Add `torch.nn.functional as F` to flake8-import-conventions aliases

[IlanCosman]

Summary

Import style	Search regex	Number of files (links are slightly borked)
from torch.nn import functional	/from torch\.nn import functional$/	1.6k
from torch.nn import functional as <SOMETHING>	/from torch\.nn import functional as .*$/	162k
from torch.nn import functional as F	/from torch\.nn import functional as F$/	129k
import torch.nn.functional as <SOMETHING>	/import torch\.nn\.functional as .*$/	1.4M
import torch.nn.functional as F	/import torch\.nn\.functional as F$/	1.2M

In total 1.329 / 1.562 ~= 85% of the torch.nn.functional imports are as F.

Considerations

• Afaict F is nowhere explicitly recommended/mentioned by the PyTorch documentation, though it appears in every code example on https://docs.pytorch.org/docs/stable/nn.functional.html

The F alias violates rule N812, lowercase-imported-as-non-lowercase.

• This is unfortunate, but I think it's case where practicality beats purity, since the F alias is so ubiquitous in the PyTorch ecosystem.
• I think it's worth making some sort of special case for this in rule N812
  • Quick and dirty option: add functional to the default ignore-names in [tool.ruff.lint.pep8-naming]
    • This has the slight disadvantage of allowing all import of functional, regardless of whether they are from PyTorch or not, to be importable with uppercase names.
  • Cleaner option: some sort of particular exception for torch.nn.functional? Not sure tbh

[ntBre]

Just to make sure I understand, is the suggestion here to add this to the default value for lint.flake8-import-conventions.aliases? I think pytorch users could already configure this with either the aliases or extend-aliases options. Similarly, I think you could get around the N812 violation by configuring lint.pep8-naming.ignore_names, as you said.

As an aside, it looks like you have to add F to extend-ignore-names, not functional: https://play.ruff.rs/80ffcd74-1b08-4c19-bea4-f973fcff9850

[IlanCosman]

Just to make sure I understand, is the suggestion here to add this to the default value for lint.flake8-import-conventions.aliases? I think pytorch users could already configure this with either the aliases or extend-aliases options. Similarly, I think you could get around the N812 violation by configuring lint.pep8-naming.ignore_names, as you said.

Yah, everything you mentioned is correct. I think it would be good for the standard name to be enforced by default, and for it to not be a violation of N812.

As an aside, it looks like you have to add F to extend-ignore-names, not functional: https://play.ruff.rs/80ffcd74-1b08-4c19-bea4-f973fcff9850

Apparently both work, thanks for showing me that :)

---

Maybe the most reasonable way to do this is to just say that anything defined in import-conventions.aliases automatically passes N812. If a name is the standard, as defined by import-conventions.aliases, then rules should just accept that that's an allowed name.