Radial Bayesian Neural Networks

Code to accompany the paper Radial Bayesian Neural Networks: Beyond Discrete Support in Large-Scale Bayesian Deep Learning.

Using the code

Implementations of pytorch layers supporting stochastic variational inference with Radial BNNs can be found in the folder radial_layers/variational_bayes.py. These layers are compatible with the Elbo loss implemented in radial_layers/loss.py. To drop this code into yours, you need only place the radial_layers folder into your pytorch code and instantiate layer objects imported from variational_bayes.py.

Layers inherit from the SVI_Base class, which instantiates stochastic variational inference weights and biases using the reparameterization trick. At the moment, the available layers are SVI_Linear, SVI_Conv2D, SVIMaxPool2D, SVIGlobalMaxPool2D, andSVIAverageMaxPool2D. Further contributions are very welcome.

Requirements

This version of the code has been tested on pytorch version 1.4.0 with python 3.8.1 (though earlier iterations of the code worked in earlier versions of both). By default the code expects a cuda device, but can be modified to run on cpu.

Example usage

This code-base provides an example usage. train.py loads configuration files from configs and builds trainer, model, loss, and optimizer objects based on the configuration. For example, you can see config_cnn_mnist.json for an example that loads and trains a CNN on MNIST.

You can easily change the code to run standard mean-field variational inference with a multivariate Gaussian instead of Radial BNNs by changing the config file to set arch["args"]["variational_distribution"] = "gaussian" instead of radial.

For further details, issues are welcome, as are questions emailed to sebastian.farquhar@cs.ox.ac.uk.

Citing this code

If you use this code, please cite us as:

@article{farquhar_radial_2020,
    author = {Sebastian Farquhar and Michael Osborne and Yarin Gal},
    title = {Radial Bayesian Neural Networks: Beyond Discrete Support in Large-Scale Bayesian Deep Learning},
    journal = {Proceedings of the 23rtd International Conference on Artificial Intelligence and Statistics},
    year = {2020}
}

Acknowledgements

Much of the 'boilerplate' code of the example derives from the https://github.com/victoresque/pytorch-template