def converted(self) -> Tokenizer:
ot = self.original_tokenizer
vocab = ot.encoder
merges = list(ot.bpe_ranks.keys())
tokenizer = Tokenizer(
BPE(
vocab=vocab,
merges=merges,
dropout=None,
continuing_subword_prefix="",
end_of_word_suffix="",
fuse_unk=False,
))
tokenizer.pre_tokenizer = pre_tokenizers.ByteLevel(
add_prefix_space=ot.add_prefix_space)
tokenizer.decoder = decoders.ByteLevel()
tokenizer.post_processor = processors.TemplateProcessing(
single="[CLS]:0 $A:0 [SEP]:0",
pair="[CLS]:0 $A:0 [SEP]:0 $B:0 [SEP]:0",
special_tokens=[
("[CLS]",
self.original_tokenizer.convert_tokens_to_ids("[CLS]")),
("[SEP]",
self.original_tokenizer.convert_tokens_to_ids("[SEP]")),
],
)
return tokenizer
def post_processor(self):
eos = self.original_tokenizer.eos_token
special_tokens = [
(eos, self.original_tokenizer.eos_token_id),
]
return processors.TemplateProcessing(single=["$A", eos],
pair=["$A", "$B", eos],
special_tokens=special_tokens)
def post_processor(self):
return processors.TemplateProcessing(
single=["$A", "</s>"],
pair=["$A", "</s>", "$B", "</s>"],
special_tokens=[
("</s>", self.original_tokenizer.convert_tokens_to_ids("</s>")),
],
)
def post_processor(self):
return processors.TemplateProcessing(
single="$A:0 <sep>:0 <cls>:2",
pair="$A:0 <sep>:0 $B:1 <sep>:1 <cls>:2",
special_tokens=[
("<sep>", self.original_tokenizer.convert_tokens_to_ids("<sep>")),
("<cls>", self.original_tokenizer.convert_tokens_to_ids("<cls>")),
],
)
def post_processor(self):
return processors.TemplateProcessing(
single="[CLS]:0 $A:0 [SEP]:0",
pair="[CLS]:0 $A:0 [SEP]:0 $B:1 [SEP]:1",
special_tokens=[
("[CLS]", self.original_tokenizer.convert_tokens_to_ids("[CLS]")),
("[SEP]", self.original_tokenizer.convert_tokens_to_ids("[SEP]")),
],
)
def post_processor(self):
return processors.TemplateProcessing(
single="<s> $A </s>",
pair="<s> $A </s> </s> $B </s>",
special_tokens=[
("<s>", self.original_tokenizer.convert_tokens_to_ids("<s>")),
("</s>", self.original_tokenizer.convert_tokens_to_ids("</s>")),
],
)
def set_tgt_lang_special_tokens(self, lang):
self.cur_lang_code = self.convert_tokens_to_ids(lang)
self.prefix_tokens = []
self.suffix_tokens = [self.EOS, self.cur_lang_code]
prefix_tokens_str = self.convert_ids_to_tokens(self.prefix_tokens)
suffix_tokens_str = self.convert_ids_to_tokens(self.suffix_tokens)
self._tokenizer.post_processor = processors.TemplateProcessing(
single=prefix_tokens_str + ["$A"] + suffix_tokens_str,
pair=prefix_tokens_str + ["$A", "$B"] + suffix_tokens_str,
special_tokens=list(
zip(prefix_tokens_str + suffix_tokens_str,
self.prefix_tokens + self.suffix_tokens)),
)
def set_tgt_lang_special_tokens(self, lang: str) -> None:
"""Reset the special tokens to the target language setting. Prefix [tgt_lang_code], suffix =[eos]."""
self.cur_lang_code = self.convert_tokens_to_ids(lang)
self.prefix_tokens = []
self.suffix_tokens = [self.eos_token_id, self.cur_lang_code]
prefix_tokens_str = self.convert_ids_to_tokens(self.prefix_tokens)
suffix_tokens_str = self.convert_ids_to_tokens(self.suffix_tokens)
self._tokenizer.post_processor = processors.TemplateProcessing(
single=prefix_tokens_str + ["$A"] + suffix_tokens_str,
pair=prefix_tokens_str + ["$A", "$B"] + suffix_tokens_str,
special_tokens=list(zip(prefix_tokens_str + suffix_tokens_str, self.prefix_tokens + self.suffix_tokens)),
)
def set_src_lang_special_tokens(self, src_lang) -> None:
"""Reset the special tokens to the source lang setting. No prefix and suffix=[eos, cur_lang_code]."""
