Layout and vision tables

pdfspine 0.8.0 offers an opt-in ONNX backend for table regions and whole-page layout. The default find_tables() still uses strategy="lines". Choosing strategy="vision" without a backend still selects TATR; request backend="onnx" explicitly.

Runtime and model files

pip install "pdfspine[onnx]==0.8.0"

The extra installs ONNX Runtime, NumPy and Pillow, without PyTorch. Weights are not bundled. Supply PP-DocLayout and SLANet-plus files using PDFSPINE_ONNX_MODELS or explicit paths. Missing runtime/weights raise pdfspine.PdfUnsupportedError with setup guidance. These are not the OCR weights. See the release model setup for model choices and filenames; the examples below assume files already exist.

options = {
    "layout_model": "/models/layout.onnx",
    "table_model": "/models/table.onnx",
}
finder = page.find_tables(strategy="vision", backend="onnx", vision_options=options)
for table in finder:
    assert table.source == "onnx"
    print(table.to_html())  # preserves row/column spans

regions = page.find_layout(**options)
for region in regions:
    print(region.bbox, region.label, region.score, region.raw_label)
html = page.get_layout_html(**options)

pdfspine.OnnxOptions documents the supported settings; pdfspine.LayoutBlock is the returned region value type. The default detector is PP-DocLayoutV3; layout_variant="pp_doclayout_l" selects the optional faster L variant with its matching weights. providers="auto" chooses CUDA when available, otherwise CPU; it does not select CoreML.

Geometry and text have different sources

Models choose regions and cell geometry. Cell and paragraph text always comes from the native PDF text layer; no OCR is performed. Image-only scans therefore need a separate OCR workflow. get_layout_html() emits semantic headings, paragraphs, tables and figure placeholders, with unclaimed text retained in a pre element. This differs from the reconstruction returned by get_text("html").

The 0.8.0 backend is opt-in and has no ground-truth corpus accuracy claim. Its detector order and labels are predictions; validate them on your own files.