EnRefiner

This repository is the code of APSEC 2023 "Improving Code Refinement for Code Review Via Input Reconstruction and Ensemble Learning".

Dependency

• Pytorch 1.7.1
• tensorboard 2.4.1
• transformers 4.6.1
• tree-sitter 0.2.2

Download

CodeT5 Pre-trained checkpoints: https://console.cloud.google.com/storage/browser/sfr-codet5-data-research/pretrained_models

Train prompt-augmented model A

cd prompt_augmented_A\sh

python run_exp.py --model_tag codet5_base --task translate --sub_task java-cs

NOTE: modify WORKDIR in exp_with_args.sh

Train prompt-augmented model B

cd prompt_augmented_B\sh

python run_exp.py --model_tag codet5_base --task translate --sub_task java-cs

NOTE: modify WORKDIR in exp_with_args.sh

Train retrieval-augmented model

cd retrieval_augmented\sh

python run_exp.py --model_tag codet5_base --task translate --sub_task java-cs

NOTE: modify WORKDIR in exp_with_args.sh

Ensemble

use vote.py in dir output