Our used CUDA version is 12.2. The Python packages and the corresponding versions required for HyGNN are as follows:
torch==1.8.0
torchvision==0.9.0
torchaudio==0.8.0
Enter the folder hybrid_kernel and run sudo python setup.py install to compile and install the SpMM kernels of HC-SpMM.
Go back to the folder HC-SpMM and run python HC-SpMM_main.py --dataset example --model gcn to start the GCN training on the dataset example. There are 8 parameters that can be customized. The detailed information is listed below:
--dataset: the training dataset which uses the COO format to represent the graph
--dim: the embedding dimension
--num_layers: the number of layers of GNN
--hidden: the dimension of hidden layers
--classes: the number of output classes
--epochs: the number of epochs
--model: the GNN model to train (GCN and GIN are available in the current implementation)
--single_kernel: only call the SpMM kernel to achieve the multiplication of the adjacency matrix and the embedding matrix