build_tokenizer.py

from tokenizers.decoders import WordPiece as WordPieceDecoder
from tokenizers.pre_tokenizers import BertPreTokenizer
from tokenizers.normalizers import BertNormalizer
from tokenizers.trainers import WordPieceTrainer
from tokenizers.models import WordPiece as WordPieceModel
from tokenizers import Tokenizer
import itertools

from datasets import load_dataset
from datasets.utils.logging import set_verbosity_error
set_verbosity_error()

from utils import SampleBatch

def unpack_samples(
    batch: SampleBatch
):
    iterator = (
        sample.values()
        for sample in batch['translation']
    )

    return list(
        itertools.chain
        .from_iterable(iterator)
    )


def build_tokenizer(
    clean_text: bool = True,
    strip_accents: bool = True,
    lowercase: bool = True
) -> Tokenizer:
    tokenizer = Tokenizer(
        model=WordPieceModel(
            unk_token='<UNK>'
        )
    )
    tokenizer.normalizer = BertNormalizer(
        clean_text=clean_text,
        handle_chinese_chars=True,
        strip_accents=strip_accents,
        lowercase=lowercase
    )
    tokenizer.pre_tokenizer = BertPreTokenizer()
    tokenizer.decoder = WordPieceDecoder()

    return tokenizer


train_dset = load_dataset(
    path='nordmann2023',
    name='balanced',
    split='train'
)

tokenizer = build_tokenizer(
    clean_text=True,
    strip_accents=False,
    lowercase=False
)
tokenizer.train_from_iterator(
    iterator=(
        unpack_samples(batch)
        for batch in train_dset.iter(
            batch_size=10000
        )
    ),
    trainer=WordPieceTrainer(
        vocab_size=40000,
        special_tokens=[
            '<UNK>', '<CLS>', '<SEP>', '<PAD>', '<MASK>'
        ]
    ),
    length=train_dset.num_rows * 2
)
tokenizer.save(
    path='tokenizer.json'
)