bert-base-chinese模型使用教程

向量编码和向量相似度展示

import torch
from transformers import BertTokenizer, BertModel
import numpy as npmodel_name = "C:/Users/Administrator.DESKTOP-TPJL4TC/.cache/modelscope/hub/tiansz/bert-base-chinese"sentences = ['春眠不觉晓', '大梦谁先觉', '浓睡不消残酒', '东临碣石以观沧海']tokenizer = BertTokenizer.from_pretrained(model_name)
# print(type(tokenizer)) # <class 'transformers.models.bert.tokenization_bert.BertTokenizer'>model = BertModel.from_pretrained(model_name)
# print(type(model)) # <class 'transformers.models.bert.modeling_bert.BertModel'>def test_encode():input_ids = tokenizer.encode('春眠不觉晓', return_tensors='pt') # shape (1, 7)output = model(input_ids)print(output.last_hidden_state.shape)  # shape (1, 7, 768)v = torch.mean(output.last_hidden_state, dim=1)  # shape (1, 768)print(v.shape)  # shape (1, 768)print(output.pooler_output.shape)  # shape (1, 768)

根据 transformers\modeling_outputs.py:196，即 BaseModelOutputWithPoolingAndCrossAttentions 的注释：

@dataclass
class BaseModelOutputWithPoolingAndCrossAttentions(ModelOutput):"""Base class for model's outputs that also contains a pooling of the last hidden states.Args:last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):Sequence of hidden-states at the output of the last layer of the model.pooler_output (`torch.FloatTensor` of shape `(batch_size, hidden_size)`):Last layer hidden-state of the first token of the sequence (classification token) after further processingthrough the layers used for the auxiliary pretraining task. E.g. for BERT-family of models, this returnsthe classification token after processing through a linear layer and a tanh activation function. The linearlayer weights are trained from the next sentence prediction (classification) objective during pretraining.hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,sequence_length)`.Attentions weights after the attention softmax, used to compute the weighted average in the self-attentionheads.cross_attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` and `config.add_cross_attention=True` is passed or when `config.output_attentions=True`):Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,sequence_length)`.Attentions weights of the decoder's cross-attention layer, after the attention softmax, used to compute theweighted average in the cross-attention heads.past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape`(batch_size, num_heads, sequence_length, embed_size_per_head)`) and optionally if`config.is_encoder_decoder=True` 2 additional tensors of shape `(batch_size, num_heads,encoder_sequence_length, embed_size_per_head)`.Contains pre-computed hidden-states (key and values in the self-attention blocks and optionally if`config.is_encoder_decoder=True` in the cross-attention blocks) that can be used (see `past_key_values`input) to speed up sequential decoding."""last_hidden_state: torch.FloatTensor = Nonepooler_output: torch.FloatTensor = Nonehidden_states: Optional[Tuple[torch.FloatTensor, ...]] = Nonepast_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = Noneattentions: Optional[Tuple[torch.FloatTensor, ...]] = Nonecross_attentions: Optional[Tuple[torch.FloatTensor, ...]] = None

1为batch_size，我们出入的就是一个字符串所以batch_size为1。

7为sequence_length，BERT模型会为单句的输入前面加特殊字符[CLS]和后面加特殊字符[SEP]，因此为7个字符。

768为hidden_size，即每个字符被编码成768个数字组成的向量。

根据文档的说法，pooler_output向量一般不是很好的句子语义摘要，因此这里采用了将 last_hidden_state 进行池化的方法。

torch.mean(dim)函数

这是常用的池化pooling方法，降低向量的维度，便于运算。

dim不指定任何参数就是所有元素的算术平均值

dim指定为0时，求得是列的平均值。

dim指定为1时，求得是行的平均值；

也就是经过tarch.mean()之后，这句话变成了一个 1X768 的向量，即使用这一个向量来代表着句话的语义。

接下来就可以计算文本相似度得分了，目标为上面给出的四个句子。

def test_similarity():with torch.no_grad():vs = [sentence_embedding(sentence).numpy() for sentence in sentences]nvs = [v / np.linalg.norm(v) for v in vs]  # normalize each vectorm = np.array(nvs).squeeze(1)  # shape (4, 768)print(np.around(m @ m.T, decimals=2))  # pairwise cosine similaritydef sentence_embedding(sentence):input_ids = tokenizer.encode(sentence, return_tensors='pt')output = model(input_ids)return torch.mean(output.last_hidden_state, dim=1)

@符号是矩阵相乘

输出结果 4X4

[[1.   0.75 0.83 0.57][0.75 1.   0.72 0.51][0.83 0.72 1.   0.58][0.57 0.51 0.58 1.  ]]

	春眠不觉晓	大梦谁先觉	浓睡不消残酒	东临碣石以观沧海
春眠不觉晓	1	0.75	0.83	0.57
大梦谁先觉	0.75	1	0.72	0.51
浓睡不消残酒	0.83	0.72	1	0.58
东临碣石以观沧海	0.57	0.51	0.58	1

with torch.no_grad() 的作用是将模型状态置为推断（inference），即在计算过程中不进行梯度计算和反向传播操作。

关于 last_hidden_state 与 pooler_output

看bert中文文本分类任务，发现训练输出结果向量仅使用了第一个token的向量，而一句话中的第一个位置是特殊字符[CLS]，那么该如何理解呢？

BERT在第一句前会加一个[CLS]标志，CLS 就是 classification 分类之意，即 CLS 的作用就是它能代表整句话的意思，使用它就可以对整句话进行分类。用于下游的分类任务等。

为什么选它呢，因为与文本中已有的其它词相比，这个无明显语义信息的符号会更“公平”地融合文本中各个词的语义信息，从而更好的表示整句话的语义。具体来说，self-attention是用文本中的其它词来增强目标词的语义表示，但是目标词本身的语义还是会占主要部分的，因此，经过BERT的12层，每次词的embedding融合了所有词的信息，可以去更好的表示自己的语义。而[CLS]位本身没有语义，经过12层，得到的是attention后所有词的加权平均，相比其他正常词，可以更好的表征句子语义。

当然，也可以通过对最后一层所有词的embedding做pooling池化处理去表征句子语义，最常用的池化方法就是平均法mean。

从上面的打印信息我们可以知道 pooler_output 的结果也是 1X768，与 torch.mean() 的维度一样，可以理解为它是经过模型池化后的结果。pooler_output是通过应用一个线性层和一个激活函数（通常是Tanh）到last_hidden_state的第一个token（即[CLS]标记）的隐藏状态来生成的。这个输出通常用于分类任务，因为它编码了整个序列的信息。