Algorithm configuration

The ALG_CONFIG_FILE is a yaml file containing the hyper-parameters of the federated algorithm. The set of hyper-parameters are divided into two main categories: the hyper-parameters of the clients and the hyper-parameters of the server. Additionally, there is the model parameter that specifies the neural network to federate.

hyperparameters:
    client:
        # HPs of the clients

    server:
        # HPs of the server

model: # The model (neural network) to federate

Client hyper-parameters

The client hyper-parameters are divided into three main categories: the hyper-parameters of the client overall training process, the hyper-parameters of the optimizer, and the hyper-parameters of the scheduler.

Training process HPs

This set of hyper-parameters are related to the training process of the client excluding the optimizer and the scheduler. Here you can specify the batch size, the number of local epochs, the loss function and other algorithm-specific hyper-parameters.

For example:

client:
    # Batch size
    batch_size: 10
    # Number of local epochs
    local_epochs: 5
    # The loss function (loss name from torch.nn)
    loss: CrossEntropyLoss
    # Algorithm-specific hyper-parameters
    param: value
    ...

The optimizer HPs

The optimizer hyper-parameters are related to the optimizer used to train the client’s model. Here you can specify the learning rate, the momentum, the weight decay, etc.

For example:

client:
    ...
    optimizer:
        name: SGD
        lr: 0.8
        momentum: 0.9
        weight_decay: 0.0001
        ...

The name parameter is the name of the optimizer from torch.optim. If omitted, the default optimizer is SGD.

The scheduler HPs

The scheduler hyper-parameters are related to the scheduler used to adjust the learning rate during the training process. Here you can specify the gamma, the step size, the mode, etc.

For example:

client:
    ...
    scheduler:
        name: StepLR
        gamma: 0.995
        step_size: 10
        ...

The name parameter is the name of the scheduler from torch.optim.lr_scheduler. If omitted, the default scheduler is StepLR.

Server hyper-parameters

Although less common, the server hyper-parameters are related to the server-side of the federated algorithm. Here you can specify whether to weight the client’s contribution to the global model.

For example:

server:
    weighted: true

Federated model

The model parameter specifies the neural network to federate. This can be the name of a neural network included in fluke.nets or the fully qualified name of an user defined class.

If the model class accepts arguments, you can specify them in the configuration file using the key net_args.

For example:

model: MNIST_2NN
net_args:
    hidden_size: [100, 50]