feat(augment): support ShiftScaleRotate on the Kornia backend

[adhavan18]

What

Adds ShiftScaleRotate to the Kornia GPU augmentation backend. Follow-up to #1330 and #1277, same issue (#1252).

Why

ShiftScaleRotate is one of the names aug_configs documents and transforms.py accepts on the CPU path, while the Kornia registry rejects it. Backend selection is automatic, so the same aug_config trains on a CPU box and raises on a GPU box:

kornia ShiftScaleRotate -> ValueError: Unknown augmentation key 'ShiftScaleRotate' for Kornia GPU backend
albu   ShiftScaleRotate -> OK

It is safe to add for the same reason Perspective was: it preserves the output resolution. The GPU path rebuilds each batch as NestedTensor(img_aug, samples.mask), reusing the pre-augmentation padding mask, so a transform that resizes leaves the mask describing a different shape. That is why the crops stay unsupported, and it does not apply here.

I should correct something I wrote in #1330 while I am here. The doc block excluded this name along with the crops, on the grounds that the group "move boxes and masks". That reason does not hold for this one: Affine moves boxes too and has always been supported. Resolution change is the real constraint, and ShiftScaleRotate does not change resolution.

The part that is not a pass-through

Albumentations calls this "a special case of Affine transform", so an alias to _make_affine looks obvious. It would be wrong, because the limits are deltas rather than absolute ranges.

scale_limit is biased by 1 — from the Albumentations docstring:

Note that the scale_limit will be biased by 1. If scale_limit is a tuple, like (low, high), sampling will be done from the range (1 + low, 1 + high).

So the documented default (-0.1, 0.1) means a scale between 0.9 and 1.1. Kornia's scale is the absolute multiplier, so forwarding the raw value would ask it to scale the image to between a tenth of its size and nothing at all. Confirmed against the installed Albumentations rather than trusting the docstring:

>>> A.ShiftScaleRotate(scale_limit=0.1).scale_limit
(0.9, 1.1)

The pivot is applied here, the same way _make_sharpen already pivots alpha at 1.0.

Two smaller differences in the same vein:

param	Albumentations	mapped as
scale_limit	delta biased by 1	scale=(1 + low, 1 + high)
shift_limit	signed fraction range	translate, Kornia's non-negative per-axis maximum, as _make_affine already does for translate_percent
shift_limit_x / shift_limit_y	per-axis override	Kornia's translate=(tx, ty) expresses this directly
rotate_limit	degrees	degrees

All three limits also read a scalar v as the symmetric (-v, v), not the degenerate (v, v) that _as_range gives, so they go through a small _as_symmetric_range helper instead.

interpolation, border_mode, mask_interpolation, fill, fill_mask and rotate_method have no RandomAffine equivalent and are ignored with a warning, matching how ToGray and Sharpen already report their dropped parameters.

One thing worth your call

Albumentations deprecates this name. Constructing it emits:

UserWarning: ShiftScaleRotate is a special case of Affine transform. Please use Affine transform instead.

I mapped it anyway because the CPU path in this repo still accepts it, so the parity gap is real for any config already using it, and albumentations<3.0.0 is pinned. But if you would rather close this name out of aug_configs and point users at Affine — which this backend already supports — that is a reasonable call and I am happy to send that PR instead. The docs note added here says new configs should prefer Affine.

Tests

14 cases in TestShiftScaleRotateFactory, covering the delta semantics (scalar and asymmetric pair), symmetric scalar expansion, the Albumentations defaults, per-axis shifts, each ignored option warning, no spurious warning on an ordinary config, resolution preservation, and boxes moving with the image.

Reading the resolved ranges needs RandomAffine._param_generator — flags carries only the resampling options — so there is a helper that fails with an actionable message if a future Kornia release changes that shape, mirroring the existing _sharpness_sampler_range.

121 passed

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 86%. Comparing base (6674d85) to head (5dd38d2).

Additional details and impacted files

@@           Coverage Diff           @@
##           develop   #1370   +/-   ##
=======================================
  Coverage       86%     86%           
=======================================
  Files          114     114           
  Lines        14880   14896   +16     
=======================================
+ Hits         12835   12851   +16     
  Misses        2045    2045           

🚀 New features to boost your workflow:

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

[adhavan18]

Bumping this since it's been quiet for a couple weeks — just rebased onto current develop, CI green across the full matrix, no conflicts. Happy to address anything if it needs another look.

[Borda]

Hi, apologies for delay, I was off last week but I have on my radar for later this week...