STAR_KGC

This repo contains the source code of the paper accepted by WWW'2021. "Structure-Augmented Text Representation Learning for Efficient Knowledge Graph Completion"(WWW 2021).

1. Thanks

The repository is partially based on huggingface transformers, KG-BERT and RotatE.

2. Installing requirement packages

conda create -n StAR python=3.6
source activate StAR
pip install numpy torch tensorboardX tqdm boto3 requests regex sacremoses sentencepiece matplotlib

2.1 Optional package (for mixed float Computation)

git clone https://github.com/NVIDIA/apex
cd apex
pip install -v --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext" ./

3. Dataset

WN18RR, FB15k-237, UMLS

Train and test set in ./data
As validation on original dev set is costly, we validated the model on dev subset during training.
The dev subset of WN18RR is provided in ./data/WN18RR called new_dev.dict. Use below commands to get the dev subset for WN18RR (FB15k-237 is similar without the --do_lower_case) used in training process.

 CUDA_VISIBLE_DEVICES=0 \
  python get_new_dev_dict.py \
 	--model_class bert \
 	--weight_decay 0.01 \
 	--learning_rate 5e-5 \
 	--adam_epsilon 1e-6 \
 	--max_grad_norm 0. \
 	--warmup_proportion 0.05 \
 	--do_train \
 	--num_train_epochs 7 \
 	--dataset WN18RR \
 	--max_seq_length 128 \
 	--gradient_accumulation_steps 4 \
 	--train_batch_size 16 \
 	--eval_batch_size 128 \
 	--logging_steps 100 \
 	--eval_steps -1 \
 	--save_steps 2000 \
 	--model_name_or_path bert-base-uncased \
 	--do_lower_case \
 	--output_dir ./result/WN18RR_get_dev \
 	--num_worker 12 \
 	--seed 42 \

 CUDA_VISIBLE_DEVICES=0 \
  python get_new_dev_dict.py \
 	--model_class bert \
 	--weight_decay 0.01 \
 	--learning_rate 5e-5 \
 	--adam_epsilon 1e-6 \
 	--max_grad_norm 0. \
 	--warmup_proportion 0.05 \
 	--do_eval \
 	--num_train_epochs 7 \
 	--dataset WN18RR \
 	--max_seq_length 128 \
 	--gradient_accumulation_steps 4 \
 	--train_batch_size 16 \
 	--eval_batch_size 128 \
 	--logging_steps 100 \
 	--eval_steps 1000 \
 	--save_steps 2000 \
 	--model_name_or_path ./result/WN18RR_get_dev \
 	--do_lower_case \
 	--output_dir ./result/WN18RR_get_dev \
 	--num_worker 12 \
 	--seed 42 \

NELL-One
- We reformat original NELL-One as the three benchmarks above.
- Please run the below command to get the reformatted data.
```
 python reformat_nell_one.py --data_dir path_to_downloaded --output_dir ./data/NELL_standard
```

4. Training and Test (StAR)

Run the below commands for reproducing results in paper. Note, all the eval_steps is set to -1 to train w/o validation and save the last checkpoint, because standard dev is very time-consuming. This can get similar results as in the paper.

4.1 WN18RR

CUDA_VISIBLE_DEVICES=0 \
python run_link_prediction.py \
    --model_class roberta \
    --weight_decay 0.01 \
    --learning_rate 1e-5 \
    --adam_betas 0.9,0.98 \
    --adam_epsilon 1e-6 \
    --max_grad_norm 0. \
    --warmup_proportion 0.05 \
    --do_train --do_eval \
    --do_prediction \
    --num_train_epochs 7 \
    --dataset WN18RR \
    --max_seq_length 128 \
    --gradient_accumulation_steps 4 \
    --train_batch_size 16 \
    --eval_batch_size 128 \
    --logging_steps 100 \
    --eval_steps 4000 \
    --save_steps 2000 \
    --model_name_or_path roberta-large \
    --output_dir ./result/WN18RR_roberta-large \
    --num_worker 12 \
    --seed 42 \
    --cls_method cls \
    --distance_metric euclidean \

