SequenceR: Sequence-to-Sequence Learning for End-to-End Program Repair

SequenceR is a seq2seq model designed to predict bug fixes at the line level. The paper (doi:10.1109/TSE.2019.2940179) explains the approach.

If you use SequenceR for academic purposes, please cite the following publication:

@article{chen2018sequencer,
  title={SequenceR: Sequence-to-Sequence Learning for End-to-End Program Repair},
  author={Chen, Zimin and Kommrusch, Steve and Tufano, Michele and Pouchet, Louis-No{\"e}l and Poshyvanyk, Denys and Monperrus, Martin},
  journal={IEEE Transaction on Software Engineering},
  year={2019}
}

Usage

Docker

Simply run the following two commands to set up use of the SequenceR Golden model:

docker build --tag=sequencer .
docker run -it sequencer

And now all dependecies are installed (including defects4j).

Or, use our this version from the Docker Hub.

Without docker

Install dependencies

First run src/setup_env.sh to setup enviroment and clone/compile project. Please view src/setup_env.sh for more details.

All models are versioned using git-lfs, make sure to configure it and correctly fetch the models before using.

Execution

Then run src/sequencer-predict.sh with the following parameters:

./sequencer-predict.sh --model=[model path] --buggy_file=[abs path] --buggy_line=[int] --beam_size=[int] --output=[abs path]

• --model: Absolute path to the model
• --buggy_file: Absolute path to buggy file
• --buggy_line: Line number indicating where the bug is, or just want it get changed.
• --beam_size: Beam size for prediction
• --output: Output directory to store the generated patches

Experiments

CodRep experiment

The training data consists of results/Golden/src-train.txt and results/Golden/tgt-train.txt (line to line correspondence).

The CodRep4 testing data consists of results/Golden/src-test.txt and results/Golden/tgt-test.txt (line to line correspondence).

Defects4J experiment

In results/Defects4J_patches you can find all patches that are found by SequencerR. Patches that are stored in *_compiled are patches that compiled. Patches that are stored in *_passed are patches that compiled and passed the test suite. Patches that are stored in *_correct are patches that compiled, passed the test suite and are equivalent to the human patch.

To rerun our experiment of SequenceR over Defects4J. Run src/Defects4J_Experiment/Defects4J_experiment.sh, make sure you have defects4j installed.

Defects4J_oneLiner_metadata.csv contains metadata for all Defects4J bugs that we consider. src/Defects4J_Experiment/validatePatch.py contains the precedure for running Defects4J test, we have time limit on compile time (60s) and test running time (300s).

Model creation, training and use:

Prerequisites

SequenceR uses the OpenNMT library to set up program repair as a translation from buggy code to fixed code. Documentation on OpenNMT including parameter setup is at http://opennmt.net/OpenNMT-py/

Setup

Choose a directory and:

git clone https://github.com/OpenNMT/OpenNMT-py

When testing a new configuration, copy a working data directory and modify *sh files as desired.

Set up environment variables:

export CUDA_VISIBLE_DEVICES=0
export THC_CACHING_ALLOCATOR=0
export OpenNMT_py=.../OpenNMT-py
export data_path=.../results/Golden  # Or a new directory path as desired

Train

For details on model training, refer to OpenNMT documentation. To run SequenceR training:

cd src
sequencer-train.sh

Test

For details on model usage (translation), refer to OpenNMT documentation. To run SequenceR testing:

cd src
sequencer-test.sh

License

The code and data in this repository are under the MIT license.