Throw ValueError for wrong reduction in losses

[oke-aditya]

As discussed in #6457

The current problem

We do not restrict the reduction mode in losses. There are only three cases possible, ["none", "mean", "sum"]. If the user passes something like "xyz" we are not throwing any error or warning.

The solution

If users aren't passing one of the valid cases and using a random string instead. It's better to throw a ValueError suggesting them to use one of the correct choice.

We had thought of Enum based solution, but to keep better consistency with Pytorch core as discussed, we are adopting the above solution.

Edits required

Need to edit losses. which are namely
ciou_loss, diou_loss, giou_loss, and focal_loss

Also need to edit the relevant tests in test_ops.py

Additional Context

I think it's good first issue and some new contributor can take this 😃

Feel free to ping the maintainers or me if you need help. 😄

[NicolasHug]

Looking at https://github.com/pytorch/vision/pull/6457/files, it looks like this is taken care of in the original PR already? Or were you referring to other losses @oke-aditya ?

[datumbox]

I think @oke-aditya refers to all the other losses. I believe that's a good idea but if the ticket is intended for a new contributor, it's worth listing explicitly all the places were the change needs to happen and a potential idiom we want to adopt. This will help people who are not familiar with our code-base to pick it up.

[oke-aditya]

Yeah I'm referring to all the other losses. I will add further details to help new contributors 😃

[oke-aditya]

Well I hope I have edited the issue and explained it in bit more detail. I haven't given away the code of what needs to be done intentionally (as I believe some new contributor should try to figure out and interact with people while fixing)

cc @datumbox

[datumbox]

@oke-aditya Looks great to me! Thanks for helping us attract more contributors :)

[adityagandhamal]

I can work on this. Would you please assign me the task?

[oke-aditya]

@adityagandhamal
feel free to send PR and ping me as a reviewer. I can review.

Also btw how did you come to know about this issue?

[adityagandhamal]

I recently started using PyTorch and was browsing about, looking for anything to contribute to, and fortunately, I came upon this!

[datumbox]

@adityagandhamal Welcome! I've assigned the ticket to you, looking forward to your PR!

[adityagandhamal]

Yes. I've finished making the necessary changes to the code and am now working on the tests. Will send the PR soon.