1.安装transformers

pip install transformers

2.加载预训练词典和分词方法

from transformers import BertTokenizer

# 方式一：网络下载词典
tokenizer = BertTokenizer.from_pretrained(
    pretrained_model_name_or_path='bert-base-chinese',
    cache_dir=None,
    force_download=False,
)
# 方式二：本地加载
tokenizer = BertTokenizer.from_pretrained('./bert-base-chinese/BertTokenizer/vocab.txt')

# 获取字典的长度
dictionary = tokenizer.get_vocab()
len(dictionary)

3.句子编码

3.1句子编码

sents = [
    '自然语言处理',
    '第三方工具包',
    '开发的应用在青少年中颇受欢迎'
]
# 编码两个句子
out = tokenizer.encode(
    text = sents[0],
    text_pair = sents[1],
    # 当句子长度大于max_length时,截断
    truncation=True,
    # 一律补pad到max_length长度
    padding='max_length',
    add_special_tokens=True,
    max_length=30,
    return_tensors=None,
)
print(out)
tokenizer.decode(out)

3.2增强编码函数

out = tokenizer.encode_plus(
    text = sents[0],
    text_pair = sents[1],
    truncation = True,
    padding = 'max_length',
    max_length = 30,
    add_special_tokens = True,
    
    # 可取值tf,pt,np，默认为返回list
    return_tensors = None,
    # 返回token_type_ids
    return_token_type_ids = True,
    # 返回attention_mask
    return_attention_mask = True,
    # 返回special_tokens_mask特殊符号标识
    return_special_tokens_mask = True,
    
    # 返回offset_mapping标识每个词的起止位置，这个参数只能BertTokenizerFast使用
    # return_offsets_mapping=True,
    # 返回length标识长度
    return_length = True,
    
)
for k, v in out.items():
    print(k, ":", v)

3.3.批量成对编码

# 批量成对编码
out = tokenizer.batch_encode_plus(
    batch_text_or_text_pairs = [(sents[0], sents[1]),(sents[0], sents[1])],
    truncation = True,
    padding = 'max_length',
    max_length = 15,
    add_special_tokens = True,
    
    # 可取值tf,pt,np，默认为返回list
    return_tensors = None,
    # 返回token_type_ids
    return_token_type_ids = True,
    # 返回attention_mask
    return_attention_mask = True,
    # 返回special_tokens_mask特殊符号标识
    return_special_tokens_mask = True,
    
    # 返回offset_mapping标识每个词的起止位置，这个参数只能BertTokenizerFast使用
    # return_offsets_mapping=True,
    # 返回length标识长度
    return_length = True,
    
)
for k, v in out.items():
    print(k, ":", v)

4.字典操作

4.1获取字典

dictionary = tokenizer.get_vocab()
type(dictionary), len(dictionary), '月光' in dictionary

4.2添加新词符号

# 添加新词
tokenizer.add_tokens(new_tokens=['月光','希望'])

# 添加新符号
tokenizer.add_special_tokens({'eos_token':'[EOS]'})
dictionary = tokenizer.get_vocab()
type(dictionary), len(dictionary), '月光' in dictionary

4.3编码新添加的词

out = tokenizer.encode(
    text='月光的新希望[EOS]',
    text_pair=None,
    # 当句子的长度大于max_length时，截断
    truncation=True,
    #一律补pad到max_length长度
    padding='max_length',
    add_special_tokens=True,
    max_length=8,
    return_tensors=None,
)
print(out)
tokenizer.decode(out)

5.datasets加载数据集

from datasets import load_dataset

dataset = load_dataset(path='../datasets', split='train')

6.加载bert模型

from transformers import BertModel

# 方式一：网络下载
pretrained = BertModel.from_pretrained(
	pretrained_model_name_or_path='bert-base-chinese',
)
# 方式二：本地加载(pytorch_model.bin、config.json)
pretrained = BertModel.from_pretrained('./bert-base-chinese/BertModel')

7.bert模型使用

# 不训练, 不需要计算梯度
for param in pretrained.parameters():
    param.requires_grad_(False)
    
out = pretrained(input_ids=input_ids.to(config.device),
                 attention_mask=attention_mask.to(config.device),
                 token_type_ids=token_type_ids.to(config.device),
                )
print(out.last_hidden_state.shape)

8.数据加载

from torch.utils.data import DataLoader,Dataset
from datasets import load_dataset

class MyDatasets(Dataset):
    def __init__(self, path, split):
        super(MyDatasets, self).__init__()
        self.dataset = load_dataset(path=path, split=split)
        
    def __getitem__(self, index):
        context = self.dataset[index]['text'][:-2]
        label = self.dataset[index]['text'][-1]
            
        return context, int(label)

    def __len__(self):  # 数据的长度
        return len(self.dataset)

def pad_collate(batch):  # 批处理函数
    sents, labels = zip(*batch)
    # 编码
    data = tokenizer.batch_encode_plus(
        batch_text_or_text_pairs=sents,
        truncation=True,
        padding='max_length',
        max_length=config.max_length,
        return_tensors='pt',
        return_length=True,
    )  
    # input_ids: 编码之后的数字
    # attention_mask: 是补零的位置是0, 其他位置是1
    input_ids = data['input_ids']
    attention_mask = data['attention_mask']
    token_type_ids = data['token_type_ids']
    labels = torch.LongTensor(labels)
    
    return input_ids, attention_mask, token_type_ids, labels

train_loader = DataLoader(dataset=MyDatasets(config.train_datasets_path, config.train_split), 
                          batch_size=config.train_batch_size, shuffle=True,
                          collate_fn=pad_collate, 
                          drop_last=True,
                          num_workers=config.num_workers)
test_loader = DataLoader(dataset=MyDatasets(config.test_datasets_path, config.test_split), 
                         batch_size=config.test_batch_size, shuffle=True,
                         collate_fn=pad_collate, 
                         drop_last=True,
                         num_workers=config.num_workers)