wanyao1992 / code_summarization_public

source code for 'Improving automatic source code summarization via deep reinforcement learning'

Note: This repository is no longer maintained now. If you are interested in the deep learning for program analysis (e.g., code summarization, code retrieval, code completion and type inference), please refer to our new project NaturalCC (https://github.com/CGCL-codes/naturalcc).

Requirement

This repos is developed based on the environment of:

• Python 2.7
• PyTorch 0.2

Data folder structure

/media/BACKUP/ghproj_d/code_summarization/github-python/ is the folder to save all the data in this project, please replace it to your own folder. The data files are organized as follows in my computer:

|- /media/BACKUP/ghproj_d/code_summarization/github-python

|--original (used to save the raw data)

|----data_ps.declbodies data_ps.descriptions

|--processed (used to save the preprocessed data)

|----all.code all.comment

|--result (used to save the results)

|--train (get the data files before training)

You need to get these files before you starting to train our model. Here I put the original folder in the dataset foler of this project. You'd better copy them to your own folder.

Data preprocess

cd script/github

python python_process.py -train_portion 0.6 -dev_portion 0.2 > log.python_process

Training

Back to the projector folder

cd ../..

Get the data for training

python run.py preprocess

Training

python run.py train_a2c 10 30 10 hybrid 1 0

Testing

python run.py test_a2c hybrid 1 0

TODO

• To build the AST, on the data preprocessing, I parse the AST into a json and then parse the json into AST on training. This kind of approach is not elegant.
• On training, I don't know how to batchify the ASTs, so I have to put the ASTs into a list and encode them one by one. It's unefficient, making the training of one epoch takes about 2-3 hours. Please let me know if you have a better way to accelerate this process.
• On the encoder side, I am working on applying Tree-CNN and GraphCNN to represent the code in a better way.
• On the decoder side, GAN network will also be considered for the code summarization task.

Acknowledgement

This repos is based on https://github.com/khanhptnk/bandit-nmt

Please cite our paper if you use this repos.

Bibtex:
@Inproceedings{wan2018improving,
title={Improving automatic source code summarization via deep reinforcement learning},
author={Wan, Yao and Zhao, Zhou and Yang, Min and Xu, Guandong and Ying, Haochao and Wu, Jian and Yu, Philip S},
booktitle={Proceedings of the 33rd ACM/IEEE International Conference on Automated Software Engineering}
pages={397--407},
year={2018},
organization={ACM}
}