This repository is used to replicate the experiments of article "Detecting, Creating, Repairing, and Understanding Indivisible Multi-Hunk Bugs" on AlphaRepair.

The repository is modified from the paper "Less Training, More Repairing Please: Revisiting Automated Program Repair via Zero-shot Learning", which is published at ESEC/FSE 2022. And the source code is available here.

In this repository, we made the following changes:

• Add add_new_line method call.
• Separate the generation patch and verification patch parts.
• Limit the generation of up to 5000 patches per suspicious location (add_new_line and process_file count separately).
• The whole process of fixing each bug is limited to 5h.

• JDK 1.8
• Python 3.10
• Defects4J 2.0.0
• CatenaD4J
• Ubuntu 20.04
• CUDA Version: 11.6

• Timeout: 5h
• beam_width: 5
• Plausible patches number limit at one suspicious location: 5000

None

Use the Dockerfile in ./Docker to create the docker image.

docker build -t alpharepair-env .

This docker image includes Defects4J, CatenaD4J, JDK 1.8, and Python 3.10.

AlphaRepair requires the use of GPU, otherwise generating the patch part would be very slow.

docker run -it --gpus all -e NVIDIA_DRIVER_CAPABILITIES=compute,utility -e NVIDIA_VISIBLE_DEVICES=all --name alpharepair alpharepair-env /bin/bash

At the root of this container, we clone the AlphaRepair repository.

cd /
git clone https://github.com/give-to/AlphaRepairAPI.git
cd /AlphaRepairAPI && chmod +x *

After testing, we found that using pip in the Docker file might cause image creation to fail, so configuration inside the container is required.

# Preparing the environment
cd /AlphaRepairAPI/code
pip install -r requirements.txt

# Install Pytorch for CUDA 11.6
pip install torch==1.13.1+cu116 torchvision==0.14.1+cu116 torchaudio==0.13.1 --extra-index-url https://download.pytorch.org/whl/cu116

If you have another version of CUDA, you can refer https://pytorch.org/get-started/previous-versions/ to install the corresponding Pytorch version.

Make sure your GPU is available by following commands in Python terminal:

import torch
torch.cuda.is_available()

This output should be True, otherwise it will be very slow in the patch generation part.

It takes several minutes to quickly test your installation. (Note: In quick test, the ochiai.ranking.txt in Chart-18-2 only contains one location! )

cp location/ochiai/chart/18/2short location/ochiai/chart/18/2
# Generating the patches
python3 runAlphaRepair.py --start_bug_index=0 --end_bug_index=0
# Validate the patches
python3 runAlphaRepair.py --start_bug_index=0 --end_bug_index=0 --validation=True

After finishing the repair, the results are in folders: codebert_result. The plausible patch will show Success (Plausible Patch).

It may take about 22 days to finish the entire experiment. The main script is runAlphaRepair.py, it reads the list of bugs in the buglist file and fixes them sequentially. You can modify buglist to determine the bugs to be fixed.

cp location/ochiai/chart/18/2bak location/ochiai/chart/18/2
# Generating the patches
python3 runAlphaRepair.py --start_bug_index=0 --end_bug_index=104
# Validate the patches
python3 runAlphaRepair.py --start_bug_index=0 --end_bug_index=104 --validation=True

After finishing the repair, the results are in folders: codebert_result.

Note: In the patch generation phase, the patches will be saved in the store_changes folder, and the time used to generate the patches will be saved in the time_info folder. Do not delete these files to ensure the normal operation of patch validation. The results of the patch validation will be in the codebert_result folder.

 |--- README.md               :  User guidance
 |--- code                    :  Source code
 |------ C4J_results          :  Results about CatenaD4J by AlphaRepair
 |--------- codebert_result   :  Validation results for all patches
 |--------- store_changes     :  The process information of the patch generation
 |--------- time_info         :  The total time to generate the patch for each bug
 |------ codebert-base-mlm    :  The model used in AlphaRepair
 |------ location             :  Bug positions localized with GZoltar
 |--- correct_patches         :  Generated patches about Defects4J by AlphaRepair