Albumentations to DALI Transform Mapping

On this page

This page maps Albumentations transforms to the closest DALI operation when DALI has a practical direct equivalent. Use it as a capability guide, not as a performance page.

How to Read This Page

  • A named DALI operation means DALI has a built-in graph operation for the same broad operation.
  • - means DALI does not support that transform as a built-in graph augmentation primitive.
  • The table shows built-in DALI operations. Custom DALI graph code is not counted as DALI support.
  • Unsupported rows stay visible because transform coverage is part of the library choice.

Execution Model

  • Albumentations receives an already decoded OpenCV-style H,W,C channel-last array and applies a Python augmentation pipeline to one sample and its targets.
  • DALI builds a data pipeline graph that can include reading, mixed-device image decoding, augmentation, normalization, output layout conversion, prefetching, and framework handoff.
  • DALI operators define graph nodes and symbolic data nodes. They are not direct per-sample Python augmentation calls.
  • DALI pipeline performance depends on the whole graph: decode, transfer scope, prefetching, framework handoff, and GPU memory.

Direct DALI Mappings

Albumentations transformDALI operation
Resizeresize
RandomCropresize plus crop
RandomResizedCroprandom_resized_crop
HorizontalFlipflip
VerticalFlipflip
Padpad
Rotaterotate
Affinewarp_affine
RandomBrightnessContrastbrightness_contrast
ColorJittercolor_twist
HueSaturationValuehue and saturation
CLAHEclahe
Equalizeequalize
GaussianBlurgaussian_blur
GaussNoisenoise.gaussian
SaltAndPeppernoise.salt_and_pepper
Erasingerase
ImageCompressionjpeg_compression_distortion
Normalizecrop_mirror_normalize

No Direct DALI Mapping

These categories are the practical reason DALI should not be presented as a general Albumentations replacement:

CategoryExamplesWhy it matters
Non-rigid geometry beyond the mapped subsetElasticTransform, GridDistortion, ThinPlateSpline, OpticalDistortionThese policies are common for robustness testing and domain-specific augmentation.
Weather, camera, and illumination effects outside simple color changesRandomRain, RandomSnow, AtmosphericFog, LensFlare, PlasmaShadowThese are augmentation policies, not only graph preprocessing steps.
Dropout and object-aware policiesCoarseDropout, GridDropout, ConstrainedCoarseDropout, CopyAndPasteThese often need target-aware bookkeeping, not only pixel operations.
Detection-safe crops and multi-image policiesBBoxSafeRandomCrop, RandomSizedBBoxSafeCrop, MosaicAlbumentations exposes these as augmentation policies with target contracts.
OBB and keypoint policiesOBB-safe transforms, keypoint visibility/filtering workflowsDALI does not provide the same high-level target-aware policy contract.
Volumetric and broad non-RGB workflowsRandomCrop3D, CubicSymmetry, CoarseDropout3DDo not generalize RGB image operators to volume or arbitrary-channel training.