activity_regularizer is omitted from get_config/serialization

[sikchi]

simple to repro:

from keras import layers as kl
from keras import regularizers as kr
kl.Dense(units=16, kernel_regularizer=kr.l2(), activity_regularizer = kr.l2()).get_config().get("activity_regularizer") # evaluates to None
kl.Dense(units=16, kernel_regularizer=kr.l2(), activity_regularizer = kr.l2()).get_config().get("kernel_regularizer") # evaluates to a dict as expected

Noticed this issue when I saw that if you save a model in .keras format and load it again, it is missing activity_regularizer.

Let me know if this is by design/actvity_regularizer is deprecated/etc.

[fchollet]

Thanks for the report. This was an oversight. This is now fixed at HEAD.

[sikchi]

Thanks for the super fast fix!

I have a broader question - Based on this - https://keras.io/api/models/model_saving_apis/ I assume I should be able to save a partially trained model using model.save(), then load it using load_model(), and resume training. Note that the document does not explicitly say that.

If my assumption is true, then there are other issues: for e.g., for a trained model with a (non-custom) Loss, model.get_config().get("losses") and model.get_config.get("loss") both return None. Whereas model.losses and model.loss work as expected. Consequently, those properties don't roundtrip via save() and load_model().

LMK if this is by design/alternatives/..

[sikchi]

@fchollet @mehtamansi29 : not sure if you saw my previous comment.

LMK if I should open a new issue instead