1

I use tensorflow fedprox to implement federated learning.(tff.learning.algorithms.build_unweighted_fed_prox)

def model_fn():
    keras_model = create_keras_model()
    return tff.learning.from_keras_model(
        keras_model,
        input_spec=preprocessed_example_dataset.element_spec,
        loss=tf.keras.losses.SparseCategoricalCrossentropy(),
        metrics=[tf.keras.metrics.SparseCategoricalAccuracy()]
    )

iterative_process = tff.learning.algorithms.build_unweighted_fed_prox(
    model_fn, 0.001,
    client_optimizer_fn=lambda: tf.keras.optimizers.SGD(learning_rate=0.001),
    server_optimizer_fn=lambda: tf.keras.optimizers.SGD(learning_rate=1.0)
)

import nest_asyncio
nest_asyncio.apply()

state = iterative_process.initialize()

for round in range(3, 11):
    state = iterative_process.next(state.state, federated_train_data)
    print('round {:2d}, metrics={}'.format(round, state.metrics))

and the result of training is:

round 3, 'sparse_categorical_accuracy'= 0.6435834

round 4, 'sparse_categorical_accuracy'= 0.6955319

round 5, 'sparse_categorical_accuracy'= 0.74295634

round 6, 'sparse_categorical_accuracy'= 0.78176934

round 7, 'sparse_categorical_accuracy'= 0.80838746

round 8, 'sparse_categorical_accuracy'= 0.8300672

round 9, 'sparse_categorical_accuracy'= 0.8486338

round 10, 'sparse_categorical_accuracy', 0.86639416

but when I want to evaluate my model on test data I get error:

evaluation = tff.learning.build_federated_evaluation(model_fn)
test_metrics = evaluation(state.state, federated_test_data)

TypeError: Mismatched number of elements between type spec and value in `to_representation_for_type`. Type spec has 2 elements, value has 5.

How do I fix it?

Park
  • 2,446
  • 1
  • 16
  • 25
hamid ebrahimi
  • 51
  • 5

1 Answers1

1

Your evaluation method expects tff.learning.ModelWeights, but you are providing the entire state, which is a bigger structure, including the model weights under global_model_weights attribute. So, this could work:

test_metrics = evaluation(state.state.global_model_weights, federated_test_data)

Side note, assigning the return value of iterative_process.next to Python variable state can become very confusing, as it contains state of the program and metrics, which leads you to the use state.state

Jakub Konecny
  • 900
  • 5
  • 15
  • thanks a lot, it worked. But I get a terrible result, the accuracy of the model is 36%. I used FedAvg at the first and got 93% accuracy But in this case (when using FedProx) I got 36%. I don't know why the model became overfitted When I used FedProx. could you help me how to fix it? (I tried with different values of mu but it doesn't work) – hamid ebrahimi Mar 26 '22 at 21:30