关于GCU、沐曦GPGPU、MLU、0卡V100资源4月7日恢复上架的公告>>> 关于共建具身智能开源数据集的倡议>>> 关于云脑任务中统一路径访问方式的公告>>> 关于将启智集群GPU资源迁移至智算集群的公告>>>

History

honey0219 594ac0b47e [Model] LTD (#186 ) * 同步原始仓库 * 整理代码结构 * 修改了MLP的参数更新方式 * 增加教师模型训练过程，增加train_conf.yaml * 修改train_conf.yaml，温度计算MLP，统一了教师模型学生模型训练集测试集 * 整理代码 * 整理代码 * 添加readme.md文件 * 修改readme.md文件 * 修改readme.md文件 * fix docs * update doc --------- Co-authored-by: dddg617 <996179900@qq.com> Co-authored-by: dddg617 <75086617+dddg617@users.noreply.github.com>		3 months ago
..
distill.py	[Model] LTD (#186)	3 months ago

ltd_trainer.py	[Model] LTD (#186)	3 months ago

readme.md	[Model] LTD (#186)	3 months ago

teacher_trainer.py	[Model] LTD (#186)	3 months ago

train_conf.yaml	[Model] LTD (#186)	3 months ago

readme.md

Learning to Distill Graph Neural Networks (LTD)
Dataset Statics
- Raw Data Statics
- Largest Connected Component Statics
Results

Learning to Distill Graph Neural Networks (LTD)

Paper link: [https://doi.org/10.1145/3539597.3570480]
Author's code repo: [https://github.com/BUPT-GAMMA/LTD].

Dataset Statics

Raw Data Statics

Dataset	# Nodes	# Edges	# Classes
Cora	2,708	10,556	7
Citeseer	3,327	9,228	6
Pubmed	19,717	88,651	3

Refer to Planetoid.

Largest Connected Component Statics

Dataset	# Nodes	# Edges	# Classes
Cora	2,485	10,138	7
Citeseer	2,120	7,358	6
Pubmed	19,717	88,651	3

Results

# available dataset: "cora", "citeseer", "pubmed" 
TL_BACKEND="torch" python ltd_trainer.py --dataset cora --model GCN
TL_BACKEND="torch" python ltd_trainer.py --dataset cora --model GAT
TL_BACKEND="torch" python ltd_trainer.py --dataset citeseer --model GCN 
TL_BACKEND="torch" python ltd_trainer.py --dataset citeseer --model GAT
TL_BACKEND="torch" python ltd_trainer.py --dataset pubmed --model GCN 
TL_BACKEND="torch" python ltd_trainer.py --dataset pubmed --model GAT

Dataset	GNN	paper(Largest Connected Component)	Our	Our(Largest Connected Component)
cora	GCN	87.21	83.10±0.41	84.71±0.44
cora	GAT	86.56	82.40±1.38	84.35±0.42
citeseer	GCN	78.51	72.39±1.09	74.69±1.12
citeseer	GAT	77.35	72.85±0.71	76.27±0.62
pubmed	GCN	81.91	79.54±0.70	79.54±0.70
pubmed	GAT	82.74	78.77±0.90	78.77±0.90