GaussNoise broken in 1.4.9 about albumentations

[daitakahashi]

In my observation, negative noise-values of the GaussNoise are truncated.

import numpy as np
import albumentations as A
x = np.full((10, 10), 128, dtype=np.uint8)
tr = A.Compose([A.GaussNoise(p=1)])  # for 1.4.8
tr = A.Compose([A.GaussNoise(p=1, noise_scale_factor=1.0)]  # for 1.4.9

1.4.8

(tr(image=x)['image'] >= x).all()  # -> almost always False

1.4.9

(tr(image=x)['image'] >= x).all()  # -> always True

[voegtlel]

Sorry, won't be able to test this until next Monday, OOO until then. Maybe the others can confirm. Anyways, thanks for the quick fix! 🥳

[JulianJvn]

We ran into the same issue, and it broke our CI.

• I think GaussNoise is not broken as stated in the PR title, but an additional parameter was introduced with a default value that changes the behavior. Looking at the code, you should be able to get the previous behavior by setting noise_scale_factor=1.0.
• The release notes do not mention that the default behavior changed (not even how the new parameter is called). Why?
• Can Albumentations switch to semantic versioning? Changing the default behavior is a breaking change which would merit a major version release under semantic versioning, not a patch version release.

[voegtlel]

@JulianJvn thanks for this finding! I also checked the release notes and didn't find this. Absolutely agree, semver should be used and reflect that breaking change.

[ternaus]

My bad.

I will put noise_scale_factor=1.0 as default, so that it would work as before.

[ternaus]

We ran into the same issue, and it broke our CI.

• I think GaussNoise is not broken as stated in the PR title, but an additional parameter was introduced with a default value that changes the behavior. Looking at the code, you should be able to get the previous behavior by setting noise_scale_factor=1.0.
• The release notes do not mention that the default behavior changed (not even how the new parameter is called). Why?
• Can Albumentations switch to semantic versioning? Changing the default behavior is a breaking change which would merit a major version release under semantic versioning, not a patch version release.

It is possible, but will require extra effort and attention.

Would you consider becoming a sponsor?

https://github.com/sponsors/albumentations-team

[ternaus]

@daitakahashi thank you. Will look at it asap.

[ternaus]

Made scale_factor=1 as default in #1800

[daitakahashi]

Here may be the cause. The value gauss should be a mean == self.mean array (possibly with some negative values).

albumentations/albumentations/augmentations/transforms.py

Line 1676 in ec8cb70

return {"gauss": gauss.astype(image.dtype)}

>>> tr.transforms[0].get_params_dependent_on_targets({'image':x})
{'gauss': array([[253,  11, 253,   3, 255,   2, 252, 253, 251, 254],
       [251, 251, 249, 253, 255,   1,   0,   5, 252,   3],
       [255,   1,   3,   9,   6, 249,   0,   1, 254,  11],
       [251,   2,   0,   1, 249, 253, 252,   0,   0,   5],
       [241,   4,   7,   1, 248, 244,   5,   0, 247,   0],
       [  0, 255, 253,  13,   5,   1,   3, 252, 249,   1],
       [252,   5,   0,   0,   0, 246, 240,   4, 253,   2],
       [  3,   5, 253,   7,   2, 251,   2, 252, 255, 250],
       [  0,   2, 253,  16,   2,   5,   3,   3,   4,   9],
       [  0, 254,   5,  17, 249, 254,  12, 255,   7,   8]], dtype=uint8)}

[ternaus]

Yep, it is this casting.

Adding tests

[ternaus]

Fixed in #1801

@voegtlel @daitakahashi @JulianJvn

Could you please verify that things work as expected? Default behavior is generate per pixel and noise is not truncated.

If no other issues will come up with it, I will make a hotfix release 1.4.10 in couple hours.

[ternaus]

Version 1.4.10 with fix for GaussNoise released.

Thanks @voegtlel @JulianJvn and @daitakahashi for help.