Add proper __all__ in all scripts

We should add a proper __all__ definition in all scripts. This will prevent strange import behaviors.

I propose to use the following format:

# Example of the proposed __all__ format
__all__ = [
    'SupportedDataset',
    'TransformationDataset',
    'TransformationSubset',
    'TransformationConcatDataset',
    'TransformationTensorDataset',
    'concat_datasets_sequentially',
    'as_transformation_dataset',
    'train_test_transformation_datasets'
]

which means:

• One definition per line, which will make it easier to analyze diff in commits in which something has been added/removed to/from __all__.
• The first definition starts in the line after __all__ = [
• Each definition is “tabbed” by just one level
• The final ] should be in its on line
• Order of elements should follow the order in which things are defined in the script (that is, the order obtained by looking at the “Structure” tab of PyCharm).

What do you think?

[lrzpellegrini]

I thinked about a possible organization of the whole import thing. As discussed before, the main problem with cascading __init__.py is that it every sub-package symbol may get into the root namespace of each major module if we don’t define a proper module structure. Having many symbols in the namespace may be an undesirable source of confusion and will also tamper with IDEs hinting capabilities.

We should focus on keeping self-explanatory import paths based on common usage patterns + the degree of “standalone”-ness of that particular sub-module.

This is my idea for the “import paths” of the benchmark module:

• avalanche.benchmarks.classic: in this way users may import classic benchmarks the way Jeremy suggested.
• avalanche.benchmarks.datasets: i think that avalanche.benchmarks.datasets.that_dataset may be a little too specific. Also, we shouln’t add symbols contained in thisdataset_data.py to the module’s __init__.py file.
• avalanche.benchmarks.generators: it now has a single source file scenario_generators.py containing all the definitions. We should probably split it to keep source code files simpler to navigate.
• avalanche.benchmarks.scenarios.new_classes (without elements from nc_utils.py)
• avalanche.benchmarks.scenarios.new_instances (without elements from ni_utils.py)
• avalanche.benchmarks.scenarios.generic (without elements from generic_definitions.py)
• Do not let anything flow upward from utils (but still put all non-private elements in __all__, so that they can be used)

Which means making the following changes:

• Creating the avalanche.benchmarks.scenarios.generic module containing generic_cl_scenario.py and generic_scenario_creation.py
• Remove the imports from avalanche.benchmarks.__init__.py, which now imports * from “scenarios”, “classic” and “generators” (which is a little bit exaggerated).
• Split scenario_generators.py (see above)
• Maybe consider moving TransformationDataset-related elements to their own new package.
• Remove any private symbol from __all__ (example: the _get_cifar10_dataset method Martin pointed out)!

This applies to the benchmarks module, but I feel we should define a structure also for other major modules, too.

What do you think? Does this make sense to you?

[jeremyforest]

I think making it so this from avalanche.benchmarks.classic import SplitCifar10 is the standard would be a good idea. IMO it is the cleanest and easiest way for people not familiar with the code base to know where is what and not spend 10min tracking the dependencies.