Simulation API Reference

The philosophy of our API design is to reduce the number of APIs as much as possible while simultaneously maintaining the flexibility.

For Simplicity, FedML Parrot has only one line API as the following example:

import fedml

if __name__ == "__main__":
    fedml.run_simulation()

To meet the customization demands, FedML Parrot also has five lines of APIs as the following example.

import fedml
from fedml.simulation import SimulatorMPI

if __name__ == "__main__":
    # init FedML framework
    args = fedml.init()

    # init device
    device = fedml.device.get_device(args)

    # load data
    dataset, output_dim = fedml.data.load(args)

    # load model
    model = fedml.model.create(args, output_dim)

    # start training
    simulator = SimulatorMPI(args, device, dataset, model)
    simulator.run()

For newly developed features, we will try to preserve the form of these APIs and only add new arguments.

To check out the details of the latest definition of each API, the best resource is always the source code itself. Please check comments of each API at: https://github.com/FedML-AI/FedML/blob/master/python/fedml/init.py

Algorithm Reference Implementations

TensorOpera®Federate supports representative algorithms in different communication topologies (as the figure shown below), including Fedvg, FedOpt (ICLR 2021), FedNova (NeurIPS 2020), FedGKT (NeurIPS 2020), Decentralized FL, Vertical FL, Hierarchical FL, FedNAS, and Split Learning.

The API for each algorithm is as simple as follows:

import fedml
from fedml.simulation import SimulatorMPI

if __name__ == "__main__":
    # init FedML framework
    args = fedml.init()

    # init device
    device = fedml.device.get_device(args)

    # load data
    dataset, output_dim = fedml.data.load(args)

    # load model
    model = fedml.model.create(args, output_dim)

    # start training (use "SimulatorSingleProcess" for single process-based simulation)
    simulator = SimulatorMPI(args, device, dataset, model)
    simulator.run()

To specify the algorithm, please change federated_optimizer at fedml_config.yaml. For example, the following fedml_config.yaml will enable FedOpt algorithm.

train_args:
  federated_optimizer: "FedOpt"
  client_id_list: "[]"
  client_num_in_total: 1000
  client_num_per_round: 10
  comm_round: 200
  epochs: 1
  batch_size: 10
  client_optimizer: sgd
  learning_rate: 0.03
  weight_decay: 0.001

For more details, please check these examples.