• Documentation
• Tutorial

model = CalibrationFnCall( UncertaityFnCall( <keras_model> ) )

Features

Welcome to WUT?! a wrapper for uncertainty in tensorflow. This is a library for uncertainty quantification in deep neural networks implemented in TensorFlow/Keras. Networks can be wrapped in one of the WUT? classes in a single line of code. For instance,

from WUT.Ensemble import Ensemble
model = Ensemble(<keras_model>)

You can estimate both the mean and standard deviation for test inputs.

Available uncertainty quantification models are,

Model	Type of Uncertainty	Compatible with
Ensemble	Epistemic	Any network
MC Dropout	Epistemic	Any network
Stochastic Variational Inference	Epistemic and Aleatoric	Fully connected layers only
Variance Networks	Aleatoric	Fully connected layers only
Mixture of Gaussians/Mixture Density Networks	Aleatoric	Last layer

Requirements:

pip install tensorflow (tested with v.2.4.1, should work for >= v.2.2)

pip install tensorflow-probability --upgrade(tested with v.0.12.2)

pip install keras-mdn-layer

This library has a dependency on this repo. Shoutout to compercussion on github for making the mdn layer, which is extremely good.

For developers:

We welcome any contributions. Please make a pull request. Documents are generated using pdoc3. To install pdoc3 pip install pdoc3. To update the documentation pdoc --html WUT --output-dir docs. You might need to delete the subfolder WUT/docs/WUT.

Maintainers

This library is maintained by Chris Healy and Ransalu Senanayake.