def get_recurrent_tokenizer(vocab,
max_context_tokens,
unk_token,
pad_token,
device="cpu"):
"""
Return a tokenizer to be used with recurrent-based models
"""
question_tokenizer = Tokenizer(WordLevel(vocab, unk_token=unk_token))
question_tokenizer.normalizer = Sequence(
[StripAccents(), Lowercase(), Strip()])
question_tokenizer.pre_tokenizer = PreSequence(
[Whitespace(), Punctuation()])
question_tokenizer.enable_padding(direction="right",
pad_id=vocab[pad_token],
pad_type_id=1,
pad_token=pad_token)
context_tokenizer = Tokenizer(WordLevel(vocab, unk_token=unk_token))
context_tokenizer.normalizer = Sequence(
[StripAccents(), Lowercase(), Strip()])
context_tokenizer.pre_tokenizer = PreSequence(
[Whitespace(), Punctuation()])
context_tokenizer.enable_padding(
direction="right",
pad_id=vocab[pad_token],
pad_type_id=1,
pad_token=pad_token,
)
context_tokenizer.enable_truncation(max_context_tokens)
return RecurrentSquadTokenizer(question_tokenizer,
context_tokenizer,
device=device)
def __init__(self,
vocab_size=25000,
min_freq=5,
lang="en",
files=[None, None]) -> None:
"""
Args:
vocab_size: (int)
min_freq: minimum frequency
lang:
files: (List[str]) ["vocab.json", "merge.txt"]
"""
super(BPETokenizer, self).__init__()
self.tokenizer = Tokenizer(BPE(files[0], files[1]))
self.lang = lang
self.trainer = BpeTrainer(vocab_size=vocab_size,
min_frequency=min_freq,
special_tokens=["[PAD]", "[SEP]"],
initial_alphabet=ByteLevel.alphabet())
# https://huggingface.co/docs/tokenizers/python/latest/components.html#normalizers
self.tokenizer.normalizer = Sequence([NFKC(), Lowercase()])
# https://huggingface.co/docs/tokenizers/python/latest/components.html#pre-tokenizers
self.tokenizer.pre_tokenizer = ByteLevel()
self.tokenizer.decoder = ByteLevelDecoder()
def main() -> None:
args = parse_args()
special_tokens = list(SPECIAL_TOKENS)
if args.reserved < len(special_tokens):
raise AssertionError(
f"number of reserved tokens should be more than number of f{len(special_tokens)}")
for i in range(len(special_tokens), args.reserved):
special_tokens.append(f"[unused{i:03d}]")
all_filenames = get_all_filenames(args.input)
# "C:\Users\demianmedich\data\wiki\20191120.en\pp_cased/"
tokenizer = Tokenizer(get_model(args.model))
tokenizer.normalizer = normalizers.Sequence([
NFKC(), StripAccents(), Lowercase()
])
tokenizer.pre_tokenizer = Whitespace()
trainer = WordPieceTrainer(
vocab_size=args.vocab_size,
special_tokens=special_tokens)
tokenizer.train(trainer, all_filenames)
model_files = tokenizer.model.save()
sys.exit(0)
def __init__(
self,
load_from: str = None,
vocab_size: int = 10000,
max_example_len: int = 128,
batch_size: int = 16,
num_stopwords: int = 250,
mask_output_len: int = 4,
):
self.char_dict: Dict[str, int] = {}
self.char_rev: Dict[int, str] = {}
self.token_dict: Dict[str, int] = {}
self.token_rev: Dict[str, int] = {}
self.vocab_size = vocab_size
self.max_example_len = max_example_len
self.batch_size = batch_size
self.num_stopwords = num_stopwords
self.mask_output_len = mask_output_len
self.tokenizer_fit = False
self.tokenizer = Tokenizer(BPE(unk_token="[UNK]"))
self.tokenizer.pre_tokenizer = Whitespace()
self.tokenizer.normalizer = Sequence(
[NFD(), Lowercase(), StripAccents()])
self.tok_trainer = BpeTrainer(special_tokens=["[UNK]", "[MASK]"],
vocab_size=self.vocab_size)
if load_from:
self._load(load_from)
def test_normalize(self):
tokenizer = Tokenizer(BPE())
tokenizer.add_tokens(["my", "name", "is", "john", "pair"])
tokenizer.normalizer = Lowercase()
output = tokenizer.normalize("My Name Is John")
assert output == "my name is john"
def __init__(self,
vocab_file: Optional[str] = None,
merges_file: Optional[str] = None,
unk_token: Optional[str] = "<unk>",
suffix: Optional[str] = "</w>",
dropout: Optional[float] = None):
if vocab_file is not None and merges_file is not None:
tokenizer = Tokenizer(
BPE.from_files(vocab_file,
merges_file,
dropout=dropout,
unk_token=unk_token,
end_of_word_suffix=suffix))
else:
tokenizer = Tokenizer(BPE.empty())
tokenizer.normalizer = Sequence.new([NFKC.new(), Lowercase.new()])
tokenizer.pre_tokenizer = pre_tokenizers.Whitespace.new()
tokenizer.decoder = decoders.BPEDecoder.new(suffix=suffix)
parameters = {
"model": "BPE",
"unk_token": unk_token,
"suffix": suffix,
"dropout": dropout,
}
super().__init__(tokenizer, parameters)
def train_tokenizer(langs, dataset, vocab_size):
"""Train a tokenizer on given list of languages.
