Copyright (c) 2018-2019 Department of Software Engineering, University of Szeged, Hungary.
DeepBugHunter is an Python framework whose goal is to make it easier to experiment with deep learning and compare its capabilities with more "traditional" machine learning algorithms. The name comes from its original purpose (testing deep neural networks in combination with static source code metrics to predict bugs) but its design is intentionally more general and extensible so it can be a helpful resource for anyone looking for a higher abstraction level tool to perform machine learning studies and model building.
The user should prepare a Python 3 environment (3.6 and up) with the following packages installed:
- numpy 1.14.3 and up
- scikit-learn 0.19.2 and up
- tensorflow 1.8.0 and up
- pandas 0.22.0 and up
Running DeepBugHunter consists of calling the main executable (dbh.py) with a previously cleaned and valid CSV dataset and a list of machine learning strategies to try on it. After an optional preprocessing phase, the script automatically partitions the input dataset into training, development, and testing sets, applies the strategies, and calculates a score (we concentrated on F-measure, but others are easy to add) based on a cross validation. Most current strategies are thin wrappers around existing sklearn algorithms, while sdnnc and cdnnc are the simple and complex implementations of a Deep Neural Network (DNN) using Tensorflow.
In order to not duplicate the list of command line arguments (thereby avoiding the danger of getting out of sync with the source code), we rely instead on the verbose and human-readable nature of the argparse Python module and refer the reader to the "Arguments" section of dbh.py, listing the possible switches relating to the overall infrastructure.
We additionally note that the second half of the --strategy switch expects a whole other set of independent arguments corresponding to the machine learning strategy selected in the first half. These arguments will be parsed using the same argparse module, only one layer lower. Their definition can be found in the beginning of the individual strategy files withing the strategies folder.
If you would like to contribute, contact Rudolf Ferenc and Dénes Bán, and we'll figure something out. PRs welcome.
DeepBugHunter is licensed under the Apache License 2.0 (see the LICENSE file).
If you find this tool useful, please cite the following article and drop us a message.
Rudolf Ferenc, Dénes Bán, Tamás Grósz, and Tibor Gyimóthy:
Deep learning in static, metric-based bug prediction