Method configuration

In this page, we detail the configuration values related to the FL algorithm and its hyperparameters. The configuration groups or values in method are:

• _target_: the class corresponding to the FL algorithm to apply;

• hyperparameters: the hyperparameters of the FL algorithm, across different config groups;

  • client: for the client hyperparameters;

  • server: for the server hyperparameters;

  • model: for the model hyperparameters;

  • and additional config groups depending on the chosen FL class.

Client hyperparameters

The client field includes all the hyperparameters that are related to the client-side of the federated algorithm. It contains:

• batch_size: batch size;

• local_epochs: number of local epochs performed during each FL round;

• loss: config group that contains the information about the loss function. Users should specify the _target_ config value, which represents the loss function to instantiate (e.g., torch.nn.BCELoss), with the required additional parameters;

• optimizer: config group that contains the hyperparameters related to the optimizer. Users should specify the _target_ config value, which represents the optimizer to instantiate (e.g., Adam), with the required additional parameters (e.g., the learning rate or the weight decay);

• scheduler: config group that contains the hyperparameters related to the scheduler. Users should specify the _target_ config value, which represents the scheduler to instantiate (e.g., StepLR), with the required additional parameters (e.g., the gamma or the step size);

• and additional config values or groups depending on the chosen FL algorithm.

Server hyperparameters

The server field includes all the hyperparameters that are related to the server-side of the federated algorithm. It contains:

• loss: config group that contains the information about the loss function (the same as the client loss config group), which will be used only for the server evaluation;

• time_to_accuracy_target: value between 0 and 1 representing the accuracy target to achieve (used for calculating time to accuracy);

• weighted: boolean specifying whether to weight the client’s contribution to the global model;

• and additional config values or groups depending on the chosen FL algorithm.

Model hyperparameters

The model field specifies the machine learning model to federate. It contains:

• _target_: the class corresponding to the ML model to instantiate (e.g., armlet.utils.net.LogRegression);

• input_size: the input size of the ML model (automatically loaded in ARMLET according to the data properties);

• num_classes: the number of output classes of the ML model (automatically loaded in ARMLET according to the data properties);

• and additional config values or groups depending on the chosen ML model.

Note that an ML model is usually linked to some hyperparameters located in the client and server config groups (such as the loss function). Consequently, the preconfigured options for the model config group in ARMLET could override other config values that are originally included in the client and server config groups.