---

name: translation description: "Operate TLLib image/domain translation and style-transfer components: CycleGAN, FDA, CyCADA, and SPGAN." disable-model-invocation: true metadata: disco-role: operating license: MIT

Translation Sub-skill

Use this sub-skill when a task involves TLLib image/domain translation or style transfer components rather than classifier-only adaptation: CycleGAN generators/discriminators/losses, the Translation transform, FDA/Fourier amplitude transfer, CyCADA semantic consistency, or SPGAN Siamese/contrastive components.

Route first

• Need unsupervised or supervised domain adaptation training after translated images/features exist? Route to ../domain-adaptation/SKILL.md.
• Need image-list formats, dataset roots, segmentation/re-id/detection dataset conventions, transforms, model factories, or metrics? Route to ../vision-data-models/SKILL.md.
• Need only component API signatures, tensor shapes, and checkpoint-safe loading? Use references/api-reference.md.
• Need a safe workflow for choosing CycleGAN vs FDA vs CyCADA vs SPGAN? Use references/translation-workflows.md.
• Need to debug amplitude caches, PIL/tensor ranges, generator checkpoints, full-training side effects, or GPU/data requirements? Use references/troubleshooting.md.

Safe operating sequence

1. Confirm tllib imports successfully in an environment compatible with TLLib 0.4-era PyTorch/TorchVision.

2. Run the bundled component smoke before editing a user workflow:

   python scripts/tllib_translation_smoke.py
   

   It uses tiny synthetic tensors/images, creates only temporary files, and does not download data.

3. Choose the smallest translation mechanism:

   • FDA for lightweight target-style amplitude transfer, especially segmentation preprocessing.
   • CycleGAN Translation for applying an already trained generator to PIL images or dataset items.
   • CyCADA SemanticConsistency when semantic labels/predictions must constrain image translation.
   • SPGAN Siamese/contrastive pieces when preserving person re-id identity similarity matters.

4. Keep benchmark-scale CycleGAN/FDA/SPGAN training as optional, data/GPU-heavy work. Do not imply it was verified by the smoke script.

What this sub-skill does not claim

• It does not provide dataset download instructions or benchmark reproduction guarantees.
• It does not validate Detectron2/MMCV object-detection stacks, segmentation datasets, re-id datasets, or CUDA training throughput.
• It does not require opening the original repository examples at runtime; the references here are self-contained distilled usage notes.