Reserves a special token for each language which is
[LANG] where LANG is the language tag. These are assigned
to tokens 5, 6, ..., len(langs) + 4.
"""
# Byte-pair encoding
tokenizer = Tokenizer(BPE(unk_token='[UNK]'))
# trainer
lang_tokens = ['[' + lang + ']' for lang in langs]
special_tokens = ['[MASK]', '[CLS]', '[SEP]', '[PAD]', '[UNK]'] + lang_tokens
trainer = BpeTrainer(
special_tokens=special_tokens,
vocab_size=vocab_size)
# normalise and pre tokenize
tokenizer.normalizer = Sequence([NFD(), Lowercase(), StripAccents()])
tokenizer.pre_tokenizer = pre_tokenizers.ByteLevel()
tokenizer.decoder = decoders.ByteLevel()
# create iterator and train
iterator = _MultilingualIterator(dataset, langs)
tokenizer.train_from_iterator(iterator, trainer)
# post process start/end tokens
tokenizer.post_processor = TemplateProcessing(
single="[CLS] $A [SEP]",
pair="[CLS] $A [SEP] $B:1 [SEP]:1",
special_tokens=[
("[CLS]", tokenizer.token_to_id("[CLS]")),
("[SEP]", tokenizer.token_to_id("[SEP]")),
], )
return tokenizer
def get_daily_dialog_tokenizer(tokenizer_location=None):
'''
Get the daily dialog tokenizer. Trains a new one if no location is provided
:param tokenizer_location: Json containing information about the tokenizer.
:return:
'''
if tokenizer_location:
tokenizer = Tokenizer.from_file(tokenizer_location, )
tokenizer.enable_padding()
return tokenizer
else:
dataset_train = datasets.load_dataset("daily_dialog", split="train", )
utterances = [special_tokens["sep_token"].join(dialogue["dialog"]) for dialogue in dataset_train]
trainer = WordPieceTrainer(
vocab_size = 2048,
special_tokens = token_utils.special_tokens.values()
)
custom_tokenizer = Tokenizer(WordPiece(unk_token=special_tokens["unk_token"], ))
custom_tokenizer.normalizer = normalizers.Sequence([NFD(), Lowercase(), StripAccents()])
custom_tokenizer.pre_tokenizer = Whitespace()
custom_tokenizer.train_from_iterator(utterances, trainer, )
custom_tokenizer.enable_padding()
# Write every dialogue to file
location = './daily_dialog/'
custom_tokenizer.save(location + "tokenizer.json")
return custom_tokenizer
def __init__(
self,
path_src,
path_tgt,
path_tokenizer,
path_root: Optional[str] = '',
):
self.path_src = path_root + path_src
self.path_tgt = path_root + path_tgt
self.len = 0
self.max_len = 512
self.tokenizer = Tokenizer(
BPE(
path_root + path_tokenizer + 'vocab.json',
path_root + path_tokenizer + 'merges.txt',
))
self.tokenizer.normalizer = Sequence([NFKC(), Lowercase()])
with open(self.path_src, 'r+') as f:
lines_src = f.readlines()
with open(self.path_tgt, 'r+') as f:
lines_tgt = f.readlines()
self.len = len(lines_src)
self.example = list(zip(lines_src, lines_tgt))
def train():
"""Source: https://huggingface.co/docs/tokenizers/pipeline"""
base = os.environ['DATA_ROOT']
corpus_path = base + 'MimicIII/Encounters/Text/'
bert_tokenizer = Tokenizer(WordPiece(unk_token="[UNK]"))
# input to tokenizer.encode() goes through this pipeline:
# normalization, pre-tokenization, model, post-processing
bert_tokenizer.normalizer = normalizers.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)])
files = [str(file) for file in Path(corpus_path).glob('*.txt')]
trainer = WordPieceTrainer(
vocab_size=30522,
show_progress=True,
