This codebase contains PyTorch implementation of the paper:
Hypernetwork Knowledge Graph Embeddings. Ivana Balažević, Carl Allen, and Timothy M. Hospedales. International Conference on Artificial Neural Networks, 2019. [Paper]
To run the model, execute the following command:
CUDA_VISIBLE_DEVICES=0 python hyper.py --algorithm HypER --dataset FB15k-237
Available algorithms are:
HypER
HypE
DistMult
ComplEx
ConvE
Available datasets are:
FB15k-237
WN18RR
FB15k
WN18
To reproduce the results from the paper, use the following combinations of hyperparameters with batch_size=128, ent_vec_dim=200 and rel_vec_dim=200:
| dataset | lr | dr | input_dropout | feature_map_dropout | hidden_dropout | label_smoothing |
|---|---|---|---|---|---|---|
| FB15k | 0.005 | 0.995 | 0.2 | 0.2 | 0.3 | 0. |
| WN18 | 0.001 | 1.0 | 0.2 | 0.2 | 0.3 | 0.1 |
| FB15k-237 | 0.0001 | 0.995 | 0.3 | 0.2 | 0.3 | 0.1 |
| WN18RR | 0.005 | 1.0 | 0.2 | 0.2 | 0.3 | 0.1 |
The codebase is implemented in Python 3.6.6. Required packages are:
numpy 1.14.5
pytorch 0.4.0
If you found this codebase useful, please cite:
@inproceedings{balazevic2019hypernetwork,
title={Hypernetwork Knowledge Graph Embeddings},
author={Bala\v{z}evi\'c, Ivana and Allen, Carl and Hospedales, Timothy M},
booktitle={International Conference on Artificial Neural Networks},
year={2019}
}