self.cur_lang_code = self.convert_tokens_to_ids(src_lang)
self.prefix_tokens = []
self.suffix_tokens = [self.eos_token_id, self.cur_lang_code]
prefix_tokens_str = self.convert_ids_to_tokens(self.prefix_tokens)
suffix_tokens_str = self.convert_ids_to_tokens(self.suffix_tokens)
self._tokenizer.post_processor = processors.TemplateProcessing(
single=prefix_tokens_str + ["$A"] + suffix_tokens_str,
pair=prefix_tokens_str + ["$A", "$B"] + suffix_tokens_str,
special_tokens=list(zip(prefix_tokens_str + suffix_tokens_str, self.prefix_tokens + self.suffix_tokens)),
)
def converted(self) -> Tokenizer:
vocab = self.original_tokenizer.vocab
tokenizer = Tokenizer(
WordPiece(vocab, unk_token=str(self.original_tokenizer.unk_token)))
# # Let the tokenizer know about special tokens if they are part of the vocab
# if tokenizer.token_to_id(str(self.original_tokenizer.unk_token)) is not None:
# tokenizer.add_special_tokens([str(self.original_tokenizer.unk_token)])
# if tokenizer.token_to_id(str(self.original_tokenizer.sep_token)) is not None:
# tokenizer.add_special_tokens([str(self.original_tokenizer.sep_token)])
# if tokenizer.token_to_id(str(self.original_tokenizer.cls_token)) is not None:
# tokenizer.add_special_tokens([str(self.original_tokenizer.cls_token)])
# if tokenizer.token_to_id(str(self.original_tokenizer.pad_token)) is not None:
# tokenizer.add_special_tokens([str(self.original_tokenizer.pad_token)])
# if tokenizer.token_to_id(str(self.original_tokenizer.mask_token)) is not None:
# tokenizer.add_special_tokens([str(self.original_tokenizer.mask_token)])
tokenize_chinese_chars = False
strip_accents = False
do_lower_case = False
if hasattr(self.original_tokenizer, "basic_tokenizer"):
tokenize_chinese_chars = self.original_tokenizer.basic_tokenizer.tokenize_chinese_chars
strip_accents = self.original_tokenizer.basic_tokenizer.strip_accents
do_lower_case = self.original_tokenizer.basic_tokenizer.do_lower_case
tokenizer.normalizer = normalizers.BertNormalizer(
clean_text=True,
handle_chinese_chars=tokenize_chinese_chars,
strip_accents=strip_accents,
lowercase=do_lower_case,
)
tokenizer.pre_tokenizer = pre_tokenizers.BertPreTokenizer()
cls = str(self.original_tokenizer.cls_token)
sep = str(self.original_tokenizer.sep_token)
cls_token_id = self.original_tokenizer.cls_token_id
sep_token_id = self.original_tokenizer.sep_token_id
tokenizer.post_processor = processors.TemplateProcessing(
single=
f"{cls}:2 $A:0 {sep}:0", # token_type_id is 2 for Funnel transformer
pair=f"{cls}:2 $A:0 {sep}:0 $B:1 {sep}:1",
special_tokens=[
(cls, cls_token_id),
(sep, sep_token_id),
],
)
tokenizer.decoder = decoders.WordPiece(prefix="##")
return tokenizer
def converted(self) -> Tokenizer:
vocab = self.original_tokenizer.vocab
tokenizer = Tokenizer(
WordPiece(vocab, unk_token=str(self.original_tokenizer.unk_token)))
tokenize_chinese_chars = False
strip_accents = False
do_lower_case = False
if hasattr(self.original_tokenizer, "basic_tokenizer"):
tokenize_chinese_chars = self.original_tokenizer.basic_tokenizer.tokenize_chinese_chars
strip_accents = self.original_tokenizer.basic_tokenizer.strip_accents
do_lower_case = self.original_tokenizer.basic_tokenizer.do_lower_case
tokenizer.normalizer = normalizers.BertNormalizer(
clean_text=True,
handle_chinese_chars=tokenize_chinese_chars,
strip_accents=strip_accents,
lowercase=do_lower_case,
)
tokenizer.pre_tokenizer = pre_tokenizers.BertPreTokenizer()
cls = str(self.original_tokenizer.cls_token)
sep = str(self.original_tokenizer.sep_token)
question = str(self.original_tokenizer.question_token)
dot = "."
