This is a PyTorch implementation for generating the Spirograph dataset as in Improving Transformation Invariance in Contrastive Representation Learning. Spirograph produces beautiful images from a fully differentiable generative process using a family of curves called hypotrochoids.
The script generate_spirograph.py generates two tensor datasets called spirograph_train.pth and spirograph_test.pth of size 5000 and 2000 respectively. Each element of the dataset is a pair of (3,32,32) Spirograph images and its 10 corresponding generative and nuisance transformation parameters.
The input of the class DrawSpirograph contains generative and transformations parameters that we can change. For example, the default setting is as follows
spirograph = DrawSpirograph(['m', 'b', 'sigma', 'rfore'], ['h', 'rback', 'gfore', 'gback', 'bfore', 'bback'])
Each parameter is sampled from a uniform distribution, m ~ U(2, 5), b ~ U(0.1, 1.1), h ~ U(0.5, 2.5), sigma ~ U(0.25, 1), foreground_colour ~ U(0.4, 1), background ~ U(0, 0.6).
The generating process of the Spirograph dataset is fully differentiable. When we call a function spirograph.dataset() we draw only the generative parameters from the mentioned distribution. We also generate transformations parameters by calling spirograph.sample_random_numbers(). Now we can generate the spirograph image by passing these parameters to the function spirograph(gen_param, trans_param) to get the generated spirograph image. The final step produces Spirograph sample images with pixel values that are fully differentiable in the input tensors.
If you use the Spirograph dataset, please consider citing the following paper
@article{foster2020improving,
title={Improving Transformation Invariance in Contrastive Representation Learning},
author={Foster, Adam and Pukdee, Rattana and Rainforth, Tom},
journal={arXiv preprint arXiv:2010.09515},
year={2020}
}