This code reproduces results in Canonical Tensor Decomposition for Knowledge Base Completion to appear at ICML 2018.
Create a conda environment with pytorch cython and scikit-learn :
conda create --name kbc_env python=3.6
source activate kbc_env
conda install --file requirements.txt -c pytorch
Then install the kbc package to this environment (this requires cython to build the library)
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/datasets/process_datasets.py
This will create the files required to compute the filtered metrics.
To reproduce results, use learning.learn as follows
python kbc/learning/learn.py --dataset FB15K --model ComplEx --rank 2000 --optimizer Adagrad --learning_rate 1e-2 --batch_size 100 --regularizer L3ComplEx --reg 5e-3 --learn_inverse_rels 1 --max_epochs 100 --valid 1
learn_inverse_rels corresponds to the Reciprocal setting described in the paper. To reproduce results in this setting, use the following hyper-parameters (model ComplEx, optimizer Adagrad, regularizer L3ComplEx, learn_inverse_rels 1):
| Dataset | rank | lr | reg | batch_size | Time |
|---|---|---|---|---|---|
| WN18 | 2000 | 1e-1 | 1e-1 | 100 | 150s/b |
| WN18RR | 2000 | 1e-1 | 1e-1 | 100 | 93s/b |
| FB15K | 2000 | 1e-2 | 5e-3 | 100 | 225s/b |
| FB15K-237 | 2000 | 1e-1 | 1e-1 | 100 | 115s/b |
| YAGO3-10 | 1000 | 1e-1 | 1e-2 | 1500 | 485s/b |
Start on learning/learn.py to understand how the model is built.
The evaluation procedure is in datasets/big.py (do_eval method). It calls models.cpp getRanking() to obtain the filtered rankings.
To understand the forward and backward pass, assume that factor == RIGHT (will be the case for learn_inverse_rels == 1).
The python library is mostly used for reading the datasets. The C++ library is organized around 4 objects :
- Models : Forward and backward pass for the model. Define the parameters.
- Loss : Forward and backward pass for the loss.
- Regularizer : Forward and backward pass for the regularizer.
- Optimizer : Call the forward and backward passes, applies the gradient step.
Any extension needs to be added in bindings.pyx.
kbc is CC-BY-NC licensed, as found in the LICENSE file.