Refactored and fixed flaky resize tests

[vfdev-5]

Addresses #3906 (comment)

Description:

• Refactored and fixed flaky resize tests

I refactored the test using pytest and could find seed value with MAE larger 8. The case when it is failing looks like to be related to max_size=33. I suppose that odd values are not handled in the same way between PIL and pytorch. If I use max_size=34 I could not find a seed value between 0 to 2000 such that MAE > 8.0.

[pmeier]

Minor comment below. Otherwise LGTM! Thanks Victor.

[NicolasHug]

Thanks @vfdev-5 ! Some minor comments but LGTM mostly

[NicolasHug]

do we have to set the seed? If yes maybe add a comment with a potential FIXME?

[vfdev-5]

I think we should create deterministic and random input instead purely random one

[NicolasHug]

OK, we generally don't do that, but I agree with you that it's better practice when we know the test aren't flaky.

Another option is to actually parametrize the seed, which we can do if the tests are very fast.

[vfdev-5]

Otherwise I would put it module-wise. Seed as parameter is something new to me :)

[NicolasHug]

I think this is already setting the seed module-wise, at least for the tests that will be run after this one.

In the long term we'll use something local like rng = np.random.RandomState(0) (surely there's a pytorch equivalent?) and generate stuff from the local rng variable, but for now this is good enough

[vfdev-5]

In the long term we'll use something local like rng = np.random.RandomState(0) (surely there's a pytorch equivalent?) and generate stuff from the local rng variable, but for now this is good enough

Yes, something like that : https://pytorch.org/docs/stable/generated/torch.Generator.html#torch.Generator.manual_seed

[pmeier]

For all tests that do not perform a loop and thus only need one seed per test invocation, we could use an autouse fixture that is defined in the root conftest.py

import pytest

@pytest.fixture(scope="function", autouse=True)
def random_seed():
    torch.manual_seed(12)

[datumbox]

LGTM! Thanks for your quick fixes.