Fix serialization error with EfficientNet

[RocketDelivery]

Fixes keras-team/tf-keras#383. The argument to layer.Rescaling is passed as EagerTensor which is not serializable, because tf.math.sqrt() may return EagerTensor.
It is functionally equivalent to the solution suggested by @hctomkins in keras-team/tf-keras#383

[google-cla]

Thanks for your pull request! It looks like this may be your first contribution to a Google open source project. Before we can look at your pull request, you'll need to sign a Contributor License Agreement (CLA).

View this failed invocation of the CLA check for more information.

For the most up to date status, view the checks section at the bottom of the pull request.

[jeromemassot]

Please merge the fix ASAP as we are legions in need of saving and restoring EfficientNet implementations without altering the official EfficientNet code. Thanks.

[fchollet]

LGTM

[gpiosenka]

encountered the same issue with any of the EfficientNet models. Saving the trained model works for tensorflow version 2.9.2 but does not work for version 2.11.0. Is there a fix for this?

[Berganzio]

Unable to serialize [2.0896919 2.1128857 2.1081853] to JSON. Unrecognized type .

Still facing the same issue with efficientnetB0 today.
Is there a fix for this?

[mpolinowski]

The same error after:

tf.__version__
# '2.11.0'

base_model_v2 = tf.keras.applications.efficientnet_v2.EfficientNetV2M(include_top=False)

...

model.save('../saved_models/food101_env2m_30epochs')

see gist

EDIT: I re-ran a 1 epoch training and this time saving worked 🤔

EDIT2: I just finished a full re-run and was unable to reproduce the error. All that I did after the first failed attempt was stopping and re-opening the notebook and I also tried to update Tensorflow (but it was already the latest version).