cls_token_id = self.original_tokenizer.cls_token_id
sep_token_id = self.original_tokenizer.sep_token_id
question_token_id = self.original_tokenizer.question_token_id
dot_token_id = self.original_tokenizer.convert_tokens_to_ids(".")
if self.original_tokenizer.padding_side == "right":
pair = f"{cls}:0 $A:0 {question} {dot} {sep}:0 $B:1 {sep}:1"
else:
pair = f"{cls}:0 $A:0 {sep}:0 $B:1 {question} {dot} {sep}:1"
tokenizer.post_processor = processors.TemplateProcessing(
single=f"{cls}:0 $A:0 {sep}:0",
pair=pair,
special_tokens=[
(cls, cls_token_id),
(sep, sep_token_id),
(question, question_token_id),
(dot, dot_token_id),
],
)
tokenizer.decoder = decoders.WordPiece(prefix="##")
return tokenizer
def train_tokenizer() -> Tuple[tokenizers.Tokenizer, Generator, int]:
tokenizer = tokenizers.Tokenizer(models.WordPiece(unk_token="<unk>"))
tokenizer.decoder = decoders.WordPiece()
tokenizer.normalizer = normalizers.Sequence([
normalizers.NFD(), # NFD unicode normalizer
normalizers.Lowercase(),
normalizers.StripAccents()
])
tokenizer.pre_tokenizer = tokenizers.pre_tokenizers.Sequence([
pre_tokenizers.Whitespace(),
pre_tokenizers.Digits(individual_digits=False)
])
tokenizer.post_processor = processors.TemplateProcessing(
single="$A </s>",
pair="$A </s> [SEP] <s> $B:1",
special_tokens=[("[SEP]", 1), ("<s>", 2), ("</s>", 3)])
dataset = datasets.load_dataset("wikitext",
"wikitext-103-raw-v1",
split="test")
def batch_iterator(batch_size=1000):
for i in range(0, len(dataset), batch_size):
yield dataset[i:i + batch_size]["text"]
tokenizer.train_from_iterator(
batch_iterator(),
trainer=trainers.WordPieceTrainer(
vocab_size=10000, special_tokens=["<unk>", "[SEP]", "<s>",
"</s>"]))
def generator():
for record in dataset:
if record['text'].strip() != '':
for sentence in sent_tokenizer(record['text']):
yield sentence
data = tf.data.Dataset.from_generator(generator,
output_signature=(tf.TensorSpec(
shape=(None), dtype=tf.string)))
data = data.map(tf.strings.strip,
num_parallel_calls=tf.data.experimental.AUTOTUNE)
return tokenizer, data
def train_tokenizer() -> tokenizers.Tokenizer:
tokenizer = tokenizers.Tokenizer(models.WordPiece(unk_token="<unk>"))
tokenizer.decoder = decoders.WordPiece()
tokenizer.normalizer = normalizers.Sequence([
normalizers.NFD(), # NFD unicode normalizer
normalizers.Lowercase(),
normalizers.StripAccents()
])
tokenizer.pre_tokenizer = tokenizers.pre_tokenizers.Sequence([
pre_tokenizers.Whitespace(),
pre_tokenizers.Digits(individual_digits=False)
])
tokenizer.post_processor = processors.TemplateProcessing(
single="$A </s>",
pair="$A </s> [SEP] <s> $B:1",
special_tokens=[("[SEP]", 1), ("<s>", 2), ("</s>", 3)])
# dataset = datasets.load_dataset("wikitext", "wikitext-103-raw-v1", split="train+test+validation")
dataset = datasets.load_dataset("wikitext",
"wikitext-103-raw-v1",
split="validation")
def batch_iterator(batch_size=1000):
for i in range(0, len(dataset), batch_size):
yield dataset[i:i + batch_size]["text"]
tokenizer.train_from_iterator(
batch_iterator(),
trainer=trainers.WordPieceTrainer(
vocab_size=10000, special_tokens=["<unk>", "[SEP]", "<s>",
"</s>"]))
def generator():
for record in dataset:
if record['text'].strip() != '':
yield record['text']
return tokenizer, generator
def converted(self) -> Tokenizer:
vocab = self.original_tokenizer.vocab
tokenizer = Tokenizer(
WordPiece(vocab, unk_token=str(self.original_tokenizer.unk_token)))
tokenize_chinese_chars = False
strip_accents = False
do_lower_case = False
