We propose GypSum, a new deep learning model that learns hybrid representations using graph neural networks and a pre-trained programming and natural language model. GypSum uses two encoders to learn from the AST-based graph and the token sequence of source code, respectively, and modifies the encoder-decoder sublayer in the Transformer's decoder to fuse the representations.

This repos is developed based on the environment of:

Python 3.7

PyTorch 1.7.0

1. How to construct graph for python and java program?

   you can use the function get_graph_from_source from proprocess/java(python)_graph_construction.py After processing the code, save the data as pkl into data folder.

2. How to run?

   ./run $GPU_ID$ $DATASET$ $TASK$ For example, if you want to train java model from scrach with cuda:0, you can just run command ./run 0 java train

Gypsum-main
├─- README.md
├── c2nl
│   ├── __init__.py
│   ├── __pycache__
│   ├── config.py
│   ├── decoders
│   ├── encoders
│   ├── eval
│   ├── inputters
│   ├── models
│   ├── modules
│   ├── objects
│   ├── tokenizers
│   ├── translator
│   └── utils
├── config
│   ├── general_config.yml
│   ├── java_xxx_xxx.yml
│   ├── ...
├── data
│   ├── java
│   └── python
├── evaluation
│   ├── bleu
│   ├── evaluate.py
│   ├── meteor
│   └── rouge
├── gypsum
│   ├── __pycache__
│   ├── data
│   ├── metor.ipynb
│   ├── model.py
│   ├── modules
│   ├── predict.py
│   ├── train.py
│   └── utils
├── modules
│   ├── __pycache__
│   └── attention_zoo.py
├── preprocess
│   ├── generate_java_graph.ipynb
│   ├── java_graph_construct.py
│   ├── python_ast.ipynb
│   └── python_graph.py
└── run

We have uploaded our dataset on google drive: Dataset For Experiment

We borrowed and modified code from DrQA, OpenNMT. We would like to expresse our gratitdue for the authors of these repositeries.