This repository is the official implementation of On the Expressiveness of Approximate Inference in Bayesian Neural Networks. The code in the repository was jointly written by Andrew Y. K. Foong and David R. Burt.
To install requirements:
python setup.py install
Additionally, in order to run the active learning experiments, you will need to install Bayesian benchmarks by cloning the git repository and running
python setup.py install
in the root of the bayesian_benchmarks repository.
To recreate the experiments from the paper:
To minimise squared loss in experiments/approximator run:
bash run_approximator.sh
in order to recreate figures 2 and 4 in the main text.
To run the 2d experiments with 1HL networks (Figure 3) in experiments/2d_experiments run:
bash run_2d.sh
To run the 2d experiments with deeper networks and generate the boxplots (Figure 5 in the main text, Figures 9 and 10 in the supplement) in experiments/2d_experiments run:
bash run_box_plots.sh
After which boxplots can be generated using python box_plots.py (run python box_plots.py --help to view syntax for arguments to be passed to the plotting script).
Samples from HMC, generated using pyro are provided in experiments/2d_experiments/HMC.
To run the active learning experiments (Table 1 and Figure 6 in the main body, Figures 12-14 in the supplement) in experiments/active_learning run (this will take a while!):
bash active.sh
After this is complete, the t-sne plots (Figs 6, Figs 12-14 supplement) can be made by running:
bash make_tsne_plots.sh
and the results in table 1 can be collected by running,
python collect_results.py
Use python collect_results.py --help to view syntax for arguments to be passed to this script.
To run experiments with inputs generated in pairs of clusters on a 5-dimensional sphere (Figures 6 and 7 in the supplement), in experiments/random_clusters run:
bash run_random_clusters.sh