KGE-Literals

Link prediction with literals in PyTorch.

Getting Started

1. Install miniconda http://conda.pydata.org/miniconda.html
2. Do conda env create
3. Enter the env source activate kga
4. Install Pytorch v0.3+

Preparing new experiments

1. Create your model in kga/models/literals.py.
2. Create experiment script at experiments/{fb15k,yago,ml} dir. See baseline files run_multitask_{fb15k,yago,ml} for reference and use it as template.
   • The script has to implements all of the variables as in the run_multitask_{fb15k,yago,ml}'s argument parsers.
   • It is very important to use proper model name when saving the model file. Consult the reference above.
3. Run the experiment, e.g. nohup python -u experiments/yago3-10/run_multitask_yago.py --use_gpu --log_interval -1 --nepoch 300 --lr_decay_every 100 --k 100 --mbsize 200 --lr 1e-3 --weight_decay 5e-4 &> mtkgnn_yago_lr1e-3_wd5e-4 &
   • In the example above, tt will save a model in models/yago/mtkgnn_yago_lr0.001_wd0.0001
4. After finished training, test the resulting model, e.g. python experiments/yago3-10/run_multitask_yago.py --use_gpu --k 100 --test --test_model mtkgnn_yago_lr0.001_wd0.0001
   • This will load the models from the previous training.
   • It will print all of the evaluation metrics over the test set.

Dependencies

1. Python 3.5+
2. PyTorch 0.2+
3. Numpy
4. Scikit-Learn
5. Pandas