SEbox4DL : A Modular Software Engineering Toolbox for Deep Learning Models

Deep learning (DL) models are widely used in software applications. Novel DL models and datasets are published from time to time. Developers may also tempt to apply new software engineering (SE) techniques on their DL models. However, no existing work supports the applications of software testing and debugging techniques on new DL models and their datasets without modifying the code. Developers should manually write code to glue every combination of models, datasets, and SE technique and chain them together.

We propose SEbox4DL, a novel and modular toolbox that automatically integrates models, datasets, and SE techniques into SE pipelines seen in developing DL models. SEbox4DL exemplifies six SE pipelines and can be extended with ease. Each user-defined task in the pipelines is to implement a SE technique within a function with a unified interface so that the whole design of SEbox4DL is generic, modular, and extensible. We have implemented several SE techniques as user-defined tasks to make SEbox4DL off-the-shelf. Our experiments demonstrate that SEbox4DL can simplify the applications of software testing and repair techniques on the latest or popular DL models and datasets.