dsprites-noisy.py

import datasets
import numpy as np
from PIL import Image

_DSPRITES_URL = "https://github.com/google-deepmind/dsprites-dataset/raw/refs/heads/master/dsprites_ndarray_co1sh3sc6or40x32y32_64x64.npz"

class DSprites(datasets.GeneratorBasedBuilder):
    """dSprites dataset: 3x6x40x32x32 factor combinations, 64x64 binary images."""

    VERSION = datasets.Version("1.0.0")

    def _info(self):
        return datasets.DatasetInfo(
            description=(
                "Noisy dSprites dataset: procedurally generated 2D shapes dataset with known ground-truth factors, "
                "commonly used for disentangled representation learning. "
                "This is the NoisyDSprites variant, where each background is randomly noised, "
                "while background remains black. "
                "Factors: color (1), shape (3), scale (6), orientation (40), position X (32), position Y (32). "
                "Images are 64x64 RGB."
            ),
            features=datasets.Features(
                {
                    "image": datasets.Image(),  # (64, 64), grayscale
                    "index": datasets.Value("int32"),  # index of the image
                    "label": datasets.Sequence(datasets.Value("int32")),  # 6 factor indices (classes)
                    "label_values": datasets.Sequence(datasets.Value("float32")),  # 6 factor continuous values
                    "color": datasets.Value("int32"),  # color index (always 0)
                    "shape": datasets.Value("int32"),  # shape index (0-2)
                    "scale": datasets.Value("int32"),  # scale index (0-5)
                    "orientation": datasets.Value("int32"),  # orientation index (0-39)
                    "posX": datasets.Value("int32"),  # posX index (0-31)
                    "posY": datasets.Value("int32"),  # posY index (0-31)
                    "colorValue": datasets.Value("float64"),  # color index (always 0)
                    "shapeValue": datasets.Value("float64"),  # shape index (0-2)
                    "scaleValue": datasets.Value("float64"),  # scale index (0-5)
                    "orientationValue": datasets.Value("float64"),  # orientation index (0-39)
                    "posXValue": datasets.Value("float64"),  # posX index (0-31)
                    "posYValue": datasets.Value("float64"),  # posY index (0-31)
                }
            ),
            supervised_keys=("image", "label"),
            homepage="https://github.com/google-research/disentanglement_lib/tree/master",
            license="apache-2.0",
            citation="""@inproceedings{locatello2019challenging,
  title={Challenging Common Assumptions in the Unsupervised Learning of Disentangled Representations},
  author={Locatello, Francesco and Bauer, Stefan and Lucic, Mario and Raetsch, Gunnar and Gelly, Sylvain and Sch{\"o}lkopf, Bernhard and Bachem, Olivier},
  booktitle={International Conference on Machine Learning},
  pages={4114--4124},
  year={2019}
}""",
        )

    def _split_generators(self, dl_manager):
        npz_path = dl_manager.download(_DSPRITES_URL)

        return [
            datasets.SplitGenerator(
                name=datasets.Split.TRAIN,
                gen_kwargs={"npz_path": npz_path},
            ),
        ]

    def _generate_examples(self, npz_path):
        # Load npz
        data = np.load(npz_path, allow_pickle=True)
        images = data["imgs"]  # shape: (737280, 64, 64), uint8
        latents_classes = data["latents_classes"]  # shape: (737280, 6), int64
        latents_values = data["latents_values"]    # shape: (737280, 6), float64

        # Iterate over images
        for idx in range(len(images)):
            img = images[idx]  # (64, 64), uint8
            img = img.astype(np.float32) / 1.0
            noise = np.random.uniform(0, 1, size=(64, 64, 3))
            img_rgb = np.minimum(img[..., None] + noise, 1.0) * 255
            img_pil = Image.fromarray(img_rgb.astype(np.uint8), mode="RGB")

            factors_classes = latents_classes[idx].tolist()  # [color_idx, shape_idx, scale_idx, orientation_idx, posX_idx, posY_idx]
            factors_values = latents_values[idx].tolist()

            yield idx, {
                "image": img_pil,
                "index": idx,
                "label": factors_classes,
                "label_values": factors_values,
                "color": factors_classes[0],        # always 0
                "shape": factors_classes[1],
                "scale": factors_classes[2],
                "orientation": factors_classes[3],
                "posX": factors_classes[4],
                "posY": factors_classes[5],
                "colorValue": factors_values[0],    # always 0.0
                "shapeValue": factors_values[1],
                "scaleValue": factors_values[2],
                "orientationValue": factors_values[3],
                "posXValue": factors_values[4],
                "posYValue": factors_values[5],
            }