def train_tokenizer(sentences: List[str],
serialize_path: str = "",
vocab_size: int = 6000) -> Tokenizer:
bert_tokenizer = Tokenizer(WordPiece(unk_token="[UNK]"))
bert_tokenizer.normalizer = Sequence([NFD(), Lowercase(), StripAccents()])
bert_tokenizer.pre_tokenizer = Whitespace()
bert_tokenizer.post_processor = TemplateProcessing(
single="[CLS] $A [SEP]",
pair="[CLS] $A [SEP] $B:1 [SEP]:1",
special_tokens=[
("[CLS]", 1),
("[SEP]", 2),
],
)
trainer = WordPieceTrainer(
vocab_size=vocab_size,
special_tokens=["[UNK]", "[CLS]", "[SEP]", "[PAD]", "[MASK]"])
bert_tokenizer.train_from_iterator(sentences, trainer=trainer)
if serialize_path:
bert_tokenizer.save_model(serialize_path)
return bert_tokenizer
print("Decoded string: {}".format(decoded))
# %%
from tokenizers import ByteLevelBPETokenizer
# tokenizer提供了一些经典tokenization算法的高级封装
# 譬如可以用`ByteLevelBPETokenizer`简单地重写上面的内容
#
tokenizer = ByteLevelBPETokenizer()
tokenizer.train(files=['data/big.txt'], vocab_size=25000, show_progress=True)
SAVE_PATH = Path('tokenizers')
PATH = SAVE_PATH / 'bytelevel-bpe-tokenizer-model'
if not PATH.exists():
PATH.mkdir(parents=True, exist_ok=True)
tokenizer.save_model(str(PATH))
tokenizer = ByteLevelBPETokenizer(vocab_file=str(PATH / 'vocab.json'),
merges_file=str(PATH / 'merges.txt'))
encoded = \
tokenizer.encode("This is a simple input to be tokenized.")
print("Encoded string: {}".format(encoded.tokens))
decoded = \
tokenizer.decode(encoded.ids)
print("Decoded string: {}".format(decoded))
# %%
# 与transformers搭配使用时，Encoding structure中有用的properties
# - normalized_str