special_tokens=["[UNK]", "[CLS]", "[SEP]", "[PAD]", "[MASK]"])
bert_tokenizer.train(files, trainer)
os.mkdir('./Tokenizer')
bert_tokenizer.save("Tokenizer/tokenizer.json")
def __init__(
self,
vocab: Optional[Union[str, Dict[str, int]]] = None,
merges: Optional[Union[str, Dict[Tuple[int, int], Tuple[int,
int]]]] = None,
add_prefix_space: bool = False,
lowercase: bool = False,
dropout: Optional[float] = None,
unicode_normalizer: Optional[str] = None,
continuing_subword_prefix: Optional[str] = None,
end_of_word_suffix: Optional[str] = None,
trim_offsets: bool = False,
):
if vocab is not None and merges is not None:
tokenizer = Tokenizer(
BPE(
vocab,
merges,
dropout=dropout,
continuing_subword_prefix=continuing_subword_prefix or "",
end_of_word_suffix=end_of_word_suffix or "",
))
else:
tokenizer = Tokenizer(BPE())
# Check for Unicode normalization first (before everything else)
normalizers = []
if unicode_normalizer:
normalizers += [unicode_normalizer_from_str(unicode_normalizer)]
if lowercase:
normalizers += [Lowercase()]
# Create the normalizer structure
if len(normalizers) > 0:
if len(normalizers) > 1:
tokenizer.normalizer = Sequence(normalizers)
else:
tokenizer.normalizer = normalizers[0]
tokenizer.pre_tokenizer = pre_tokenizers.ByteLevel(
add_prefix_space=add_prefix_space)
tokenizer.decoder = decoders.ByteLevel()
tokenizer.post_processor = processors.ByteLevel(
trim_offsets=trim_offsets)
parameters = {
"model": "ByteLevelBPE",
"add_prefix_space": add_prefix_space,
"lowercase": lowercase,
"dropout": dropout,
"unicode_normalizer": unicode_normalizer,
"continuing_subword_prefix": continuing_subword_prefix,
"end_of_word_suffix": end_of_word_suffix,
"trim_offsets": trim_offsets,
}
super().__init__(tokenizer, parameters)
def __init__(
self,
vocab_file,
delimiter,
lowercase,
unk_token,
eos_token,
add_eos=False,
add_double_eos=False,
normalization: Optional[str] = None,
):
try:
tokenizer = WordLevel(vocab_file, unk_token=unk_token)
tokenizer = Tokenizer(tokenizer)
except Exception:
raise ValueError(
"Unable to parse file {}. Unknown format. "
"If you tried to load a model saved through TransfoXLTokenizer,"
"please note they are not compatible.".format(vocab_file))
# Create the correct normalization path
normalizer = []
# Include unicode normalization
if normalization:
normalizer += [unicode_normalizer_from_str(normalization)]
# Include case normalization
if lowercase:
normalizer += [Lowercase()]
# Strip normalizer at the end
normalizer += [Strip(left=True, right=True)]
if len(normalizer) > 0:
tokenizer.normalizer = Sequence(
normalizer) if len(normalizer) > 1 else normalizer[0]
# Setup the splitter
tokenizer.pre_tokenizer = CharDelimiterSplit(
delimiter) if delimiter else WhitespaceSplit()
if add_double_eos:
tokenizer.post_processor = BertProcessing(
(eos_token, tokenizer.token_to_id(eos_token)),
(eos_token, tokenizer.token_to_id(eos_token)))
parameters = {
"model": "TransfoXLModel",
"add_eos": add_eos,
"add_double_eos": add_double_eos,
"unk_token": unk_token,
"eos_token": eos_token,
"delimiter": delimiter,
"lowercase": lowercase,
}
super().__init__(tokenizer, parameters)
def __init__(
self,
vocab_file: Optional[str] = None,
merges_file: Optional[str] = None,
add_prefix_space: bool = False,
lowercase: bool = False,
