To conduct federated domain generalization

I am contributing domain generalization dataset to FS, where each client holds a domain-specific dataset. The model is learned from, e.g., client 1-3. and evaluated on client 4. As for validation, current best practice is to validate the model on the source domains and report the performance on target domains.

It is convenient to hold out the target domains as unseen clients. However, how to validate the model on the validation sets of the clients that are responsible for model training?

It would be very helpful to support this scenario as ood generalization has been advocated for FL scenarios in recent literature.