Comparison between SLR method and sparse training method on GNN dataset.
we use the SLR method for train and prune.
Folder SLR_Link_Pred
is for SLR training for link prediction, which follows dense training -> reweight training -> sparse training procedure.
Folder SLR_Node_Class
is for SLR training for node classification, which follows dense training -> reweight training -> sparse training procedure.
Folder SLR_GCN_Node_Class
is for SLR training for node classification, which follows dense training -> reweight training -> sparse training procedure. The training and prune experiment is done on 3 dataset: Cora, Pubmed, and CiteSeer
we follow same experiment setup as RigL paper (rigging the lottery: making all tickets winners), using weight magnitute for drop and weight gradient for grow.
If you use our code in your design, please cite our ICCD'22 paper:
@inproceedings{PengSpsGNN,
title={Towards Sparsification of Graph Neural Networks},
author={Peng, Hongwu and Gurevin, Deniz and Huang, Shaoyi and Geng, Tong and Jiang, Weiwen and Khan, Omer and Ding, Caiwen},
booktitle={2022 IEEE 40th International Conference on Computer Design (ICCD)},
year={2022}
}