CUDA_VISIBLE_DEVICES=2 \
python run_link_prediction.py \
    --model_class bert \
    --weight_decay 0.01 \
    --learning_rate 5e-5 \
    --adam_betas 0.9,0.98 \
    --adam_epsilon 1e-6 \
    --max_grad_norm 0. \
    --warmup_proportion 0.05 \
    --do_train --do_eval \
    --do_prediction \
    --num_train_epochs 7 \
    --dataset WN18RR \
    --max_seq_length 128 \
    --gradient_accumulation_steps 4 \
    --train_batch_size 16 \
    --eval_batch_size 128 \
    --logging_steps 100 \
    --eval_steps 4000 \
    --save_steps 2000 \
    --model_name_or_path bert-base-uncased \
    --do_lower_case \
    --output_dir ./result/WN18RR_bert \
    --num_worker 12 \
    --seed 42 \
    --cls_method cls \
    --distance_metric euclidean \

4.2 FB15k-237

CUDA_VISIBLE_DEVICES=0 \
python run_link_prediction.py \
    --model_class roberta \
    --weight_decay 0.01 \
    --learning_rate 1e-5 \
    --adam_betas 0.9,0.98 \
    --adam_epsilon 1e-6 \
    --max_grad_norm 0. \
    --warmup_proportion 0.05 \
    --do_train --do_eval \
    --do_prediction \
    --num_train_epochs 7. \
    --dataset FB15k-237 \
    --max_seq_length 100 \
    --gradient_accumulation_steps 4 \
    --train_batch_size 16 \
    --eval_batch_size 128 \
    --logging_steps 100 \
    --eval_steps -1 \
    --save_steps 2000 \
    --model_name_or_path roberta-large \
    --output_dir ./result/FB15k-237_roberta-large \
    --num_worker 12 \
    --seed 42 \
    --fp16 \
    --cls_method cls \
    --distance_metric euclidean \

4.3 UMLS

CUDA_VISIBLE_DEVICES=0 \
python run_link_prediction.py \
    --model_class roberta \
    --weight_decay 0.01 \
    --learning_rate 1e-5 \
    --adam_betas 0.9,0.98 \
    --adam_epsilon 1e-6 \
    --max_grad_norm 0. \
    --warmup_proportion 0.05 \
    --do_train --do_eval \
    --do_prediction \
    --num_train_epochs 20 \
    --dataset UMLS \
    --max_seq_length 16 \
    --gradient_accumulation_steps 1 \
    --train_batch_size 16 \
    --eval_batch_size 128 \
    --logging_steps 100 \
    --eval_steps -1 \
    --save_steps 200 \
    --model_name_or_path roberta-large \
    --output_dir ./result/UMLS_model \
    --num_worker 12 \
    --seed 42 \
    --cls_method cls \
    --distance_metric euclidean

4.4 NELL-One

CUDA_VISIBLE_DEVICES=0 \
python run_link_prediction.py \
    --model_class bert \
    --do_train --do_eval \usepacka--do_prediction \
    --warmup_proportion 0.1 \
    --learning_rate 5e-5 \
    --num_train_epochs 8. \
    --dataset NELL_standard \
    --max_seq_length 32 \
    --gradient_accumulation_steps 1 \
    --train_batch_size 16 \
    --eval_batch_size 128 \
    --logging_steps 100 \
    --eval_steps -1 \
    --save_steps 2000 \
    --model_name_or_path bert-base-uncased \
    --do_lower_case \
    --output_dir ./result/NELL_model \
    --num_worker 12 \
    --seed 42 \
    --fp16 \
    --cls_method cls \
    --distance_metric euclidean

5. StAR_Self-Adp

5.1 Data preprocessing

Get the trained model of RotatE, more details please refer to RotatE.

Run the below commands sequentially to get the training dataset of StAR_Self-Adp.

