PathE: Leveraging Entity-Agnostic Paths for Parameter-Efficient Knowledge Graph Embeddings

Model overview

Instructions:

1. Install required dependencies by running pip install -r requirements.txt in the terminal

Training

1. FB15k-237
   • python runner.py train pathe --log_dir [YOUR LOG DIR] --expname [YUR EXP NAME] --wandb_project fb15k237 --train_paths [PATH TO TRAIN DATA] --valid_paths [PATH TO VALID DATA] --test_paths [PATH TO TEST DATA] --max_ppt 8 --batch_size 4096 --nhead 2 --num_encoder_layers 1 --dim_feedforward 256 --embedding_dim 64 --num_negatives 99 --monitor valid_link_mrr --patience 5 --dropout 0.1 --lrate 1e-3 --device cuda --num_devices 1 --val_num_negatives 99 --num_workers 8 --ent_aggregation transformer --lp_loss_fn ce --accumulate_gradient 8 --label_smoothing 0.01 --node_projector dummy --path_setup 20_10 --num_agg_layers 1 --seed 7

Full Evaluation

1. FB15k-237
   • python runner.py full_eval pathe --log_dir [YOUR LOG DIR] --expname [YUR EXP NAME] --wandb_project fb15k237 --train_paths [PATH TO TRAIN DATA] --valid_paths [PATH TO VALID DATA] --test_paths [PATH TO TEST DATA] --max_ppt 8 --batch_size 4096 --nhead 2 --num_encoder_layers 1 --dim_feedforward 256 --embedding_dim 64 --num_negatives 99 --monitor valid_link_mrr --patience 5 --dropout 0.1 --lrate 1e-3 --device cuda --num_devices 1 --val_num_negatives 99 --num_workers 8 --ent_aggregation transformer --lp_loss_fn ce --accumulate_gradient 8 --label_smoothing 0.01 --node_projector dummy --path_setup 20_10 --num_agg_layers 1 --seed 7 --checkpoint [PATH TO MODEL CHECKPOINT]