tokenizer.py

import os
import re
import pandas as pd
import numpy as np
import datasets

logger = datasets.logging.get_logger(__name__)

import datasets

URL = "https://huggingface.co/datasets/thewall/tokenizer/resolve/main"


class TokenizerConfig(datasets.BuilderConfig):
    def __init__(self, **kwargs):
        super(TokenizerConfig, self).__init__(**kwargs)

class Tokenizer(datasets.GeneratorBasedBuilder):
    BUILDER_CONFIGS = [
        TokenizerConfig(name=key) for key in ["esm", "progen2", "raptgen", "aptamer"]
    ]

    DEFAULT_CONFIG_NAME = "esm"

    def _info(self):
        return datasets.DatasetInfo(
            features=datasets.Features(
                {
                    "id": datasets.Value("int32"),
                    "name": datasets.Value("string"),
                    "path": datasets.Value("string"),
                    "tokenizer": datasets.Value("string"),
                    "special_tokens_map": datasets.Value("string"),
                }
            ),
        )

    def _split_generators(self, dl_manager):
        file = dl_manager.download_and_extract(f"{URL}/{self.config.name}.tar.gz")
        return [
            datasets.SplitGenerator(name=datasets.Split.TRAIN, gen_kwargs={"filepath": file}),
        ]

    def _generate_examples(self, filepath):
        """This function returns the examples in the raw (text) form."""
        logger.info("generating examples from = %s", filepath)
        name = self.config.name
        tokenizer_file = os.path.join(filepath, name, "tokenizer.json")
        tokens_map_file = os.path.join(filepath, name, "special_tokens_map.json")
        with open(tokenizer_file) as f:
            tokenizer = "".join(f.readlines())
        with open(tokens_map_file) as f:
            special_tokens = "".join(f.readlines())
        yield 0, {"id": 0,
                  "path": os.path.join(filepath, name),
                  "name": name,
                  "tokenizer": tokenizer,
                  "special_tokens_map": special_tokens,}


if __name__=="__main__":
    from datasets import load_dataset
    from tokenizers import Tokenizer
    from transformers import PreTrainedTokenizerFast
    import json
    dataset = load_dataset("tokenizer.py", split="all")
    tokenizer = Tokenizer.from_str(dataset[0]['tokenizer'])
    token_map = json.loads(dataset[0]['special_tokens_map'])
    fast_tokenizer = PreTrainedTokenizerFast(tokenizer_object=tokenizer, **token_map)