pykeen/pykeen

What it solves

PyKEEN simplifies the process of training and evaluating knowledge graph embedding models. It provides a standardized framework to handle the complexities of knowledge graph data, model implementation, and performance evaluation, including support for multi-modal information.

How it works

PyKEEN offers a high-level pipeline function that allows users to quickly train and evaluate models (such as TransE) on specific datasets (such as Nations) with minimal code. It is built on PyTorch Lightning and integrates with Optuna for hyperparameter optimization. The library is designed to be extensible, providing consistent APIs for models, training loops (like the stochastic local closed world assumption), and triple factories for custom data ingestion.

Who it’s for

It is designed for researchers and developers working with knowledge graphs who need a robust tool for embedding entities and relations to perform tasks like link prediction.

Highlights

• Includes 37 built-in datasets and 5 inductive datasets.
• Implements 40 different embedding models.
• Provides a unified API for swapping models and training loops.
• Supports custom dataset integration via TriplesFactory.