Ralutet/PP-DocLayout-L-onnx-int8

PP-DocLayout-L, ONNX with int8 weights

Document layout detector PaddlePaddle/PP-DocLayout-L (RT-DETR, 23 region classes: text, titles, figures, tables, formulas, captions, headers, footers, ...), converted to ONNX so that it runs in a web browser with ONNX Runtime Web.

The Paddle model was exported with paddle2onnx 2.1. Its float weights are stored as int8 with a scale per output channel and expanded back to float32 by Cast/Mul nodes, which ONNX Runtime folds once when the session starts (129 MB -> 34 MB, same speed, the same regions on our sample papers).

file
pp-doclayout-l.onnx	the detector
labels.json	class names, input size (640x640) and the score threshold used

Conversion script: tools/convert_layout.py of taiyaku. Used by taiyaku, which translates English PDFs into Japanese in the browser.

License

Apache License 2.0, as the original model (LICENSE.txt). Original: PaddlePaddle/PP-DocLayout-L by PaddlePaddle; changes: conversion to ONNX and int8 weight storage.