抱脸分词器无法正确加载文件

Question

我正在尝试使用 HuggingFace 的 BartModel 架构从 sratch 训练翻译模型。我正在使用 ByteLevelBPETokenizer 来标记事物。

我面临的问题是，当我在训练后保存分词器时，它没有正确加载，即使它显然创建了正确的 vocab.json 和 merges.txt 文件。

from tokenizers import ByteLevelBPETokenizer

# Initialize a tokenizer
tokenizer = ByteLevelBPETokenizer()

# Customize training
tokenizer.train(files=['/work/vaibhav/iwslt14/de-en/train.en', '/work/vaibhav/iwslt14/de-en/train.de'], vocab_size=8000, min_frequency=2, special_tokens=[
    "<s>",
    "<pad>",
    "</s>",
    "<unk>",
    "<mask>",
])

# Save files to disk
print(tokenizer)
tokenizer.save_model("tokens")

这就是我训练和保存分词器的方式。打印语句打印如下：

Tokenizer(vocabulary_size=8000, model=ByteLevelBPE, add_prefix_space=False, lowercase=False, dropout=None, unicode_normalizer=None, continuing_subword_prefix=None, end_of_word_suffix=None, trim_offsets=False)

但是，当我尝试通过以下代码行训练我的模型时加载分词器时：

#get the tokenizer
tokenizer = ByteLevelBPETokenizer()
tokenizer.from_file('tokens/vocab.json', 'tokens/merges.txt')
print(tokenizer)
return tokenizer

然后打印语句打印如下：

Tokenizer(vocabulary_size=0, model=ByteLevelBPE, add_prefix_space=False, lowercase=False, dropout=None, unicode_normalizer=None, continuing_subword_prefix=None, end_of_word_suffix=None, trim_offsets=False)

现在这对我来说很奇怪，因为 vocab_size 应该是 8000 而不是零，因此它基本上停止工作。如果我重新训练并直接使用它而不保存和加载它那么它可以工作但是效率不高。

这是 vocab.json（截断）的视图。

{"<s>":0,"<pad>":1,"</s>":2,"<unk>":3,"<mask>":4,"!":5,"\"":6,"#":7,"$":8,"%":9,"&":10,"'":11,"(":12,")":13,"*":14,"+":15,",":16,"-":17,".":18,"/":19,"0":20,"1":21,"2":22,"3":23,"4":24,"5":25,"6":26,"7":27,"8":28,"9":29,":":30,";":31,"<":32,"=":33,">":34,"?":35,"@":36,"A":37,"B":38,"C":39,"D":40,"E":41,"F":42,"G":43,"H":44,"I":45,"J":46,"K":47,"L":48,"M":49,"N":50,"O":51,"P":52,"Q":53,"R":54,"S":55,"T":56,"U":57,"V":58,"W":59,"X":60,"Y":61,"Z":62,"[":63,"\":64,"]":65,"^":66,"_":67,"`":68,"a":69,"b":70,"c":71,"d":72,"e":73,"f":74,"g":75,"h":76,"i":77,"j":78,"k":79,"l":80,"m":81,"n":82,"o":83,"p":84,"q":85,"r":86,"s":87,"t":88,"u":89,"v":90,"w":91,"x":92,"y":93,"z":94,"{":95,"|":96,"}":97,"~":98,"¡":99,"¢":100,"£":101,"¤":102,"¥":103,"¦":104,"§":105,"¨":106,"©":107,"ª":108,"«":109,"¬":110,"®":111,"¯":112,"°":113,"±":114,"²":115,"³":116,"´":117,"µ":118,"¶":119,"·":120,"¸":121,"¹":122,"º":123,"»":124,"¼":125,"½":126,"¾":127,"¿":128,"À":129,"Á":130,"Â":131,"Ã":132,"Ä":133,"Å":134,"Æ":135,"Ç":136,"È":137,"É":138,"Ê":139,"Ë":140,"Ì":141,"Í":142,"Î":143,"Ï":144,"Ð":145,"Ñ":146,"Ò":147,"Ó":148,"Ô":149,"Õ":150,"Ö":151,"×":152,"Ø":153,"Ù":154,"Ú":155,"Û":156,"Ü":157,"Ý":158,"Þ":159,"ß":160,"à":161,"á":162,"â":163,"ã":164,"ä":165,"å":166,"æ":167,"ç":168,"è":169,"é":170,"ê":171,"ë":172,"ì":173,"í":174,"î":175,"ï":176,"ð":177,"ñ":178,"ò":179,"ó":180,"ô":181,"õ":182,"ö":183,"÷":184,"ø":185,"ù":186,"ú":187,"û":188,"ü":189,"ý":190,"þ":191,"ÿ":192,"Ā":193,"ā":194,"Ă":195,"ă":196,"Ą":197,"ą":198,"Ć":199,"ć":200,"Ĉ":201,"ĉ":202,"Ċ":203,"ċ":204,"Č":205,"č":206,"Ď":207,"ď":208,"Đ":209,"đ":210,"Ē":211,"ē":212,"Ĕ":213,"ĕ":214,"Ė":215,"ė":216,"Ę":217,"ę":218,"Ě":219,"ě":220,"Ĝ":221,"ĝ":222,"Ğ":223,"ğ":224,"Ġ":225,"ġ":226,"Ģ":227,"ģ":228,"Ĥ":229,"ĥ":230,"Ħ":231,"ħ":232,"Ĩ":233,"ĩ":234,"Ī":235,"ī":236,"Ĭ":237,"ĭ":238,"Į":239,"į":240,"İ":241,"ı":242,"Ĳ":243,"ĳ":244,"Ĵ":245,"ĵ":246,"Ķ":247,"ķ":248,"ĸ":249,"Ĺ":250,"

这是 merges.txt（截断）的视图。

#version: 0.2 - Trained by `huggingface/tokenizers`
e n
e r
i n
Ġ t
c h
Ġ a
Ġ d
Ġ w
Ġ s
Ġt h
n d
i e
e s
i s
a t
o n
i t
Ġ m
a s
a n
r e
Ġth e
Ġ h

可以看到文件是正常的。对此问题的任何帮助将不胜感激。

Answer 1

所以我终于自己弄明白了！ huggingface 代码中有一些错误，所以我像这样加载了分词器并且它起作用了。

tokenizer = ByteLevelBPETokenizer('tokens/vocab.json', 'tokens/merges.txt')

抱脸分词器无法正确加载文件

Hugging face tokenizer cannot load files properly

python

nlp

huggingface-tokenizers