Knowledge Base Completion (kbc)
This code reproduces results in Canonical Tensor Decomposition for Knowledge Base Completion (ICML 2018).
Installation
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
Datasets
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.
Update notes
In moving from cpp to pytorch, a hidden regularization weight constant was removed. This changes the settings required to reproduce the results in the paper. As a rule of thumb, and before I re-run a grid-search to find optimal parameters, multiplying the regularization strength in the table below by a factor of two should give reasonable results.
The "reciprocal" parameter was also lost in a temporary effort to simplify the code.
Reproducing results (only valid before pytorch update)
To reproduce results, use learning.learn as follows
python kbc/learn.py --dataset FB15K --model ComplEx --rank 2000 --optimizer Adagrad --learning_rate 1e-2 --batch_size 100 --regularizer N3 --reg 5e-3 --reciprocals 1 --max_epochs 10 --valid 1
To reproduce results in this setting, use the following hyper-parameters (model ComplEx, optimizer Adagrad, regularizer N3, reciprocals 1):
| Dataset | rank | lr | reg | batch_size | Time |
|---|---|---|---|---|---|
| WN18 | 2000 | 1e-1 | 1e-1 | 100 | 150s/epoch |
| WN18RR | 2000 | 1e-1 | 1e-1 | 100 | 93s/epoch |
| FB15K | 2000 | 1e-2 | 5e-3 | 100 | 225s/epoch |
| FB15K-237 | 2000 | 1e-1 | 1e-1 | 100 | 115s/epoch |
| YAGO3-10 | 1000 | 1e-1 | 1e-2 | 1500 | 485s/epoch |
License
kbc is CC-BY-NC licensed, as found in the LICENSE file.