Personalized Federated Learning of Probabilistic Models: A PAC-Bayesian Approach

This repository provides source code corresponding the paper Personalized Federated Learning of Probabilistic Models: A PAC-Bayesian Approach.

A demo using the synthetic toy dataset with our method, PAC-PFL, and all baseline methods is provided in experiments/Toy_dataset.

The server package contains the code ran by the server for different methods.

The baselines directory contains the implementations of MAML, MTL, and pooled baselines for all versions of the PV dataset. PAC-PFL, Vanilla, and pFedGP have a similar implementation, and are evaluated in experiments/PV/train_full_gp. It is possible to first train the prior mean, using experiments/PV/train_full_gp, and then train a zero-mean GP on the residuals, by running experiments/PV/train_only_kernel.

New clients are simulated and evaluated by running experiments/PV/new_clients.

Structure search for all NNs and hyper-parameter tuning is performed using the functions provided in the utils folder.

Installation

To install full dependencies, please run in the main directory of this repository

pip install -r requirements.txt

Usage

The GPU is disabled by default. To enable the GPU, set disable_gpu = False in config.py.

The demo.ipynb notebook can be ran from experiments/Toy_dataset without downloading further files.

To conduct the PV experiments, please follow these steps:

• 1. Download the dataset files from the PV_dataset folder in the provided Google Drive link.
• 2. Unzip the downloaded files and place them in the third_party/Synthetic_PV_Profiles/saved_results folder. Ensure that there are four files in the folder, each ending with _env.
• 3. Download the pre-trained models for the PV-EW (150) and PV-EW (610) datasets from the pre_trained folder in the provided Google Drive link.
• 4. Unzip the downloaded models and place them in the experiments/PV/saved_results folder. The folders PV_BiModal and PV_BiModal_NewClients should appear directly in the experiments/PV/saved_results folder.
• 5. Run the experiments/PV/paper_figs.ipynb file. Please note that executing this file may take some time. By following these instructions, you should be able to reproduce the boxplots shown in Figure 2 of the paper.