fix(augment): align Kornia GaussianBlur/GaussNoise defaults with albumentations

[adhavan18]

What

Aligns two Kornia augmentation-factory defaults with the Albumentations defaults they stand in for. Closes #1351.

Why

For the same aug_config, switching augmentation_backend silently changed augmentation strength, because two Kornia defaults did not match Albumentations':

GaussianBlur sigma. AUG_INDUSTRIAL sets {"GaussianBlur": {"blur_limit": 3, "p": 0.3}} with no sigma, so both backends fall back to their own default:

GPU (kornia)         sigma = (0.1, 2.0)   <- Kornia's convention
CPU (albumentations) sigma = (0.5, 3.0)   <- what the config actually means

The call-site comment claimed it "Match[es] the CPU albumentations default", which was the opposite of what it did.

GaussNoise std_range. Defaulted to rf-detr's own (0.01, 0.05) against the Albumentations default (0.2, 0.44) on the same 0-1 image scale, 4-9x weaker. The divergence warning also reported rf-detr's own [0.01, 0.05] as if it were Albumentations', hiding the gap.

Verified against the pinned albumentations==2.0.8:

>>> A.GaussianBlur().sigma_limit
(0.5, 3.0)
>>> A.GaussNoise().std_range
(0.2, 0.44)

Changes

• _make_gaussian_blur default sigma (0.1, 2.0) -> (0.5, 3.0), comment corrected.
• _make_gauss_noise default std_range (0.01, 0.05) -> (0.2, 0.44). An explicit range still wins, and the warning text is unchanged since it is still correct for a non-degenerate explicit range.
• Doc table in aug_configs.py updated.

This changes only the default used when a config omits the value; any config that sets sigma/std_range explicitly is unaffected. It follows the same "align to Albumentations" direction accepted for the sibling scalar-clip_limit fix (#1350).

Testing

Three regression tests pin both defaults to the live Albumentations values. The two that assert the defaults fail on develop and pass here:

develop:     unspecified GaussianBlur sigma is (0.1, 2.0) ... FAILED
             unspecified GaussNoise std is 0.05 ...          FAILED
this branch: 8 passed (Gauss* subset)

Full augmentation suite: 55 passed. ruff check and ruff format --check clean.

[Copilot]

Pull request overview

Aligns Kornia Gaussian augmentation defaults with Albumentations while preserving explicit configuration values.

Changes:

• Updates GaussianBlur and GaussNoise defaults.
• Adds regression tests against live Albumentations defaults.
• Updates GPU augmentation documentation.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 2 comments.

File	Description
src/rfdetr/datasets/kornia_transforms.py	Aligns Gaussian defaults.
src/rfdetr/datasets/aug_configs.py	Documents the new blur default.
tests/datasets/test_kornia_transforms.py	Tests default alignment and warning behavior.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 86%. Comparing base (207a0bb) to head (4a3e10a).

Additional details and impacted files

@@           Coverage Diff           @@
##           develop   #1395   +/-   ##
=======================================
  Coverage       86%     86%           
=======================================
  Files          112     112           
  Lines        14664   14669    +5     
=======================================
+ Hits         12623   12628    +5     
  Misses        2041    2041           

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[adhavan18]

Thanks for picking this up and polishing it @Borda, the reworded GaussNoise note is exactly right: bounds aligned, but Kornia's fixed std is not the CPU path's sampled range, so "same default, not same distribution" is the precise claim.

Happy to take anything else you'd like on it, otherwise I'll leave it in your hands.