Would it be possible (or encouraged) to specify different hyperparameters (e.g. number of training rounds) for different federated scenarios
scenarioXX? The main motivation is that the best performing settings may depend on the number of clients, client dataset sizes, etc. From the log files of federated runs, it seems that we have some ideas about, say, the number of clients; is there any way this information can be accessed at run time?
Setting different hyperparameters for different federation scenarios is permitted.
The number of partitions should be apparent, but here they are for reference.
Track A: Financial Crime Prevention
- SWIFT + 2 bank partitions
- SWIFT + 4 bank partitions
- SWIFT + 9 bank partitions
Track B: Pandemic Forecasting
- 2 partitions
- 5 partitions
- 10 partitions
As documented, partitioning is done by taking the full dataset (that is available to the centralized evaluation), and dividing it up among the partitions in a fairly even way, subject to partitioning boundary constraints specific to each track. These partitioning scenarios are the same between the evaluation dataset and the smoke test dataset (i.e., the data from the centralized case is divided up into the same numbers of partitions for each scenario).