image_dataset.py

import os
import json
import random
from datasets import (
    BuilderConfig,
    DatasetInfo,
    DownloadManager,
    GeneratorBasedBuilder,
    SplitGenerator,
    Split,
    Features,
    Image,
    Sequence,
    Value,
)

class ImageSubsetConfig(BuilderConfig):
    """BuilderConfig for small vs. full dataset."""
    def __init__(self, name, sample_size=None, **kwargs):
        super().__init__(
            name=name,
            version="1.0.0",
            description=kwargs.get("description", "")
        )
        self.sample_size = sample_size

class MyImageDataset(GeneratorBasedBuilder):
    """A dataset of images with actuated_angle from metadata.json."""
    BUILDER_CONFIGS = [
        ImageSubsetConfig(
            name="full",
            sample_size=None,
            description="All images (≈100 GB). Use with streaming=True."
        ),
        ImageSubsetConfig(
            name="small",
            sample_size=2,
            description="Random subset of 10 000 images for quick iteration."
        ),
    ]
    DEFAULT_CONFIG_NAME = "smgit all"

    def _info(self):
        return DatasetInfo(
            description="Images with a 2-D actuated_angle from metadata.json",
            features=Features({
                "image": Image(),                      # the image file
                "actuated_angle": Sequence(Value("int32")),  # [angle0, angle1]
            }),
            supervised_keys=None,
        )

    def _split_generators(self, dl_manager: DownloadManager):
        data_dir = dl_manager.manual_dir or os.path.abspath(os.path.dirname(__file__))
        images_dir = os.path.join(data_dir, "images")
        meta_path = os.path.join(data_dir, "metadata.json")

        # load your filename → {"0":…, "1":…} mapping
        with open(meta_path, "r", encoding="utf-8") as f:
            metadata = json.load(f)

        # sample if in "small" config
        if self.config.sample_size:
            all_fnames = list(metadata.keys())
            random.seed(42)
            selected = set(random.sample(all_fnames, self.config.sample_size))
        else:
            selected = set(metadata.keys())

        return [
            SplitGenerator(
                name=Split.TRAIN,
                gen_kwargs={
                    "images_dir": images_dir,
                    "metadata": metadata,
                    "selected": selected,
                },
            ),
        ]

    def _generate_examples(self, images_dir, metadata, selected):
        """Yields examples of the form { "image": path, "actuated_angle": [int, int] }"""
        for idx, fname in enumerate(selected):
            if fname not in metadata:
                continue
            image_path = os.path.join(images_dir, fname)
            if not os.path.exists(image_path):
                continue
            meta = metadata[fname]
            # extract the two angle values (keys "0" and "1")
            angles = [ int(meta.get("0", 0)), int(meta.get("1", 0)) ]
            yield idx, {
                "image": image_path,
                "actuated_angle": angles,
            }