Invert some keys

[rijobro]

Invert some but not all transforms.

To skip the inverse of RandAdjustContrastd, for example, this can be done with:

d_fwd = transforms(d)
d_fwd_inv = transforms.inverse_with_omissions(d, "RandAdjustContrastd")

Skipped transforms will be added to a list in the dictionary e.g., image_skipped_inverses.

Status

Ready

Types of changes

• Non-breaking change (fix or new feature that would not break existing functionality).
• New tests added to cover the changes.
• Integration tests passed locally by running ./runtests.sh -f -u --net --coverage.
• Quick tests passed locally by running ./runtests.sh --quick --unittests.
• In-line docstrings updated.
• Documentation updated, tested make html command in the docs/ folder.

[Nic-Ma]

Hi @rijobro ,

Thanks for the quick PR.
I have several comments here:

1. Maybe we should support more complicated cases, not only skip. So I would suggest to enhance the inverse() API directly instead of bringing new APIs. And as we are still not clear about all the use cases, I think we can provide the basic feature to just specify a list of transforms to invert by class name or index. If no list, invert all.
2. Could you please also help enhance the Invertd post transform? Almost all the users invert by this API.
3. As Invertd is heavily used now, the AdjustContrast transform will cause invert ing error, we must fix this blocking issue ASAP.

@wyli @ericspod What do you think?

Thanks.

[wyli]

shall we revert #2217 temporarily? we could include that after properly merging this PR.

[rijobro]

Sure, we can revert #2217, although I don't fully understand why we can't just temporarily manipulate image_transforms and label_transforms. If you remove AdjustContrastd from these lists, then they won't be inverted.

[rijobro]

@Nic-Ma I thought about modifying the inverse method, but the problem is that the interface should be the same for all InvertibleTransform classes, including Compose and all the others. The idea of skipping, however, only really makes sense for the Compose class, which is why I created a new method for it. Happy to discuss this tomorrow, though.

[rijobro]

Also @Nic-Ma I can't really modify Invertd until we make a decision here, since Invertd calls BatchInverseTransform, which eventually calls the inverse method. So first we need to agree on the inverse API before we can do Invertd.

[Nic-Ma]

Sure, let's discuss it online.

Thanks.

[madil90]

/build

[madil90]

/build

[madil90]

/build

[wyli]

this is addressed by Metatensor and the array level inverse.