if hasattr(self.original_tokenizer, "basic_tokenizer"):
tokenize_chinese_chars = self.original_tokenizer.basic_tokenizer.tokenize_chinese_chars
strip_accents = self.original_tokenizer.basic_tokenizer.strip_accents
do_lower_case = self.original_tokenizer.basic_tokenizer.do_lower_case
tokenizer.normalizer = normalizers.BertNormalizer(
clean_text=True,
handle_chinese_chars=tokenize_chinese_chars,
strip_accents=strip_accents,
lowercase=do_lower_case,
)
tokenizer.pre_tokenizer = pre_tokenizers.BertPreTokenizer()
cls = str(self.original_tokenizer.cls_token)
sep = str(self.original_tokenizer.sep_token)
cls_token_id = self.original_tokenizer.cls_token_id
sep_token_id = self.original_tokenizer.sep_token_id
tokenizer.post_processor = processors.TemplateProcessing(
single=f"{cls}:0 $A:0 {sep}:0",
pair=
f"{cls}:0 $A:0 {sep}:0 {sep}:0 $B:1 {sep}:1", # MPNet uses two [SEP] tokens
special_tokens=[
(cls, cls_token_id),
(sep, sep_token_id),
],
)
tokenizer.decoder = decoders.WordPiece(prefix="##")
return tokenizer
def converted(self) -> Tokenizer:
from .models.roformer.tokenization_utils import JiebaPreTokenizer
vocab = self.original_tokenizer.vocab
tokenizer = Tokenizer(
WordPiece(vocab, unk_token=str(self.original_tokenizer.unk_token)))
strip_accents = False
do_lower_case = False
if hasattr(self.original_tokenizer, "basic_tokenizer"):
strip_accents = self.original_tokenizer.basic_tokenizer.strip_accents
do_lower_case = self.original_tokenizer.basic_tokenizer.do_lower_case
tokenizer.normalizer = normalizers.BertNormalizer(
clean_text=True,
handle_chinese_chars=False,
strip_accents=strip_accents,
lowercase=do_lower_case,
)
tokenizer.pre_tokenizer = pre_tokenizers.PreTokenizer.custom(
JiebaPreTokenizer(vocab))
cls = str(self.original_tokenizer.cls_token)
sep = str(self.original_tokenizer.sep_token)
cls_token_id = self.original_tokenizer.cls_token_id
sep_token_id = self.original_tokenizer.sep_token_id
tokenizer.post_processor = processors.TemplateProcessing(
single=f"{cls}:0 $A:0 {sep}:0",
pair=f"{cls}:0 $A:0 {sep}:0 $B:1 {sep}:1",
special_tokens=[
(cls, cls_token_id),
(sep, sep_token_id),
],
)
tokenizer.decoder = decoders.WordPiece(prefix="##")
return tokenizer
def train_custom_tokenizer(dataset,
token_model,
tknzr_file,
vocab_size,
vocab=None,
pretrain_fast=False,
max_input_chars_per_word=None,
eos_token=None,
bos_token=None,
pad_token=None,
mask_token=None,
unk_token=None):
"""
Building a Tokenizer using HuggingFace library. The pipeline seems to be:
- Model : algorithm that tokenizes, it is a mandatory
component. There are only 4 models implemented
(BPE, Unigram, WordLevel, WordPiece)
- Normalizer : some preprocessing that could happen before, but
doesn't necessarily have to
- Pre-Tokenizer : splitting the input according to some rules
- Post-Processing : needing to add some tokens/input after (mostly seems
to be eos, bos tokens)
- Decoder : certain previous pipeline steps need to be reversed
for proper decoding
- Trainer : The corresponding training algorithm for the model
Note : Some pre-processing might need to happen beforehand in previous
functions (might be easier using pandas before)
Input
token_model (str) : algorithm to use for tokenization
dataset (class) : a python iterator that goes through the data
to be used for training
token_dir (str) : directory with tokenizers
vocab_size (int) : size of the vocabulary to use
tokenFilename (str) : filename of particular token we want to
train. Will overwrite previously save files.
