[ENH] save/load aka serialization/deserialization for estimators

[fkiraly]

This adds a prototype design for serializing and deserializing estimators.
This follows the current (Aug 19) draft STEP 27 sktime/enhancement-proposals#27.

This is done via save and two load methods which can be overridden by class, and a generic load (class agnostic), in the BaseObject module.

Also changes the pickle/unpickle test to a generic serialize/deserialize test which allows custom, class specific implementations of serialize and deserialize.

FYI @AurumnPegasus - design alternative to #3128 which avoids the workflow where we have to construct an object first when we load.

[AurumnPegasus]

I see,
I do not know if serialisation/deserialisation works for keras models (will check it out once).
I was thinking something along the lines of (better explained in the step document I was working on, will push that soon):

1. Creating a save function which does 2 things:
   a. Creates pickle of class dictionary (EXCLUDING optimizer)
   b. Saves the network using keras.save
   In this case, we will be able to solve both the problems my design was having, firstly being able to save estimator parameters, and secondly, there would be no pickle error since we are excluding the optimizer from the state dict.
2. A generic load function would then first load from the pickle file, then load in the network using keras.load

[fkiraly]

with this design, you don't have to implement separate functions or logic, just overload the three new BaseObject methods for neural networks.

[AurumnPegasus]

Hey @fkiraly ,
Can you please explain a doubt I had wrt this implementation?

if isinstance(serial, tuple):
        cls = serial[0]
        stored = serial[1]
        return cls.load_from_serial(stored)
    elif isinstance(serial, (str, ZipFile)):

        if isinstance(serial, str):
            zipfile = ZipFile(serial)

        with zipfile.open("metadata", mode="r") as metadata:
            cls = metadata.read()
        with zipfile.open("object", mode="r") as object:
            return cls.load_from_path(object.read())

Here, I am assuming cls would refer to the particular estimator.
What is this line doing? cls.load_from_path(object.read()) (is the object.read a pickle file?)

[fkiraly]

What is this line doing? cls.load_from_path(object.read()) (is the object.read a pickle file?)

The idea is that object.read() can be anything, but the contract is that cls.load_from_path must convert it into an instance of cls.

In the default implementation, object.read is a pickle file, but you could override load_from_path (a class method) if the object us unpickleable.

[achieveordie]

Could you confirm if this works for DL Classifiers, like CNNClassifier()? I altered CNNClassifier directly to serialize contents(save(), load_from_serial(), __getstate__() etc) and I got error for the line meta_file.write(type(self)):

TypeError: object of type 'ABCMeta' has no len()

[fkiraly]

Could you confirm if this works for DL Classifiers

@achieveordie, in the tests, this is currently skipped for all DL classifiers, since they are not compatible with pickle (skip is in the _config file).

Re your proposed change, is there a draft PR? I could look at it.

Your error message might be related to BaseClassifier inheriting from ABC. Perhaps it works if you remove the inheritane from ABC?

[fkiraly]

Could you confirm if this works for DL Classifiers

@achieveordie, I suppose I have misunderstood your question.

Did you want to know whether the default in this PR works for DL classifiers?

The answer is "no", and I know that it doesn't work.

However, having the abstract methods in place, it should be possible to override this in the deep learning base class, with something that works - as you have probably tried to (this was the idea, I just don't know what precisely would work).