Learning Graph Normalization for Graph Neural Networks ArXiv

note1: Our implementation is based on graphdeeplearning/benchmarking-gnns, thanks for their great work!

note2: For some business reasons, the released code may be a little different from our original code. If you find any problem, feel free to contact us.

Sep 28, 2020

• add Softmax United Norm

Sep 24, 2020

• First release of the project.

Follow these instructions to install the benchmark and setup the environment.

Proceed as follows to download the benchmark datasets.

Use this page to run the codes and reproduce the published results.

Node-wise Normalization: equivalent to Layer Normalization

Adjance-wise Normalization: adjance_norm.py

Graph-wise Normalization: graph_norm.py

Batch-wise normalization: equivalent to Batch Normalization

United Normalization: united_norm.py

Modify the value of norm in config.json or add one kind of norm after --norm.

Run the following command：

python main_SBMs_node_classification.py --dataset CLUSTER --gpu_id 3 --seed 41 --config 
'configs/SBMs_node_clustering_GatedGCN_CLUSTER_100k.json' --norm GraphNorm

The choices of norm consist of "NodeNorm", "AdjanceNorm", "GraphNorm", "BatchNorm", "UnitedNorm","UnitedNormSoftmax"

For a receipt, each text bbox can be viewed as a node of a graph. Its positions, the attributes of bounding box, and the corresponding text are used as the node feature. Our goal is to label each node (text bounding box) with five different classes, including Company, Date, Address, Total and Other. Sample images are shown below:

SROIE Dataset Download: Dropbox, BaiduPan: u4tm

cd sroie
python train.py

@misc{chen2020learning,
    title={Learning Graph Normalization for Graph Neural Networks},
    author={Yihao Chen and Xin Tang and Xianbiao Qi and Chun-Guang Li and Rong Xiao},
    year={2020},
    eprint={2009.11746},
    archivePrefix={arXiv},
    primaryClass={cs.LG}
}

This project is licensed under the MIT License. See LICENSE for more details.