This is the source code repository for the paper CoreGen: Contextualized Code Representation Learning for Commit Message Generation.

If there is any suggestion or question, feel free to fire an issue here.

All required packages and versions can be found in the environment configuration file environment.yml, or you may simply build an identical conda environment like this:

conda env create -f environment.yml
conda activate coregen

Download the data here and unzip the dataset into the folder.

Simply run the script ./run_all.sh with layer number and head number specified. You may easily adjust other parameters, such as batch size, epoch number and mask rate, by modifying the script file. The results reported in paper can be reproduced by setting the layer number to 2 and head number to 6 while keeping other parameter settings in the script unchanged.

./run_all.sh LAYER_NUM HEAD_NUM

Compute the BLUE-4, METEOR and ROUGE score by running the script ./score.sh with the reference file path and the inference result path specified.

./score.sh REFERENCE PREDICTION

CoreGen significantly outperforms previous state-of-the-art models with at least 28.18% improvement on BLEU-4 score.

In our experiment comparing vanilla Transformer and CoreGen on model’s convergence along the fine-tuning procedure, CoreGen also converges faster to achieve equivalent generation quality as the vanilla Transformer model at 25 training epochs ahead.

More details about the performance of CoreGen are presented and analyzed in the paper.