This code reproduces results in Canonical Tensor Decomposition for Knowledge Base Completion (ICML 2018).

Create a conda environment with pytorch cython and scikit-learn :

conda create --name kbc_env python=3.7
source activate kbc_env
conda install --file requirements.txt -c pytorch

Then install the kbc package to this environment

python setup.py install

To download the datasets, go to the kbc/scripts folder and run:

chmod +x download_data.sh
./download_data.sh

Once the datasets are download, add them to the package data folder by running :

python kbc/process_datasets.py

This will create the files required to compute the filtered metrics.

Reproduce the results below with the following command :

python kbc/learn.py --dataset FB15K --model ComplEx --rank 500 --optimizer
Adagrad --learning_rate 1e-1 --batch_size 1000 --regularizer N3 --reg 1e-2
 --max_epochs 100 --valid 5

In addition to the results in the paper, here are the performances of ComplEx regularized with the weighted N3 on several datasets, for several dimensions. We use an init scale of 1e-3, a learning rate of 0.1, a batch size of 1000 and 100 max epochs unless specified otherwise. We use the Adagrad optimizer.

For rank 2000 : learning rate 1e-2, batch-size 100, max epochs 200.

Max Epochs : 20

Batch Size : 100 (1000 for rank 1000)

Batch Size : 100 (1000 for rank 8)

kbc is CC-BY-NC licensed, as found in the LICENSE file.