---

name: transform-catalog description: "Select and parameterize Albumentations 2D transform families, including pixel, crop, resize, geometric, dropout, mixing/domain adaptation, spectrogram, text, target support, interpolation, and fill behavior." disable-model-invocation: true metadata: disco-role: operating license: MIT

Transform Catalog

Use this sub-skill when an agent needs to choose Albumentations 2D transforms, translate older parameter names to Albumentations 2.x, or decide whether a transform is safe for images, masks, segmentation, detection, keypoints, spectrograms, text rendering, or domain-adaptation workflows.

Route by task

• Use references/transform-reference.md for transform families, representative APIs, target support, v2 parameter names, interpolation, fill/fill_mask, and metadata requirements.
• Use references/recipes.md for copy-adaptable pipelines: classification, segmentation-safe geometry, detection/keypoints routing, normalization, dropout, domain adaptation, spectrograms, and text overlays.
• Use references/troubleshooting.md when constructor validation fails, old v1 names appear, masks get corrupted, dtype/range assumptions break, metadata is missing, text rendering fails, or OpenCV interpolation/fill constraints are unclear.
• Use scripts/transform_probe.py to inspect an installed Albumentations transform signature, map common deprecated parameter names, or run a tiny image/mask smoke probe.

Boundaries

• For Compose, ReplayCompose, probabilities across groups, strict validation, seeds, additional targets, and pipeline operator edits, route to ../pipeline-composition/.
• For bbox/keypoint formats, label fields, coordinate normalization, 3D volumes, and target validation, route to ../targets-and-formats/.
• For PyTorch tensor conversion, ToTensorV2, ToTensor3D, dataset placement, and channel order, route to ../framework-integration/.
• For saving/loading pipelines, replay dictionaries, and non-serializable transforms, route to ../serialization-and-reproducibility/.

Quick defaults

• For segmentation masks, prefer geometry transforms that declare mask support and keep masks nearest-neighbor: mask_interpolation=cv2.INTER_NEAREST and class-safe fill_mask values.
• For pixel-only color/noise/weather transforms, apply only to images; do not expect masks, bboxes, or keypoints to change.
• In Albumentations 2.x, use fill and fill_mask instead of old value and mask_value, and use tuple range parameters such as num_holes_range, hole_height_range, and hole_width_range instead of old min/max dropout names.
• For domain adaptation and mixing transforms, pass required per-sample metadata arrays through the transform call, not source-repo paths or lazy file handles.