ocrspine-models

The default PP-OCRv5 ONNX model weights for ocrspine's pure-Rust PaddleOCR engine — the shared data companion for the spine document family (pdfspine / pptspine / docspine).

This is a pure-data distribution. The host wheels already contain the OCR code (compiled in) but ship no models; this one package supplies the ~28 MB of weights, so the family keeps exactly one git-tracked copy of the models (in the ocrspine repo) and one published copy of the data wheel.

It is published on PyPI — pip install ocrspine-models — though you normally do not install it directly: it is a hard dependency of the hosts, so pip install pdfspine (or pptspine / docspine) pulls it in automatically. The host then resolves the models at runtime by reading ocrspine_models.models_dir() and exporting it as the engine's OCRSPINE_MODELS environment variable. Everything is offline — no model download at runtime.

import ocrspine_models

print(ocrspine_models.models_dir())  # dir holding the 3 ONNX files
print(ocrspine_models.det_path())    # .../ppocrv5_det.onnx
print(ocrspine_models.rec_path())    # .../ppocrv5_rec.onnx
print(ocrspine_models.cls_path())    # .../ppocrv5_cls.onnx
print(ocrspine_models.keys_path())   # .../ppocr_keys_v5.txt

What ships here

Only the default zh/en/ja model set ships here. Language-evaluation artifacts that live in the ocrspine repo (e.g. the Thai recognition model) are out of scope.

License

Apache-2.0. The redistributed PP-OCR model weights are Copyright (c) PaddlePaddle Authors (Apache-2.0), converted to ONNX via Paddle2ONNX. See ocrspine_models/NOTICE and ocrspine_models/PROVENANCE.md.