port tests for container transforms #8012
Conversation
| @@ -396,6 +396,8 @@ def check_transform(transform, input, check_v1_compatibility=True, check_sample_ | |||
| if check_v1_compatibility: | |||
| _check_transform_v1_compatibility(transform, input, **_to_tolerances(check_v1_compatibility)) | |||
|
|
|||
| return output | |||
There was a problem hiding this comment.
We already compute the output in check_transform. By returning it, we don't need to recompute in case the test performs additional checks.
| @@ -100,14 +100,15 @@ def _extract_params_for_v1_transform(self) -> Dict[str, Any]: | |||
| return {"transforms": self.transforms, "p": self.p} | |||
|
|
|||
| def forward(self, *inputs: Any) -> Any: | |||
| sample = inputs if len(inputs) > 1 else inputs[0] | |||
| needs_unpacking = len(inputs) > 1 | |||
There was a problem hiding this comment.
This got the same treatment as transforms.Compose in #7758. TL;DR: RandomApply now has the same UX as Compose in terms of passing packed or unpacked inputs.
| @@ -173,8 +174,9 @@ def __init__(self, transforms: Sequence[Callable]) -> None: | |||
| self.transforms = transforms | |||
|
|
|||
| def forward(self, *inputs: Any) -> Any: | |||
| sample = inputs if len(inputs) > 1 else inputs[0] | |||
| needs_unpacking = len(inputs) > 1 | |||
| # horizontal and vertical flip are commutative. Meaning, although the order in the transform is indeed random, | ||
| # we don't need to care here. | ||
| expected = F.vertical_flip(F.horizontal_flip(input)) | ||
|
|
||
| assert_equal(actual, expected) |
There was a problem hiding this comment.
Unfortunately it also means that this test doesn't check that the order of the transforms is indeed random. At best it checks that the input transforms are applied. I think we should at least acknowledge that in the comment instead of saying "we don't need to care".
There was a problem hiding this comment.
Fair enough. All ears if you have an idea to check whether the transforms are actually applied in random order.
There was a problem hiding this comment.
a pair of non-commutative transforms, with 2 hand-chosen seeds, asserting that the results are different. What makes it a bit harder is that the transforms themselves must be non-random. Worst case scenario we could just define Plus2 and Times2.
|
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 |
| input = make_image() | ||
|
|
||
| actual = check_transform(transform, input) | ||
| # We can't really check whether the transforms are actually applied in random order. However, horizontal and |
There was a problem hiding this comment.
I might be too pedantic but I wouldn't say we "can't", we just choose not to.
Reviewed By: vmoens Differential Revision: D50789102 fbshipit-source-id: d95eea91a6574a13c8d53023ab4b2fb65be62ea3
cc @vfdev-5