dataset.py

import os
import json
from datasets import GeneratorBasedBuilder, DatasetInfo, SplitGenerator, Split, Features, Value, Sequence, Image as HFImage


class LineGraphicsDataset(GeneratorBasedBuilder):
    def _info(self):
        return DatasetInfo(
            features=Features({
                "image_name": Value("string"),
                "image": HFImage(),
                "width": Value("int32"),
                "height": Value("int32"),
                "instances": Sequence({
                    "category_id": Value("int32"),
                    "mask": Sequence(Sequence(Value("float32")))  # List of polygons (each polygon = list of floats)
                })
            }),
            description="Line Graphics Dataset with polygon instance masks in COCO format.",
        )

    def _split_generators(self, dl_manager):
        data_dir = dl_manager.download_and_extract(self.config.data_dir or ".")

        return [
            SplitGenerator(
                name=Split.TRAIN,
                gen_kwargs={"split_dir": os.path.join(data_dir, "train")},
            ),
            SplitGenerator(
                name=Split.VALIDATION,
                gen_kwargs={"split_dir": os.path.join(data_dir, "val")},
            ),
            SplitGenerator(
                name=Split.TEST,
                gen_kwargs={"split_dir": os.path.join(data_dir, "test")},
            ),
        ]

    def _generate_examples(self, split_dir):
        # Load annotations
        annotation_file = os.path.join(split_dir, "annotations.json")
        with open(annotation_file, "r") as f:
            coco = json.load(f)

        # Map image_id to metadata
        image_id_to_info = {img["id"]: img for img in coco["images"]}

        # Group annotations by image
        from collections import defaultdict
        anns_by_image = defaultdict(list)
        for ann in coco["annotations"]:
            anns_by_image[ann["image_id"]].append(ann)

        # Generate examples
        for image_id, image_info in image_id_to_info.items():
            file_name = image_info["file_name"]
            image_path = os.path.join(split_dir, "images", file_name)

            instances = []
            for ann in anns_by_image[image_id]:
                # Each mask is a list of polygons
                segmentation = ann["segmentation"]
                if not isinstance(segmentation, list):
                    continue  # skip if malformed
                instances.append({
                    "category_id": ann["category_id"],
                    "mask": segmentation
                })

            yield file_name, {
                "image_name": file_name,
                "image": image_path,
                "width": image_info["width"],
                "height": image_info["height"],
                "instances": instances
            }