randall-lab/dsprites

dSprites数据集是一个由6个独立的基础潜在因素（颜色、形状、大小、旋转角度、x轴位置、y轴位置）生成的二维形状数据集。每个潜在因素的组合都恰好出现一次，共生成737,280张图像。该数据集被广泛用于研究机器学习中的解耦表示。每个样本包含一个64×64的灰度图像以及形状、大小、颜色、位置等属性。数据集分为训练集和测试集，分别占总数据的70%和30%。

The dSprites dataset is a set of 2D shapes procedurally generated from 6 ground truth independent latent factors: color, shape, scale, rotation, x-position, and y-position of a sprite. Each combination of these latents appears exactly once, resulting in a total of 737,280 images. The dataset is widely used for studying disentangled representations in machine learning. Each sample contains a 64x64 grayscale image along with attributes such as shape, scale, color, and position. The dataset is split into a training set and a test set, accounting for 70% and 30% of the total data, respectively.