---

name: augmentation-pipelines description: "Use Kornia augmentation pipelines for random or deterministic image, mask, box, keypoint, class, patch, and video transforms." disable-model-invocation: true metadata: disco-role: operating license: NOASSERTION

Kornia Augmentation Pipelines

Use this sub-skill when a task asks for Kornia kornia.augmentation random or deterministic transforms, synchronized image/mask/box/keypoint/class augmentation, deterministic replay of random parameters, augmentation transform matrices, inverse augmentation, or augmentation containers such as AugmentationSequential, ImageSequential, PatchSequential, and VideoSequential.

This skill assumes Kornia 0.9.0rc1 with Python >=3.11, PyTorch >=2.0, numpy, packaging, and kornia-rs available. Optional ONNX, transformers, diffusers, and ivy integrations may be absent. Augmentation image tensors should be float PyTorch tensors in [0, 1], usually C,H,W or B,C,H,W.

Read first

• For stable call shapes, constructor signatures, and selected classes, read references/api-reference.md.
• For task recipes, read references/workflows.md.
• For data_keys, input order, masks, boxes, keypoints, class labels, transform matrices, and inverse behavior, read references/data-keys-and-matrices.md.
• For common failures and repair actions, read references/troubleshooting.md.
• To verify a no-download Kornia augmentation runtime, run scripts/augmentation_smoke.py.

Fast routing

Choose this sub-skill for:

• AugmentationSequential pipelines with data_keys such as "input", "mask", "bbox", "bbox_xyxy", "bbox_xywh", "keypoints", "class", or "label".
• Random geometric image transforms such as RandomAffine, RandomPerspective, RandomResizedCrop, RandomHorizontalFlip, and RandomVerticalFlip.
• Random intensity transforms in an augmentation chain, especially ColorJiggle and related image-only color/photometric perturbations.
• Container behavior: same_on_batch, keepdim, random_apply, random_apply_weights, deterministic replay with cached or supplied params, transform matrices, and inverse.
• Patch and video augmentation routing through PatchSequential and VideoSequential.

Route away when the main problem is not augmentation orchestration:

• Low-level deterministic resize, affine/perspective warp construction, homographies, camera geometry, or coordinate-system reasoning: geometry-vision.
• General non-random image processing, filters, color conversions, morphology, enhancement, or tensor I/O conventions: image-processing.
• Feature detection/matching, learned local features, LoFTR/LightGlue/DISK, or model weights: features-and-matching.
• Losses or metrics for training/evaluation: losses-and-metrics.

Operating rules

1. Keep image tensors floating point in [0, 1] before augmentation. Convert uint8 images with division by 255.0; do not pass raw integer images to random color or geometric augmentations.
2. Use B,C,H,W for ordinary image batches. Single C,H,W images are accepted by some modules, but B,C,H,W is safer for multi-target pipelines and matrix checks.
3. In AugmentationSequential, make the number and order of positional inputs exactly match data_keys; the first key must be image/input when random parameters are generated from the image.
4. Prefer string keys in user-facing examples: data_keys=["input", "mask"]. "image" and "input" are aliases; "class" and "label" are aliases.
5. Use p=1.0 for deterministic smokes and required transforms; set explicit ranges such as degrees=0.0, scale=(1.0, 1.0), or fixed crop sizes when a test must be reproducible.
6. Use same_on_batch=True only when every batch item must receive the same sampled parameters. Leave it False for independent per-item stochastic augmentation.
7. Use keepdim=True when downstream code expects the same rank/shape as the input; leave defaults only when the target API accepts Kornia's batch-form normalization.
8. Use extra_args only for per-data-key overrides. Masks default to nearest interpolation semantics in AugmentationSequential, which is usually what segmentation masks need.
9. Treat .transform_matrix and cached params as per-call eager state. They are useful for diagnostics, replay, and inverse, but not a stable serialization format for deployment.
10. For export/deployment, prefer deterministic preprocessing transforms and single-input image pipelines. Random multi-target propagation is best kept in eager training or data-loading code unless the target exporter is proven with a focused smoke.