This repository evaluates multi-order interaction strengths of timm backbones for visual representation learning, which is based on Discovering and Explaining the Representation Bottleneck of DNNs (ICLR'2022) and the official implementation (code).

This repository works with PyTorch 1.8 or higher and timm. There are installation steps with the latest PyTorch:

conda create -n bottleneck python=3.8 pytorch=1.12 cudatoolkit=11.3 torchvision -c pytorch -y
conda activate bottleneck
pip install -r requirements.txt

Then, please download datasets and place them under './datasets'. CIFAR-10 will be automatically downloaded, while ImageNet should be downloaded and unziped manually.

We only support the evaluation of pre-trained models. Please download released pre-trained models from timm and place them in './timm_hub'. We provide examples on ImageNet in 'interaction_in1k.sh':

bash interaction.sh

You can uncomment the setting (the model name and ckeckpoints) you want to run on top of the script. The results will be saved in the results directory by default.

Please cite the relevant papers if you find this repository useful in your research.

@article{Deng2022Discovering,
  title={Discovering and Explaining the Representation Bottleneck of DNNs},
  author={Huiqi Deng and Qihan Ren and Xu Chen and Hao Zhang and Jie Ren and Quanshi Zhang},
  journal={ArXiv},
  year={2022},
  volume={abs/2111.06236}
}

@article{Wu2022Discovering,
  title={Discovering the Representation Bottleneck of Graph Neural Networks from Multi-order Interactions},
  author={Fang Wu and Siyuan Li and Lirong Wu and Stan Z. Li and Dragomir Radev and Q. Zhang},
  journal={ArXiv},
  year={2022},
  volume={abs/2205.07266}
}