Knowledge graph embedding research has mainly focused on learning continuous representations of knowledge graphs towards the link prediction problem. Recently developed frameworks can be effectively applied in a wide range of research-related applications. Yet, using these frameworks in real-world applications becomes more challenging as the size of the knowledge graph grows.
We developed the DICE Embeddings framework to compute embeddings for large-scale knowledge graphs in a hardware-agnostic manner. To achieve this goal, we rely on
- Pandas & DASK to use parallelism while preprocessing a large input knowledge graph,
- PyTorch & PytorchLightning to learn knowledge graph embeddings via multi-CPUs, GPUs, TPUs or computing cluster, and
- Gradio to ease the deployment of pre-trained models.
Why Pandas & DASK ? Pandas allows us to read, preprocess (e.g., removing literals) and index an input knowledge graph in parallel. Through Parquet within pandas, billions of triples can be read in parallel fashion. Importantly, Dask allows us to perform all necessary computations on a wide array of hardware configurations ranging from a single CPU to a cluster of computers.
Why PyTorch & PytorchLightning ? PyTorch is one of the most popular machine learning frameworks available at the time of writing. PytorchLightning facilitates scaling the training procedure of PyTorch without boilerplate. In our framework, we combine PyTorch & PytorchLightning. By this, we are able to train large knowledge graph embedding models with billions of parameters. PytorchLightning allows us to use state-of-the-art model parallelism techniques (e.g. Fully Sharded Training, FairScale, or DeepSpeed) without extra effort. With our framework, practitioners can directly use PytorchLightning for model parallelism to train gigantic embedding models.
Why Hugging-face Gradio? Deploy a pre-trained embedding model without writing a single line of code.
Clone the repository:
git clone https://github.com/dice-group/dice-embeddings.git
To install dependencies:
# python=3.10 with torch cuda nncl https://discuss.pytorch.org/t/issues-on-using-nn-dataparallel-with-python-3-10-and-pytorch-1-11/146745/13
conda create -n dice python=3.9.12
conda activate dice
pip3 install pandas==1.5.0
pip3 install swifter==1.1.2 # we can remove it later
pip3 install torch --extra-index-url https://download.pytorch.org/whl/cu113
pip3 install pytorch-lightning==1.6.4
pip3 install "dask[complete]"==2022.6.0
pip3 install scikit-learn==1.1.1
pip3 install pytest==6.2.5
pip3 install gradio==3.0.17
pip3 install pyarrow==8.0.0
To test the Installation
wget https://hobbitdata.informatik.uni-leipzig.de/KG/KGs.zip
unzip KGs.zip
pytest -p no:warnings -x # it takes circa 15 minutes
pytest -p no:warnings --lf # run only the last failed test
pytest -p no:warnings --ff # to run the failures first and then the rest of the tests.
Please contact: caglar.demir@upb.de or caglardemir8@gmail.com , if you lack hardware resources to obtain embeddings of a specific knowledge Graph.
- DBpedia version: 06-2022 Embeddings:
- Models: ConEx, QMult
- YAGO3-10 ConEx embeddings
- FB15K-237 ConEx embeddings
- WN18RR ConEx embeddings
- For more please look at Hobbit Data
A knowledge graph embedding model can be trained via different strategies (e.g. 1vsAll, KvsAll or Negative Sampling). For details, we refer to
documents/training_techniques.