Run the run_get_ensemble_data.py in ./StAR

 CUDA_VISIBLE_DEVICES=0 python run_get_ensemble_data.py \
 	--dataset WN18RR \
 	--model_class roberta \
 	--model_name_or_path ./result/WN18RR_roberta-large \
 	--output_dir ./result/WN18RR_roberta-large \
 	--seed 42 \
 	--fp16

Run the ./codes/run.py in rotate. (please replace the TRAINED_MODEL_PATH with your own trained model's path)

 CUDA_VISIBLE_DEVICES=3 python ./codes/run.py \
 	--cuda --init ./models/RotatE_wn18rr_0 \
 	--test_batch_size 16 \
 	--star_info_path /home/wangbo/workspace/StAR_KGC-master/StAR/result/WN18RR_roberta-large \
 	--get_scores --get_model_dataset

5.2 Train and Test

Run the run.py in ./StAR/ensemble. Note the --mode should be alternate in head and tail, and perform a average operation to get the final results.
Note: Please replace YOUR_OUTPUT_DIR, TRAINED_MODEL_PATH and StAR_FILE_PATH in ./StAR/peach/common.py with your own paths to run the command and code.

CUDA_VISIBLE_DEVICES=2 python run.py \
--do_train --do_eval --do_prediction --seen_feature \
--mode tail \
--learning_rate 1e-3 \
--feature_method mix \
--neg_times 5 \
--num_train_epochs 3 \
--hinge_loss_margin 0.6 \
--train_batch_size 32 \
--test_batch_size 64 \
--logging_steps 100 \
--save_steps 2000 \
--eval_steps -1 \
--warmup_proportion 0 \
--output_dir /home/wangbo/workspace/StAR_KGC-master/StAR/result/WN18RR_roberta-large_ensemble  \
--dataset_dir /home/wangbo/workspace/StAR_KGC-master/StAR/result/WN18RR_roberta-large \
--context_score_path /home/wangbo/workspace/StAR_KGC-master/StAR/result/WN18RR_roberta-large \
--translation_score_path /home/wangbo/workspace/StAR_KGC-master/rotate/models/RotatE_wn18rr_0  \
--seed 42

"Structure-Augmented Text Representation Learning for Efficient Knowledge Graph Completion"(WWW 2021)

Related tags

Overview

STAR_KGC

1. Thanks

2. Installing requirement packages

2.1 Optional package (for mixed float Computation)

3. Dataset

4. Training and Test (StAR)

4.1 WN18RR

4.2 FB15k-237

4.3 UMLS

4.4 NELL-One

5. StAR_Self-Adp

5.1 Data preprocessing

5.2 Train and Test

Owner

Bo Wang

Weighted K Nearest Neighbors (kNN) algorithm implemented on python from scratch.

Official PyTorch implementation of the Fishr regularization for out-of-distribution generalization

PyTorch implementation of PSPNet segmentation network

Implementation of CVPR'21: RfD-Net: Point Scene Understanding by Semantic Instance Reconstruction

Neural network for recognizing the gender of people in photos

Learning kernels to maximize the power of MMD tests

Civsim is a basic civilisation simulation and modelling system built in Python 3.8.

MoveNetを用いたPythonでの姿勢推定のデモ

Exploring Classification Equilibrium in Long-Tailed Object Detection, ICCV2021

A curated list of awesome Active Learning

tsai is an open-source deep learning package built on top of Pytorch & fastai focused on state-of-the-art techniques for time series classification, regression and forecasting.

This repository is for Contrastive Embedding Distribution Refinement and Entropy-Aware Attention Network (CEDR)

Unofficial PyTorch implementation of the Adaptive Convolution architecture for image style transfer

UnpNet - Rethinking 3-D LiDAR Point Cloud Segmentation(IEEE TNNLS)

Predicting Axillary Lymph Node Metastasis in Early Breast Cancer Using Deep Learning on Primary Tumor Biopsy Slides

DALL-Eval: Probing the Reasoning Skills and Social Biases of Text-to-Image Generative Transformers

Denoising Diffusion Probabilistic Models

A Flow-based Generative Network for Speech Synthesis

Official code for "End-to-End Optimization of Scene Layout" -- including VAE, Diff Render, SPADE for colorization (CVPR 2020 Oral)

FLVIS: Feedback Loop Based Visual Initial SLAM