This repository contains our implementation of the HyperKG: Hyperbolic Knowledge Graph Embeddings for Knowledge Base Completion.

This code is partially based on code from the following repositories:

• OpenKE-PyTorch (old)
• poincare-embeddings

Every source code file written exclusively by the author of this repo is licensed under Apache License Version 2.0. For more information, please refer to the license.

• Prerequisites :

  • Python, C, C++.
  • Python Libraries: NumPy, SciPy, pytorch (with CUDA support).

• Run the code:

  1. Compile the C, C++ code using: sh make.sh
  2. To analyze HyperKG's performance on a dataset, please run:

     python example_train_poincare.py
     

     All parameters/hyperparameters can be altered by directly modifying the example_train_poincare.py file.
  3. Known issues: There is a portability issue with the original C code provided by OpenKE-PyTorch (old). As a quick workaround, I added a Datatype control variable in the original Config Class. If a segmentation fault occurs after Step 2, then this command con.set_int_type('int64') should be commented out in both example_train_poincare.py and example_test_poincare.py files.

The folder res/saved_models contains saved models for the experiments WN18RR and FB15k-237.

• prodromos DOT kolyvakis AT epfl DOT ch