pipeline-composition

# Albumentations Pipeline Composition Use this sub-skill when an agent needs to assemble or modify an Albumentations pipeline, reason about pipeline-level validation, or debug pipeline randomness and target routing. ## Route Here For - Building `A.Compose` / `A.ReplayCompose` pipelines around existing transforms. - Choosing composition blocks: `OneOf`, `SomeOf`, `RandomOrder`, `Sequential`, `OneOrOther`, and `SelectiveChannelTransform`. - Enabling `strict`, `is_check_shapes`, `additional_targets`, `mask_interpolation`, `seed`, or `save_applied_params` on a pipeline. - Editing a pipeline with `+`, left-`+`, or `-` while preserving compose configuration. - Debugging invalid keys, shape mismatches, bad operator operands, seed expectations, label-field routing, or mask interpolation surprises. ## Route Elsewhere - Transform family choice and transform parameter recipes: `../transform-catalog/`. - Bbox, keypoint, label, volume, and 3D target formats: `../targets-and-formats/`. - Replay dictionaries, save/load, JSON/YAML, and reproducibility artifacts: `../serialization-and-reproducibility/`. - Tensor conversion and dataloader placement: `../framework-integration/`. ## References And Helpers - `references/composition-api.md`: API signatures, composition semantics, validation options, seed behavior, target routing, operator edits, and worked examples. - `references/troubleshooting.md`: Symptom-driven fixes for invalid keys, mismatched shapes, single-transform warnings, missing labels, seed confusion, operator failures, and mask interpolation issues. - `scripts/check_pipeline_contract.py`: A tiny self-contained checker that exercises strict keys, shape checks, additional depth targets, mask interpolation propagation, operator metadata preservation, and seed independence from global NumPy seeding. ## Default Composition Pattern Prefer an explicit, strict top-level `Compose` while developing, then relax only the checks that are intentionally incompatible with the data: ```python import albumentations as A pipeline = A.Compose( [ A.Resize(256, 256, p=1), A.OneOf([A.HorizontalFlip(p=1), A.VerticalFlip(p=1)], p=0.5), ], additional_targets={"depth": "mask"}, is_check_shapes=True, strict=True, mask_interpolation=0, seed=137, save_applied_params=True, ) result = pipeline(image=image, mask=mask, depth=depth) ``` Keep `Compose` calls keyword-only: use `pipeline(image=image, mask=mask)`, not `pipeline(image)`.