Skip to content

Latest commit

 

History

History
75 lines (56 loc) · 3.67 KB

domain-constraints.md

File metadata and controls

75 lines (56 loc) · 3.67 KB
Raw

Reproducing Integration of Domain Constraints Results

First, set up the project and download and preprocess the ogbl-biokg data set as showed in README.md. The section Training the Models below lists the commands that trains ComplEx, ComplEx2 and ComplEx2 with domain constraints that are used in this experiments

Then, running the script shell/eval_consistency.sh will print out the semantic consistency scores at k (Sem@k) (see our paper). Note that you need to specify the directory containing the models and the device, as showed in the following example.

MODELS_PATH=path/to/saved/models DEVICE=cuda:0 bash shell/eval_consistency.sh

Furthermore, to plot the line graph showing how the test Mean Reciprocal Rank (MRR) changes by increasing the embedding size, run the script shell/eval_rank, as showed in the following example.

MODELS_PATH=path/to/saved/models DEVICE=cuda:0 bash shell/eval_rank.sh

Training the Models

ComplEx

python -m kbc.experiment --experiment_id PLL --dataset ogbl-biokg --model ComplEx --rank 10 --optimizer Adam --batch_size 5000 --learning_rate 0.001 --score_lhs True --score_rel True --score_rhs True
python -m kbc.experiment --experiment_id PLL --dataset ogbl-biokg --model ComplEx --rank 50 --optimizer Adam --batch_size 5000 --learning_rate 0.001 --score_lhs True --score_rel True --score_rhs True
python -m kbc.experiment --experiment_id PLL --dataset ogbl-biokg --model ComplEx --rank 200 --optimizer Adam --batch_size 5000 --learning_rate 0.001 --score_lhs True --score_rel True --score_rhs True
python -m kbc.experiment --experiment_id PLL --dataset ogbl-biokg --model ComplEx --rank 1000 --optimizer Adam --batch_size 5000 --learning_rate 0.001 --score_lhs True --score_rel True --score_rhs True

ComplEx2

python -m kbc.experiment --experiment_id PLL --dataset ogbl-biokg --model SquaredComplEx --rank 10 --optimizer Adam --batch_size 1000 --learning_rate 0.001 --score_lhs True --score_rel True --score_rhs True
python -m kbc.experiment --experiment_id PLL --dataset ogbl-biokg --model SquaredComplEx --rank 50 --optimizer Adam --batch_size 2000 --learning_rate 0.001 --score_lhs True --score_rel True --score_rhs True
python -m kbc.experiment --experiment_id PLL --dataset ogbl-biokg --model SquaredComplEx --rank 200 --optimizer Adam --batch_size 5000 --learning_rate 0.001 --score_lhs True --score_rel True --score_rhs True
python -m kbc.experiment --experiment_id PLL --dataset ogbl-biokg --model SquaredComplEx --rank 1000 --optimizer Adam --batch_size 5000 --learning_rate 0.001 --score_lhs True --score_rel True --score_rhs True

ComplEx2 with Domain Constraints

python -m kbc.experiment --experiment_id PLL --dataset ogbl-biokg --model TypedSquaredComplEx --rank 10 --optimizer Adam --batch_size 5000 --learning_rate 0.001 --score_lhs True --score_rel True --score_rhs True
python -m kbc.experiment --experiment_id PLL --dataset ogbl-biokg --model TypedSquaredComplEx --rank 50 --optimizer Adam --batch_size 2000 --learning_rate 0.001 --score_lhs True --score_rel True --score_rhs True
python -m kbc.experiment --experiment_id PLL --dataset ogbl-biokg --model TypedSquaredComplEx --rank 200 --optimizer Adam --batch_size 2000 --learning_rate 0.001 --score_lhs True --score_rel True --score_rhs True
python -m kbc.experiment --experiment_id PLL --dataset ogbl-biokg --model TypedSquaredComplEx --rank 1000 --optimizer Adam --batch_size 2000 --learning_rate 0.001 --score_lhs True --score_rel True --score_rhs True