prevent feature wrapping if the feature is not the primary operand #6095
Conversation
| assert label_to.categories is label.categories | ||
|
|
||
|
|
||
| def test_to_feature_reference(): |
There was a problem hiding this comment.
This is the test case that makes sure we actually fix #6094.
| if not isinstance(args[0], cls): | ||
| return output |
There was a problem hiding this comment.
Although the behavior should be kept for all ops where we might wrap, right now this only applies to torch.Tensor.to. torch.Tensor.clone does not accept any inputs and thus this method will only be invoked in case the data to clone is actually a feature. All other ops will be handled by the else branch which performs no wrapping.
Thus, we could also merge this into the torch.Tensor.to branch for now, but will need to put it back where it is right now as soon as we special case another op that takes tensors as inputs. I would keep it were it is, but won't oppose putting it elsewhere for now.
|
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 |
…perand (#6095) Summary: * prevent feature wrapping if the feature is not the primary operand * explicitly add feature tests to CI Reviewed By: datumbox Differential Revision: D40138743 fbshipit-source-id: b5e523ce612b7380f8f9d11565f39b0fb6ef7b22
Fixes #6094.
Apart from the fix, this PR also adds a few minimal tests that I copied from the design doc, to ensure I'm not breaking anything else.