clyly / ccrep

The replication package of "CCRep: Learning Code Change Representations via Pre-Trained Code Model and Query Back""

This is the replication package for CCRep model titled "CCRep: Learning Code Change Representations via Pre-Trained Code Model and Query Back" including both code and datasets.

Requirements

Python Package Dependencies

You should al least install following packages to run our code:

• PyTorch: 1.8.0+cu111
• allennlp: 2.8.0
• allennlp_models: 2.8.0
• nltk: 3.5
• NumPy: 1.22.2
• jsonnet: 0.18.0
• ...

The full list of dependencies is listed in requirements.txt.

Run CCRep

CCRep is evaluated on three different code-change-related tasks:

• Commit Message Generation (CMG)
• Automated Patch Correctness Assessment (APCA)
• Just-in-time Defect Prediction (JIT-DP)

Data Preparation

Due to size limit, we archive our data in the Google Drive and you can download the data from this link: CCRep-data.zip. Unzip this file and move the data folder to the root of this project to finish preparation of data.

Before Running

To ensure some scripts will work well, you have to do two things first:

1. Open "base_global.py" and check the path of Python interpreter. If you are not using the default "python", you should configure the right python interpreter path here.
2. Make sure you are running all the code at the root directory of the CCRep project, this is important.

APCA task

Execute follow command at the root of the project to run the apca task:

python tasks/apca/apca_cv_train_helper.py -model {token/line/hybrid} -dataset {Small/Large} -cuda {(your cuda device id)}

• model: Using which variant of CCRep: "token", "line" or "hybrid"
• dataset: Run CCRep on which dataset: "Large" or "Small" (case-sensitive!)
• cuda: ID of CUDA device you use

The script will automatically do cross-validation training, testing and report the final performance.

CMG task

Execute following command at the root of the project to run the cmg task:

python tasks/cmg/cmg_train_from_config.py -dataset {corec/fira} -model {token/line/hybrid} -cuda {(your cuda device id)}

• dataset: Running on which dataset: "fira" or "corec"
• model: Using which variant of CCRep: "token", "line" or "hybrid"
• cuda: ID of CUDA device you use

The script will automatically do training, validation and testing, and report the final performance.

JIT-DP task

Execute following command at the root of the project to run the jit-dp task:

python tasks/jitdp/jitdp_train_from_config.py -model {token/line/hybrid} -project {(which project to run)} -cuda {(your cuda device id)}

• model: Using which variant of CCRep: "token", "line" or "hybrid"
• project: Run CCRep on which project: "gerrit", "jdt", "go", "platform", "openstack"
• cuda: ID of CUDA device you use

Also, this script will automatically do training, validation and testing, and report the final performance.

Citation

If you use this repository, please consider citing our paper:

@inproceedings{liu2023ccrep,
  title={CCRep: Learning Code Change Representations via Pre-Trained Code Model and Query Back},
  author={Liu, Zhongxin and Tang, Zhijie and Xia, Xin and Yang, Xiaohu},
  booktitle={Proceedings of the 2023 IEEE/ACM 45th International Conference on Software Engineering},
  pages={1--13},
  year={2023}
}

QA

1. • Q: My GPU memory is not enough to run the code and always encounter "CUDA out of memory" error:
   • A:
     • If you are encountering this problem during training: Try to modify the "batch_size" in allennlp config file to solve it. For example, if you are running jitdp task using "token" model on "gerrit" project, you should open tasks/jitdp/configs/token/gerrit_train.jsonnet and decrease the data_loader.batch_size. However, to keep the batch_size consistent with us, you should also modify trainer.num_gradient_accumulation_steps. The real_batch_size = batch_size * num_gradient_accumulation_steps, thus when you decrese data_loader.batch_size, you should correspondingly increase trainer.num_gradient_accumulation_steps.
     • If you are encountering this problem during testing: Try to decrease the batch_size of the testing script (e.g., cmg_predict_to_file.py, apca_evaluate.py and jitdp_evaluate.py). It uses default batch_size=32, which may be somehow large for some models.
2. • Q: "FileNotFoundError" during running code.
   • A: First make sure you are running all the commands at the root of this project. Then try to explore the path that file located.
3. • Q: Where is the data and my dumped models?
   • A: In the data and models folders in the root of the project directory.

Note

• For the CoReC dataset of CMG, our code reads the raw data provided by NNGen and does filtering during reading them by dataset reader, instead of directly removing filtered items from this dataset in-place.