Repeat augmentations with exactly same parameters on two different images

[abhishekaich27]

Hi,

How can I repeat the same transformation (with exactly the same parameters) on two different tensors/images?

For example, if I use aug = kornia.augmentation.RandomPerspective(0.2, p=1.0) then, the perspective on two different images img1 and img2 should look the same.

[edgarriba]

as described in this link https://kornia.readthedocs.io/en/latest/applications/image_augmentations.html#customization , you can do:

img1 = torch.rand(1,3,224,224)
img2 = torch.rand(1,3,224,224)

aug = K.RandomAffine(360, return_transform=True, same_on_batch=True)
img3, trans = aug(img1)
img4 = aug(img2, trans)

[abhishekaich27]

This doesn't seem to work on AugmentationSequential. I did the following:

from kornia.augmentation import AugmentationSequential

aug_forms = AugmentationSequential(
    kornia.augmentation.ColorJitter(0.1, 0.1, 0.1, 0.1, p=1.0),
    kornia.augmentation.RandomPerspective(0.2, p=1.0),
    return_transform=True,
    same_on_batch=False)

aug_img, trans = aug_forms(img_1)
aug_img_2 = aug_forms(img_2, trans)

This threw the following error:
AssertionError: The number of inputs must align with the number of data_keys. Got 2 and 1.

EDIT: Corrected a typo.

[edgarriba]

where is aug_list defined ? Notice that in this upcoming release we are deprecating return_transform and you will get it with aug_forms.transform_matrix. Checkout master for that. /cc @shijianjian

[abhishekaich27]

Corrected the typo. It should be aug_forms.

Notice that in this upcoming release we are deprecating return_transform and you will get it with aug_forms.transform_matrix
Thanks. But I have been using the old version with my codebase. Is there a way I can still use the return_transform and address the error I got?

[shijianjian]

Reproduce the augmentation is not about transformation matrix. The trans mat is only for geometric transforms. You need the whole set of parameters. Below code shall work.

from kornia.augmentation import AugmentationSequential

aug_forms = AugmentationSequential(
    kornia.augmentation.ColorJitter(0.1, 0.1, 0.1, 0.1, p=1.0),
    kornia.augmentation.RandomPerspective(0.2, p=1.0),
    same_on_batch=False)

aug_img = aug_forms(img_1)
aug_img_2 = aug_forms(img_2, params=aug_forms._params)

[domef]

Any news about this?

[edgarriba]

@abhishekaich27 is the solution from @shijianjian working for you ?