Targets that should be ignored

[daitakahashi]

Feature description

We need a way to declare a target that will be ignored by Compose-wrapped transforms.

Motivation and context

Let say, we have a following data and Compose-wrapped transforms.

data = {'image': image, 'mask': mask, 'text_description': 'image description'}
transform = A.Compose([A.Blur()])  # note: A.Blur() is image only

What we want to do is applying the transformation on image and mask but text_description without manipulating the data.

transformed = transform(**data)
transformed['text_description'] == 'image description'  # unchanged

Before #1706, Compose simply ignored all unknown keys. As a result, the above code worked "as expected" (as a byproduct of an issue solved by the pr).

However, current transforms check args by default.

>>> transformed = transform(**data)
Traceback (most recent call last):
...
ValueError: Key text_description is not in available keys.

I expected that we have a way to declare the text_description being ignored, e.g., transform.add_targets({'text_description': 'ignore'}). But in the current albumentations, this kind of declaration does not work and raises ValueError when it is called.

>>> transform.add_targets({'text_description': 'ignore'})  # or anything other than 'image'
>>> transformed = transform(**data)
Traceback (most recent call last):
...
ValueError: Key text_description is not in available keys.

Alternatives

A currently available workaround is setting is_check_args of the Compose object to be False. The workaround skips all argument checks added by #1706.

[ternaus]

Thanks.

Applying transforms, without wrapping with Compose, typically leads to enexpected bugs => better to always wrap with compose.

At the same time adding is_check_args parameter to transforms looks quite doable, and if user knows that they are doing - why not have such functionality?

Will work on this.

[ayasyrev]

Do you use single transform or inside Compose?
Single transform don`t check keys - just ignores unused keys.

[daitakahashi]

Ayasyrev is right. I found that I have always wrapped transforms with Compose. So the situation I described in this issue is limited to transforms that are wrapped with Compose. I will update this issue to make sure it.

[daitakahashi]

By the way, this is a bit unintuitive for me (albumentations 1.4.8).

>>> tr = A.Compose([A.Blur()])
>>> tr.add_targets({'mask2': 'mask'})
>>> out = tr(image=np.zeros((2, 2, 3)), mask2=np.zeros((2, 2, 3)))
Traceback (most recent call last):
...
ValueError: Key mask2 is not in available keys.
>>> tr = A.Compose([A.Flip()])
>>> tr.add_targets({'mask2': 'mask'})
>>> out = tr(image=np.zeros((2, 2, 3)), mask2=np.zeros((2, 2, 3)))
>>>

I encountered it when I compiling an example case of this issue.

[ternaus]

It does not look right to me eigher.

[daitakahashi]

I feel that checking against _available_keys seems to be too restrictive as _available_keys depends on transforms inside and key-permissive nature of those transforms.

[ayasyrev]

I encountered it when I compiling an example case of this issue.

This is not OK. Will fix it,

[ayasyrev]

By the way, this is a bit unintuitive for me (albumentations 1.4.8).

>>> tr = A.Compose([A.Blur()])
>>> tr.add_targets({'mask2': 'mask'})
>>> out = tr(image=np.zeros((2, 2, 3)), mask2=np.zeros((2, 2, 3)))
Traceback (most recent call last):
...
ValueError: Key mask2 is not in available keys.
>>> tr = A.Compose([A.Flip()])
>>> tr.add_targets({'mask2': 'mask'})
>>> out = tr(image=np.zeros((2, 2, 3)), mask2=np.zeros((2, 2, 3)))
>>>

I encountered it when I compiling an example case of this issue.

The reason why you get error here - Blur is ImageOnly transform, so masks not processed here and add some additional key for mask make not many sense. But as mask is special case, I'll add this possibility.

[daitakahashi]

@ayasyrev Yes, Blur() is imageonly, I understand it. I'm concerning a case of trying out several combinations of transformations to a predefined dataset (I often do such experiments, so my concern may be biased). Anyway, I want the transforms not to crash regardless of the data format and algorithms inside.
A.Compose([A.Flip(p=0), A.Blur()]) will work for most of data if we declare key-target pairs by add_targets(), because Flip knows mask, bboxes, and keypoints. But this doesn't look nice.

[daitakahashi]

I mean the example in the documentation may not work if we use an ImageOnlyTransform, e.g., GaussianBlur, instead of the HorizontalFlip (note: I haven't tried yet).