Support for federated evaluation

[Mansi Sharma]

Add an example of federated evaluation using aggregator based workflow or the director based workflow. Federated evaluation refers to model evaluation on the client/collaborator side. Once the global model is sent to the collaborators, they could perform one pass of local evaluation. Evaluation wouldn't need any training/construction of optimizers/gradient descent.

For aggregator based workflow:

• An existing workspace like torch_cnn_mnist could be taken as a base example from the list of supported templates for OpenFL to create a new template for federated evaluation.

• Changes to the plan/plan.yaml will be needed:

  • Since there's no training involved, parameter rounds_to_train should be set to 1
  • Task aggregator_model_validation should be assigned instead of the default tasks.

• Changes to the src/pt_cnn.py will be needed:

  • No optimizer initialization needed.

[Beverly Klemme]

Could you supply more detail on what you are looking for?

[Wang Wenjie]

OpenFL_Team4 is working on this