dropout: Optional[float] = None,
unicode_normalizer: Optional[str] = None,
continuing_subword_prefix: Optional[str] = None,
end_of_word_suffix: Optional[str] = None,
):
if vocab_file is not None and merges_file is not None:
tokenizer = Tokenizer(
BPE.from_files(
vocab_file,
merges_file,
dropout=dropout,
continuing_subword_prefix=continuing_subword_prefix or "",
end_of_word_suffix=end_of_word_suffix or "",
))
else:
tokenizer = Tokenizer(BPE.empty())
# Check for Unicode normalization first (before everything else)
normalizers = []
if unicode_normalizer:
normalizers += [unicode_normalizer_from_str(unicode_normalizer)]
if lowercase:
normalizers += [Lowercase()]
# Create the normalizer structure
if len(normalizers) > 0:
if len(normalizers) > 1:
tokenizer.normalizer = Sequence(normalizers)
else:
tokenizer.normalizer = normalizers[0]
tokenizer.pre_tokenizer = pre_tokenizers.ByteLevel(
add_prefix_space=add_prefix_space)
tokenizer.decoder = decoders.ByteLevel()
parameters = {
"model": "ByteLevelBPE",
"add_prefix_space": add_prefix_space,
"lowercase": lowercase,
"dropout": dropout,
"unicode_normalizer": unicode_normalizer,
"continuing_subword_prefix": continuing_subword_prefix,
"end_of_word_suffix": end_of_word_suffix,
}
super().__init__(tokenizer, parameters)
def normalizer(self, proto):
normalizers = [Replace("``", '"'), Replace("''", '"')]
if not self.original_tokenizer.keep_accents:
normalizers.append(NFKD())
normalizers.append(StripAccents())
if self.original_tokenizer.do_lower_case:
normalizers.append(Lowercase())
precompiled_charsmap = proto.normalizer_spec.precompiled_charsmap
normalizers.append(Precompiled(precompiled_charsmap))
return Sequence(normalizers)
def tokenizer_pipeline():
"""
specific pipeline for Cebuano Corpus tokenization
- Uses a Byte pair encoding (BPE) tokenizer
"""
tokenizer = Tokenizer(BPE())
# string normalization
tokenizer.normalizer = Sequence([NFD(), StripAccents(), Lowercase()])
tokenizer.pre_tokenizer = ByteLevel()
tokenizer.decoder = ByteLevelDecoder()
return tokenizer
def __init__(
self,
vocab_file: Optional[str] = None,
unk_token: Union[str, AddedToken] = "[UNK]",
pad_token: Union[str, AddedToken] = "[PAD]",
mask_token: Union[str, AddedToken] = "[MASK]",
lowercase: bool = False,
unicode_normalizer: Optional[str] = None,
):
if vocab_file is not None:
logging.info(f"Initiating tokenizer at {vocab_file}")
tokenizer = Tokenizer(
WordLevel(vocab=vocab_file, unk_token=unk_token))
else:
tokenizer = Tokenizer(WordLevel(unk_token=unk_token))
# Let the tokenizer know about special tokens if they are part of the vocab
if tokenizer.token_to_id(str(unk_token)) is not None:
tokenizer.add_special_tokens([str(unk_token)])
if tokenizer.token_to_id(str(pad_token)) is not None:
tokenizer.add_special_tokens([str(pad_token)])
if tokenizer.token_to_id(str(mask_token)) is not None:
tokenizer.add_special_tokens([str(mask_token)])
# Check for Unicode normalization first (before everything else)
normalizers = []
if unicode_normalizer:
normalizers += [unicode_normalizer_from_str(unicode_normalizer)]
if lowercase:
normalizers += [Lowercase()]
# Create the normalizer structure
if len(normalizers) > 0:
if len(normalizers) > 1:
tokenizer.normalizer = Sequence(normalizers)
else:
