🔄 Workflows

Workflows in P2PFL define the sequence of operations performed by each node during a federated learning experiment. They provide a structured approach to the training process, ensuring that all nodes follow a consistent set of steps. Currently, P2PFL implements a single core workflow, LearningWorkflow, which governs the lifecycle of a federated learning experiment.

Learning Workflow

The LearningWorkflow orchestrates the training process, coordinating actions such as model initialization, training set selection, local training, model aggregation, and evaluation. It uses a series of stages, each responsible for a specific part of the workflow. The workflow’s progression is illustrated in the following diagram:

graph LR A(StartLearningStage) --> B(VoteTrainSetStage) B -- Node in trainset? --> C(TrainStage) B -- Node not in trainset? --> D(WaitAggregatedModelsStage) C --> E(GossipModelStage) D --> E E --> F(RoundFinishedStage) F -- No more rounds? --> Finished F -- More rounds? --> B

Stages

  1. StartLearningStage: Initializes the federated learning process. This includes setting up the experiment, initializing the model, and gossiping the initial model parameters to all nodes.

  2. VoteTrainSetStage: Nodes vote to select a subset of nodes (train_set) that will participate in the current training round. This stage ensures that not all nodes need to participate in every round, which can improve efficiency and scalability.

  3. TrainStage: Nodes in the train set perform local training on their datasets, evaluate their local models, and contribute their updates to the aggregation process.

  4. WaitAggregatedModelsStage: Nodes not participating in the current training round wait for the aggregated model from their neighbors.

  5. GossipModelStage: All nodes gossip their models (either locally trained or aggregated) to their neighbors. This dissemination of model updates ensures eventual convergence across the decentralized network.

  6. RoundFinishedStage: Marks the end of a training round. If more rounds are remaining, the workflow loops back to the VoteTrainSetStage. Otherwise, the experiment concludes, and final evaluation metrics are calculated.

This workflow ensures a structured and coordinated training process across all nodes in the decentralized network. The use of stages and the voting mechanism for training set selection provide flexibility and scalability.