This repo provides the source code & data of our paper: Break-It-Fix-It: Unsupervised Learning for Program Repair (ICML 2021).

@InProceedings{yasunaga2021break,
  author =  {Michihiro Yasunaga and Percy Liang},
  title =   {Break-It-Fix-It: Unsupervised Learning for Program Repair},
  year =    {2021},  
  booktitle = {International Conference on Machine Learning (ICML)},  
}

Problem: Repair Task

Our approach: BIFI

Run the following commands to create a conda environment (assuming CUDA10.1):

conda create -n BIFI python=3.7.7
conda activate BIFI
pip install tqdm
pip install torch==1.4.0 torchvision==0.5.0
cd utils/fairseq
pip install -e .
pip install numpy==1.20.1 editdistance

Alternatively, you can use the Dockerfile in the docker folder of this repo to set up the environment.

Download all the data from here (data.zip) and unzip it (note: 67GB when compressed, 400GB when decompressed). This includes the GitHub-Python dataset, and all the processed training data and trained models associated with BIFI. If you only want the original GitHub-Python dataset, you can download it from here (data_minimal.zip; 1GB). After unzipping the data.zip, the resulting file structure will look like:

.
├── README.md
└── data/
    ├── orig_bad_code/       (GitHub-Python dataset's bad code)
    ├── orig_good_code/      (GitHub-Python dataset's good code)
    └── round0/
        ├── data_paired      (paired data used to train fixer in round0)
        └── model-fixer      (fixer trained in round0)
    ├── round1-BIFI-part1/
        ├── data_paired      (paired data used to train breaker in BIFI round1)
        └── model-breaker    (breaker trained in BIFI round1)
    ├── round1-BIFI-part2/
        ├── data_paired      (paired data used to train fixer in BIFI round1)
        └── model-fixer      (fixer trained in BIFI round1)
    ├── ...

We collected 3 million Python3 snippets from GitHub. Using the critic (Python AST parser), the code snippets are split into a set of bad code (with AST parse errors) and a set of good code (with no errors). The set of bad code is located at data/orig_bad_code/orig.bad.json and good code at data/orig_good_code/orig.good.json. Each entry of orig.bad.json or orig.good.json is a dictionary consisting of

• "code_string": raw code in the string format
• "code_toks_joined": the raw code is split into tokens by Python tokenizer, anonymized (string/number is replaced with special tokens <STRING>/<NUMBER>), and then joined by whitespace. The tokenization was done by utils/code_utils.py: tokenize_python_code()
• "anonymize_dict": mapping betweens raw string/number and <STRING>/<NUMBER> so that "code_string" can be recovered from "code_toks_joined". This recovery can be done by utils/code_utils.py: code_toks_to_code_string()
• "err_obj": type of the error caught by the critic (e.g. unbalanced parentheses, indentation error). This is only applicable to orig.bad.json.

The bad code snippets in orig.bad.json are split into 5 chunks (orig.0.bad to orig.4.bad in data/orig_bad_code/), where 3,4 is heldout as the test set and 0,1,2 is made available for BIFI training. This splitting was done by scripts/split_orig_bad_and_good.py

First, train the initial fixer by running commands in src/run-round0.py one by one. We then consider three training algorithms on top of it: BIFI (our proposed method), FixerOnly (BIFI without breaker), and BackTranslation (BT; our baseline). For each algorithm,

• BIFI: run commands in src/run-BIFI.py one by one
• FixerOnly: run commands in src/run-FixerOnly.py one by one
• BT: run commands in src/run-BT.py one by one

Below is an illustration for the case of BIFI.

run-round0.sh

export PYTHONPATH=.

#Train initial fixer on synthetic paired data
python src/c001__train_fixer.py --round_name round0 --gpu_id 0 --max_epoch 2

#Run the trained fixer on the bad code (chunk 0-4) and check the outputs by critic
python src/c003__run_fixer.py   --round_name round0 --gpu_ids '0,1,2,3,4'

#Evaluate the fixer outputs on the test set (chunk 3,4)
python src/c005__eval_fixer.py  --round_name round0

run-BIFI.sh (round 1)

#Use the fixer outputs on the bad code (chunk 0,1,2) to get new paired data (Equation 6 in the paper)
python src/c006__generate_paired_data_from_fixer.py --round_name round0 --out_round_name round1-BIFI-part1

#Train breaker on the new paired data (Equation 7 in the paper)
python src/c002__train_breaker.py --round_name round1-BIFI-part1 --gpu_id 0 --max_epoch 3

#Run the trained breaker on the good code and get new paired data (Equation 8 in the paper)
python src/c004__run_breaker.py   --round_name round1-BIFI-part1 --gpu_ids '0,1,2,3,4'
python src/c007__generate_paired_data_from_breaker.py --round_name round1-BIFI-part1 --out_round_name round1-BIFI-part2

#Train fixer on the new paired data (Equation 9 in the paper)
python src/c001__train_fixer.py --round_name round1-BIFI-part2 --gpu_id 0 --max_epoch 2 --continue_from 'data/round0/model-fixer/checkpoint.pt'

#Run the trained fixer on the bad code (chunk 0-4) and check the outputs by critic
python src/c003__run_fixer.py   --round_name round1-BIFI-part2 --gpu_ids '0,1,2,3,4'

#Evaluate the fixer outputs on the test set (chunk 3,4)
python src/c005__eval_fixer.py  --round_name round1-BIFI-part2

This is repeated similarly for round 2.