Awesome

<h1 align="center"> InteractE </h1> <h4 align="center">Improving Convolution-based Knowledge Graph Embeddings by Increasing Feature Interactions</h4> <p align="center"> <a href="https://aaai.org/Conferences/AAAI-20/"><img src="http://img.shields.io/badge/AAAI-2020-4b44ce.svg"></a> <a href="https://arxiv.org/abs/1911.00219"><img src="http://img.shields.io/badge/Paper-PDF-red.svg"></a> <a href="https://shikhar-vashishth.github.io/assets/pdf/interacte_supp.pdf"><img src="http://img.shields.io/badge/Supplementary-PDF-green.svg"></a> <a href="https://github.com/malllabiisc/InteractE/blob/master/LICENSE"> <img src="https://img.shields.io/badge/License-Apache%202.0-blue.svg"> </a> </p> <h2 align="center"> Overview of InteractE <img align="center" src="./overview.png" alt="..."> </h2>

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.*

Dependencies

Compatible with PyTorch 1.0 and Python 3.x.
Dependencies can be installed using requirements.txt.

Dataset:

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

Training model from scratch:

Install all the requirements from requirements.txt.
Execute sh preprocess.sh for extracting the datasets and setting up the environment.

To start training InteractE run:

# FB15k-237
python interacte.py --data FB15k-237 --gpu 0 --name fb15k_237_run

# WN18RR
python 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

# YAGO03-10
python 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 are WN18RR and YAGO3-10.
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.
Execute python interacte.py --help for listing all the available options.

Evaluating Pre-trained model:

Execute sh preprocess.sh for extracting the datasets and setting up the environment.
Download the pre-trained model from here and place in torch_saved directory.

To restore and evaluate run:

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

Citation:

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 contact Shikhar.