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

core-ocr-and-kie

core-ocr-and-kie

Run docTR end-to-end OCR and KIE inference with predictor APIs, rotation/layout/table options, batching, devices, and output checks.

Independent GitHub snapshot · Automated classification · Original documentation governs

Documentation de la Skill

GitHub

name: core-ocr-and-kie description: "Run docTR end-to-end OCR and KIE inference with predictor APIs, rotation/layout/table options, batching, devices, and output checks." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

core-ocr-and-kie

Use this sub-skill when the task is to run or debug docTR's Python end-to-end inference pipelines:

  • doctr.io.DocumentFile or a list of RGB NumPy pages -> doctr.models.ocr_predictor(...) -> Document.
  • doctr.io.DocumentFile or a list of RGB NumPy pages -> doctr.models.kie_predictor(...) -> KIEDocument.
  • Predictor-level flags for rotated/skewed pages, orientation/language detection, layout detection, table detection, ignored layout regions, batching, and device movement.
  • Sanity-checking output object shapes, page counts, geometry conventions, and the no-download/random-weight caveat.

Do not use this sub-skill for raw file loading/export format details, model zoo catalogs/custom weights/ONNX export, CLI wrappers, training, datasets, or evaluation scripts. Route those tasks to sibling sub-skills:

Read first

  1. For API recipes and option interactions, read references/ocr-kie-workflows.md.
  2. For failure diagnosis, read references/troubleshooting.md.
  3. For a safe import/factory/optional-forward smoke check, run or adapt scripts/ocr_api_smoke.py.

Operating rules

  • Prefer explicit det_arch, reco_arch, pretrained, pretrained_backbone, assume_straight_pages, and batch-size choices when reproducibility matters.
  • Treat pretrained=False as an API/shape smoke mode only; randomly initialized detection and recognition models do not produce useful OCR.
  • Be explicit about network/cache behavior: pretrained model weights and some orientation helpers may download on first use unless already cached or disabled.
  • Validate outputs before downstream use: Document.pages for OCR, KIEDocument.pages[*].predictions for KIE, optional page.layout, optional page.tables, and relative geometries.
Skills enregistrées