cyberzhg / keras-lookahead Goto Github PK

Describe the Bug
I use tf.keras,Radam project is OK,but Lookahead goes wrong,detailed information is below:

/usr/local/lib/python3.6/dist-packages/tensorflow/python/keras/backend.py in get_value(x)
2796 elif ops.inside_function():
2797 raise RuntimeError('Cannot get value inside Tensorflow graph function.')
-> 2798 return x.eval(session=get_session())
2799
2800

AttributeError: 'ReplicatedVariable' object has no attribute 'eval'

Version Info

• The version I used is: <tensorflow 1.13.1>

Minimal Codes To Reproduce

from keras_lookahead import Lookahead
model.compile(loss='categorical_crossentropy',
                  optimizer = Lookahead('adam', sync_period=5, slow_step=0.5),
                  metrics=['accuracy'])


model.fit(X_train, X_train_label,
                  validation_data=(X_val, X_val_label),
                  epochs=100, batch_size=64,
                  shuffle=True,
                  callbacks=[early_stopping])

Optimizer weights are now stored in model.optimizer.optimizer, and model.optimizer.weights==[] - hence, default saving and loading methods will not work. Existing code can account for this as follows:

if "Lookahead" in str(model.optimizer):
    optimizer = model.optimizer.optimizer
else:
    optimizer = model.optimizer

optimizer.get_weights(...)
optimizer.set_weights(...)

NOTE: unsure if above accounts for all differences. Packing weights into model.optimizer directly will render this redundant.

Per my understanding, TF_KERAS=True in backend.py imports keras from tensorflow.python - and, in optimizers.py, if TF_KERAS gets evaluated, and TF_KERAS=True is set. Though TF_KERAS=False by default, it always ends up =True - so else is never executed.

Is this accurate? (if so, why an if-statement?) Lastly, does the code somehow depend on tensorflow.python.keras as opposed to simply keras? Thanks

Could you please extend your work to have a Lookahead optimizer for Tensorflow as well? Also it would be great if one could use your Tensorflow RAdam implementation with the TF version of the Lookahead optimizer.

Thank you for implementing this package btw!

How can I use TF.Keras?

Thanks a lot for this implementation.

0.6.0 is committed on GitHub, but not present on PyPi. It is useful for TF 2.0 and Keras 2.3 users. Do you plan to push it in a near future?

When used with RAdam, how to set the value of "sync_period", and "slow_step" ?