DeepM

A deep learning-based approach to measuring code maintainability

This repository contains materials on our paper "Measuring Code Maintainability with Deep Neural Networks".

In /Datasets/, we provide:

• two datasets:
  • TrainingData: an automatically constructed large dataset, which contains 1,394,514 Java classes.
  • TestingData: an manually constructed small dataset, which contains 240 Java classes.

In /Parser/, we provide:

• source code (written in Java) that is used to extract featurs from Java classes.
  • Based on the code, you can extract features for a single Java class or a directory that contains a number of Java classes quickly because of multithreading.

In /Model/, we provide:

• source code (/Model/Model/*.py):

  • /Model/Model/GenerateDataset.py generates TFRecords file in TensorFlow based on the training data.
  • /Model/Model/TrainOrTest.py trains and tests our deep learning-based model.
  • /Model/Model/DeepM.py computes the maintainability index of a Java class using the trained model.
  • /Model/Model/Model.py contains the model of DeepM, and models for ablation study.

• baselines

  • /Model/CHC/contains the code for the baseline CHC, which is slightly modified from KEEL for evaluation
  • /Model/LRR/contains the code for the baseline LRR, which is slightly modified from KEEL for evaluation

• a trained model (in /Model/Model/all/)

• /Model/requirements.txt, which describes the requirements of running our code in /Model/Model/.

• /Model/runDeepM.sh, which computes the maintainability indexes of Java classes in a directory using the traned model.

  • usage:

  sh runDeepM.sh [The directory containing /Model/Model/DeepM.py] [The directory containing Java classes] [The file saving results]

In /Results, we provide:

• the maintainability indexes (computed by the trained model) of Java classes in the testing data

Contact

If you have questions, please contact me directly: ymhu@bit.edu.cn. Thank you!