---

name: domain-adaptation description: "Router for Transfer-Learning-Library domain adaptation workflows, losses, and optional benchmark stacks." disable-model-invocation: true metadata: disco-role: operating license: MIT

Domain Adaptation

Use this sub-skill for TLLib domain-adaptation workflows, loss wiring, and benchmark recipe selection.

What it covers

• Closed-set feature alignment: DANN, CDAN, ADDA, DAN, JAN, CORAL, BSP, MCD, MDD.
• Partial and open-set adaptation: PADA, IWAN, OSBP, class-weight reweighting, unknown-class thresholding.
• Special DA variants: AFN, MCC, RegDA keypoint, D-adapt object detection, and WILDS DA examples.

Route elsewhere when

• Datasets, ImageList formats, model factories, or transforms are the main question: ../vision-data-models/SKILL.md
• Translation, CycleGAN, FDA, or style-transfer is the main question: ../translation/SKILL.md
• SSL, pseudo-labeling, or teacher/student training is the main question: ../self-training/SKILL.md
• DG, fine-tuning, or standalone transfer regularization is the main question: ../task-generalization/SKILL.md

How to use

1. Identify the transfer setting: closed-set, partial, open-set, regression, keypoint, detection, or WILDS.
2. Pick the matching API or workflow reference in references/.
3. For CPU-safe validation, use scripts/tllib_domain_adaptation_smoke.py.
4. Treat object-detection and WILDS coverage as optional-stack guidance unless the user explicitly has those dependencies.

Notes

• AFN appears here because it is used in DA workflows, but its standalone module mechanics live in the task-generalization / normalization path.
• MCC appears here because it is used in DA workflows, but its standalone semi-supervised mechanics live in the self-training path.
• Do not point users back to the source checkout or benchmark launchers; keep them on the bundled references and smoke helper.
• Keep benchmark-scale training, downloads, and optional dependency setup out of the smoke path.

Reference map

• references/api-reference.md - public modules, losses, shapes, and common traps.
• references/domain-adaptation-workflows.md - classifier, regression, keypoint, and WILDS workflow patterns.
• references/object-detection-adaptation.md - D-adapt, Detectron2, proposal and feedback flow, optional stack notes.
• references/wilds-workflows.md - WILDS-specific CLI and dependency patterns.
• references/troubleshooting.md - shape, NumPy, TorchVision, CUDA, Detectron2, and WILDS issues.
• scripts/tllib_domain_adaptation_smoke.py - tiny CPU tensor smoke using installed tllib.