cleanup legacy transforms tests

[pmeier]

This is here to give us a better overview of what is left to port after #8011 and #8012. There are only a few test cases left in two files:

• test/test_transforms_v2_consistency.py: We have consistency tests against v1 as part of our regular v2 tests. Meaning, the only things left in here are transforms that are not part of v1. These are the transforms that are part of our reference scripts and became part of the v2 API. The tests for them are

  • vision/test/test_transforms_v2_consistency.py

    Line 441 in ee28bb3

    class TestRefDetTransforms:

  • vision/test/test_transforms_v2_consistency.py

    Line 546 in ee28bb3

    class TestRefSegTransforms:

  As for all the other consistency tests, these tests had the most value during the heavy development to avoid breaking consistency. Now that the API is stable and we need to keep BC anyway and the tests have been running smoothly for almost a year, their value is fairly low.

  Plus, Add --use-v2 support to classification references  #7724 and the sister PRs enabled the v2 transforms in the references. In the long term, we'll likely drop support for transforms v1 in the reference scripts.

  My vote is out for removing them completely.

• test/test_transforms_v2: A few "transform only", i.e. no dedicated functionals, testcases remain. They have to be ported, but that should be quick.

cc @vfdev-5

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/vision/8013

• 📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

✅ You can merge normally! (6 Unrelated Failures)

As of commit c7f6efb with merge base af3077e ():

FLAKY - The following jobs failed but were likely due to flakiness present on trunk:

• Tests / unittests-macos (3.8, macos-m1-12) / macos-job (gh)
• Tests / unittests-windows (3.11, windows.4xlarge, cpu) / windows-job (gh)
• Tests / unittests-windows (3.8, windows.4xlarge, cpu) / windows-job (gh)
• Tests / unittests-windows (3.9, windows.4xlarge, cpu) / windows-job (gh)

BROKEN TRUNK - The following jobs failed but were present on the merge base:

👉 Rebase onto the `viable/strict` branch to avoid these failures

• Build M1 Conda / pytorch/vision / conda-py3_8-cpu (gh)
• Build Macos Conda / pytorch/vision / conda-py3_8-cpu (gh)

This comment was automatically generated by Dr. CI and updates every 15 minutes.

[pmeier]

I grew a little tired of all the porting. The last commit is 95% copy-pasting what we had before. I'll highlight a few things that might stand out. Otherwise, there is nothing new in this PR. Meaning, unless there is something that you absolutely cannot stand in our previous tests, let's please address this in follow-up PRs.

After all tests were ported, I acted on

vision/test/transforms_v2_legacy_utils.py

Lines 6 to 13 in ddfee23

	The following legacy modules depend on this module
	- test_transforms_v2_functional.py
	- test_transforms_v2_consistency.py
	- test_transforms_v2.py
	When all the logic is ported from the files above to test_transforms_v2_refactored.py, delete
	all the legacy modules including this one and drop the _refactored prefix from the name.

and this is the cause of this massive diff.

[pmeier]

Here

[pmeier]

The only change here is that I merged three individual test functions into one class. The tests itself are copy-pasted.

[pmeier]

This was the only change I needed to make to be able to drop the whole transforms_v2_legacy_utils module. I'm ok with that trade, since we either won't touch this file or might kill them as well (see #8013 (comment)).

[pmeier]

Ok, I reverted this in 8d04cb2 since I forgot about the prototype tests. We need to port them of the legacy utils as well at some point, since these are not really legacy tests. But this for another PR.

[NicolasHug]

Thanks @pmeier ,

LGTM. Do we also need to add common checks (check_transforms, etc.) for all these test classes? Looks like we forgot to do it after TestLinearTransform?

[pmeier]

Do we also need to add common checks (check_transforms, etc.) for all these test classes? Looks like we forgot to do it after TestLinearTransform?

We should where applicable, yes. I didn't forget it, it is just not applicable to the transforms that came before this PR.

• check_transform unconditionally checks that the output type matches the input type, which by design doesn't work for the type conversion transforms
• check_transform unconditionally checks that the transform is serializable. That is not the case for the local function that we use to test transforms.Lambda.

For the tests ported in this PR we should be able to add them except for RandomIoUCrop. It requires a "sample input", i.e. at least an image and bounding boxes, but check_transform expects a single input. Will check that in a follow-up.

[github-actions]

Hey @pmeier!

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