# To download a pretrained ConEx
mkdir ConEx && cd ConEx && wget -r -nd -np https://hobbitdata.informatik.uni-leipzig.de/KGE/DBpedia/ConEx/ && cd ..from core import KGE
pre_trained_kge = KGE(path_of_pretrained_model_dir='ConEx')
pre_trained_kge.triple_score(head_entity=["http://dbpedia.org/resource/Albert_Einstein"],relation=["http://dbpedia.org/ontology/birthPlace"],tail_entity=["http://dbpedia.org/resource/Ulm"]) # tensor([0.9309])
pre_trained_kge.triple_score(head_entity=["http://dbpedia.org/resource/Albert_Einstein"],relation=["http://dbpedia.org/ontology/birthPlace"],tail_entity=["http://dbpedia.org/resource/German_Empire"]) # tensor([0.9981])
pre_trained_kge.triple_score(head_entity=["http://dbpedia.org/resource/Albert_Einstein"],relation=["http://dbpedia.org/ontology/birthPlace"],tail_entity=["http://dbpedia.org/resource/Kingdom_of_Württemberg"]) # tensor([0.9994])
pre_trained_kge.triple_score(head_entity=["http://dbpedia.org/resource/Albert_Einstein"],relation=["http://dbpedia.org/ontology/birthPlace"],tail_entity=["http://dbpedia.org/resource/Germany"]) # tensor([0.9498])
pre_trained_kge.triple_score(head_entity=["http://dbpedia.org/resource/Albert_Einstein"],relation=["http://dbpedia.org/ontology/birthPlace"],tail_entity=["http://dbpedia.org/resource/France"]) # very low
pre_trained_kge.triple_score(head_entity=["http://dbpedia.org/resource/Albert_Einstein"],relation=["http://dbpedia.org/ontology/birthPlace"],tail_entity=["http://dbpedia.org/resource/Italy"]) # very low
pre_trained_kge.predict_topk(head_entity=["http://dbpedia.org/resource/Albert_Einstein"],relation=["http://dbpedia.org/ontology/birthPlace"]) # needs more memory than simple triple eval.
# ...Any pretrained model can be deployed with an ease. Moreover, anyone on the internet can use the pretrained model with --share parameter.
python deploy.py --path_of_experiment_folder 'ConEx' --share True
Loading Model...
Model is loaded!
Running on local URL: http://127.0.0.1:7860/
Running on public URL: https://54886.gradio.app
This share link expires in 72 hours. For free permanent hosting, check out Spaces (https://huggingface.co/spaces)
In documents folder, we explained many details about knowledge graphs, knowledge graph embeddings, training strategies and many more background knowledge. We continuously work on documenting each and every step to increase the readability of our code.
Currently, we are working on our manuscript describing our framework. If you really like our work and want to cite it now, feel free to chose one :)
# DICE Embedding Framework
@article{demir2022hardware,
title={Hardware-agnostic computation for large-scale knowledge graph embeddings},
author={Demir, Caglar and Ngomo, Axel-Cyrille Ngonga},
journal={Software Impacts},
year={2022},
publisher={Elsevier}
}
# KronE
@article{demir2022kronecker,
title={Kronecker Decomposition for Knowledge Graph Embeddings},
author={Demir, Caglar and Lienen, Julian and Ngomo, Axel-Cyrille Ngonga},
journal={arXiv preprint arXiv:2205.06560},
year={2022}
}
# QMult, OMult, ConvQ, ConvO
@InProceedings{pmlr-v157-demir21a,
title = {Convolutional Hypercomplex Embeddings for Link Prediction},
author = {Demir, Caglar and Moussallem, Diego and Heindorf, Stefan and Ngonga Ngomo, Axel-Cyrille},
booktitle = {Proceedings of The 13th Asian Conference on Machine Learning},
pages = {656--671},
year = {2021},
editor = {Balasubramanian, Vineeth N. and Tsang, Ivor},
volume = {157},
series = {Proceedings of Machine Learning Research},
month = {17--19 Nov},
publisher = {PMLR},
pdf = {https://proceedings.mlr.press/v157/demir21a/demir21a.pdf},
url = {https://proceedings.mlr.press/v157/demir21a.html},
}
# ConEx
@inproceedings{demir2021convolutional,
title={Convolutional Complex Knowledge Graph Embeddings},
author={Caglar Demir and Axel-Cyrille Ngonga Ngomo},
booktitle={Eighteenth Extended Semantic Web Conference - Research Track},
year={2021},
url={https://openreview.net/forum?id=6T45-4TFqaX}}
# Shallom
@inproceedings{demir2021shallow,
title={A shallow neural model for relation prediction},
author={Demir, Caglar and Moussallem, Diego and Ngomo, Axel-Cyrille Ngonga},
booktitle={2021 IEEE 15th International Conference on Semantic Computing (ICSC)},
pages={179--182},
year={2021},
organization={IEEE}
For any questions or wishes, please contact: caglar.demir@upb.de or caglardemir8@gmail.com