vocab (list of str) : models other than BPE can use non-mandatory
vocab as input
max_input_chars_per_word : used for WordPiece
Output
tokenizer : huggingFace Tokenizer object, our fully
trainer tokenizer
"""
special_token_lst = [
pad_token, bos_token, eos_token, mask_token, unk_token
]
# NFKC
normalizer_lst = []
pre_tokenizer_lst = [Whitespace, ByteLevel]
decoder_lst = []
bos_idx = special_token_lst.index(bos_token)
eos_idx = special_token_lst.index(eos_token)
if token_model == 'BPE':
model = BPE(unk_token=unk_token)
Trainer = BpeTrainer
elif token_model == 'Unigram':
model = Unigram(vocab=vocab)
Trainer = UnigramTrainer
elif token_model == 'WordLevel':
model = WordLevel(unk_token=unk_token, vocab=vocab)
Trainer = WordLevelTrainer
elif token_model == 'WordPiece':
model = WordPiece(unk_token=unk_token,
vocab=vocab,
max_input_chars_per_word=max_input_chars_per_word)
Trainer = WordPieceTrainer
else:
error_msg = f'Error: token_model ({token_model}) not an algorithm in%s' \
% VALID_TOKENIZATIONS
raise SystemExit(error_msg)
# instantiation
tokenizer = Tokenizer(model)
# Select a tokenization trainer
if vocab_size is None:
trainer = Trainer(show_progress=True, special_tokens=special_token_lst)
else:
trainer = Trainer(vocab_size=vocab_size,
show_progress=True,
special_tokens=special_token_lst)
# Set the normalizer
tokenizer.normalizer = normalizers.Sequence(
[fcn() for fcn in normalizer_lst])
# Set the pre-tokenizer
tokenizer.pre_tokenizer = pre_tokenizers.Sequence(
[fcn() for fcn in pre_tokenizer_lst])
# Set the post-processing
tokenizer.post_processor = processors.TemplateProcessing(
single=bos_token + " $A " + eos_token,
special_tokens=[(bos_token, bos_idx), (eos_token, eos_idx)],
# pair=bos_token+" $A "+eos_token" $B:1 "+eos_token+":1",
)
# Set the decoder
if ByteLevel in pre_tokenizer_lst:
tokenizer.decoder = decoders.ByteLevel()
if Metaspace in pre_tokenizer_lst:
tokenizer.decoder = decoders.Metaspace()
if token_model == 'WordPiece':
tokenizer.decoder = decoders.WordPiece()
# creating iterator
def batch_iterator():
for i in np.arange(0, len(dataset)):
yield dataset[i]
# train call
tokenizer.train_from_iterator(trainer=trainer,
iterator=batch_iterator(),
length=len(dataset))
if Path(tknzr_file).exists():
print(f"Warning : overwriting previously save tokenizer with\
same filename ( {tknzr_file} ).")
tokenizer.save(tknzr_file)
if pretrain_fast:
tokenizer = PreTrainedTokenizerFast(tokenizer_file=tknzr_file)
else:
tokenizer = PreTrainedTokenizer(tokenizer_file=tknzr_file)
tokenizer.pad_token = pad_token
tokenizer.mask_token = mask_token
return tokenizer
