Dataset Card for dSprites

Dataset Details

Dataset 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-position, and y-position of a sprite. All possible combinations of these latents are present exactly once, generating 737,280 total images. The dataset is widely used for studying disentangled representations in machine learning.

Dataset Sources

• Homepage: https://github.com/google-deepmind/dsprites-dataset
• Paper: Matthey, L., Higgins, I., Hassabis, D., & Lerchner, A. (2017, May). dsprites: Disentanglement testing sprites dataset.

Dataset Structure

Each sample in the dataset contains:

• image: A 64×64 grayscale image

• orientation: A float representing the rotation angle of the shape

• shape: A categorical label representing the shape type (square, ellipse, heart)

• scale: A float representing the size of the shape

• color: A categorical label representing the color (only 'white' in this dataset)

• position_x: A float representing the x-position of the shape

• position_y: A float representing the y-position of the shape

Total images: 737,280

Classes: 3 (square, ellipse, heart)

Splits:

• Train: 70% of total dataset

• Test: 30% of total dataset

Image specs: PNG format, 64×64 pixels, grayscale

Example Usage

Below is a quick example of how to load this dataset via the Hugging Face Datasets library.

from datasets import load_dataset  

# Load the dataset  
dataset = load_dataset("../../aidatasets/images/dsprites.py", split="train", trust_remote_code=True)
# dataset = load_dataset("../../aidatasets/images/dsprites.py", split="test", trust_remote_code=True)  

# Access a sample from the dataset  
example = dataset[0]  
image = example["image"]  
shape = example["shape"]  

image.show()  # Display the image  
print(f"Shape: {shape}")

Citation

BibTeX:

@misc{matthey2017dsprites, title={dsprites: Disentanglement testing sprites dataset}, author={Matthey, Loic and Higgins, Irina and Hassabis, Demis and Lerchner, Alexander}, year={2017} }