hankniu01 / transe-pytorch

TransE-PyTorch

Implementation of TransE [1] model in PyTorch.

Table of Contents

1. Results
   1. Datasets
      1. FB15k
2. Usage
   1. Training
      1. Options
   2. Unit tests
3. References

Results

Datasets

FB15k

Source/Metric	Hits@1 (raw)	Hits@3 (raw)	Hits@10 (raw)	MRR (raw)
Paper [1]	X	X	34.9	X
TransE-PyTorch	11.1	25.33	46.53	22.29

python3 main.py --dataset_path=<path_to_fb15k_dataset> --epochs=50000 --batch_size=128

Negative sampling impact over time

X axis - epoch id

Y axis - % of samples with nonzero loss

Usage

Synthetic data

For fast debugging/feedback loop use synthetic data from synth_data directory.

python3 main.py --nouse_gpu

Training

python3 main.py --nouse_gpu --dataset_path=<path_to_your_dataset>

Options

To see possible configuration options run help

python3 main.py --help

Unit tests

python3 -m unittest discover -p "*_test.py"

References

[1] Bordes et al., "Translating embeddings for modeling multi- relational data," in Adv. Neural Inf. Process. Syst., 2013