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
