KGATE is a knowledge graph embedding library bridging the encoders from Pytorch Geometric and the decoders from TorchKGE.
This tool relies heavily on the performances of TorchKGE and its numerous implemented modules for link prediction, negative sampling and model evaluation. The main goal here is to address the lack of encoders in the original library, who is unfortunately not maintained anymore.
It is recommended to download the configuration template alongside your installation (see Usage below).
pip install kgateClone this repository and install it in a virtual environment like so:
git clone git@github.com:BAUDOTlab/KGATE.git
python -m venv kge_env
source kge_env/bin/activateKGATE is developed using Poetry. If you want to contribute to KGATE or make your own modifications, follow these steps:
pip install poetrygit clone git@github.com:BAUDOTlab/KGATE.gitcd KGATE
poetry install(usage)=
KGATE is meant to be a self-sufficient training environment for knowledge graph embedding that requires very little code to work but can easily be expanded or modified. Everything stems from the Architect class, which holds all the necessary attributes and methods to fully train and test a KGE model following the autoencoder architecture, as well as run inference.
The configuration file lets you iterate quickly without changing your code. See the template to learn what the different options do.
At the very least, KGATE expects the Knowledge Graph to be given as a pandas dataframe or a CSV file with the columns "from", "to" and "rel", corresponding respectively to the head nodes, tail nodes and relation typesof the triplets, with one triplet per row. Any extra columns are ignored. In addition, a metadata dataframe can be submitted (can also be a CSV) to map each node with their type, requiring the columns "id" and "type". Extra columns are likewise ignored.
from kgate import Architect
config_path = "/path/to/your/config.toml"
architect = Architect(config_path = config_path)
# Train the model using KG and hyperparameters specified in the configuration
architect.train_model()
# Test the trained model, using the best checkpoint
architect.test()
# Run KG completion task, the empty list is the element that will be predicted
known_heads = []
known_relations = []
known_tails = []
architect.infer(known_heads, known_relations, known_tails)For a more detailed example and specific methods that are available in the package, see the upcoming readthedocs documentation.
This project is licensed under the MIT License. See the LICENSE file for details.