Skip to content

Hop-Wise Graph Attention for Scalable and Generalizable Learning on Circuits

License

Notifications You must be signed in to change notification settings

cornell-zhang/HOGA

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

17 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

HOGA

arXiv

HOGA is an attention model for scalable and generalizable learning on circuits. By leveraging a novel gated attention module on hop-wise features, HOGA not only outperforms prior graph learning models on challenging circuit problems, but is also friendly to distributed training by mitigating communication overhead caused by graph dependencies. This renders HOGA applicable to industrial-scale circuit applications. More details are available in our paper.

HOGA.png
Figure1: An overview of HOGA and gated attention module.

Requirements

  • python 3.9
  • pytorch 1.12 (CUDA 11.3)
  • torch_geometric 2.1

Datasets

Pre-processed CSA and Booth Multipliers (for Gamora experiments)

Check at: https://huggingface.co/datasets/yucx0626/Gamora-CSA-Multiplier/tree/main

Pre-processed OpenABC-D benchmark (for OpenABC-D experiments)

Check at: https://zenodo.org/records/6399454#.YkTglzwpA5k

Note

The implementation of hop-wise feature generation is available here. The model (i.e., hop-wise gated attention) implementation is available here. You can adjust them for your own tasks.

Citation

If you use HOGA in your research, please cite our work published in DAC'24.

@inproceedings{deng2024hoga,
  title={Less is More: Hop-Wise Graph Attention for Scalable and Generalizable Learning on Circuits},
  author={Chenhui Deng and Zichao Yue and Cunxi Yu and Gokce Sarar and Ryan Carey and Rajeev Jain and Zhiru Zhang},
  booktitle={DAC},
  year={2024},
}