Add logging to torchvision ops

[fmassa]

This PR adds logging to the public ops in torchvision.

I haven't tried making it use inspect to automatically find the name of the function as there is no chance this is supported in torchscript.
This function isn't supported in torchscript anyway, so let's make it a no-op in torchscript for now.

A quick benchmark shows ~900ns of overhead, which should be fine

%timeit torchvision.utils._log_api_usage_once('testing')
914 ns ± 17.6 ns per loop (mean ± std. dev. of 7 runs, 1000000 loops each)

[facebook-github-bot]

💊 CI failures summary and remediations

As of commit 693ea25 (more details on the Dr. CI page):

---

💚 💚 Looks good so far! There are no failures yet. 💚 💚

---

This comment was automatically generated by Dr. CI (expand for details).

Please report bugs/suggestions to the (internal) Dr. CI Users group.

Click here to manually regenerate this comment.

[fmassa]

Great, need to fix lint :-)

[fmassa]

Ok, I'll have to fight with torchscript then :-)

[datumbox]

LGTM, thanks!

We could consider adding also on LastLevelMaxPool, LastLevelP6P7, ExtraFPNBlock but it's also OK without it. Can be merged once a scriptable operator for the logging is added as discussed offline.

[oke-aditya]

May I ask, why are we logging functions? I see such addition in many PRs by Francisco.

(If the reason is confidential to FB, please ignore)

Edit
Reason why I wanted to know is whether this some software development practice?

[fmassa]

Hi @oke-aditya

This function provides similar functionality as logging from the Python stdlib.

We log which methods are called for debugging purposes and it's used only in FB infra.

By no means this collects any data from our open-source users, and is pretty-much a no-op for external users.

This PR in PyTorch introduced this logging functionality, and has more information on the lower-levels of it pytorch/pytorch#20745

[datumbox]

Maybe @torch.jit.unused
and using
if not torch.jit.is_scripting():

[fmassa]

Yeah, I'm changing it now to use torch.jit.is_scripting() or torch.jit.is_tracing()

[oke-aditya]

Thank you for the information :)

My additional doubt was if this would cause additional overhead. But benchmarks indicate it does not affect (as per the PR you linked).

[NicolasHug]

it's used only in FB infra.
By no means this collects any data from our open-source users, and is pretty-much a no-op for external users.

Maybe we should make this crystal clear in the docstring of _log_api_usage_once, considering fb Meta's press coverage...

[github-actions]

Hey @fmassa!

You merged this PR, but no labels were added. The list of valid labels is available at https://github.com/pytorch/vision/blob/main/.github/process_commit.py