afl-bench

Test-bench to test aggregation strategies for asynchronous federated learning

Development installation

1. Create and import a conda environconda with required dependencies: conda env create -f env.yaml --name <YOUR_ENV_NAME>
2. Activate new conda environment: conda activate <YOUR_ENV_NAME>
3. Install this repo package as a local module symbolically (assuming cwd is this repository): pip install -e .Set up wandb via running wandb login.

TODOs

1. Should an agent pull immediately from the global model if it has trained the same version? I think not?
2. Supporting non-IID data-loaders and sampling.