pytorch-metric-learning 0.9.90

pytorch 1.6.0+

dgl 0.6.0

a) For wordnet:

Step 1: Graph contrastive learning to learn both entity embedding and relation embedding

---- Key parameter setting (need to tune) --> refer to configs/train.wn18rr.json python3 ckge_train.py --config_file configs/train.wn18rr.json --run_type graph_contrastive_train

Step 2: Inference the entity embedding and relation embedding and save as numpy data

---- Note: need the same config setting, however, for sub-graph sample (all neighorhood information is used) ---- The saved entity embedding and relation embedding are 'entity.npy' and 'relation.npy' python3 ckge_train.py --config_file configs/train.wn18rr.json --run_type graph_contrastive_train

Step 3: initialize the KGE model by the inferred entity and relation embeddings for KGE training

---- The parameter in 'init_checkpoint' stores the folder name of 'entity.npy' and 'relation.npy' python3 ckge_train.py --config_file configs/train.wn18rr.json --run_type KGE_train

b) For FreedBase: python3 ckge_train.py --config_file configs/train.fb15k237.json --run_type graph_contrastive_train python3 ckge_train.py --config_file configs/train.fb15k237.json --run_type graph_contrastive_train python3 ckge_train.py --config_file configs/train.fb15k237.json --run_type KGE_train