GCN-NAS

PyTorch Source code for "Learning Graph Convolutional Network for Skeleton-based Human Action Recognition by Neural Searching", AAAI2020

Requirements

• python packages
  • pytorch = 0.4.1
  • torchvision>=0.2.1

Data Preparation

• Download the raw data from NTU-RGB+D and Skeleton-Kinetics. And pre-processes the data.

• Preprocess the data with

  python data_gen/ntu_gendata.py

  python data_gen/kinetics-gendata.py.

• Generate the bone data with:

  python data_gen/gen_bone_data.py

Model Training

• Here, you can train the model searched by our method.

• Configure the config file for different settings. For example, training model under the corss-view protocal:

  python main.py --config ./config/nturgbd-cross-view/train_joint.yaml

Model Evaluation

• Change the config file for corresponding dateset with its protocal.

  python main.py --config ./config/nturgbd-cross-view/test_joint.yaml

Model searching

• Devide the training data into trainging set and searching parts, as it is for a general NAS.
• Run the gcn_search.py with corresponding configuration.
• Here, we search on the NTU RGB+D dataset under the cross-view evaluation.

python gcn_search.py --config ./config/nturgbd-cross-view/search_joint.yaml

Acknowledgement

• Our work is inpired by Two-Stream AGCN and CEM-RL, we thank the authors for sharing their codes.

BibTeX

@article{peng2020learning,
  title={Learning Graph Convolutional Network for Skeleton-based Human Action Recognition by Neural Searching},
  author={Peng, Wei and Hong, Xiaopeng and Chen, Haoyu and Zhao, Guoying},
  journal={The Thirty-Fourth AAAI Conference on Artificial Intelligence, AAAI},
  year={2020}
}

License

All materials in this repository are released under the Apache License 2.0.