Weird behavior of LongestMaxSize

[Optimox]

Describe the bug

Hello me again,

I might be doing something wrong with the way I use kornia augmentations, please let me know if it is the case.

I was expecting LongestMaxSize in kornia to perform similarily as the albumentation implementation. Meaning that I can throw any images with different shapes to the the transformation function and get an image with different shapes but similar ratios. The largest size being equal to the value given to LongestMaxSize.

See bellow a small code sample that disturbs me.

Reproduction steps

import kornia.augmentation as K
a = torch.ones((512, 256))
b = torch.ones((512, 756))

print("first try")
transfo = K.LongestMaxSize(max_size=256, p=1.)

print(transfo(a).shape)
print(transfo(b).shape)

print("second try")

a = torch.ones((512, 256))
b = torch.ones((512, 756))

transfo = K.LongestMaxSize(max_size=256, p=1.)
print(transfo(b).shape)
print(transfo(a).shape)

Outputs:
first try
torch.Size([1, 1, 256, 128])
torch.Size([1, 1, 256, 128])
second try
torch.Size([1, 1, 173, 256])
torch.Size([1, 1, 173, 256])

Expected behavior

I would expect to have the same values for the transformations no matter the order of the elements.

ie transfo(a).shape == torch.Size([1, 1, 256, 128]) and transfo(b).shape ==torch.Size([1, 1, 173, 256])

Am I missing something here ?

Environment

kornia='0.6.9'
torch='1.12.1+cu113'

Additional context

No response

[Optimox]

Seems like Resize behave similarly when size is an integer.

[Optimox]

@johnnv1 @shijianjian Is this the expected behavior ?

[shijianjian]

@twsl Can you take a look?

[twsl]

Unfortunately I don't have much time right now, but a quick peek at the source code makes me wonder if the resize generator in the resize augmentation does include any checks to recalculate the parameters

[shijianjian]

@johnnv1 Can you help fix it if you got some time?

[johnnv1]

as @twsl commented, seems for this case we need to recompute the params -- i'm not sure of the best way to try avoid generating the params in each call