AttributeError: module 'tensorflow_federated.python.learning' has no attribute 'assign_weights_to_keras_model'

Question

eval_model = None
for round_num in range(1, 51):
    state, tff_metrics = iterative_process.next(state, federated_train_data)
    eval_model = create_keras_model()
    eval_model.compile(optimizer=optimizers.SGD(learning_rate=0.3),loss=losses.SparseCategoricalCrossentropy(),metrics=[metrics.SparseCategoricalAccuracy()])
    tff.learning.assign_weights_to_keras_model(eval_model, state.model)
    ev_result = eval_model.evaluate(x_test, y_test, verbose=0)
    print('round {:2d}, metrics={}'.format(round_num, tff_metrics))
    print(f"Eval loss : {ev_result[0]} and Eval accuracy : {ev_result[1]}")
    tff_train_acc.append(float(tff_metrics.sparse_categorical_accuracy))
    tff_val_acc.append(ev_result[1])
    tff_train_loss.append(float(tff_metrics.loss))
    tff_val_loss.append(ev_result[0])

It throws an error that we cannot assign weights. Earlier it was working.

Please format your code and include the full traceback error. — ewokx
– ewokx, Commented Apr 27, 2022 at 2:39

user11989081 · Accepted Answer · 2022-07-02 16:03:01Z

1

tff.learning.assign_weights_to_keras_model was removed in version 0.17.0 and was replaced by tff.learning.ModelWeights.assign_weights_to

Try replacing

tff.learning.assign_weights_to_keras_model(eval_model, state.model)

with

state.model.assign_weights_to(eval_model)

edited Jul 2, 2022 at 16:03

user11989081

8,75211 gold badges34 silver badges44 bronze badges

answered Apr 27, 2022 at 12:05

Zachary Garrett

2,93115 silver badges23 bronze badges

Sign up to request clarification or add additional context in comments.

4 Comments

Amandeep Singh Over a year ago

I tried but did not work. AttributeError: The tuple of length 2 does not have named field "assign_weights_to". Fields (up to first 10): ['trainable', 'non_trainable'] keras_model = create_keras_model() state = tff.learning.state_with_new_model_weights( state, trainable_weights=[v.numpy() for v in keras_model.trainable_weights], non_trainable_weights=[ v.numpy() for v in keras_model.non_trainable_weights ])

Zachary Garrett Over a year ago

This then depends on how iterative_process was created and possibly what version of TFF is being used. Could the question be extended to include this information?

Dushi Fdz Over a year ago

Hi @ZacharyGarrett, I have the same issue. state.model.assign_weights_to(eval_model) worked in tff 0.27 but does not work in 0.28. Has it been changed?

S M Sarwar Over a year ago

@ZacharyGarrett I faced same problem in 0.57.0 'state.model.assign_weights_to(eval_model)' didn't work.

S M Sarwar · Accepted Answer · 2023-05-12 05:50:48Z

0

You can follow this TFF documentation

model_weights = iterative_process.get_model_weights(state)
model_weights.assign_weights_to(eval_model)

However, need to change tff_metrics.sparse_categorical_accuracy to tff_metrics['client_work']['train']['sparse_categorical_accuracy'] and tff_metrics.loss to tff_metrics['client_work']['train']['loss'].

answered May 12, 2023 at 5:50

S M Sarwar

531 silver badge9 bronze badges

Collectives™ on Stack Overflow

AttributeError: module 'tensorflow_federated.python.learning' has no attribute 'assign_weights_to_keras_model'

2 Answers 2

4 Comments

Comments

Your Answer

Hot Network Questions

Collectives™ on Stack Overflow

2 Answers 2

4 Comments

Comments

Your Answer

Sign up or log in

Post as a guest

Related