dsprites.py

import numpy as np
import datasets
from sklearn.model_selection import train_test_split


class DSprites(datasets.GeneratorBasedBuilder):
    """TODO: Short description of my dataset."""

    VERSION = datasets.Version("1.0.0")

    def _info(self):
        features = datasets.Features(
            {
                "image": datasets.Image(),
                "orientation": datasets.Value("float"),
                "shape": datasets.ClassLabel(names=["square", "ellipse", "heart"]),
                "scale": datasets.Value("float"),
                "color": datasets.ClassLabel(names=["white"]),
                "position_x": datasets.Value("float"),
                "position_y": datasets.Value("float"),
            }
        )

        homepage = "https://github.com/deepmind/dsprites-dataset"
        license = "zlib/libpng"
        return datasets.DatasetInfo(
            description="""dSprites is a dataset of 2D shapes procedurally generated from 6 ground truth independent latent factors. These factors are color, shape, scale, rotation, x and y positions of a sprite.
All possible combinations of these latents are present exactly once, generating N = 737280 total images.""",
            features=features,
            supervised_keys=("image", "shape"),
            homepage=homepage,
            license=license,
            citation="""@misc{dsprites17,
                        author = {Loic Matthey and Irina Higgins and Demis Hassabis and Alexander Lerchner},
                        title = {dSprites: Disentanglement testing Sprites dataset},
                        howpublished= {https://github.com/deepmind/dsprites-dataset/},
                        year = "2017"}""",
        )

    def _split_generators(self, dl_manager):
        archive = dl_manager.download(
            "https://github.com/google-deepmind/dsprites-dataset/raw/refs/heads/master/dsprites_ndarray_co1sh3sc6or40x32y32_64x64.npz"
        )
        return [
            datasets.SplitGenerator(
                name=datasets.Split.TRAIN,
                gen_kwargs={"archive": archive, "split": "train"},
            ),
            datasets.SplitGenerator(
                name=datasets.Split.TEST,
                gen_kwargs={"archive": archive, "split": "test"},
            ),
        ]

    # method parameters are unpacked from `gen_kwargs` as given in `_split_generators`
    def _generate_examples(self, archive, split):
        dataset_zip = np.load(archive, allow_pickle=True)
        images = dataset_zip["imgs"]
        latents_values = dataset_zip["latents_values"]

        # Split the indices for train and test
        indices = np.arange(len(images))
        train_indices, test_indices = train_test_split(indices, test_size=0.3, random_state=42)

        if split == "train":
            selected_indices = train_indices
        elif split == "test":
            selected_indices = test_indices

        for key in selected_indices:
            yield int(key), {  # Ensure the key is a Python native int
                "image": images[key],
                "color": int(latents_values[key, 0]) - 1,
                "shape": int(latents_values[key, 1]) - 1,
                "scale": latents_values[key, 2],
                "orientation": latents_values[key, 3],
                "position_x": latents_values[key, 4],
                "position_y": latents_values[key, 5],
            }