Convert protobuf model to native format

[psfoley]

This PR:

• Adds a new command, fx model save, that takes as input the .pbuf model file produced by a federated experiment and converts it to the native PyTorch / TensorFlow model representation for future use.

This PR is WIP. The core functionality to be refactored by @igor-davidyuk into the utilities folder, so this can also be used via API.

closes #552

[igor-davidyuk]

We want to introduce an additional piece of functionality in two forms:

• CLI: fx model save command called from the workspace must save the model on disk.
• Python API: openfl.model_save(path) must return the model object.

We start with model.pbuf path to the model snapshot that we need to decompress and load weights into the model object to save or return it

1. To save the model in a native format we need save_native method implemented in TaskRunner
   This method is not implemented in all our template TaskRunners but this is fine as we may ask users to implement it.
2. To return the model as a Python object we need to access this object
   Models are stored in different ways in our template TaskRunners. Moreover, there is a particular problem with TF1 examples. We may need to introduce an additional get_model method to all of our template TaskRunners.
3. To get access to the model we need to initialize TaskRunner
   Model initialization happens in TaskRunner code and often requires feature_shape information to build a model.
4. To initialize TaskRunner we need the plan.yaml path and DataLoader
   Again, model initialization is coupled with DataLoader initialization.
5. To initialize DataLoader we need data.yaml and cols.yaml
   Here we use the first collaborator's name to get its data_path and load data to get the feature_shape. It works for our templates, which use synthetic or downloaded datasets, but in real-world usecases, users often may not have access to any data from their machine. A way to overcome this obstacle may be to define feature_shape in plan.yaml or hardcode it inside DataLoader.

With all said above, at this point, we can not guarantee model save API functioning to all users in the intended way.

[igor-davidyuk]

As a solution to the abovementioned problems I propose the following approach:

1. For the CLI command we will rely on TaskRunner's save_native method that users can implement or change if needed.
2. For the Python API part, we will return a TaskRunner object, thus allowing users to interact with the linked model.

These signatures allow us to mitigate issues caused by the diversity of model definition forms in our TuskRunners and yet deliver the expected functionality. At the same time, the data problem mentioned under point 5 is ignored, as we know it is already solved prior to executing fx plan initialize which follows a similar path.

Work left in this PR:

• Update the docs
• Write tests

[itrushkin]

We could require this option as well for users to make sure they provide the path for the saved model. In case of someone forgot to specify it, it would be easier to notify them about the wrong command call and ask for the path. Otherwise, they have to search the save path in the console output.

[igor-davidyuk]

I understand the intention, but I would argue that we should make arguments required only if they are literally required.

[itrushkin]

output_filepath is required by the _save function for calling TaskRunner.save_native.

For example, torch.save and tf.keras.Model.save_weights do require a file argument.

[igor-davidyuk]

Yes, but those functions are python function calls, in our case we have a CLI command which is called from a specific place, namely, an experiment workspace.
Honestly, I am not against making this argument required. @mansishr, what do you think?

[aleksandr-mokrov]

I believe that here is used good approach with default value. It is checked that file already exists and is asked user for the confirmation to rewrite. At the end, it is written to the output in witch path the model is saved.

[mansishr]

I agree, specifying a default path helps for our CLI command.

[mansishr]

Looks good to me