Use run.py with an a appropriate virtual environment to run with the 5 folds
given in datasets/5folds.
The output is given in coauthor_results.txt.
A link prediction is treated as true iff the most probable completion
matches the true completion.
This is the implementation of Neural Logic Programming, proposed in the following paper:
Differentiable Learning of Logical Rules for Knowledge Base Reasoning. Fan Yang, Zhilin Yang, William W. Cohen. NIPS 2017.
- Python 2.7
- Numpy
- Tensorflow 1.0.1
- Scikit-learn 0.20.4
The following command starts training a dataset about family relations, and stores the experiment results in the folder exps/demo/.
python src/main.py --datadir=datasets/family --exps_dir=exps/ --exp_name=demo
Wait for around 8 minutes, navigate to exps/demo/, there is rules.txt that contains learned logical rules.
To evaluate the prediction results, follow the steps below. The first two steps is preparation so that we can compute filtered ranks (see TransE for details).
We use the experiment from Quick Start as an example. Change the folder names (datasets/family, exps/dev) for other experiments.
. eval/collect_all_facts.sh datasets/family
python eval/get_truths.py datasets/family
python eval/evaluate.py --preds=exps/demo/test_predictions.txt --truths=datasets/family/truths.pckl