Gelato

Bayesian dessert for Lasagne

Recent results in Bayesian statistics for constructing robust neural networks have proved that it is one of the best ways to deal with uncertainty, overfitting but still having good performance. Gelato will help to use bayes for neural networks. Library heavily relies on Theano, Lasagne and PyMC3.

Installation

• from github (assumes bleeding edge pymc3 installed)

  # pip install git+git://github.com/pymc-devs/pymc3.git
  pip install git+https://github.com/ferrine/gelato.git

• from source

  git clone https://github.com/ferrine/gelato
  pip install -r gelato/requirements.txt
  pip install -e gelato

Usage

I use generic approach for decorating all Lasagne at once. Thus, for using Gelato you need to replace import statements for layers only. For constructing a network you need to be the in pm.Model context environment.

Warning

• lasagne.layers.noise is not supported
• lasagne.layers.normalization is not supported (theano problems with default updates)
• functions from lasagne.layers are hidden in gelato as they use Lasagne classes. Some exceptions are done for lasagne.layers.helpers. I'll try to solve the problem generically in future.

Examples

For comprehensive example of using Gelato you can reference this notebook

Life Hack

Any spec class can be used standalone so feel free to use it everywhere

References

Charles Blundell et al: "Weight Uncertainty in Neural Networks" (arXiv preprint arXiv:1505.05424)