Uncertainty wizard is a plugin on top of tensorflow.keras,
allowing to easily and efficiently create uncertainty-aware deep neural networks:
- Plain Keras Syntax: Use the layers and APIs you know and love.
- Conversion from keras: Convert existing keras models into uncertainty aware models.
- Smart Randomness: Use the same model for point predictions and sampling based inference.
- Fast ensembles: Train and evaluate deep ensembles lazily loaded and using parallel processing.
- Super easy setup: Pip installable. Only tensorflow as dependency.
Installation
It's as easy as pip install uncertainty-wizard
Requirements
uncertainty-wizard is tested on python 3.8 and a recent tensorflow versions. Earlier versions (python 3.6+ and tensorflow 2.3+) should mostly work as well, but may require some mild tweaks.
Note: There's currently a known issue regarding multi-gpu usage and tensorflow>=2.10. For using Multi-Device GPU config, please
fall back to tensorflow==2.9.0 and stay tuned for changes in issue #75.
Documentation
Our documentation is deployed to uncertainty-wizard.readthedocs.io. In addition, as uncertainty wizard has a 100% docstring coverage on public method and classes, your IDE will be able to provide you with a good amount of docs out of the box.
You may also want to check out the technical tool paper (preprint),
describing uncertainty wizard functionality and api as of version v0.1.0.
Examples
A set of small and easy examples, perfect to get started can be found in the models user guide and the quantifiers user guide. Larger and examples are also provided - and you can run them in colab right away. You can find them here: Jupyter examples.
Authors and Papers
Uncertainty wizard was developed by Michael Weiss and Paolo Tonella at USI (Lugano, Switzerland). If you use it for your research, please cite these papers:
@inproceedings{Weiss2021FailSafe,
title={Fail-safe execution of deep learning based systems through uncertainty monitoring},
author={Weiss, Michael and Tonella, Paolo},
booktitle={2021 14th IEEE Conference on Software Testing, Verification and Validation (ICST)},
pages={24--35},
year={2021},
organization={IEEE}
}
@inproceedings{Weiss2021UncertaintyWizard,
title={Uncertainty-wizard: Fast and user-friendly neural network uncertainty quantification},
author={Weiss, Michael and Tonella, Paolo},
booktitle={2021 14th IEEE Conference on Software Testing, Verification and Validation (ICST)},
pages={436--441},
year={2021},
organization={IEEE}
}
The first paper (preprint) provides an empricial study comparing the approaches implemented in uncertainty wizard, and a list of lessons learned useful for reasearchers working with uncertainty wizard. The second paper (preprint) is a technical tool paper, providing a more detailed discussion of uncertainty wizards api and implementation.
References to the original work introducing the techniques implemented in uncertainty wizard are provided in the papers listed above.
Contributing
Issues and PRs are welcome! Before investing a lot of time for a PR, please open an issue first, describing your contribution. This way, we can make sure that the contribution fits well into this repository. We also mark issues which are great to start contributing as as good first issues. If you want to implement an existing issue, don't forget to comment on it s.t. everyone knows that you are working on it.