zhuyiche / torchuq

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

This repo is still under developement. Please check back in a few weeks.


TorchUQ is your one-stop solution for uncertainty quantification (UQ). At its core, TorchUQ supports various representations of uncertainty (including probability, quantile, particle, ensemble, set, etc), and

  1. converts between different representations of uncertainty

  2. adjusts any uncertainty representation so it becomes calibrated/valid/multi-accuarate.....

  3. evaluates and visualizes the quality of uncertainty quantification

TorchUQ is built on pytorch, and supports auto-differentiation for most of its functions. Current version only support a pytorch interface (i.e. all arrays have to be pytorch arrays). Support for numpy interface to come.

Why use torchuq?

  1. Torchuq is a one-stop solution. You can use one toolbox with a single consistent interface to evaluate predictions, or run popular calibration, conformal inference and online learning algorithms. Each function can be accomplished with 1-5 lines of code.

  2. Torchuq is based on pytorch, so all functions support GPU acceleration with no overhead (If a function receives GPU tensors as input, then it is automatically computed on GPU). Most functions are also end-to-end differentiable and can be incorporated into a deep learning pipeline.

  3. Torchuq includes a full set of tutorials to illustrate popular algorithms and evaluation metrics for uncertainty quantification.

  4. A large set of benchmark datasets used in recent UQ papers with a one-line interface to retrieve these datasets

  5. Torchuq is high quality. As you can see in the dev folder, there are a large number of unit tests to ensure correctness.



Language:Jupyter Notebook 96.8%Language:Python 3.2%