tokenizer.normalizer = normalizers[0]
tokenizer.pre_tokenizer = pre_tokenizers.WhitespaceSplit()
parameters = {
"model": "WordLevel",
"unk_token": unk_token,
"pad_token": pad_token,
"mask_token": mask_token,
"lowercase": lowercase,
"unicode_normalizer": unicode_normalizer,
}
super().__init__(tokenizer, parameters)
def __init__(
self,
vocab_file,
delimiter,
lowercase,
unk_token,
eos_token,
add_eos=False,
add_double_eos=False,
normalization: Optional[str] = None,
):
tokenizer = WordLevel.from_files(vocab_file, unk_token=unk_token)
tokenizer = Tokenizer(tokenizer)
# Create the correct normalization path
normalizer = []
# Include unicode normalization
if normalization:
normalizer += [unicode_normalizer_from_str(normalization)]
# Include case normalization
if lowercase:
normalizer += [Lowercase()]
if len(normalizer) > 0:
tokenizer.normalizer = Sequence(
normalizer) if len(normalizer) > 1 else normalizer[0]
# Setup the splitter
tokenizer.pre_tokenizer = CharDelimiterSplit(
delimiter) if delimiter else WhitespaceSplit()
if add_double_eos:
tokenizer.post_processor = BertProcessing(
(eos_token, tokenizer.token_to_id(eos_token)),
(eos_token, tokenizer.token_to_id(eos_token)))
parameters = {
"model": "TransfoXLModel",
"add_eos": add_eos,
"add_double_eos": add_double_eos,
"unk_token": unk_token,
"eos_token": eos_token,
"delimiter": delimiter,
"lowercase": lowercase,
}
super().__init__(tokenizer, parameters)
def _prepare_pipeline(self):
self.tokenizer.normalizer = normalizers.Sequence(
[NFD(), Lowercase(), StripAccents()])
self.tokenizer.pre_tokenizer = Whitespace()
self.tokenizer.post_processor = TemplateProcessing(
single="[CLS] $A [SEP]",
pair="[CLS] $A [SEP] $B:1 [SEP]:1",
special_tokens=[
("[CLS]", 1),
("[SEP]", 2),
],
)
self.tokenizer.enable_padding(
pad_id=self.__class__.SPECIAL_TOKENS.index("[PAD]"),
pad_token="[PAD]")
def main(args):
# from tokenizers import BertWordPieceTokenizer
from tokenizers import Tokenizer
from tokenizers.models import WordPiece
bert_tokenizer = Tokenizer(WordPiece())
# bert_tokenizer = Tokenizer(MBartTokenizer())
from tokenizers import normalizers
from tokenizers.normalizers import Lowercase, NFD, StripAccents
bert_tokenizer.normalizer = normalizers.Sequence(
[NFD(), Lowercase(), StripAccents()])
from tokenizers.pre_tokenizers import Whitespace
bert_tokenizer.pre_tokenizer = Whitespace()
# from tokenizers.processors import TemplateProcessing
#
# bert_tokenizer.post_processor = TemplateProcessing(
# single="[CLS] $A [SEP]",
# pair="[CLS] $A [SEP] $B:1 [SEP]:1",
# special_tokens=[
# ("[CLS]", 1),
# ("[SEP]", 2),
# ],
# )
from tokenizers.trainers import WordPieceTrainer
trainer = WordPieceTrainer(
vocab_size=10000,
special_tokens=["[UNK]", "[CLS]", "[PAD]",
"[MASK]"] # "[SEP]", "[PAD]", "[MASK]"]
)
files = glob.glob(args.text_raw_files_pattern)
bert_tokenizer.train(trainer, files)
os.makedirs(args.output_dir, exist_ok=True)
model_files = bert_tokenizer.model.save(args.output_dir,
"bert-tokenizer-kr")
bert_tokenizer.model = WordPiece.from_file(*model_files, unk_token="[UNK]")
bert_tokenizer.save(os.path.join(args.output_dir,
"bert-tokenizer-kr.json"))
def configure_tokenizers(self, padding, truncation, max_length, lower):
# Settings
pad_length = None
