Cambridge-Uni-MLMI-Advanced-ML-21-22-G7/BBB

BBB

Environment Setup/Installation

Create a fresh Python 3.9 environment using venv, pyenv or conda (as preferred). Then, execute setup.sh to install the required python libraries and versions.

For mac users, the following would be appropriate.

brew install python@3.9
python3.9 -m venv .venv
source .venv/bin/activate
./setup.sh

M1 Torch Issue

If running on M1 you will need to ensure the following library versions are being used. These should already be set in requirements.txt.

pip install --upgrade torch==1.9.0
pip install --upgrade torchvision==0.10.0

brew install hdf5
pip3 install --no-binary=h5py h5py
export HDF5_DIR=/opt/homebrew/opt/hdf5

python3 -m pip install tensorflow-macos==2.8.0
python3 -m pip install tensorflow-metal

Execution

Once installed bbb can be run from the command line:

bbb [MODEL TYPE] [-d]

where the -d command is used to indicate whether to run the model deterministically (i.e., non-Bayesian approaches.).

For example, to run classification using BBB use the command:

bbb class

Whereas, classification can be run deterministically using:

bbb class -d

Class Inheritance

The following diagram outlines the class inheritance structure of the implemented models.

Citations

The following repositories were used as sources and for inspiration when implementing this project, and when debugging:

• Bayesian Neural Networks (1)
• Bayesian Neural Networks (2)
• Weight Uncertainty
• PyTorch Bayesian CNN
• Weight Uncertainty in Neural Networks