We're testing for the presence of the model (self.model_ is not None should evaluate to True) as per the current implementation. This, in my opinion, is implicit in the documentation we have built during serialization (by mentioning that keras/ subdirectory is present in the zipped file).

I also have trouble thinking of a use case for saving an estimator that lacks the most important element. Perhaps I've probably used the wrong kind of error (NotFittedError) to give you a false impression (I probably did that since there was no NotBuiltError?).

In any case, I'm making a commit that actually lets you save the estimator even when self.model_ = None for DL Classifiers. Let's discuss if this is actually what we want so we can change the documentation/other parts of the code.

I don't fully understand why the current failure (in 3.10, macOS-11) is occurring. I also highly doubt this failure has anything to do with the current PR's work since the problem is setting up sktime.

One possibly relevant piece of information is that grpcio is downloaded directly as a wheel in Windows whereas it is downloaded as a .tar.gz file for Mac and it is during the building stage of this package where the error is occurring.