openbiolink-2021-embedding-baseline

Installation

pip install torch openbiolink dgl==0.4.3

Also from folder dglke-0.1.2 run

pip install -e .

Training

CLI-commands for each model can be found in run.sh. Each command should be run from the project root.

Hyperparameter

Hyperparameter used to achieve these results:

	learning rate	embedding size	regularization coefficient	gamma	iterations
RESCAL	0.05	300	3.00E-07		350000 on 2 GPUs
TransR	0.1	220	1.00E-08	12	550000 on 2 GPUs
ComplEx	0.1	380	2.00E-06		360000 on 8 GPUs
DistMult	0.1	380	4.00E-07		950000 on 2 GPUs
RotatE	0.05	128	1.00E-07	12	550000 on 2 GPUs
TransE	0.1	360	3.00E-09	8	550000 on 2 GPUs

Evaluation

Run from project root:

python3 save_test_submission.py
  --model_path MODEL_PATH
                        The path of the directory where models are saved.
  --batch_size_eval BATCH_SIZE_EVAL
                        The batch size used for evaluation.
  --gpu GPU             gpu id to be used, e.g. 0, -1 means only cpu is used

F.e. for evaluating TransE:

python3 save_test_submission.py --model_path ./ckpts/DistMult_OBL2021_0 --batch_size_eval 100 --gpu 0

Attribution

This code is based on dgl-ke