Skip to content

ishaan-mehta/KnowledgeGraphEmbedding

 
 

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

18 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

RotatE: Knowledge Graph Embedding by Relational Rotation in Complex Space

Introduction

This is the PyTorch implementation of the RotatE model for knowledge graph embedding (KGE). We provide a toolkit that gives state-of-the-art performance of several popular KGE models. The toolkit is quite efficient, which is able to train a large KGE model within a few hours on a single GPU.

A faster multi-GPU implementation of RotatE and other KGE models is available in GraphVite.

Implemented features

Models:

  • RotatE
  • pRotatE
  • TransE
  • ComplEx
  • DistMult

Evaluation Metrics:

  • MRR, MR, HITS@1, HITS@3, HITS@10 (filtered)
  • AUC-PR (for Countries data sets)

Loss Function:

  • Uniform Negative Sampling
  • Self-Adversarial Negative Sampling

Usage

Knowledge Graph Data:

  • entities.dict: a dictionary map entities to unique ids
  • relations.dict: a dictionary map relations to unique ids
  • train.txt: the KGE model is trained to fit this data set
  • valid.txt: create a blank file if no validation data is available
  • test.txt: the KGE model is evaluated on this data set

Train

For example, this command train a RotatE model on FB15k dataset with GPU 0.

CUDA_VISIBLE_DEVICES=0 python -u codes/run.py --do_train \
 --cuda \
 --do_valid \
 --do_test \
 --data_path data/FB15k \
 --model RotatE \
 -n 256 -b 1024 -d 1000 \
 -g 24.0 -a 1.0 -adv \
 -lr 0.0001 --max_steps 150000 \
 -save models/RotatE_FB15k_0 --test_batch_size 16 -de

Check argparse configuration at codes/run.py for more arguments and more details.

Test

CUDA_VISIBLE_DEVICES=$GPU_DEVICE python -u $CODE_PATH/run.py --do_test --cuda -init $SAVE

Reproducing the best results

To reprocude the results in the ICLR 2019 paper RotatE: Knowledge Graph Embedding by Relational Rotation in Complex Space, you can run the bash commands in best_config.sh to get the best performance of RotatE, TransE, and ComplEx on five widely used datasets (FB15k, FB15k-237, wn18, wn18rr, Countries).

The run.sh script provides an easy way to search hyper-parameters:

bash run.sh train RotatE FB15k 0 0 1024 256 1000 24.0 1.0 0.0001 200000 16 -de

Speed

The KGE models usually take about half an hour to run 10000 steps on a single GeForce GTX 1080 Ti GPU with default configuration. And these models need different max_steps to converge on different data sets:

Dataset FB15k FB15k-237 wn18 wn18rr Countries S*
MAX_STEPS 150000 100000 80000 80000 40000
TIME 9 h 6 h 4 h 4 h 2 h

Results of the RotatE model

Dataset FB15k FB15k-237 wn18 wn18rr
MRR .797 ± .001 .337 ± .001 .949 ± .000 .477 ± .001
MR 40 177 309 3340
HITS@1 .746 .241 .944 .428
HITS@3 .830 .375 .952 .492
HITS@10 .884 .533 .959 .571

Using the library

The python libarary is organized around 3 objects:

  • TrainDataset (dataloader.py): prepare data stream for training
  • TestDataSet (dataloader.py): prepare data stream for evluation
  • KGEModel (model.py): calculate triple score and provide train/test API

The run.py file contains the main function, which parses arguments, reads data, initilize the model and provides the training loop.

Add your own model to model.py like:

def TransE(self, head, relation, tail, mode):
    if mode == 'head-batch':
        score = head + (relation - tail)
    else:
        score = (head + relation) - tail

    score = self.gamma.item() - torch.norm(score, p=1, dim=2)
    return score

Citation

If you use the codes, please cite the following paper:

@inproceedings{
 sun2018rotate,
 title={RotatE: Knowledge Graph Embedding by Relational Rotation in Complex Space},
 author={Zhiqing Sun and Zhi-Hong Deng and Jian-Yun Nie and Jian Tang},
 booktitle={International Conference on Learning Representations},
 year={2019},
 url={https://openreview.net/forum?id=HkgEQnRqYQ},
}

About

No description, website, or topics provided.

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Python 88.7%
  • Shell 11.3%