A Neural Architecture for Generating Natural Language Descriptions from Source Code Changes

• Requirements

  • Torch, Cutorch (http://torch.ch/docs/getting-started.html)
  • Python packages unidiff, pygments: pip install unidiff pygments

• Setup environment

  1. Clone this repositoty: cd ~ git clone https://github.com/epochx/commitgen-dev.git
  2. Create data path: mkdir ~/data/preprocessing
  3. Export env variable: export env WORK_DIR=~/data (without trailing slash!)

• Download our paper data:

  1. Get the raw commit data used in our paper from https://osf.io/67kyc/?view_only=ad588fe5d1a14dd795553fb4951b5bf9 (click on "OSF Storage" and then on "Download as zip".) Unzip the file where convenient.
  2. Unzip the desired dataset zip and move the resulting folder to ~/data.

• Pre-process data

  1. Parse and filter commits and messages: cd ~/commitgen python ./preprocess.py FOLDER_NAME --language LANGUAGE, where FOLDER_NAME is the name of the folder from the previous step. Add the '--atomic' flag to keep only atomic commits. This will generate a pre-processed version of the dataset in a pickle file in ~/data/preprocessing. Try python ./preprocess.py --help for more details on additional pre-processing parameters.
  2. Generate training data: cd ~/commitgen ./buildData.sh PICKLE_FILE_NAME LANGUAGE (PICKLE_FILE_NAME with no .pickle).

• Train the model 1.- Run the model cd ~/commitgen ./run.sh PICKLE_FILE_NAME LANGUAGE (PICKLE_FILE_NAME with no .pickle)

You can also dowload additional github project data by using our crawler do cd ~/commitgen and run python crawl_commits.py --help for more details on how to do it.