BernNet: Learning Arbitrary Graph Spectral Filters via Bernstein Approximation

This repository contains a PyTorch implementation of "BernNet: Learning Arbitrary Graph Spectral Filters via Bernstein Approximation".(https://arxiv.org/abs/2106.10994)

Environment Settings

• pytorch 1.8.1
• numpy 1.18.1
• torch-geometric 1.6.3
• tqdm 4.59.0
• scipy 1.6.2
• seaborn 0.11.1
• scikit-learn 0.24.1

Learning filters from the signal(./LearningFilters)

We conduct an empirical analysis on 50 real images with the resolution of 100×100 from the Image Processing Toolbox in Matlab.

Datasets

We provide the processed dataset and you can run the code directly. We also provide the original images in the folder './LearningFilters/image‘ and the Matlab code './LearningFilters/preprocessing.m' for preprocessing.

Running the code

Input Parameters

• filter_type: the type of the filter applied to the spectral domain
• net: the GNN models, default='BernNet'

You can run the following script in the folder './LearningFilters' directly

sh bernnet.sh

or run the following Command

• The band-pass filter for BernNet

python training.py --filter_type band --net BernNet

• The high-pass filter for BernNet

python training.py --filter_type high --net BernNet

Node classification on real-world datasets (./NodeClassification)

We evaluate the performance of BernNet against the competitors on 10 real-world datasets.

Datasets

We provide the datasets in the folder './NodeClassification/data' and you can run the code directly, or you can choose not to download the datasets('./NodeClassification/data') here. The code will automatically build the datasets through the data loader of Pytorch Geometric.

Running the code

You can run the following script in the folder './NodeClassification' directly and this script describes the hyperparameters settings of BernNet on each dataset.

sh bernnet.sh

or run the following Command

• Pubmed

python training.py  --dataset Pubmed --Bern_lr 0.01 --dprate 0.0 --weight_decay 0.0  --net BernNet

• Texas

python training.py --dataset Texas --lr 0.05 --Bern_lr 0.002 --dprate 0.5 --net BernNet

Citation

@inproceedings{he2021bernnet,
  title={BernNet: Learning Arbitrary Graph Spectral Filters via Bernstein Approximation},
  author={He, Mingguo and Wei, Zhewei and Huang, Zengfeng and Xu, Hongteng},
  booktitle={NeurIPS},
  year={2021}
}

Contact

If you have any questions, please feel free to contact me with [email protected]