nvflare.app_common.response_processors.global_weights_initializer module

class GlobalWeightsInitializer(weights_prop_name: str = 'global_model', weight_method: str = 'first', client_name: str | None = None)

Bases: ResponseProcessor

Set global model weights based on specified weight setting method.

Parameters:

• weights_prop_name – name of the prop to be set into fl_ctx for the determined global weights

• weight_method – the method to select final weights: one of “first”, “client”

• client_name – the name of the client to be used as the weight provider

If weight_method is “first”, then use the weights reported from the first client; If weight_method is “client”, then only use the weights reported from the specified client.

create_task_data(task_name: str, fl_ctx: FLContext) → Shareable

Create the data for the task to be sent to clients to collect their weights

Parameters:

• task_name – name of the task

• fl_ctx – the FL context

Returns: task data

final_process(fl_ctx: FLContext) → bool

Perform the final check on all the received weights from the clients.

Parameters:

fl_ctx – FLContext

Returns:

boolean indicating whether the final response processing is successful. If not successful, the control flow will exit.

process_client_response(client: Client, task_name: str, response: Shareable, fl_ctx: FLContext) → bool

Process the weights submitted by a client.

Parameters:

• client – the client that submitted the response

• task_name – name of the task

• response – submitted data from the client

• fl_ctx – FLContext

Returns:

boolean to indicate if the client data is acceptable. If not acceptable, the control flow will exit.

class WeightMethod

Bases: object

CLIENT = 'client'

FIRST = 'first'