rageyu/platen2-pdf-table

SLANet+ — Core ML

PaddleOCR's SLANet_plus table-structure model, converted to Core ML for use on Apple platforms. Given an image of a table it returns the table's structure tokens (<td>, <td colspan="4">, <tr>…) and one bounding quadrilateral per cell. It reads no text. In Platen2 the cell text comes from the PDF's own text layer, so no character in a rendered table is authored by a model.

Input	1×3×488×488, BGR, ImageNet mean/std applied by position
Outputs	loc [1,501,8] cell quads · structure [1,501,50] token logits
Size	12 MB (fp32)
Speed	~42 ms per table on an M5 Max

Provenance and changes

Converted from slanet-plus.onnx as published by RapidAI/RapidTable, which is itself an ONNX export of PaddleX's SLANet_plus. The weights are unmodified; what changed is the serialisation:

• the ONNX graph was brought across as PyTorch and re-exported to Core ML
• the encoder was folded at a static 488×488 input
• SLAHead's early-exit Loop was unrolled to its own 501-step bound, because Core ML has no loop construct for it

Verified against the ONNX original on 23 real table regions: identical structure token sequences on all 23, worst cell-box delta 0.005 px.

Notes for implementers

• fp32, not fp16. Token sequences survive fp16 but the cell-box regression does not — 36.9 px worst error on the Neural Engine.
• Outputs are padded to a 64-byte row. structure reports shape [1,501,50] with strides [32064,64,1]. Read by stride, not as a tight buffer.
• Resize with a plain bilinear sample. Antialiased downscaling changes the token sequence on some tables with no other symptom.
• The 50-token vocabulary is in user_defined_metadata["structure_tokens"].

Licence

Apache License 2.0, inherited from PaddlePaddle. Copyright (c) PaddlePaddle Authors. ONNX export by RapidAI. Converted to Core ML by tekl; weights unmodified, serialisation changed as described above.