It is the official repository of "Generalized Neural Sorting Networks with Error-Free Differentiable Swap Functions," which has been presented at the 12th International Conference on Learning Representations (ICLR 2024).
For part of this repository, we modified the source code obtained from the repository linked.
You can install some required packages using the following command.
pip install .You can run experiments with src/main_multidigit.py and src/main_jigsaw.py.
For example, they are run with the following commands in the src directory.
# experiments on sorting multi-digit images
python main_multidigit.py --dataset mnist_cnn
python main_multidigit.py --dataset svhn_cnn
python main_multidigit.py --dataset mnist_transformer
python main_multidigit.py --dataset svhn_transformer
python main_multidigit.py --dataset mnist_transformer_large
python main_multidigit.py --dataset svhn_transformer_large
# experiments on sorting image fragments
python main_jigsaw.py --dataset mnist_2_2 --model cnn
python main_jigsaw.py --dataset mnist_2_2 --model transformer
python main_jigsaw.py --dataset mnist_3_3 --model cnn
python main_jigsaw.py --dataset mnist_3_3 --model transformer
python main_jigsaw.py --dataset cifar10_2_2 --model cnn
python main_jigsaw.py --dataset cifar10_2_2 --model transformer
python main_jigsaw.py --dataset cifar10_3_3 --model cnn
python main_jigsaw.py --dataset cifar10_3_3 --model transformerYou can test different learning rate and steepness using options --nloglr and --steepness, respectively.
@inproceedings{KimJ2024iclr,
title={Generalized Neural Sorting Networks with Error-Free Differentiable Swap Functions},
author={Kim, Jungtaek and Yoon, Jeongbeen and Cho, Minsu},
booktitle={Proceedings of the International Conference on Learning Representations (ICLR)},
year={2024},
address={Vienna, Austria)
}