EcclesialSchematisms.py

import json, os
from PIL import Image
import datasets
from datasets.utils.file_utils import xopen          # <─ key helper
from datasets.utils.logging import get_logger, tqdm

logger = get_logger(__name__)
logger.setLevel("INFO")

NUM_PARTS = 14
_TAR_URLS = [
    f"https://huggingface.co/datasets/artpods56/EcclesialSchematisms/"
    f"resolve/main/images/images_part{i}.tar"
    for i in range(1, NUM_PARTS + 1)
]
_LABEL_URL = "https://huggingface.co/datasets/artpods56/EcclesialSchematisms/" \
             "resolve/main/labels/train.jsonl"

class EcclesiaeSchematisms(datasets.GeneratorBasedBuilder):
    def _info(self):
        return datasets.DatasetInfo(
            features=datasets.Features({
                "image_pil": datasets.Image(decode="pil"),   # keep memory low
                "image":     datasets.Value("string"),
                "width":     datasets.Value("int32"),
                "height":    datasets.Value("int32"),
                "words":     datasets.Sequence(datasets.Value("string")),
                "bboxes":    datasets.Sequence(datasets.Sequence(datasets.Value("int32"))),
                "labels":    datasets.Sequence(datasets.Value("string")),
                "conf":      datasets.Sequence(datasets.Value("float32")),
            })
        )

    def _split_generators(self, dl_manager):
        images_tars = dl_manager.download(_TAR_URLS)
        labels_file = dl_manager.download(_LABEL_URL)  # local, small

        return [
            datasets.SplitGenerator(
                name=datasets.Split.TRAIN,
                gen_kwargs={
                    "images_tars": images_tars,
                    "labels": labels_file,
                    "dl_manager": dl_manager,          # pass it forward
                },
            )
        ]

    # NOTE the extra parameter
    def _generate_examples(self, images_tars, labels, dl_manager):
        logger.info("Loading label annotations…")
        annotations = {}
        with xopen(labels, "r") as f:                  # works in streaming
            for line in f:
                ann = json.loads(line)
                annotations[os.path.basename(ann["image_path"])] = ann
        logger.info("Loaded %d annotations.", len(annotations))

        example_id = 0
        for tar_path in images_tars:
            logger.info("Iterating %s", tar_path)
            # iter_archive streams members one-by-one
            for member_name, fobj in dl_manager.iter_archive(tar_path):
                if not member_name.endswith(".jpg"):
                    continue
                img_name = os.path.basename(member_name)
                ann = annotations.get(img_name)
                if ann is None:
                    continue

                with Image.open(fobj).convert("RGB") as image:
                    yield example_id, {
                        "image_pil": image,
                        "image": ann["image_path"],
                        "width": ann["width"],
                        "height": ann["height"],
                        "words": ann["words"],
                        "bboxes": ann["bboxes"],
                        "labels": ann["labels"],
                        "conf": ann.get("conf", []),
                    }
                example_id += 1