Uncertainty Wizard Versions Save

What's Changed

• :bug: Multi-GPU ensembles are possible again. Just extend the DeviceAllocatorContextManagerV2 context manager.
• :sparkles: Add return_alias_dict=True when calling predict_quantified and the return value will be a dict of quantifier aliases pointing to (prediction, confidence_or_uncertainty) tuples.

Minor fix regarding some quantifier aliases which were described in the docs, but not actually implemented.

There were some issues with recent tf versions, especially regarding experimental features. These are not fixed, but a corresponding warning is raised, explaining workarounds to users

New Feature:

• #92 Added DeepGini as a point-prediction based quantifier. This is part of our most recent issta paper .

Bugfix:

• #86 Added a missing alias to the PredictiveEntropy quantifier

Release Notes:

• Allow model independent tasks on ensembles, i.e., tasks where no model is loaded or persisted. This can typically be helpful for pre- or postprocessing tasks which have to be conducted for every model, but are independent of said model. (#50)
• A context manager for ensembles that runs only on CPU, even if GPUs are available. (#52)
• Tasks on ensemble methods (create, consume, predict, predict_quantified, ...) can now be executed on a selected subset of models. There are many use-cases for this. For example, it can be used if you need to re-train only one model in an ensemble, or if you have multiple distinct machines create atomic models for the same ensemble at the same time, using a shared drive. (#65)
• Minor updates to documentation and code comments (#51, #73)

[Refactor]

• Refactor broadcasting and aggregation logic of stochastic models into a separate class
  As a side effect, this allows broadcasting behavior to be overriden by users: Extend uwiz.models.stochastic_utils.Broadcaster oruwiz.models.stochastic_utils.DefaultBroadcaster and pass it as broadcaster=... when calling predict_quantified.

This release contains no functional changes, but various minor updates to the project description and docstrings.

Docs and Description Updates

Initial public release.

Note: This is not yet tested on tensorflow 2.4 (which has been released only a couple of days ago)