Pratical Sampling-based Bayesian Inference for multimodal distribution

Pytorch implementation of "Pratical Sampling-based Bayesian Inference for multimodal distribution"

Prerequisites

• Python 3.6
• Pytorch 0.4.0
• tensorboardX 1.6

Usage

Toy Example

There are two implementation, one is for toy example and the other is for training model with MNIST dataset.
for toy example, you can follow toy_test.ipynb notebook file.
Our algorithm efficiently searches modes that are far from each other. The result of the toy example can be seen below

MNIST Example

Train

for training model with MNIST dataset, you can run the code by

python main_mnist.py --gpu_number=0 --lr=1e-2 --bs=200 --threshold=0.333 --noise_pow=-0.25

tensorboard can be seen in ./runs/lr0.01_bs200_th0.333_pow-0.25/log/ directory.

Test

you can see the experiment with the 60 saved models in experiment_mnist.ipynb notebook file.
our model can express uncertainty when tested in different dataset such as fashionMNIST and notMNIST.

MNIST

fashionMNIST

notMNIST