Hephaestus

Using neural machine translation for buggy code repair

Overview

Recent approaches to automatic code repair have seen the use of neural machine translation (NMT) models to translate buggy code into fixed code. These models are trained with the buggy and fixed abstracted code token sequences directly, as in this paper by Tufano et al. Hephaestus aims to use similar models, but train them with Edit Operations instead of the traditional approach mentioned above. Thus, we train the models with the "difference" between the buggy and fixed code.

Hephaestus provides a modular framework from which researchers can reproduce or extend this work.

Installation

Hephaestus requires the following Python dependencies:

• OpenNMT-py (version == 2.0.1)

  pip install OpenNMT-py==2.0.1
  

  If the above command fails because dependencies were not able to be installed, then use:

  pip install OpenNMT-py==2.0.1 --no-deps
  

  Note that parts of the Hephaestus library call OpenNMT command line tools, and these tools are expected to be in the PATH environment variable. Pip will install the command line tools in the Python Scripts directory, so you will need to add this directory to PATH if it isn't there already.

• PyTorch (version >= 1.6.0)

  pip install torch>=1.6.0+cu111
  

  If the above command fails, then install PyTorch from source as given by the PyTorch website:

  pip3 install torch==1.8.1+cu111 torchvision==0.9.1+cu111 torchaudio===0.8.1 -f https://download.pytorch.org/whl/torch_stable.html
  

• Pandas

  pip install pandas
  

• Matplotlib

  pip install matplotlib
  

Once the dependencies have been installed, clone the repository with:

git clone https://github.com/WM-SEMERU/hephaestus.git

Your installation is now complete.

Reproduction and further research

To reproduce the results of this project, run the notebooks 06_TrainModels and 07_Experiment, in order. Due to the random nature of training the models, your results may come out slightly differently than those in the Documentation.

This project was built using nbdev. If you would like to modify the Hephaestus library for research purposes, then you must install nbdev according to the instructions in the link above.

Documentation

Find our documentation here.