AlbumentationsX

AlbumentationsX 2.6.0

Summary

• Use float16 NumPy arrays and CPU PyTorch tensors for image and volume augmentation, including batches, masks, and reference data.
• Apply MixUp and input normalization after collation through albumentations.batch. Both operations run outside autograd and return tensors that can feed a trainable model.
• Configure ChannelDropout with an eligible channel set and a dropout ratio. Update older configurations that use channel_drop_range.
• Compose releases input storage promptly after successful calls and calls that raise. Custom samplers that accessed SamplingContext.invocation need the migration below.
• The release uses Albucore 0.3.5 for image kernels and raises the NumKong minimum to 7.8.7.

Breaking changes

ChannelDropout configuration

ChannelDropout replaces channel_drop_range with channels_to_drop and dropout_ratio. Omit channels_to_drop to make all channels eligible, or pass the channel indices that may be replaced. The ratio applies to the eligible set. The transform rounds half up and retains at least one input channel. Single-channel inputs pass through. The default ratio is 0.5, so set it explicitly when migrating a policy that relied on the old default of one dropped channel.

For an RGB input, both examples below select one of three channels. An old count range that sampled multiple possible counts requires a new policy choice because the ratio produces a fixed count for a given eligible set. Update saved configurations that contain the old constructor field. #589

import albumentations as A

# 2.5.1 used A.ChannelDropout(channel_drop_range=(1, 1), fill=0, p=0.5)
rgb_dropout = A.ChannelDropout(dropout_ratio=1 / 3, fill=0, p=0.5)

Custom sampling hooks

SamplingContext.invocation and the invocation= constructor argument have been removed. Read sampling.py_random and sampling.random_generator directly. Inspect inputs and target descriptors through the hook's data and targets arguments. For standalone tests or tools, use SamplingContext.from_seed() for an isolated stream or SamplingContext.from_owner() to draw from a configured transform. This keeps sampling state from holding input buffers after the call ends. Migration guide · #602

# Before
value = sampling.invocation.py_random.uniform(0, 1)

# After
value = sampling.py_random.uniform(0, 1)

Image type annotations and wavelength validation

• albumentations.core.type_definitions.ImageType has been removed. In custom transforms, annotate the dtypes that the operation accepts. For a NumPy operation that accepts all three image dtypes, use NDArray[np.uint8] | NDArray[np.float16] | NDArray[np.float32]. The pinned Albucore release also removes its ImageType alias and exposes convert_dtype for explicit conversions. #614
• WaterRefraction validates both endpoints of wavelength_range at construction. Fractions must lie in [0.001, 1.0]. Update older configurations outside this range. #612

New features

Batch operations for PyTorch training

The new albumentations.batch module operates on collated tensors such as (N, C, H, W) images and (N, C, D, H, W) volumes. It runs separately from per-sample Compose pipelines. [Batch Augmentations guide](https://github.com/albumentations-team/albumentations-ai-docs/blob/main/docs/4-advanced-guides/

AlbumentationsX 2.5.1

Summary

• Faster MRI augmentation with MotionArtifact, RicianNoise, KSpaceSpikeNoise, and BiasField.
• Lower temporary-array usage for eligible Rician noise and bias-field reconstruction, with channel-independent scratch buffers for shared k-space spikes.
• Regression coverage for MRI replay, strided arrays, backend selection, and direct volume batches.
• One source for development dependencies and automated license-review synchronization for eligible Dependabot lockfile updates.

Breaking changes

No public API or minimum Python-version changes; Python 3.11 or higher remains required. Optimized MRI kernels can introduce floating-point roundoff differences. Motion and Rician reconstruction can differ from 2.5.0 by one intensity level for uint8 outputs; seeded replay remains deterministic within this version.

For contributors, requirements-dev.txt has been removed. Use the groups in pyproject.toml:

uv sync --locked --group dev --inexact

Install the PyTorch build appropriate for your environment. For the CPU-only CI profile:

uv sync --locked --group dev --group ci-torch-cpu --inexact

The pip fallback is python -m pip install -e . --group dev with pip 25.1 or higher. See development setup.

New features

MRI performance

• MotionArtifact computes FFTs along the acquisition axis, eliminating two spatial FFT axes. Large writable single-channel volumes use CPU PyTorch FFTs; other inputs use SciPy. #577
• RicianNoise fuses magnitude reconstruction for eligible large float32 arrays and removes a full-size imaginary-square temporary. #577
• KSpaceSpikeNoise combines separable coefficients for up to five shared spikes, using published Albucore 0.2.21 operations. Coefficient storage is independent of channel count, and each allocating addition is bounded to one XY plane. #578
• BiasField interpolates and applies gains one plane at a time for eligible float32 volumes, avoiding a full-volume gain field. Other shapes, dtypes, and field configurations retain their established routes. #578

Measured through the allocating Compose route on Apple M4 Max, exactly one CPU thread, NumPy DHWC float32 volumes of shape 32 × 256 × 256 × 1:

These are the paired measurements reported in #577 and #578, which include exact revisions, parameters, methods, and complete comparison matrices. They u

AlbumentationsX 2.5.0

Summary

• Adds four medical imaging transforms: BiasField, GhostingArtifact, GibbsRinging, and MotionArtifact.
• Raises the minimum supported Python version to 3.11.
• Fixes RandomGamma so gamma 1 preserves every uint8 value.
• Cuts FFT work and temporary memory for selected large MRI volume cases.

Breaking changes

• Python 3.10 is no longer supported. Install and run AlbumentationsX with Python 3.11 or newer.
• Medical transform implementations and functional helpers now live in albumentations.augmentations.medical.transforms and albumentations.augmentations.medical.functional. Code importing implementation modules should update its import path. For example:

# Before
from albumentations.augmentations.pixel.color_advanced import HEStain

# After
from albumentations.augmentations.medical.transforms import HEStain

Top-level exports such as A.HEStain retain their constructor parameters and serialization identifiers.

New features

• BiasField simulates smooth multiplicative intensity variation in images and 3D volumes, with shared or per-channel fields.
• GhostingArtifact simulates periodic MRI ghost replicas in real 3D magnitude volumes.
• GibbsRinging applies rectangular k-space truncation to 2D images and 3D volumes.
• MotionArtifact simulates motion between MRI acquisition segments in 3D volumes.
• The new medical package groups these transforms with Anisotropy3D, HEStain, KSpaceSpikeNoise, and RicianNoise. Their top-level exports, target behavior, and serialization remain intact.

Bug fixes

• RandomGamma now builds its uint8 lookup table from exact integer pixel values. With gamma 1, all 256 uint8 values remain unchanged across images and volumes.
• Corrects the documented gamma equation and brightening/darkening directions, and adds a runnable numeric example.

Misc

• For large 3D volumes, KSpaceSpikeNoise, GibbsRinging, and BiasField reduce FFT computation or temporary memory in selected parameter, shape, and dtype ranges. Other inputs keep the existing execution paths.
• Updates the supported Python and operating-system matrix, contributor guidance, release metadata, and development dependencies.
• Keeps release-core ASV discovery compatible with 2.4.13 and corrects pinned artifact action version comments in CI workflows (#571, #573).

Commits

What's Changed

• fix(release): pin verified release tags and bump to 2.4.13 by @ternaus in https://github.com/albumentations-team/AlbumentationsX/pull/556

Full Changelog: https://github.com/albumentations-team/AlbumentationsX/compare/2.4.12...2.4.13