Under each folder:
-
gnn.py and conv.py are basic gnn(GCN,SAGE,PNA,Deepergcn...) implementation codes
-
Splearner.py implements a node-granular graph pruner
-
MoE.py implements a model that mixes pruners with different sparsities
-
MoG.py bridges the connection between MoE and GNN
For the ogbn_proteins dataset, please cd ogbn_proteins first.
python main.py --use_gpu --conv_encode_edge --use_one_hot_encoding --num_layers 28 --block res+ --gcn_aggr max --k_list 1 --expert_select 1
python main.py --use_gpu --conv_encode_edge --use_one_hot_encoding --num_layers 28 --block res+ --gcn_aggr max --k_list 0.9 0.7 0.5 --expert_select 2
python main.py --use_gpu --conv_encode_edge --use_one_hot_encoding --num_layers 28 --block res+ --gcn_aggr max --k_list 0.7 0.5 0.3 --expert_select 2
python main.py --use_gpu --conv_encode_edge --use_one_hot_encoding --num_layers 28 --block res+ --gcn_aggr max --k_list 0.4 0.3 0.2 --expert_select 2
For the MNIST dataset, please cd MNIST first.
python main.py --k_list 1 --expert_select 1
python main.py --k_list 0.8 0.5 0.4 --expert_select 2
python main.py --k_list 0.6 0.3 0.2 --expert_select 2
python main.py --k_list 0.35 0.1 0.1 --expert_select 2