P-INT

Requirements

• Python 3.7.9
• PyTorch 1.7.0

Datasets

We conduct our experiments on two datasets — NELL-One and FB15k237-One.

You can find original datasets(NELL-One) from here.

You can download datasets used in this work from here.

Pre-trained embeddings

• NELL-One
• FB15k237-One (We use this repository to get the embeddings used in this work.)

How to run

Training

• For NELL-One: python train.py --dataset "NELL-One" --few n --max_batches 200000
• For FB15k237-One: python train.py --dataset "FB15k237-One" --few n --max_batches 10000

In this work, we set n=1 for one-shot and n=5 for five-shot.

You can set --max_batches smaller.

After training, the checkpoints will be saved in ./models and the corresponding results will be printed.

You can also download the checkpoints from here.

Test

• For NELL-One: python train.py --dataset "NELL-One" --few n --test
• For FB15k237-One: python train.py --dataset "FB15k237-One" --few n --test

(The checkpoints are for your reference only. We find some errors in net.py, which slightly affect the effect after correction.)