Set up Python packaging

[akx]

Sibling of Stability-AI/stablediffusion#269 but for the new fancy repo (and congrats on shipping)!

This makes it possible to build a standard Python .whl from the repo and use it instead of having to e.g. clone the entire repository.

Features

• Adds __version__ to sgm/ (copying 0.0.1 from setup.py)
• Sets up packaging using hatch (removing the defunct setup.py)
  • configs/ gets packaged as sgm/configs when using a packaged version
• Adds sgm.get_configs_path() to make it easy for downstream users to find the configs path

Caveats

• Unlike #269, I haven't tested this with a downstream package (because I can't summon weights from thin air 😁)
• Requirements aren't yet handled here (because the current requirements* files are a bit messy), but I would recommend an approach where inference and training requirements are strictly separated, and training reqs would be installed with optional requirements.

The package seems to be installable and the configs are correctly shipped:

$ docker run -it -v (pwd):/a:ro python:3.11 bash
root@a8a617f480ff:/# pip install /a/dist/sgm-0.0.1-py3-none-any.whl
Processing /a/dist/sgm-0.0.1-py3-none-any.whl
Successfully installed sgm-0.0.1
root@a8a617f480ff:/# python
>>> import sgm
...
ModuleNotFoundError: No module named 'torchdata'
>>>
root@a8a617f480ff:/# ls -la /usr/local/lib/python3.11/site-packages/sgm/configs/
total 16
drwxr-xr-x 4 root root 4096 Jun 27 08:53 .
drwxr-xr-x 7 root root 4096 Jun 27 08:53 ..
drwxr-xr-x 4 root root 4096 Jun 27 08:53 example_training
drwxr-xr-x 2 root root 4096 Jun 27 08:53 inference

[benjaminaubin]

Hello @akx ,
I am starting reviewing your PR

[akx]

@benjaminaubin Cool - let me know if you have any questions.

[benjaminaubin]

@akx thank you for this. Tested your branch, the package installed properly from the hatch wheel.

Could you

• add in the documentation a section Packaging with commands pip install hatch hatch build . and pip3 install xxxx.whl
• add **.pyc to gitignore
• [Optional] split requirements in train / inference as suggested

[akx]

@benjaminaubin Thank you 🙌

• Rebased
• Added *.py[cod] to .gitignore (and sorted the file 🧹)
• Added a section on packaging with a note that dependencies are not yet specified for the built package
• I'll take a look at the requirements in a subsequent PR, if that's okay? It will also change the general installation instructions (as it will become something like pip install -e .[train]).

[benjaminaubin]

Looks great to me.
Agreed requirements split can be done in another PR

[akx]

@benjaminaubin Awesome, thanks! Can you take a look at #44? It's a requirement for being able to split requirements.

[tubylec8]

How to install it?