common_formulas.py

from io import BytesIO
from PIL import Image
from pathlib import Path
import datasets
import json


RAW_METADATA_URL = r'/home/lhy/arxiv_crawler/common_formulas/raw_formulas.jsonl'

DIR_URL = r'/home/lhy/arxiv_crawler/common_formulas'


class LatexFormulasConfig(datasets.BuilderConfig):
    def __init__(self, data_url, **kwargs):
        super().__init__(**kwargs)
        self.data_url = data_url


class LatexFormulas(datasets.GeneratorBasedBuilder):
    BUILDER_CONFIGS = [
        LatexFormulasConfig(
            name="raw_formulas",
            data_url=RAW_METADATA_URL
        ),
        LatexFormulasConfig(
            name="tokenized_formulas",
            data_url=DIR_URL
        )
    ]

    def _info(self):
        if self.config.name == "raw_formulas":
            return datasets.DatasetInfo(
                features=datasets.Features({
                    "latex_formula": datasets.Value("string")
                })
            )
        if self.config.name == "tokenized_formulas":
            return datasets.DatasetInfo(
                features=datasets.Features({
                    "image": datasets.Image(),
                    "latex_formula": datasets.Value("string")
                })
            )

    def _split_generators(self, dl_manager: datasets.DownloadManager):
        data_path = dl_manager.download(self.config.data_url)
        if self.config.name == 'raw_formulas':
            return [
                datasets.SplitGenerator(
                    name=datasets.Split.TRAIN,
                    gen_kwargs={
                        "data_path": data_path
                    }
                )
            ]

        if self.config.name == "tokenized_formulas":
            dir_path = Path(data_path)
            assert dir_path.is_dir()

            return [
                datasets.SplitGenerator(
                    name=datasets.Split.TRAIN,
                    gen_kwargs={
                        'dir_path': dir_path,
                        'dl_manager': dl_manager
                    }
                )
            ]

    def _generate_examples(self, data_path=None, dir_path: Path=None, dl_manager=None):
        if self.config.name == 'tokenized_formulas':
            idx = 0
            for directory in dir_path.iterdir():
                if not directory.is_dir():
                    continue
                if not directory.name.startswith('process'):
                    continue
                image_path = str(directory / "compressed_img.tar.gz")
                metadata_path = str(directory / "tokenized_finally.jsonl")
                images = dl_manager.iter_archive(image_path)

                img_formula_pair = {}
                with open(metadata_path, 'r', encoding='utf-8') as f:
                    for line in f:
                        single_json = json.loads(line)
                        img_formula_pair[single_json['id']] = single_json['formula']

                for img_path, img_obj in images:
                    img_name = img_path.split('/')[-1]

                    if img_name in img_formula_pair:
                        idx += 1
                        # yield idx, {
                        yield str(directory) + img_path, {
                            "image": {"path": img_path, "bytes": img_obj.read()},
                            "latex_formula": img_formula_pair[img_name]
                        }

        if self.config.name == 'raw_formulas':
            assert data_path is not None
            with open(data_path, 'r', encoding="utf-8") as f:
                for idx, line in enumerate(f):
                    yield idx, {
                        "latex_formula": json.loads(line)["formula"]
                    }