Afterglow Documentation¶
Afterglow is an uncertainty estimation extension library for PyTorch.
Developed at GSK.ai, it allows users to equip their deep learning models with state-of-the-art uncertainty estimation with minimal effort and minimal changes to their codebase. With afterglow, the following line will change code that trains a normal neural network into code that trains an uncertainty-estimating model:
enable_swag(model, start_iteration=100, update_period_in_iters=10, max_cols=20)
The core method used by afterglow, known as SWAG, has been shown to produce high-quality uncertainty estimates across a range of models and tasks. It works by keeping track of the model’s parameters during training, and so has minimal overhead in terms of time complexity.
The package has been tested against a wide variety of models and supports distributed training.
To get up-and-running with afterglow, see the quickstart tutorial. We also provide a computer vision example application here.
Contents
Package Reference