/variational-dropout

TensorFlow implementation of the method from Variational Dropout Sparsifies Deep Neural Networks, Molchanov et al. (2017)

Primary LanguagePythonGNU General Public License v3.0GPL-3.0

Variational Dropout Sparsifies Deep Neural Networks

Molchanov et al. (2017)

Implementation for TensorFlow, based on the Theano/Lasagne version by the authors here. You can read the original paper here.

Requirements

  • TensorFlow 1.1
  • Python 2.7
  • Numpy
  • Scikit-learn
  • Keras

This is a work in progress. The repo contains fully connected and convolutional layers with variational dropout.

Currently runs a simple CNN or larger VGG-like network on MNIST, CIFAR-10 or CIFAR-100.