Predicting source code bugs using graph neural networks

About

This repository contains the code for training the graph neural network used in my master's thesis Predicting source code bugs using graph neural networks. The thesis itself as well as the dataset are available under Releases.

Note

The model does not train successfully on the problem, only reaching accuracies barely above randomness. For a discussion of this, I recommend reading my thesis. The code and dataset are provided in the context of the thesis for future research.

Usage

Installation

This project requires Poetry and a GPU with CUDA setup for TensorFlow is recommended. The version of TensorFlow used in this project works best with Python 3.7. Upgrading Python or TensorFlow may break things.

Run poetry install to install the projects Python dependencies.

Training

The graph neural network can be trained by running the command python main.py -m [MONGO_DB_URI] in the poetry environment. The program requires a MongoDB instance running at MONGO_DB_URI. Graph samples are expected to be in a collection bugFixSamples in a database data. The dataset provided in the 'Releases' section can be imported with the command:

gunzip -c bug_graph_samples.json.gz | mongoimport --uri=[MONGO_DB_URI] --db data --collection bugFixSamples