Repository files navigation

Given entity and relation embeddings, InteractE generates multiple permutations of these embeddings and reshapes them using a "Chequered" reshaping function. Depthwise circular convolution is employed to convolve each of the reshaped permutations, which are then fed to a fully-connected layer to compute scores. Please refer to Section 6 of the paper for details.*

• Compatible with PyTorch 1.0 and Python 3.x.
• Dependencies can be installed usingrequirements.txt.

• We use FB15k-237, WN18RR and YAGO3-10 datasets for evaluation.
• FB15k-237, WN18RR are included in the repo. YAGO3-10 can be downloaded fromhere.

• Install all the requirements fromrequirements.txt.

• Executesh preprocess.sh for extracting the datasets and setting up the environment.

• To start trainingInteractE run:

  # FB15k-237python interacte.py --data FB15k-237 --gpu 0 --name fb15k_237_run# WN18RRpython interacte.py --data WN18RR --batch 256 --train_strategy one_to_n --feat_drop 0.2 --hid_drop 0.3 --perm 4 --ker_sz 11 --lr 0.001# YAGO03-10python interacte.py --data YAGO3-10 --train_strategy one_to_n  --feat_drop 0.2 --hid_drop 0.3 --ker_sz 7 --num_filt 64 --perm 2

  • data indicates the dataset used for training the model. Other options areWN18RR andYAGO3-10.
  • gpu is the GPU used for training the model.
  • name is the provided name of the run which can be later used for restoring the model.
  • Executepython interacte.py --help for listing all the available options.

• Executesh preprocess.sh for extracting the datasets and setting up the environment.

• Download the pre-trained model fromhere and place intorch_saved directory.

• To restore and evaluate run:

  python interacte.py --data FB15k-237 --gpu 0 --name fb15k_237_pretrained --restore --epoch 0

Please cite the following paper if you use this code in your work.

@inproceedings{interacte2020,title={InteractE: Improving Convolution-based Knowledge Graph Embeddings by Increasing Feature Interactions},author={Vashishth, Shikhar and Sanyal, Soumya and Nitin, Vikram and Agrawal, Nilesh and Talukdar, Partha},booktitle={Proceedings of the 34th AAAI Conference on Artificial Intelligence},pages={3009--3016},publisher={AAAI Press},url={https://aaai.org/ojs/index.php/AAAI/article/view/5694},year={2020}}

For any clarification, comments, or suggestions please create an issue or contactShikhar.