笔记
Transformers==4.0.0 由于每次调用bert等模型都需要使用模型的tokenizer,所以写个笔记,方便自己以及后人查阅,(其实看官方文档也可以,但是英文的看着头疼。有错误也请留言吧。
base : PreTrainedTokenizerBase
作为基类,该类有着所有tokenizer都具有的方法还有属性。
PreTrainedTokenizer
使用multiprocess 加速tokenize
仿照这个就完事了。partial可以固定函数中的参数,简直是专门为多进程准备的。
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
def
global
tokenizer = tokenizer_for_convert
features = []
threads = min
with
annotate_ = partial(
tokenzier.encode_plus, # batch_encode_plus 不知道为什么不work
max_seq_length=max_seq_length,
doc_stride=doc_stride,
max_query_length=max_query_length,
padding_strategy=padding_strategy,
is_training=is_training,
)
features = list
tqdm(
p.imap(annotate_, examples, chunksize=32
total=len
desc="convert squad examples to features"
disable=not
)
)
HuggingFace PretrainTokenizer学习笔记
作者 CheaSim
发布于 2020-12-11
更新于 2020-12-11
许可协议