/apd_public

Code for "Adversarial Distillation of Bayesian Neural Network Posteriors" https://arxiv.org/abs/1806.10317

Primary LanguagePython

Adversarial Posterior Distillation (APD)

Requirements

  • Python 3.6.3
  • PyTorch 0.3.1.post2
  • Tensorflow 1.4.0

Environment Setup

Here is an example that shows how to set up a conda environment with the appropriate versions of the frameworks:

conda create -n apd-env python=3.6
source activate apd-env
conda install pytorch=0.3.1 cuda80 -c soumith
conda install torchvision -c pytorch
pip install -r requirements.txt

To reproduce results run the following from a CUDA enabled cluster.

MNIST fcNN1 (784-100-10)

MC-Dropout (p=0.5)

python train_new.py model/config/fc1-mnist-100-drop-50.yaml opt/config/sgd-mnist.yaml mnist-50000 --cuda --mc_dropout_passes 200

python train_new.py model/config/fc1-svhn-100-drop-50.yaml opt/config/sgd-svhn.yaml svhn-50000 --cuda --mc_dropout_passes 200

SGLD

python train_new.py model/config/fc1-mnist-100.yaml opt/config/sgld-mnist-1-1.yaml mnist-50000 --cuda

python train_new.py model/config/fc1-svhn-100.yaml opt/config/sgld-svhn-1-1.yaml svhn-50000 --cuda

APD

python gan.py fc1-mnist-100-X-sgld-mnist-1-1-X-mnist-50000@DATE opt/gan-config/gan1.yaml

python gan.py fc1-svhn-100-X-sgld-svhn-1-1-X-svhn-50000@DATE opt/gan-config/gan1.yaml