Skip to content
A Python library for learning and evaluating knowledge graph embeddings
Jupyter Notebook Python
Branch: master
Clone or download
Latest commit d420aed Sep 5, 2019


PyKEEN Build Status Coverage Status on CodeCov Documentation Status Zenodo DOI

PyKEEN (Python KnowlEdge EmbeddiNgs) is a package for training and evaluating knowledge graph embeddings. Currently, it provides implementations of 10 knowledge graph emebddings models, and can be run in training mode in which users provide their own set of hyper-parameter values, or in hyper-parameter optimization mode to find suitable hyper-parameter values from set of user defined values. PyKEEN can also be run without having experience in programing by using its interactive command line interface that can be started with the command pykeen from a terminal.


We are currently working on PyKEEN 1.0 which will provide additional features such as several negative sampling approaches and further evaluation metrics. Furthermore, we are integrating additional KGE models.

We are developing a new package for training and evaluating multimodal KGE models which will be later integrated into PyKEEN.


If you find PyKEEN useful in your work, please consider citing:

[1]Ali, M., et al.: The KEEN Universe: An ecosystem for knowledge graph embeddings with a focus on reproducibility and transferability. In: International Semantic Web Conference (2019).

Share Your Experimental Artifacts

You can share you trained KGE models along the other experimental artifacts through the KEEN Model Zoo.

Installation Current version on PyPI Supported Python Versions: 3.6 and 3.7 MIT License

To install pykeen, Python 3.6+ is required, and we recommend to install it on Linux or Mac OS systems. Please run following command:

pip install pykeen

Alternatively, it can be installed from the source for development with:

$ git clone pykeen
$ cd pykeen
$ pip install -e .

However, GPU acceleration is limited to Linux systems with the appropriate graphics cards as described in the PyTorch documentation.

Installing Extras with Pip

PyKEEN uses pip's extras functionality to allow some non-essential features to be skipped. They can be installed with the following:

  1. pip install pykeen[ndex] enables support for loading networks from NDEx. They can be added to the training file paths by prefixing files with ndex:


Contributions, whether filing an issue, making a pull request, or forking, are appreciated. See CONTRIBUTING.rst for more information on getting involved.


Code examples can be found in the notebooks directory.

Further tutorials are available in our documentation.

CLI Usage - Set Up Your Experiment within 60 seconds

To start the PyKEEN CLI, run the following command:


then the command line interface will assist you to configure your experiments.

To start PyKEEN with an existing configuration file, run:

pykeen -c /path/to/config.json

then the command line interface won't be called, instead the pipeline will be started immediately.

Starting the Prediction Pipeline

To make prediction based on a trained model, run:

pykeen-predict -m /path/to/model/directory -d /path/to/data/directory

where the value for the argument -m is the directory containing the model, in more detail following files must be contained in the directory:

  • configuration.json
  • entities_to_embeddings.json
  • relations_to_embeddings.json
  • trained_model.pkl

These files are automatically created after model is trained (and evaluated) and exported in your specified output directory.

The value for the argument -d is the directory containing the data for which inference should be applied, and it needs to contain following files:

  • entities.tsv
  • relations.tsv

where entities.tsv contains all entities of interest, and relations.tsv all relations. Both files should contain a single column containing all the entities/relations. Based on these files, PyKEEN will create all triple permutations, and computes the predictions for them, and saves them in data directory in predictions.tsv. Note: the model- and the data-directory can be the same directory as long as all required files are provided.

Optionally, a set of triples can be provided that should be exluded from the prediction, e.g. all the triples contained in the training set:

pykeen-predict -m /path/to/model/directory -d /path/to/data/directory -t /path/to/triples.tsv

Hence, it is easily possible to compute plausibility scores for all triples that are not contained in the training set.

Summarize the Results of All Experiments

To summarize the results of all experiments, please provide the path to parent directory containing all the experiments as sub-directories, and the path to the output file:

pykeen-summarize -d /path/to/experiments/directory -o /path/to/output/file.csv
You can’t perform that action at this time.