3

I am running the code in tensorflow federated learning (10 epochs, 100 rounds), but the accuracy is not increasing and it's around 0.5. Part of my code is as follows:

def create_compiled_keras_model():
    layer1 = tf.keras.layers.GlobalAveragePooling2D()(output)
    layer1 = tf.keras.layers.Dense(units=256)(output)
    model_output = tf.keras.layers.Dense(units=2, activation='relu')(layer1)
    model = tf.keras.Model(model.input, model_output)
    return model

def model_fn():
     keras_model = create_compiled_keras_model()
     return tff.learning.from_keras_model(keras_model, sample_batch, loss=tf.keras.losses.CategoricalCrossentropy(),metrics=[tf.keras.metrics.CategoricalAccuracy()])

And,

for round_num in range(1, NUM_ROUNDS+1):
    state, tff_metrics = iterative_process.next(state, federated_train_data) 
    print('round {:2d}, metrics{}'.format(round_num,tff_metrics['train'].items()))
    
    eval_model = create_keras_model()
    eval_model.compile(optimizer=optimizers.Adam(learning_rate=client_lr),
                       loss=losses.BinaryCrossentropy(),
                       metrics=[tf.keras.metrics.Accuracy()])
    
    state.model.assign_weights_to(eval_model)
    
    ev_result = eval_model.evaluate(x_val, y_val, verbose=2)
    train_metrics = tff_metrics['train']
      for name, value in tff_metrics['train'].items():
            tf.summary.scalar(name,value, step=round_num)

I saw few questions on stackoverflow on this, but couldn't find the solution anywhere.I would appreciate any help.

Dushi Fdz
  • 161
  • 3
  • 22

1 Answers1

0

It seems that the number of federated rounds is insufficient for federated learning, you may need a few thousands of rounds to converge and for the model to perform better. Check out this answer

Eden
  • 325
  • 3
  • 13