Release notes

Summary

This release adds custom apply targets for transforms, eight new transforms (explore.albumentations.ai), and performance improvements for PiecewiseAffine (vectorization), Affine and ISONoise on videos/volumes.

New features

Custom apply methods (apply_to_X)

Transforms can now define custom data targets that are automatically wired into Compose and ReplayCompose.

• Mixin: Use CustomTransformsApplyMixin before your base class (e.g. class MyAug(CustomTransformsApplyMixin, A.DualTransform)).
• Convention: Implement methods named apply_to_<key> (e.g. apply_to_label, apply_to_heatmap). They are discovered at init and registered for the corresponding data key.
• Behavior: Custom targets receive the same get_params output, respect p=, and work with replay. No need to subclass Compose or use ad-hoc pipelines.

Example:

class Rotate90WithLabel(A.CustomTransformsApplyMixin, A.DualTransform):
    def apply(self, img, k=0, **p):
        return np.rot90(img, k)
    def apply_to_mask(self, mask, k=0, **p):
        return np.rot90(mask, k)
    def apply_to_label(self, label, k=0, **p):
        return (label + k) % 4

aug = A.Compose([Rotate90WithLabel(p=1)], seed=42)
out = aug(image=img, mask=mask, label=my_label)  # label handled automatically

by @batiamida

New transforms

All support images and (where applicable) volumes/masks; see docstrings and explore.albumentations.ai for parameters and examples.

by @ternaus

Performance

Benchmarks below: baseline = commit 9aa0241 (pre–speedup release), same machine, back-to-back runs.

PiecewiseAffine (vectorization)

• Change: PiecewiseAffine pipeline was vectorized/optimized for single-image and batch usage.
• Measured speedup: ~200× for single-image apply (512×512×3, 10 iterations: 15.34 s → 0.076 s).

Affine on videos (volumes)

• Change: Affine implements apply_to_images (and thus efficient apply_to_volume/apply_to_volumes) with a single matrix for the whole batch and a pre-allocated loop.
• Measured: ~1.02× vs baseline (no regression; comparable).

ISONoise on videos (volumes)

• Change: ISONoise gained a vectorized apply_to_images (and hence efficient volume handling) so noise is applied in batch instead of per-frame.

• Measured speedups (volume 16×H×W×3, 30 iterations):

  Size	Baseline	Current	Speedup
  256²	0.83 s	0.28 s	2.9×
  512²	3.59 s	0.96 s	3.8×
  1024²	11.95 s	5.54 s	2.2×

by @Dipet

Other changes

• Docstring and documentation improvements.