README.md

NePTuNe: Neural Powered Tucker Network.

This codebase contains PyTorch implementation of the paper:

NePTuNe: Neural Powered Tucker Network. Shashank Sonkar, Arzoo Katiyar, and Richard G. Baraniuk. [Paper]

${\color{green}{\textrm{This paper won the best short paper award at 10th IJCKG'21.}}}$

The codebase is inspired from TuckER's github repository. TuckER is the PyTorch implementation of the paper:

TuckER: Tensor Factorization for Knowledge Graph Completion. Ivana Balažević, Carl Allen, and Timothy M. Hospedales. Empirical Methods in Natural Language Processing (EMNLP), 2019. [Paper]

Link Prediction Results

Dataset	MRR	Hits@1	Hits@3	Hits@10
FB15k-237	0.366	0.272	0.404	0.547
WN18RR	0.491	0.455	0.507	0.557

Running the NePTuNE model

To run the model on FB15k-237 dataset, execute the following command:

 cd neptune-fb15k237/
 CUDA_VISIBLE_DEVICES=0 python -u main.py 
                --dataset FB15k-237 
                --num_iterations 1500 
                --batch_size 128 
                --lr 0.0005 --dr 1.0 
                --edim 200 --rdim 150 
                --input_dropout 0.3
                --hidden_dropout1 0.2
                --hidden_dropout2 0.4 
                --label_smoothing 0.1

To run the model on WN18RR dataset, execute the following command:

cd neptune-wn18rr/
CUDA_VISIBLE_DEVICES=0 python -u main.py
                --dataset WN18RR 
                --num_iterations 2500
                --batch_size 128
                --lr 0.003 --dr 1.0
                --edim 500 --rdim 30
                --input_dropout 0.1
                --hidden_dropout1 0.2
                --hidden_dropout2 0.7
                --label_smoothing 0.1

Note that WN18RR uses inverse relations, while FB15k-237 does not.

Requirements

The codebase is implemented in Python 3.6.6. Required packages are:

numpy      1.15.1
pytorch    1.0.1