if padding in {True, "longest"}:
pass
elif padding in {"max_length"}:
pad_length = max_length
elif padding in {False, "do_not_pad"}:
pass
else:
raise ValueError("Unknown padding type")
# SRC tokenizer
tok_normalizers = [NFD(), Strip()]
if lower:
tok_normalizers += [Lowercase()]
self.tokenizer = Tokenizer(tok_model()) # unk_token=... not working
self.tokenizer.add_special_tokens(self.special_tokens)
self.tokenizer.pre_tokenizer = pre_tokenizers.Sequence(
[WhitespaceSplit()])
self.tokenizer.normalizer = normalizers.Sequence(
tok_normalizers) # StripAccents requires NFD
self.tokenizer.decoder = tok_decoder()
# Define template (Needed for the sos/eos tokens)
basic_template = TemplateProcessing(
single=f"{self.SOS_WORD} $A {self.EOS_WORD}",
pair=
f"{self.SOS_WORD} $A {self.EOS_WORD} {self.SOS_WORD} $B {self.EOS_WORD}",
special_tokens=[
(self.SOS_WORD, self.tokenizer.token_to_id(self.SOS_WORD)),
(self.EOS_WORD, self.tokenizer.token_to_id(self.EOS_WORD))
],
)
self.tokenizer.post_processor = basic_template
if padding:
self.tokenizer.enable_padding(pad_id=self.tokenizer.token_to_id(
self.PAD_WORD),
pad_token=self.PAD_WORD,
length=pad_length)
if truncation:
self.tokenizer.enable_truncation(max_length,
stride=0,
strategy='longest_first')
def train_tokenizer(sentences: List[str], serialize_path: str = "", vocab_size: int = 8000) -> Tokenizer:
bert_tokenizer = Tokenizer(WordPiece(unk_token="[UNK]"))
bert_tokenizer.normalizer = normalizers.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(serialize_path)
return bert_tokenizer
def train_wordpiece_bert():
"""
Sample code from: https://huggingface.co/docs/tokenizers/python/latest/pipeline.html
"""
from tokenizers.models import WordPiece
bert_tokenizer = Tokenizer(WordPiece(unk_token="[UNK]"))
from tokenizers import normalizers
from tokenizers.normalizers import Lowercase, NFD, StripAccents
bert_tokenizer.normalizer = normalizers.Sequence(
[NFD(), Lowercase(), StripAccents()])
from tokenizers.pre_tokenizers import Whitespace
bert_tokenizer.pre_tokenizer = Whitespace()
from tokenizers.processors import TemplateProcessing
bert_tokenizer.post_processor = TemplateProcessing(
single="[CLS] $A [SEP]",
pair="[CLS] $A [SEP] $B:1 [SEP]:1",
special_tokens=[
("[CLS]", 1),
("[SEP]", 2),
],
)
bert_tokenizer.decoder = decoders.WordPiece()
from tokenizers.trainers import WordPieceTrainer
trainer = WordPieceTrainer(
vocab_size=30522,
special_tokens=["[UNK]", "[CLS]", "[SEP]", "[PAD]", "[MASK]"])
files = [
DIR_DATA + os.sep + 'wikitext-103' + os.sep + 'wiki.%s.raw' % a
for a in ["test", "train", "valid"]
]
bert_tokenizer.train(files, trainer)
bert_tokenizer.save(DIR_TOKENIZERS + os.sep + 'bert_wiki.json')
return bert_tokenizer
def __init__(self, tokenizer: PreTrainedTokenizerFast, cased: bool,
target_vocab_size: int):
"""
Args:
tokenizer: A Rust-based 🤗 Tokenizer
cased: If False, ignore uppercases in corpus
target_vocab_size: Size of augmented vocabulary
Raises:
ValueError: If :obj:`target_vocab_size` is larger or equal to the existing vocabulary of :obj:`tokenizer`
RuntimeError: If :obj:`tokenizer` uses an unsupported tokenization model
"""
if target_vocab_size <= tokenizer.vocab_size:
raise ValueError(
f"Ensure that `target_vocab_size` is larger than tokenizer's vocab size."
