How the centralized server model is updated with aggregated client metrics in TensorflowFederated

user_18428250 0
Viewed 61

I have designed the Federated Learning model with TensorFlow Federated framework. Defined the iterative process as below,

iterative_process = tff.learning.build_federated_averaging_process(
    model_fn,
    client_optimizer_fn=lambda: tf.keras.optimizers.SGD(learning_rate=1.0),
    server_optimizer_fn=lambda: tf.keras.optimizers.SGD(learning_rate=0.9))

I have 2 remote workers running the tffruntime remote executor service and the context for running computation is defined as tff.backends.native.set_remote_python_execution_context(channels). When the model is broadcasted to the client with iterative_process.next(state, train_data), how can we identify that the client metrics is aggregated and applied to the server model. Is the single api build_federated_averaging_process is enough to get the metrics from clients, aggregate and then update the server model? If means how can we identify that the server model is updated? Can anyone please help me to understand this.

federated-learning tensorflow-federated tensorflow machine-learning
1 Answers

The build_federated_averaging_process API builds an iterative process of the full federated learning steps. If you want to verify that the server model is updated, you can print state.model after each iterative_process.next(state, train_data).

user_45459950
Related