Minimize re-exports from __init__ files

[akx]

Quoting the README:

General Philosophy

Modularity is king.

This allows importing parts of the package without having to import practically everything (since importing a package will import its parents' __init__s, etc).

The one alias I didn't touch is sgm.modules.GeneralConditioner since it is currently being referred to using that name in the YAML config files. I suppose the configs could be changed to do sgm.models.encoders.modules.GeneralConditioner too, but that didn't seem in scope for this.

[benjaminaubin]

Why was it an issue to import sgm as a whole ?
If not issue we will keep it as it is

[akx]

@benjaminaubin It is very much a problem.

This ties into the dependency discussion we had.

sgm.data imports a bunch of modules that are only required for training, not inference.

If you try to import, say, sgm.helpers.inference after #56 lands, it means sgm.helpers is imported, which means sgm is imported, and it imports sgm.data, and so on, and you'll end up needing a bunch of dependencies you shouldn't – not to mention it has an impact on import time too.

In short, you can't even import DiffusionEngine on main without having e.g. torchdata, webdataset, and then it'll complain about stable-datasets, which is supposed to be a submodule?

(main) $ python -c 'from sgm.models.diffusion import DiffusionEngine' 2>&1
Traceback (most recent call last):
  File "<string>", line 1, in <module>
  File "generative-models/sgm/__init__.py", line 1, in <module>
    from .data import StableDataModuleFromConfig
  File "generative-models/sgm/data/__init__.py", line 1, in <module>
    from .dataset import StableDataModuleFromConfig
  File "generative-models/sgm/data/dataset.py", line 3, in <module>
    import torchdata.datapipes.iter
ModuleNotFoundError: No module named 'torchdata'
(main) $ pip install torchdata
(Installed torchdata and about 20 other deps)
(main) $ python -c 'from sgm.models.diffusion import DiffusionEngine' 2>&1
Traceback (most recent call last):
  File "<string>", line 1, in <module>
  File "generative-models/sgm/__init__.py", line 1, in <module>
    from .data import StableDataModuleFromConfig
  File "generative-models/sgm/data/__init__.py", line 1, in <module>
    from .dataset import StableDataModuleFromConfig
  File "generative-models/sgm/data/dataset.py", line 4, in <module>
    import webdataset as wds
ModuleNotFoundError: No module named 'webdataset'
(main) $ pip install webdataset
(Installed webdataset and another dependency)
(main) $ python -c 'from sgm.models.diffusion import DiffusionEngine' 2>&1
######################################################################
Datasets not yet available
to enable, we need to add stable-datasets as a submodule
please use ``git submodule update --init --recursive``
and do ``pip install -e stable-datasets/`` from the root of this repo
######################################################################

Whereas on this PR's branch:

(minimize-re-exports) $ python -c 'from sgm.models.diffusion import DiffusionEngine' 2>&1
no module 'xformers'. Processing without...
no module 'xformers'. Processing without...
(minimize-re-exports) $

[akx]

Thanks ❤️