This repository contains the Tetrominoes dataset, used to assess the generalization performance of variational autodencoders.
If you use this dataset in your work, please cite it as follows:
@misc{tetrominoes19,
author = {Alican Bozkurt and Babak Esmaeili and Jennifer Dy and Dana Brooks and Jan-Willem van de Meent},
title = {Tetrominoes dataset},
howpublished= {https://github.com/neu-pml/tetrominoes/},
year = "2019",
}
Tetrominoes is a dataset of 2D shapes procedurally generated from 6 ground truth independent latent factors. These factors are rotation, color, scale, x and y positions, and shape.
To generate and save the Tetromino dataset, run:
python generate_tetromino.py <data_path>
Where data_path is the path of the directory where the data will be downloaded. This will create two files: id_tetrominos.pkl and ood_tetrominos.pkl corresponding to in-domain and out-of-domain settings with 50% split.
You can pass training ratio values as additional arguments in order to generate in-domain Tetrominos with different ratios of training-test splits.
python generate_tetromino.py <data_path> 0.1 0.3 0.6
where [0.1, 0.3, 0.6] are ratios.