Codes in this repository consist of GIG-GatedGCN and GIG-CTR-GCN. The original GatedGCN can be found from here and orginal CTR-GCN can be found from here. If you find our research is supportive, please cite:

@misc{wang2024graphgraphneuralnetwork,
      title={Graph in Graph Neural Network}, 
      author={Jiongshu Wang and Jing Yang and Jiankang Deng and Hatice Gunes and Siyang Song},
      year={2024},
      eprint={2407.00696},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2407.00696}, 
}

Requirements

• Python 3.8

• PyTorch 1.9.0

• CUDA 11.1

• DGL 0.9.0

• Check the required python packages in ga_dependency.txt right here for graph analysis tasks.

pip install -r ga_dependency.txt

• Check the required python packages in sk_dependency.txt right here for action recognition tasks.

pip install -r sk_dependency.txt

Training

For general graph analysis tasks (clicking the link will directly take you to the target folder):

• MNIST

1. Execute the code blocks in 'prepare_superpixels_MNIST.ipynb' to download dataset and complete preprocessing
2. Execute the 'MNIST.ipynb' to start training

• CIFAR10

1. Execute the code blocks in 'prepare_superpixels_CIFAR.ipynb' to download dataset and complete preprocessing
2. Execute the 'CIFAR10.ipynb' to start training

• CLUSTER

1. Execute the code blocks in 'prepare_SBM_CLUSTER.ipynb' to download dataset and complete preprocessing
2. Execute the 'CLUSTER.ipynb' to start training

• PATTERN

1. Execute the code blocks in 'prepare_SBM_PATTERN.ipynb' to download dataset and complete preprocessing
2. Execute the 'PATTERN.ipynb' to start training

• ZINC

1. Execute the code blocks in 'prepare_molecules.ipynb' to download dataset and complete preprocessing
2. Execute the 'ZINC.ipynb' to start training

• ZINC-full

1. Execute the code blocks in 'prepare_molecules_ZINC_full.ipynb' to download dataset and complete preprocessing
2. Execute the 'ZINC-full.ipynb' to start training

• AQSOL

1. Execute the code blocks in 'prepare_molecules.ipynb' to download dataset and complete preprocessing
2. Execute the 'AQSOL.ipynb' to start training

• TSP

1. Execute the code blocks in 'prepare_TSP.ipynb' to download dataset and complete preprocessing
2. Execute the 'TSP.ipynb' to start training

• TU

1. Dataset will be automatically downloaded and processed. Nonthing needs to be done right here.
2. Execute the 'TU-GIG-Codes.ipynb' to start training

• OGBG-PPA

1. Dataset will be automatically downloaded and processed. Nonthing needs to be done right here.
2. Execute the 'OGBG-PPA.ipynb' to start training

For human-skeleton action recognition tasks, the raw data should be downloaded from CTR-GCN (https://github.com/Uason-Chen/CTR-GCN). Then data preprocessing should be run then as following:

Directory Structure

• Put downloaded data into the following directory structure:

- data/
  - NW-UCLA/
    - all_sqe
      ... # raw data of NW-UCLA
  - ntu/
  - ntu120/
  - nturgbd_raw/
    - nturgb+d_skeletons/     # from `nturgbd_skeletons_s001_to_s017.zip`
      ...
    - nturgb+d_skeletons120/  # from `nturgbd_skeletons_s018_to_s032.zip`
      ...

Generating Data

• Generate NTU RGB+D 60 or NTU RGB+D 120 dataset:

 cd ./data/ntu # or cd ./data/ntu120
 # Get skeleton of each performer
 python get_raw_skes_data.py
 # Remove the bad skeleton 
 python get_raw_denoised_data.py
 # Transform the skeleton to the center of the first frame
 python seq_transformation.py

Executing training

• NTURGB+D X-Sub

  • Using command python main.py --config config/nturgbd-cross-subject/default.yaml --work-dir work_dir/ntu/csub/ctrgcn --device 0

• NTURGB+D X-View

  • Using command python main.py --config config/nturgbd-cross-view/default.yaml --work-dir work_dir/ntu/cview/ctrgcn --device 0