dSprites - Disentanglement testing Sprites dataset

Dataset Description

dSprites is a dataset for evaluating the disentanglement properties of unsupervised learning methods, containing 2D shape images generated from six basic independent latent factors. These factors include color, shape, scale, rotation, x position, and y position.

Latent Factor Values

• Color: white
• Shape: square, ellipse, heart
• Scale: 6 values linearly spaced between [0.5, 1]
• Rotation: 40 values ranging from [0, 2π]
• x position: 32 values ranging from [0, 1]
• y position: 32 values ranging from [0, 1]

Dataset Structure

• imgs: 737,280 images, each 64×64 pixels, uint8 format
• latents_values: 737,280 × 6 floating‑point matrix representing latent factor values
• latents_classes: 737,280 × 6 integer matrix representing indices of latent factor values
• metadata: additional information such as possible latent factor values

Dataset Usage

The dataset is used to test the ability of unsupervised models to recover the aforementioned basic latent factors and to evaluate the quality of disentanglement.