[ICML 2021] Break-It-Fix-It: Learning to Repair Programs from Unlabeled Data

Overview

Break-It-Fix-It: Learning to Repair Programs from Unlabeled Data

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

0. Dependencies

Specifically, 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 numpy==1.20.1 editdistance

1. Download Data

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)
    ├── ...

About the GitHub-Python dataset

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

2. Training and Evaluation

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.

You might also like...
[ICML 2021] “ Self-Damaging Contrastive Learning”, Ziyu Jiang, Tianlong Chen, Bobak Mortazavi, Zhangyang Wang
[ICML 2021] “ Self-Damaging Contrastive Learning”, Ziyu Jiang, Tianlong Chen, Bobak Mortazavi, Zhangyang Wang

Self-Damaging Contrastive Learning Introduction The recent breakthrough achieved by contrastive learning accelerates the pace for deploying unsupervis

[ICML 2021]
[ICML 2021] "Graph Contrastive Learning Automated" by Yuning You, Tianlong Chen, Yang Shen, Zhangyang Wang

Graph Contrastive Learning Automated PyTorch implementation for Graph Contrastive Learning Automated [talk] [poster] [appendix] Yuning You, Tianlong C

Implementation of Learning Gradient Fields for Molecular Conformation Generation (ICML 2021).
Implementation of Learning Gradient Fields for Molecular Conformation Generation (ICML 2021).

[PDF] | [Slides] The official implementation of Learning Gradient Fields for Molecular Conformation Generation (ICML 2021 Long talk) Installation Inst

Implementation of Self-supervised Graph-level Representation Learning with Local and Global Structure (ICML 2021).
Implementation of Self-supervised Graph-level Representation Learning with Local and Global Structure (ICML 2021).

Self-supervised Graph-level Representation Learning with Local and Global Structure Introduction This project is an implementation of ``Self-supervise

This repo contains the implementation of the algorithm proposed in Off-Belief Learning, ICML 2021.

Off-Belief Learning Introduction This repo contains the implementation of the algorithm proposed in Off-Belief Learning, ICML 2021. Environment Setup

Space robot - (Course Project) Using the space robot to capture the target satellite that is disabled and spinning, then stabilize and fix it up

Space robot - (Course Project) Using the space robot to capture the target satellite that is disabled and spinning, then stabilize and fix it up

Code for the ICML 2021 paper:
Code for the ICML 2021 paper: "ViLT: Vision-and-Language Transformer Without Convolution or Region Supervision"

ViLT Code for the paper: "ViLT: Vision-and-Language Transformer Without Convolution or Region Supervision" Install pip install -r requirements.txt pip

Code for the ICML 2021 paper:
Code for the ICML 2021 paper: "ViLT: Vision-and-Language Transformer Without Convolution or Region Supervision"

ViLT Code for the paper: "ViLT: Vision-and-Language Transformer Without Convolution or Region Supervision" Install pip install -r requirements.txt pip

[ICML 2021, Long Talk] Delving into Deep Imbalanced Regression
[ICML 2021, Long Talk] Delving into Deep Imbalanced Regression

Delving into Deep Imbalanced Regression This repository contains the implementation code for paper: Delving into Deep Imbalanced Regression Yuzhe Yang

Comments
Owner
Michihiro Yasunaga
PhD Student in Computer Science
Michihiro Yasunaga
code for paper "Not All Unlabeled Data are Equal: Learning to Weight Data in Semi-supervised Learning" by Zhongzheng Ren*, Raymond A. Yeh*, Alexander G. Schwing.

Not All Unlabeled Data are Equal: Learning to Weight Data in Semi-supervised Learning Overview This code is for paper: Not All Unlabeled Data are Equa

Jason Ren 22 Nov 23, 2022
Code for ICML 2021 paper: How could Neural Networks understand Programs?

OSCAR This repository contains the source code of our ICML 2021 paper How could Neural Networks understand Programs?. Environment Run following comman

Dinglan Peng 115 Dec 17, 2022
Source code for our paper "Learning to Break Deep Perceptual Hashing: The Use Case NeuralHash"

Learning to Break Deep Perceptual Hashing: The Use Case NeuralHash Abstract: Apple recently revealed its deep perceptual hashing system NeuralHash to

ml-research@TUDarmstadt 11 Dec 3, 2022
This repo is a C++ version of yolov5_deepsort_tensorrt. Packing all C++ programs into .so files, using Python script to call C++ programs further.

yolov5_deepsort_tensorrt_cpp Introduction This repo is a C++ version of yolov5_deepsort_tensorrt. And packing all C++ programs into .so files, using P

null 41 Dec 27, 2022
[ICCV2021] Learning to Track Objects from Unlabeled Videos

Unsupervised Single Object Tracking (USOT) ?? Learning to Track Objects from Unlabeled Videos Jilai Zheng, Chao Ma, Houwen Peng and Xiaokang Yang 2021

null 53 Dec 28, 2022
Code release for "Self-Tuning for Data-Efficient Deep Learning" (ICML 2021)

Self-Tuning for Data-Efficient Deep Learning This repository contains the implementation code for paper: Self-Tuning for Data-Efficient Deep Learning

THUML @ Tsinghua University 101 Dec 11, 2022
Ranking Models in Unlabeled New Environments (iccv21)

Ranking Models in Unlabeled New Environments Prerequisites This code uses the following libraries Python 3.7 NumPy PyTorch 1.7.0 + torchivision 0.8.1

null 14 Dec 17, 2021
Code for the ICML 2021 paper "Bridging Multi-Task Learning and Meta-Learning: Towards Efficient Training and Effective Adaptation", Haoxiang Wang, Han Zhao, Bo Li.

Bridging Multi-Task Learning and Meta-Learning Code for the ICML 2021 paper "Bridging Multi-Task Learning and Meta-Learning: Towards Efficient Trainin

AI Secure 57 Dec 15, 2022
Neural-Pull: Learning Signed Distance Functions from Point Clouds by Learning to Pull Space onto Surfaces(ICML 2021)

Neural-Pull: Learning Signed Distance Functions from Point Clouds by Learning to Pull Space onto Surfaces(ICML 2021) This repository contains the code

null 149 Dec 15, 2022
[ICML 2021] DouZero: Mastering DouDizhu with Self-Play Deep Reinforcement Learning | 斗地主AI

[ICML 2021] DouZero: Mastering DouDizhu with Self-Play Deep Reinforcement Learning DouZero is a reinforcement learning framework for DouDizhu (斗地主), t

Kwai Inc. 3.1k Jan 4, 2023