How-To Guides

Installation

To install, first clone the repository and move into the folder fastkg-icicle. Then run: pip install -e .

CPU/GPU Testing

To test fb15k dataset:

cd ./tests

python trans_e.py

Speedup Comparison

The ./tests/comparison compares the speedup with TorchKGE. To run comparison, do the following:

1. pip install torchkge

2. cd ./tests/comparison

3. python trans_e_fastkg.py

4. python trans_e_torchkge.py

Custom Dataset

FastKG has in-built FB15k and PPOD dataset. To use your own dataset, modify line 35 of the file ./tests/trans_e.py. Replace the line

df = load_fb15k('train')

with the following snippet:

from fastkg.utils import read_csv

df = read_csv('your/tsv/train/file')

The tsv train file should have three columns separated by \t and sequenced in head-rel-tail order. The first line of the file should not contain any header column.

Saving and loading trained model

The model_sparse in ./tests/trans_e.py is a standard PyTorch model. It can be saved and loaded using the PyTorch functions.

Saving

torch.save(model_sparse, "model.pth")

Loading

model_sparse = torch.load("model.pth")

Evaluation

To evaluate if a triplet is a probable KG candidate, use the classify_triplet function from the model_sparse when training is done.

Append this after ./tests/trans_e.py

print(model_sparse.classify_triplet(h_idx=0, r_idx=0, t_idx=1, threshold=1.5))

This checks if the 1st entity, the 1st relation, and the 2nd entity make up a plaussible triplet for the KG or not. The threshold is typicall the loss margin used in training (1.5).