Skip to content

Dataset to assess the disentanglement properties of unsupervised learning methods

License

Notifications You must be signed in to change notification settings

supershiye/dsprites-dataset

 
 

Repository files navigation

dSprites - Disentanglement testing Sprites dataset

This repository contains the dSprites dataset, used to assess the disentanglement properties of unsupervised learning methods.

If you use this dataset in your work, please cite it as follows:

Bibtex

@misc{dsprites17,
author = {Loic Matthey and Irina Higgins and Demis Hassabis and Alexander Lerchner},
title = {dSprites: Disentanglement testing Sprites dataset},
howpublished= {https://github.com/deepmind/dsprites-dataset/},
year = "2017",
}

Description

dsprite_gif

dSprites is a dataset of 2D shapes procedurally generated from 6 ground truth independent latent factors. These factors are color, shape, scale, rotation, x and y positions of a sprite.

All possible combinations of these latents are present exactly once, generating N = 737280 total images.

Latent factor values

  • Color: white
  • Shape: square, ellipse, heart
  • Scale: 6 values linearly spaced in [0.5, 1]
  • Orientation: 40 values in [0, 2 pi]
  • Position X: 32 values in [0, 1]
  • Position Y: 32 values in [0, 1]

We varied one latent at a time (starting from Position Y, then Position X, etc), and sequentially stored the images in fixed order. Hence the order along the first dimension is fixed and allows you to map back to the value of the latents corresponding to that image.

We chose the latents values deliberately to have the smallest step changes while ensuring that all pixel outputs were different. No noise was added.

The data is a NPZ NumPy archive with the following fields:

  • imgs: (737280 x 64 x 64, uint8) Images in black and white.
  • latents_values: (737280 x 6, float64) Values of the latent factors.
  • latents_classes: (737280 x 6, int64) Integer index of the latent factor values. Useful as classification targets.
  • metadata: some additional information, including the possible latent values.

Alternatively, a HDF5 version is also available, containing the same data, packed as Groups and Datasets.

Disentanglement metric

This dataset was created as a unit test of disentanglement properties of unsupervised models. It can be used to determine how well models recover the ground truth latents presented above.

You find our proposed disentanglement metric assessing the disentanglement quality of a model (along with an example usage of this dataset) in:

Higgins, Irina, Loic Matthey, Arka Pal, Christopher Burgess, Xavier Glorot, Matthew Botvinick, Shakir Mohamed, and Alexander Lerchner. "beta-VAE: Learning basic visual concepts with a constrained variational framework." In Proceedings of the International Conference on Learning Representations (ICLR). 2017.

Disclaimers

This is not an official Google product.

The images were generated using the LOVE framework, which is licenced under zlib/libpng licence:

LOVE is Copyright (c) 2006-2016 LOVE Development Team

This software is provided 'as-is', without any express or implied
warranty. In no event will the authors be held liable for any damages
arising from the use of this software.

Permission is granted to anyone to use this software for any purpose,
including commercial applications, and to alter it and redistribute it
freely, subject to the following restrictions:

1. The origin of this software must not be misrepresented; you must not
claim that you wrote the original software. If you use this software
in a product, an acknowledgment in the product documentation would be
appreciated but is not required.

2. Altered source versions must be plainly marked as such, and must not be
misrepresented as being the original software.

3. This notice may not be removed or altered from any source
distribution.

About

Dataset to assess the disentanglement properties of unsupervised learning methods

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Jupyter Notebook 100.0%