This is the source code for the paper Variational Bayes for Federated Continual Learning.

FedBNN utilizes the Bayesian Neural Network in the Federated Learning setting to address a novel and realistic problem of Federated Continual Learning.

The program is based on PyTorch, an open source machine learning framework.

The Bayesian Neural Network (BNN) implementation used in this program is based on @Harry24k's project bayesian-neural-network-pytorch, and contains modification to the original implementation.

We would like to thank the community of PyTorch and the authors of torchbnn in providing easy-to-use, function-rich and well-documented open-source projects.

The licenses used in this project is attached in /licenses directory.

• Our project is subject to MIT license.
• The torchbnn library is subject to MIT license.
• The PyTorch library is subject to PyTorch license.
• The other modules may be subject to certain licenses and can be found in their project pages.

1. Install the environment. The file /fedbnn.yml provides the requirements in anaconda format. Please download and install anaconda or miniconda according to their tutorials.

2. The experiment settings are configured in configs.py as a dict. Configure if necessary.

3. Prepare the datasets. The datasets should be placed in the /data directory. Due to possible legal and ethical issues, the datasets are not included in the archive. Some of the datasets can be loaded automatically by the torchvision in our script (see the code in /fcl_data_simulator/single_dataset.py), but some of the datasets should be manually downloaded and placed in the directory.

4. Run the experiment scripts. The experiment scripts are located in /experiments directory. You can follow the instructions in the directory.

The directory structure of the project is described as follows.