FAST is a framework that can integrate almost all first-order Federated Learning algorithms. It provides multiple client participation patterns that we call arbitrary client participation.
- Run regular FAST:
python fast_main.py --dataset=mnist --train-batch-size=128 --test-batch-size=256 --lr=0.01 --sampling-type=uniform_beta --local-update=5 --num-clients=100 --round=1000 --q=0.5 --alpha=0.1 --seed=365 --algo=fedavg --log-to-tensorboard=mnist01 --eval-iterations=10 - Run adaptive FAST:
python fast_main.py --dataset=mnist --train-batch-size=128 --test-batch-size=256 --lr=0.01 --sampling-type=uniform_beta --local-update=5 --num-clients=100 --round=1000 --q=0.5 --gamma=0.7 --alpha=0.1 --seed=365 --algo=fedavg --adaptive=1 --log-to-tensorboard=mnist01 --eval-iterations=10 - Run regular FL (Non-FAST algorithm):
python fast_main.py --dataset=mnist --train-batch-size=128 --test-batch-size=256 --lr=0.01 --sampling-type=uniform --local-update=5 --num-clients=100 --round=1000 --seed=365 --algo=fedavg --log-to-tensorboard=mnist01 --eval-iterations=10
| Hyper-parameter | Description |
|---|---|
| sampling-type | The way to sample clients in each round. The optional values include uniform, beta, gamma, weibull. If you consider using FAST algorithm, you can set --sampling-type=uniform_A where A can be chosen from the optional values. |
| num-clients | The total number of clients in the FL system. Default participation rate is 10%. If you want to change the rate, you can go to client_sampling.py to modify participation_ratio. |
| q | The probability of snapshots. |
| alpha | The concentration parameter of Dirichlet distribution. We use it to control the Non-IID degree of data. Generally, a smaller alpha is corresponding to a more Non-IID data distribution across clients. |
| gamma | The hyper-parameter for adaptive FAST algorithm. |
| local-update | The number of local update steps in FL. |
This framework has been finished and maintained by Zhe Li(RIT), who was at the University of Southern California, Bicheng Ying(Google), and advised by Prof. Haibo Yang(RIT).
