L’interface est localisée ; les présentations et documents des Skills sont traduits automatiquement par IA dans votre langue. Le code, les commandes, chemins et termes techniques restent en version originale, et le document de l’auteur fait foi.
Explorer les Skills
Productivité · SKILL.md

irpe

irpe

Routes iRPE relative-position-encoding integration for DeiT and DETR workflows.

Independent GitHub snapshot · Automated classification · Original documentation governs

Documentation de la Skill

GitHub

name: irpe description: "Routes iRPE relative-position-encoding integration for DeiT and DETR workflows." disable-model-invocation: true metadata: disco-role: operating license: MIT

iRPE

Use this sub-skill when the user wants to equip a model with iRPE or work with the bundled DeiT-with-iRPE / DETR-with-iRPE implementations.

What this route owns

  • iRPE configuration generation and build_rpe wiring.
  • DeiT-with-iRPE classification workflows.
  • DETR-with-iRPE detection workflows.
  • Optional custom-op build guidance for the rpe_ops extension.

When to use it

Choose this route for prompts like:

  • "add iRPE to DeiT"
  • "run DETR with iRPE"
  • "build the RPE config"
  • "fix the relative-position encoding settings"
  • "check whether the custom RPE ops are available"

What to read next

  • references/api-reference.md for the verified get_rpe_config, build_rpe, and transformer entry points.
  • references/workflows.md for the integration steps and launcher shapes.
  • references/troubleshooting.md for rpe_ops, shape, and dataset issues.
  • scripts/build_irpe_config.py to print a ready-to-copy Python snippet for the RPE config.
  • scripts/build_irpe_command.py to print safe command templates.
  • ../../scripts/check_custom_ops.py to report whether the optional compiled extension is present.

Important boundaries

  • Do not route MiniViT here just because it uses iRPE internally; MiniViT has its own route.
  • Keep the custom-op guidance optional. The Python path is enough for inspection even when rpe_ops is absent.
  • Do not depend on the original checkout for runtime instructions; the generated skill must stand on its own.

Working pattern

  1. Identify whether the target is DeiT or DETR.
  2. Read the API reference for the model and transformer entry points.
  3. Use the config builder to generate the get_rpe_config snippet.
  4. Use the command builder to print the launcher template and then adapt it to the user's environment.

Common signals

  • get_rpe_config and build_rpe are the core integration helpers.
  • --enc_rpe2d is the DETR flag that carries the encoding choice.
  • rpe_ops missing is a warning for speed, not a complete failure of the Python path.
Skills enregistrées