torch-adf 22.1.0

Assumed Density Filtering (ADF) Probabilistic Networks

torch-adf provides implementations for probabilistic PyTorch neural network layers, which are based on assumed density filtering. Assumed density filtering (ADF) is a general concept from Bayesian inference, but in the case of feed-forward neural networks that we consider here it is a way to approximately propagate a random distribution through the neural network.

The layers in this package have the same names and arguments as their corresponding PyTorch versions. We use Gaussian distributions for our ADF approximations, which are described by their means and (co-)variances. So unlike the standard PyTorch layers, each torch-adf layer takes two inputs and produces two outputs (one for the means and one for the (co-)variances).

torch-adf layers can be used exactly like the corresponding PyTorch layers within a model. However, as mentioned above, ADF layers take two inputs and produce two outputs instead of one, so it is not possible to simply mix ADF and standard layers within the same model.

from torch.nn import Sequential
from torchadf.nn import Linear

in_dim, out_dim = 64, 32
adflayer = Linear(in_dim, out_dim)
model = Sequential(adflayer)

The Overview and Examples sections of our documentation provide more realistic and complete examples.

Project Information

torch-adf is released under the MIT license, its documentation lives at Read the Docs, the code on GitHub, and the latest release can be found on PyPI. It’s tested on Python 3.6+.

If you’d like to contribute to torch-adf you’re most welcome. We have written a short guide to help you get you started!

Further Reading

Additional information on the algorithmic aspects of torch-adf can be found in the following works:

• Jochen Gast, Stefan Roth, “Lightweight Probabilistic Deep Networks”, 2018

• Jan Macdonald, Stephan Wäldchen, Sascha Hauch, Gitta Kutyniok, “A Rate-Distortion Framework for Explaining Neural Network Decisions”, 2019

Acknowledgments

During the setup of this project we were heavily influenced and inspired by the works of Hynek Schlawack and in particular his attrs package and blog posts on testing and packaing and deploying to PyPI. Thank you for sharing your experiences and insights.