To run the pretraining, please using the following command:
python main.py --run_type pretrain \
--pretrain_single_graph_data ${SINGLE_GRAPH_DATASETS_LIST} \
--pretrain_multi_graph_data ${MULTI_GRAPH_DATASETS_LIST}You need to replace ${SINGLE_GRAPH_DATASETS_LIST}
and ${MULTI_GRAPH_DATASETS_LIST}
with lists of graph dataset names.
For instance, [ogbn-arxiv, Reddit, FB15k_237] and [PROTEINS, HIV].
To run the adaptation for few-shot transferring, please using the following command:
python main.py --run_type adapt \
--pretrain_single_graph_data ${SINGLE_GRAPH_DATASETS_LIST} \
--pretrain_multi_graph_data ${MULTI_GRAPH_DATASETS_LIST} \
--pretrained_checkpoint checkpoints/pretrain/${SINGLE_GRAPH_DATASETS_LIST}_${MULTI_GRAPH_DATASETS_LIST}/${MODEL_NAME}.pth \
--data_name ${DATA_NAME} \
--task_type ${TASK_TYPE} \
--metric ${METRIC}$ \
--k_shot $K_SHOT$You need to replace ${MODEL_NAME}
with the file name of checkpoint that needed to use;
${DATA_NAME} with the dataset name for transfer, e.g., Computers;
${TASK_TYPE} with node_cls, graph_cls, link_cls;
${METRIC}$ with acc or auc;
$K_SHOT$ with 1, 5 or other number you want to transfer.
Figure 1. An Illustration of GRAPHGLUE Framework
Figure 2. Visualization of the pre-trained manifold from 6 datasets.