Fix model saving issues

[jalabort]

Models seem to be unable to be loaded up when using a lambda Layer for dealing with the backbone specific preprocessing.

[jalabort]

This has been partially addressed. The solution was to create Layer subclasses instead of Lambda layers. In theory, I believe that Lambda layers should work just fine but in our case they do not.

There seems to be something wrong with the serialisation process of the function inside the Lambda layer. At some point, this function tries to get called but the variable supposed to contain a reference to the function definition contains its name as a string instead (strings are obviously not callables). This might be something we are doing wrong or perhaps a bug in TF2.0