Expose misc ops at package level

[jdsgomes]

cc @datumbox @NicolasHug

[facebook-github-bot]

💊 CI failures summary and remediations

As of commit 89c1c0b (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.

[datumbox]

LGTM, thanks!

[NicolasHug]

Thanks for the PR @jdsgomes ,

If we're going to expose these in the torchvision.ops namespace, I think we should also expose them in the docs in ops.rst. These ops seem to lack docstring though, so maybe we should write some

[oke-aditya]

We were going to expose using torchvision.nn ? Or is it decided with ops.misc.py ?

[datumbox]

@oke-aditya The reason we are not doing it at the moment is because there is no consensus for renaming/moving stuff to nn. This PR is just addressing the fact that the classes are public but not exposed directly via ops. Though the RFC you send is still valid, I would recommend not to link the two so that we don't block fixing outstanding issues.

[oke-aditya]

Can understand it :) thanks for the clarification.

[jdsgomes]

Thanks for the PR @jdsgomes ,

If we're going to expose these in the torchvision.ops namespace, I think we should also expose them in the docs in ops.rst. These ops seem to lack docstring though, so maybe we should write some

Do you suggest adding the docstrings in this PR as well or that is out of scope and we should create a ticket for it?

[NicolasHug]

It's a small PR so I would suggest to do it here. It's fine to do it in a separate PR, as long it's not a "open an issue for it and forget" thing :)

[datumbox]

@jdsgomes if you feel you don't have context to write the docs, feel free to open a ticket and I'll help.

Here are some info in case you want to give it a try:

The ConvNormActivation is a common configurable block used for Convolution-Normalzation-Activation blocks. We should document the default values along with what happens if None are passed for bn and activation (the layers are skipped).

The SqueezeExcitation is another common block first introduced at https://arxiv.org/abs/1709.01507. For this we just need to talk about the default values.

[datumbox]

Overall looks good. Just one comment on the stride size type. Let me know what you think.