---

name: pytorch-cycle-gan-and-pix2pix description: "Guide CycleGAN and pix2pix image-to-image translation, colorization, dataset preparation, checkpoint inference, and custom model or dataset extensions in this PyTorch repository." disable-model-invocation: true metadata: disco-role: operating license: NOASSERTION

pytorch-CycleGAN-and-pix2pix

Use this repo skill when a task involves the PyTorch CycleGAN/pix2pix repository, paired or unpaired image-to-image translation, colorization, train.py/test.py, dataset modes such as aligned or unaligned, pretrained generator checkpoints, or custom model/dataset templates.

Install and inspect first

This is a checkout-oriented project, not a pip distribution. Use the documented Python 3.11/PyTorch 2.4 baseline in references/installation.md, then run the bundled environment check against the target checkout:

python scripts/check_env.py --repo-root TARGET_CHECKOUT

The required verified scope is CPU-compatible. CUDA/DDP, Caffe Cityscapes evaluation, and HED edge extraction are optional external surfaces; do not treat a CPU import as GPU verification.

Read references/repo-provenance.md before deciding whether this skill matches the current repository state. Read references/troubleshooting.md for cross-cutting dependency, path, backend, and source caveats.

Route by task

• Train/test/apply models, resume training, checkpoints, W&B/HTML, CPU/GPU/DDP: read sub-skills/translation-workflows/SKILL.md.
• Prepare/download/validate data, combine A/B pairs, convert Cityscapes: read sub-skills/data-preparation/SKILL.md.
• Add a custom model or dataset, edit templates, debug registry naming or option injection: read sub-skills/customization/SKILL.md.

Core decision points

1. Choose the data mode before writing a command: CycleGAN uses unaligned, pix2pix uses aligned, one-sided generator application uses single, and colorization uses colorization.
2. Validate the data root with the data-preparation helper before a real train/test run.
3. Keep model architecture, normalization, channel counts, direction, and dropout/checkpoint suffixes aligned between training and testing.
4. The current parser has no --gpu_ids option: it auto-selects CPU when CUDA is unavailable and cuda:0 otherwise. Use a CPU-only PyTorch environment or prefix a Linux command with CUDA_VISIBLE_DEVICES= to hide GPUs. Use torchrun only after a bounded DDP setup smoke; the current source has a synchronized-normalization naming/guard inconsistency.
5. Treat network downloads, external Caffe/MATLAB workflows, and long training as explicit user-approved operations, not automatic verification steps.

Shared helper

scripts/check_env.py checks dependency imports, optional CUDA availability, repo module imports, and a tiny CPU generator smoke. It does not download assets, load checkpoints, train, or write result files.

For a self-contained command builder, use sub-skills/translation-workflows/scripts/build_command.py. For safe data preparation, use the scripts linked from sub-skills/data-preparation/SKILL.md.