Home

Awesome

SONGs: Self-Organizing Neural Graphs

song

The official implementation of the SONG model described in the paper.

Dependencies

The implementation is done in Python 3.8. The packages used to run the code:

Usage

Training

Use the following commands to train the models:

The folder ./pretrained_models should contain the following files:


Evaluation

Models can be evaluated using the following commands:

The variable path2model should contain the path to a model generated by the training procedure. An exemplary path:

path2model=./results/checkpoint/CIFAR100_nodes-256_leaves-100_N-20_graphs-5_lr-0.001_bs-128_pretrained_frozen-resnet_loss-BCELoss_mixupData_tau-1.0-dirty_seed-5007_2021-05-11_021407/models-epoch244.pth

Cite our paper

If you use our work, please cite our paper:

@inproceedings{struski2023songs,
  title={SONGs: Self-Organizing Neural Graphs},
  author={Struski, {\L}ukasz and Danel, Tomasz and {\'S}mieja, Marek and Tabor, Jacek and Zieli{\'n}ski, Bartosz},
  booktitle={Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision},
  year={2023}
}