Hi @rijobro ,

I have a concern about this PR:
If we use AdjustContrastd in the transform chain for image during inference, and want to invert pred based on the transform_info of image. How to avoid inverting this transform but invert other spatial transforms?

Thanks in advance.

The applied transforms are stored in a list, so we could simply manipulate the list as we need. Untested example might be like this:

def remove_transforms(transforms: list, to_be_removed: list):
    return [t for t in transforms if t["class"] not in to_be_removed]

transforms = remove_transforms(data["image_transforms"], ["AdjustContrastd"])

We could potentially formalise this using a withable or putting it in the Inverter logic:

ts = data["image_transforms"]
with InvertibleTransform.skipInverses(ts, ["AdjustContrastd"]):
    ts.inverse(data)

Hi @rijobro ,

Thanks for your proposal, but is there any use case to avoid only 1 of same transforms?
For example:

Resized(keys="img", spatial_size=[128, 128, 128]),
RandSpatialCrop(keys="img", random_size=True),
Resized(keys="img", spatial_size=[64, 64, 64]),

And only skip the first Resized.
So in summary, I am not sure we should skip transforms by class name or instance unique ID.
@wyli @ericspod , what do you think? Or any other ideas?
We also need to think about how to enhance the Invertd transform for this feature request.

Thanks.

I think there could be lots of different reasons for wanting to do the inverse of some, but not all of the applied transforms, so there won't be any "one size fits all" solution.

I don't think we often use multiple instances of the same transform, but If the user wanted to skip the first Resized, then they could just do it via indexing: d["image_transforms] = d["image_transforms][1:]. Random transforms are appended even if do_transform was false (below threshold probability), meaning the length of invertible transforms will remain constant and indexing can be safely used.

Hi @rijobro ,

Could you please help create a PR to enhance the Invertd transform for this issue?
Inverting is heavily used in our internal projects now, once users use AdjustContrastd transform, it will cause issue..

Thanks in advance.

Glad to hear that the inverse functionality is being used! Sure, I'll create a draft PR and we can discuss if the solution is acceptable there.

Cool, sounds good!
Thanks.

perhaps could have a SpatialTransform interface? because in many cases we want to invert the spatial ones only.

Similar to my suggestion here: #2213 (comment).

I wonder if we're getting a little too complicated -- what about transforms that are a mix of spatial and intensity, what about invertible transforms that are neither spatial or intensity (we don't have any yet, but may in the future).

Would it be worth scheduling a meeting for our transforms? We may want to unify certain things to make them more future-proof.

sure @rijobro let's have a Teams chat soon...

Hi @rijobro and @ericspod , sure, we certainly need to have an online discussion about how to make the inverting progress more flexible.

Thanks.