This code try to reproduce the result in this paper(https://arxiv.org/abs/1704.02798). In order to run the experienment just run:
python main.py --BBB --cuda
with posterior sharpening:
python main.py --BBB --cuda --sharpen
For the baseline result, run
python main.py --cuda
The baseline result is consistent with the paper, but the improvement of bayesian backprop is minor.