SONGs: Self-Organizing Neural Graphs

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:

• numpy
• pandas
• torch (>= 1.8)
• torchvision
• tabulate
• tqdm
• matplotlib
• seaborn

Usage

Training

Use the following commands to train the models:

• for shallow models:
  • MNIST

    bash ./scripts/run_shallow.sh mnist use_tqdm 1 "" "" "--dirty_tau --prob_leaves_rate 1" ""
    

  • LETTER

    bash ./scripts/run_shallow.sh letter use_tqdm 1 "" "" "--dirty_tau --prob_leaves_rate 1" ""
    

  • CONNECT4

    bash ./scripts/run_shallow.sh connect4 use_tqdm 1 "" "" "--dirty_tau --prob_leaves_rate 1" ""
    

• for deep model:
  • MNIST

    bash ./scripts/run_shallow.sh mnist use_tqdm 1 "representation" "./pretrained_models" "--dirty_tau --prob_leaves_rate 1" ""
    

  • CIFAR10

    bash ./scripts/run_deep.sh c10 use_tqdm 1 "resnet" "./pretrained_models" "--dirty_tau --prob_leaves_rate 1" ""
    

  • CIFAR100

    bash ./scripts/run_deep.sh c100 use_tqdm 1 "resnet" "./pretrained_models" "--dirty_tau --use_mixup_data --use_balanced_train_data --tau_regularization 1.05 --scale_nodes_loss 1e-4 --prob_leaves_rate 1" ""
    

  • TinyImagenet200

    bash ./scripts/run_deep.sh TinyImagenet200 use_tqdm 1 "resnet" "./pretrained_models" "--dirty_tau --use_mixup_data --use_balanced_train_data --tau_regularization 1.05 --scale_nodes_loss 1e-4 --prob_leaves_rate 1" ""
    

The folder ./pretrained_models should contain the following files:

• mnist_cnn.pt
• ckpt-CIFAR10-ResNet18.pth
• ckpt-CIFAR100-ResNet18.pth
• ckpt-TinyImagenet200-ResNet18.pth

---

Evaluation

Models can be evaluated using the following commands:

• shallow models:
  • MNIST

    path2model=?
    export PYTHONPATH=.
    python ./run/main_shallow.py --root_data ./dataset --data_type mnist --resume --path_resume ${path2model} --layers_nodes 784 --batch_size 64 --lr 1e-3 --use_tqdm --evaluate
    

  • LETTER

    path2model=?
    export PYTHONPATH=.
    python ./run/main_shallow.py --root_data ./dataset --data_type letter --resume --path_resume ${path2model} --layers_nodes 16 --batch_size 64 --lr 1e-3 --use_tqdm --evaluate
    

  • CONNECT4

    path2model=?
    export PYTHONPATH=.
    python ./run/main_shallow.py --root_data ./dataset --data_type connect4 --resume --path_resume ${path2model} --layers_nodes 126 --batch_size 64 --lr 1e-3 --use_tqdm --evaluate
    

• for deep model:
  • MNIST

    path2model=?
    export PYTHONPATH=.
    python ./run/main_shallow.py --root_data ./dataset --data_type mnist --resume --path_resume ${path2model} --layers_nodes 50 --batch_size 64 --lr 1e-3 --use_representation true --use_tqdm --evaluate
    

  • CIFAR10

    path2model=?
    export PYTHONPATH=.
    python ./run/main_deep.py --root_data /dataset/cifar-10 --dataset CIFAR10 --arch ResNet18 --resume --path_resume ${path2model} --batch_size 64 --lr 1e-3 --use_tqdm --evaluate
    

  • CIFAR100

    path2model=?
    export PYTHONPATH=.
    python ./run/main_deep.py --root_data /dataset/cifar-100 --dataset CIFAR100 --arch ResNet18 --resume --path_resume ${path2model} --batch_size 64 --lr 1e-3 --use_tqdm --evaluate
    

  • TinyImagenet200

    path2model=?
    export PYTHONPATH=.
    python ./run/main_deep.py --root_data ./dataset --dataset TinyImagenet200 --arch ResNet18 --resume --path_resume ${path2model} --batch_size 64 --lr 1e-3 --use_tqdm --evaluate
    

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}
}