Skip to content

Neural Program Repair with Execution-based Backpropagation http://arxiv.org/pdf/2105.04123

Notifications You must be signed in to change notification settings

ASSERT-KTH/RewardRepair

Repository files navigation

RewardRepair

Neural Program Repair with Execution-based Backpropagation (ICSE 2022)

@inproceedings{Ye2021RewardRepair,
 title = {Neural Program Repair with Execution-based Backpropagation},
 year = {2022},
 author = {He Ye and Matias Martinez and Martin Monperrus},
 url = {http://arxiv.org/pdf/2105.04123},
 doi = {10.1145/3510003.3510222},
 booktitle = {Proceedings of the International Conference on Software Engineering},
}

Trained Model

We share our trained model in Zenodo:

RewardRepair trained model: https://doi.org/10.5281/zenodo.5997686

Folder Structure

├── data: csv data used for training
│ 
├── model: the trained model of RewardRepair
│
├── train.py: script to  train RewardRepair
│
├── test.py: script to test RewardRepair
│
├── result: raw experiment results
│
├── ComparisonData.csv: the file to compare with the state-of-the-art

Prerequisites

  • JDK 1.8
  • Pytorch==1.7.1
  • transformers>=4.10.0
  • OS: Linux and Mac
  • Add submodule Bears and copy the folder myscript to Bears
git submodule add https://github.com/bears-bugs/bears-benchmark.git
cp -rf myscripts ./bears-benchmark/
pip install transformers
pip install sentencepiece

Get the CoCoNut java dataset from here: https://github.com/lin-tan/CoCoNut-Artifact/releases/tag/training_data_1.0.0 use the script to extract data: extract_CoCoNut_data.py

Data availables here: 
https://zenodo.org/record/7009192#.YwPS4-xBxbg 

Get Megadiff dataset from here https://github.com/monperrus/megadiff use the script to extract data: extract_Megadiff_data.py

Get CodRep dataset from here https://github.com/KTH/codrep-2019

To run our script

python3 train.py

To evaluate our model on other bugs (make sure the script takes the correct test file patch)

python3 test.py

RQ2: Details reults are available in RewardRepair/ResultForRQ2/

Benchmarks Top30 Top100 Top200
QuixBugs 44.1% 35.7% 31.5%
Defects4J 45.7% 38.0% 33.6%

Note: Above are the number reported in the paper. The actual compilable rate should be higher. Because we found many uncompilable cases due to space in operators > =, < =, = =, ! =. This could be easily avoided by added tokens in training or post processing in test.py.

About

Neural Program Repair with Execution-based Backpropagation http://arxiv.org/pdf/2105.04123

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published