Implementation of AutoGCN - Towards Generic Human Activity Recognition with Neural Architecture Search.
The code is based on Python >= 3.10 and PyTorch (2.0.1). Run the following command to install all the required packages from setup.py:
pip install .
The NAS procedure and the experiments are done on the NTU RGB+D 60 & 120 datasets, which can be downloaded here.
Furthermore, the Kinetics dataset is used, which can be downloaded here.
To preprocess, the user is referred to the instructions at:
./src/data/preprocess/README.md
To run a model search, check the config file and change the following parameters:
work_dir -> path where you want to store the output of the run
dataset -> Choose the dataset and change the folders to the preprocessed npy-files
The modes can either be activated or deactivated by setting the flags to True or False - refer to the different
config run files in the logs folder for further information.
Afterward, execute:
python main.py -config /path/to/config.yaml
The results reported in our study are stored in the ./logs folder.
| Config | Iterations | Top-1 X-View | Top-1 X-Sub |
|---|---|---|---|
| 1007 | 10 | 95.3 | 85.9 |
| 1003 | 20 | 95.1 | 88.3 |
| 1004 | 30 | 95.5 | 86.4 |
The policy values figures for the different experiments are also listed in the respective folders. There are also predefined configs stored in there, which can be used :).
If you have any questions, feel free to send an email to [email protected] or open an issue on GitHub.
Please cite our paper if you use this code in your research. :)
@ARTICLE{10472042,
author={Tempel, Felix and Ihlen, Espen Alexander F. and Strümke, Inga},
journal={IEEE Access},
title={AutoGCN - Towards Generic Human Activity Recognition with Neural Architecture Search},
year={2024},
doi={10.1109/ACCESS.2024.3377103}}