含义，向量组合的实验可以证明这一点。
2. bert就是抽取器
3. 不同隐层输出的向量的使用是核心所在
4. 仔细理解文中的两幅图，和样例代码。然后就是感悟了！

引用

1. https://github.com/huggingface/transformers/issues/2986
2. https://github.com/BramVanroy/bert-for-inference/blob/master/introduction-to-bert.ipynb
3. https://www.cnblogs.com/gczr/p/11785930.html
4. https://blog.youkuaiyun.com/longxinchen_ml/article/details/86533005