ITER: Iterative Neural Repair for Multi-Location Patches

ITER is a neural program repair system with an original training and inference loop enabling advanced multi-location patches. See the paper.

@inproceedings{2304.12015,
 title = {ITER: Iterative Neural Repair for Multi-Location Patches},
 booktitle = {Proceedings of International Conference on Software Engineering},
 year = {2024},
 author = {He Ye and Martin Monperrus},
 url = {http://arxiv.org/pdf/2304.12015},
}

Repo Structure

├── repair_iteration: in this folder, you will find all repair iterations of considered bugs
│────────The folder is structured as ** repair_iteration/BugID/FL_Iteration/Ranked_Suspicious_Statement/Iterations **
│ 
├── gzoltar: in this folder, you will find the FL dependency used in ITER
│
├── Generate_Iterative_Sample.py: script to generate iterative training samples.
│
├── 1_localize_fault.py: script to obtain ranked list of bug under repair
│
├── 2_bug_representation.py: script to prepare bug representation of the ranked FL
│
├── 3_repair.py: script to iterative repair
│
├── 4_trace_patches.py:  script to trace plausible patches over iterations, this script generates patches.csv
│
├── ITER_FL.py: script to be called by 3_repair.py to re-execute FL.
│
├── utils/context.jar: tool to obtain the context of bug under repair, called by 2_bug_representation.py
│
├── run_gzoltar_fl.sh: script to execute gzoltar, used by 1_localize_fault.py
│
├── patches.csv: all plausible patches.

In addition, the models are put on Zenodo, see https://zenodo.org/records/14993858 (2.3GB).

Prerequisites

• JDK 1.8
• Pytorch==1.7.1
• transformers>=4.10.0
• pip install transformers
• pip install sentencepiece
• setup Defect4J export PATH = $PATH:your/path/defects4j/framework/bin
• configure Gzoltar: --formula "ochiai" --metric "entropy" --granularity "line" --inclPublicMethods --inclStaticConstructors --inclDeprecatedMethods

How to read repair_iteration folder

• repair_iteration/BugID/FL_Iteration/Ranked_Suspicious_Statement/Iterations (single test failure bug does not has FL iteration folder)

Example: 
>> repair_iteration/Math46/FL-2/1
indicates the second iteration (FL-2) of fault localization with 1st ranked suspicious statement (1).
each suspicious statement has its iteration_0/iteration_1/iteration_2.
each iteration has its bugs.csv, patches.csv and revert.csv(to avoid the duplicated patches).

Checkout the buggy program and produce a ranked list for the bug.

python3 1_localize_fault.py projectID bugID init
e.g., python3 1_localize_fault.py Chart 1 init

The result will be found under projects/Chart1/build/sfl/txt/ochiai.ranking.csv

Transform the FL to input representation

The suspicious_threshold by default configures to 0.1 and we consider at most top-50 ranked suspicious statements.

python3 2_bug_representation.py projectID bugID suspicious_threshold init
e.g., python3 2_bug_representation.py Chart 1 0.1 init

The result will be found under repair_iteration/Chart1/bugs.csv

Iterative Program Repair

python3 3_repair.py projectID bugID 
e.g., python3 3_repair.py Chart 1 

The result will be found under repair_iteration/Chart1/1,...,n, where n is the ranked position of suspucious statements

Trace patches

python3 4_trace_patches.py 

The result will be found in patches.csv