)
self.tokenizer = tokenizer
self.cased = cased
self.target_vocab_size = target_vocab_size
self.model_cls: Type[
BaseTokenizer] = tokenizer.backend_tokenizer.model.__class__
# Instantiate rust tokenizer
rust_tokenizer = Tokenizer(self.model_cls())
if not cased:
rust_tokenizer.normalizer = Lowercase()
rust_tokenizer.pre_tokenizer = Whitespace()
self.rust_tokenizer = rust_tokenizer
# Instantiate the appropriate Trainer based on `self.model` (i.e. BPE, WordPiece, etc)
trainer_cls = self.supported_trainers.get(self.model_cls, None)
if trainer_cls is None:
raise RuntimeError(f"{self.model_cls} is not supported")
self.trainer = trainer_cls(
vocab_size=self.target_vocab_size,
special_tokens=list(self.tokenizer.special_tokens_map.values()),
)
def test_can_make_sequences(self):
tokenizer = Tokenizer(BPE.empty())
tokenizer.normalizer = Sequence([Lowercase(), Strip()])
output = tokenizer.normalize(" HELLO ")
assert output == "hello"
# name = mention + concept
names = unique([mentions + list(concepts)], verbose=False) - stop_words
# names = unique([mentions + list(concepts)], verbose=False)
print(f"Unique names: {len(names)}\n")
name_words = {n: " ".join(split_to_words(n)) for n in names}
with open(f"{proc_path}/names.txt", "w") as f:
f.write("\n".join(list(name_words.values())))
# f.write("\n".join(words))
tokenizer = Tokenizer(BPE())
tokenizer.normalizer = Sequence([
# NFKC(),
Lowercase()
])
tokenizer.pre_tokenizer = ByteLevel()
tokenizer.decoder = ByteLevelDecoder()
trainer = BpeTrainer(vocab_size=int(vocab_size), show_progress=True)
tokenizer.train(trainer, [f"{proc_path}/names.txt"])
print("Trained vocab size: {}".format(tokenizer.get_vocab_size()))
tokenizer.model.save(proc_path)
tokenizer.model = BPE.from_file(f'{proc_path}/vocab.json',
f'{proc_path}/merges.txt')
with open(f"{proc_path}/vocab.json", "r") as f:
bpe_vocab = json.load(f)
def test_lowercase(self):
normalizer = Lowercase()
output = normalizer.normalize_str("HELLO")
assert output == "hello"
def test_instantiate(self):
assert isinstance(Lowercase(), Normalizer)
assert isinstance(Lowercase(), Lowercase)
assert isinstance(pickle.loads(pickle.dumps(Lowercase())), Lowercase)
def test_can_make_sequences(self):
normalizer = Sequence([Lowercase(), Strip()])
output = normalizer.normalize_str(" HELLO ")
assert output == "hello"
def test_lowercase(self):
tokenizer = Tokenizer(BPE.empty())
tokenizer.normalizer = Lowercase()
output = tokenizer.normalize("HELLO")
assert output == "hello"
from tokenizers.models import BPE, WordPiece, Unigram
from tokenizers.normalizers import Lowercase
from tokenizers.pre_tokenizers import Whitespace, Digits, Sequence
from tokenizers.trainers import BpeTrainer, WordPieceTrainer, UnigramTrainer
TRAIN_DATA_PATH = 'data/data_fusion_train.parquet'
OUTPUT_PATH = 'data/tokenizers/'
# Prepare data
train = pd.read_parquet(TRAIN_DATA_PATH, columns=['item_name'])
item_names = train.item_name.drop_duplicates().tolist()
# WordPiece tokenizer
tokenizer = Tokenizer(WordPiece(unk_token="[UNK]"))
tokenizer.pre_tokenizer = Sequence([Whitespace(), Digits()])
tokenizer.normalizer = Lowercase()
trainer = WordPieceTrainer(
special_tokens=["[UNK]", "[CLS]", "[SEP]", "[PAD]", "[MASK]"],
vocab_size=70000)
tokenizer.train_from_iterator(item_names, trainer)
tokenizer.save(os.path.join(OUTPUT_PATH, 'wordpiece_70k.json'))
# BPE tokenizer
tokenizer = Tokenizer(BPE(unk_token="[UNK]"))
tokenizer.pre_tokenizer = Sequence([Whitespace(), Digits()])
tokenizer.normalizer = Lowercase()
trainer = BpeTrainer(
special_tokens=["[UNK]", "[CLS]", "[SEP]", "[PAD]", "[MASK]"],
vocab_size=60000)