config.md

Key Parameters

All training parameters are in configs.

Main configuration file config.yaml contains:

• defaults:

  • model: Model architecture (e.g. resnet18). Add custom model in configs/model.
  • dataset@train_dataset: Client training data (cifar10, or custom datasets in configs/observed_data_params).
  • dataset@test_dataset: Server evaluation dataset (doesn't affect training).
  • dataset@trust_dataset: Additional sample stored on the server. Used by FLTrust for example.
  • distribution: The parameter responsible for distributing data between clients.
  • model_trainer: Entity responsible for technical training, eval, test model in different task domains
  • federated_method: Algorithm selection (fedavg, fedprox). See method.md for details.
  • client_selector: The nature of customer subset selection (e.g. FedCBS)
  • manager: Type of manager to handle multiple clients. E.g. sequential sequentially collects client updates (or custom manager type).
  • losses@loss: Client loss function, e.g. ce (cross-entropy) or custom loss in configs/losses
  • optimizer: Type of client optimizer, e.g. adam, sgd or any existing optimization scheme.
  • preaggregator: Methods for modifying client gradients before aggregation. Used in Byzantine setups.

• random_state: Sets reproducibility

• training_params

  • batch_size: client batch size
  • num_workers: amount of parallelism in Dataloader
  • device: type of computing device (cpu or cuda)
  • device_ids list of gpu indices (for the case of many GPUs on a machine and used many gpus for one experiment)

• federated_params:

  • amount_of_clients: Total number of clients
  • client_subset_size: How many clients do we select in each round
  • communication_rounds: Total number of rounds
  • local_epochs: Local client epochs
  • client_train_val_prop: Train/validation split on client-side to evaluate server model
  • print_client_metrics: Enable client metric logging
  • server_saving_metrics: Model selection metrics. Can be one of "loss", "Specificity", "Sensitivity", "G-mean", "f1-score", "fbeta2-score", "ROC-AUC", "AP", "Precision (PPV)", "NPV"
  • server_saving_agg: Aggregation method (uniform average vs weighted by dataset size)
  • clients_attack_types, prop_attack_clients, attack_scheme, prop_attack_rounds: see attack.md for details.