Huggingface简介及BERT代码浅析

Hugging face 是一家总部位于纽约的聊天机器人初创服务商，令它广为人知的是Hugging Face专注于NLP技术，拥有大型的开源社区，尤其是在github上开源的自然语言处理，预训练模型库 Transformers。最初叫 pytorch-pretrained-bert 。

安装方式

pip install transformers

当报错：‘ValueError: Connection error, and we cannot find the requested files in the cached path. Please try again or make sure your Internet connection is on’时，需要自己单独下载vocab文件和预训练模型放在对应的位置，下载链接如下：

 

• BERT-Large, Uncased (Whole Word Masking): 24-layer, 1024-hidden, 16-heads, 340M parameters
• BERT-Large, Cased (Whole Word Masking): 24-layer, 1024-hidden, 16-heads, 340M parameters
• BERT-Base, Uncased: 12-layer, 768-hidden, 12-heads, 110M parameters
• BERT-Large, Uncased: 24-layer, 1024-hidden, 16-heads, 340M parameters
• BERT-Base, Cased: 12-layer, 768-hidden, 12-heads , 110M parameters
• BERT-Large, Cased: 24-layer, 1024-hidden, 16-heads, 340M parameters
• BERT-Base, Multilingual Cased (New, recommended): 104 languages, 12-layer, 768-hidden, 12-heads, 110M parameters
• BERT-Base, Multilingual Uncased (Orig, not recommended) (Not recommended, use Multilingual Casedinstead): 102 languages, 12-layer, 768-hidden, 12-heads, 110M parameters
• BERT-Base, Chinese: Chinese Simplified and Traditional, 12-layer, 768-hidden, 12-heads, 110M parameters

PRETRAINED_VOCAB_ARCHIVE_MAP = {
    'bert-base-uncased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-uncased-vocab.txt",
    'bert-large-uncased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-uncased-vocab.txt",
    'bert-base-cased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-cased-vocab.txt",
    'bert-large-cased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-cased-vocab.txt",
    'bert-base-multilingual-uncased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-multilingual-uncased-vocab.txt",
    'bert-base-multilingual-cased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-multilingual-cased-vocab.txt",
    'bert-base-chinese': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-chinese-vocab.txt",
}

PRETRAINED_MODEL_ARCHIVE_MAP = {
    'bert-base-uncased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-uncased.tar.gz",
    'bert-large-uncased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-uncased.tar.gz",
    'bert-base-cased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-cased.tar.gz",
    'bert-large-cased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-cased.tar.gz",
    'bert-base-multilingual-uncased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-multilingual-uncased.tar.gz",
    'bert-base-multilingual-cased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-multilingual-cased.tar.gz",
    'bert-base-chinese': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-chinese.tar.gz",
}

BertTokenizer.from_pretrained(‘D:\code\EMNLP-2019-master\data\bert-base-uncased-vocab.txt’)
BertModel.from_pretrained(‘D:\code\EMNLP-2019-master\data\bert-base-uncased’)
 

加载模型运行

加载下载的模型和词表

from pytorch_pretrained_bert import BertModel, BertTokenizer
import numpy as np
​
# 加载bert的分词器
tokenizer = BertTokenizer.from_pretrained('你存放的路径/bert-base-uncased-vocab.txt')
# 加载bert模型，这个路径文件夹下有bert_config.json配置文件和model.bin模型权重文件
bert = BertModel.from_pretrained('你存放的路径/bert-base-uncased/')

输入语句

s = "I'm not sure, this can work, lol -.-"
​
tokens = tokenizer.tokenize(s)
print("\\".join(tokens))
# "i\\'\\m\\not\\sure\\,\\this\\can\\work\\,\\lo\\##l\\-\\.\\-"
# 是否需要这样做？
# tokens = ["[CLS]"] + tokens + ["[SEP]"]
​
ids = torch.tensor([tokenizer.convert_tokens_to_ids(tokens)])
print(ids.shape)
# torch.Size([1, 15])
​
result = bert(ids, output_all_encoded_layers=True)

输出结果

result = (
    [encoder_0_output, encoder_1_output, ..., encoder_11_output], 
    pool_output
)

1. 因为我选择了参数output_all_encoded_layers=True，12层Transformer的结果全返回了，存在第一个列表中，每个encoder_output的大小为[batch_size, sequence_length, hidden_size];
2. pool_out大小为[batch_size, hidden_size]，pooler层的输出在论文中描述为：
   which is the output of a classifier pretrained on top of the hidden state associated to the first character of the input (`CLS`) to train on the Next-Sentence task (see BERT's paper).
   也就是说，取了最后一层Transformer的输出结果的第一个单词[cls]的hidden states，其已经蕴含了整个input句子的信息了。
3. 如果你用不上所有encoder层的输出，output_all_encoded_layers参数设置为Fasle，那么result中的第一个元素就不是列表了，只是encoder_11_output，大小为[batch_size, sequence_length, hidden_size]的张量，可以看作bert对于这句话的表示。

当前任务上finetune

有两种方案

1. 单纯得将bert看作特征提取器，和sklearn的TfidfTransformer类似，先将你的文本分词，再丢给bert，把模型输出的结果作为你模型的input数据就ok了；
2. 上面那种方法看起来很不fancy？那么可以把bert嵌入到我们的模型中：

class CustomModel(nn.Module):
    
    def __init__(self, bert_path, n_other_features, n_hidden):
        super().__init__()
        # 加载并冻结bert模型参数
        self.bert = BertModel.from_pretrained(bert_path)
        for param in self.bert.parameters():
            param.requires_grad = False
        self.output = nn.Sequential(
            nn.Dropout(0.2),
            nn.Linear(768 + n_other_features, n_hidden),
            nn.ReLU(),
            nn.Linear(n_hidden, 1)
        )
    def forward(self, seqs, features):
        _, pooled = self.bert(seqs, output_all_encoded_layers=False)
        concat = torch.cat([pooled, features], dim=1)
        logits = self.output(concat)
        return logits

测试

s = "I'm not sure, this can work, lol -.-"
​
tokens = tokenizer.tokenize(s)
ids = torch.tensor([tokenizer.convert_tokens_to_ids(tokens)])
# print(ids)
# tensor([[1045, 1005, 1049, 2025, 2469, 1010, 2023, 2064, 2147, 1010, 8840, 2140,
#         1011, 1012, 1011]])
​
model = CustomModel('你的路径/bert-base-uncased/',10, 512)
outputs = model(ids, torch.rand(1, 10))
# print(outputs)
# tensor([[0.1127]], grad_fn=<AddmmBackward>)