How to achieve a single call to apply when having multiple image targets

[mikel-brostrom]

I recentrly created a PR with MixUp augmentation that look lite this:

class MixUp(DualTransform):
    """Generates a weighted combination of the pair of input images.

    Targets:
        image, image1, bboxes, bboxes1
    Image types:
        uint8, float32
    """

    def __init__(
        self,
        always_apply: bool = False,
        p: float = 1,
    ):
        super(MixUp, self).__init__(always_apply, p)

    def apply(self, image, image1, **params) -> np.ndarray:
        h1, w1, _ = image.shape
        h2, w2, _ = image1.shape
        if h1 != h2 or w1 != w2:
            raise TypeError("MixUp transformation expects both images to have identical shape.")
        r = np.random.beta(32.0, 32.0)  # mixup ratio, alpha=beta=32.0
        im = (image * r + image1 * (1 - r)).astype(np.uint8)
        return im
    
    def apply_to_bboxes(self, bboxes: Sequence[BoxType], bboxes1, **params) -> List[BoxType]:
        return bboxes + bboxes1

    def get_params_dependent_on_targets(self, params: Dict[str, Any]) -> Dict[str, Any]:
        return {
            "image1": params['image1'], 
            "bboxes1": params['bboxes1']
        }
    
    @property
    def targets_as_params(self) -> List[str]:
        return ["image", "image1", "bboxes", "bboxes1"]

However, apply is called twice as I have two targets (image and image1). How can I avoid this double call to apply? image and image1 are combined into a single one on the first call anyways. Is there any trick I can apply to achieve this @Dipet, @creafz ?

This would minimize the processing time and would be useful for implementing CutMix and Mosaic as well where the same is needed.

[mikel-brostrom]

Bump!

[mikel-brostrom]

I got an answer here: #1409 (comment).

Just copy pasting @i-aki-y's answer:

Albumentation has a mapping list from the argument key to the associated functions in the targets variable:

targets = {
    "image": apply_image,
    "bboxes": apply_bboxes,
    ...
}

The functions specified in the targets will be executed one by one when you apply trasform(image=image, bboxes=bboxes).

You have added new entries into the targets variable by using additional_targets

targets = {
    "image": apply_image,
    "bboxes": apply_bboxes,
    ...
    "image1": apply_image,
    ...
}

This means apply_image will be called twice; the first is for the "image", and the second is for the "image1".
I think this is the expected behavior when the additional_targets is used.

What will happen if you remove the "additional_targets"?