Paddle2ONNX — ONNX Model Export Pipeline

This is a detached fork of PaddlePaddle/Paddle2ONNX, maintained by kreuzberg-dev to provide automated ONNX model exports of PaddleOCR inference models for use with ONNX Runtime.

Purpose

PaddlePaddle publishes inference models in their native format (.json + .pdiparams). Most ML runtimes outside the Paddle ecosystem (ONNX Runtime, TensorRT, OpenVINO) require ONNX format. This repo automates the conversion pipeline:

1. Downloads latest-gen PaddleOCR models from HuggingFace
2. Exports them to ONNX using Paddle2ONNX (opset 17)
3. Applies ORT compatibility fixes (Loop node shape inference workarounds)
4. Validates every model with the ONNX checker and ONNX Runtime inference
5. Publishes validated .onnx files as GitHub release assets

Exported Models

All models are Apache 2.0 licensed, exported with ONNX opset 17.

Text Detection

Model	Size	Description
PP-OCRv5_server_det	84 MB	High-accuracy text region detection. Outputs bounding polygons around text areas. For server-side batch processing.
PP-OCRv5_mobile_det	4.5 MB	Lightweight text detection. Same function, optimized for speed and edge deployment.

Text Recognition

Model	Size	Description
PP-OCRv5_server_rec	81 MB	Multilingual text recognition (Chinese, English, Japanese, Traditional Chinese). Reads characters from cropped text regions.
PP-OCRv5_mobile_rec	16 MB	Multilingual recognition, smaller model for mobile/edge.
en_PP-OCRv5_mobile_rec	7.5 MB	English/Latin-only recognition. Smallest and fastest for English workloads.

Document Layout Analysis

Model	Size	Description
PP-DocLayoutV3	126 MB	Classifies document regions into 23 categories (title, paragraph, table, figure, header, footer, etc.). Used to understand page structure before content extraction.

Table Structure Recognition

Model	Size	Description
SLANet_plus	7.4 MB	General-purpose table structure recognition. Outputs HTML structure tokens and cell bounding boxes. Good balance of speed and accuracy.
SLANeXt_wired	348 MB	Optimized for tables with visible grid lines/borders. Higher accuracy on bordered tables.
SLANeXt_wireless	348 MB	Optimized for borderless tables (whitespace-separated columns). Higher accuracy on unbordered tables.

Table Cell Detection

Model	Size	Description
RT-DETR-L_wired_table_cell_det	123 MB	Object detection for individual cell bounding boxes in bordered tables.
RT-DETR-L_wireless_table_cell_det	123 MB	Cell bounding box detection for borderless tables.

Utility Classifiers

Model	Size	Description
PP-LCNet_x1_0_doc_ori	6.5 MB	Document orientation classifier (0/90/180/270 degrees). Run before OCR to auto-rotate pages.
PP-LCNet_x1_0_textline_ori	6.5 MB	Text line orientation classifier. For documents with mixed text directions.
PP-LCNet_x1_0_table_cls	6.5 MB	Wired vs. wireless table classifier. Used to select the appropriate SLANeXt or RT-DETR variant.

Typical OCR Pipeline

doc_ori → layout analysis → ┬─ text regions:  det → rec
                             └─ table regions: table_cls → SLANeXt + cell_det → rec

Usage

Download from Releases

Pre-built ONNX models are available as GitHub release assets. Each release includes a manifest.json with SHA256 checksums and model metadata.

Export Locally

# Install dependencies
pip install paddlepaddle huggingface-hub onnx onnxruntime numpy
pip install -e .

# Export a single model
python tools/export_hf_models.py --model SLANet_plus --output-dir ./exports

# Export all models
python tools/export_hf_models.py --all --output-dir ./exports

# Dry run (list models without exporting)
python tools/export_hf_models.py --all --output-dir ./exports --dry-run

CI/CD

The export pipeline runs automatically on release creation and uploads all .onnx files as release assets. It can also be triggered manually via workflow dispatch.

Upstream

This fork tracks PaddlePaddle/Paddle2ONNX (develop branch) with the following additions:

• tools/export_hf_models.py — automated export pipeline with ORT compatibility fixes
• .github/workflows/export_models.yml — CI workflow for model export and release publishing
• CMake and CI fixes for building on modern toolchains (Ubuntu 24.04, Python 3.13)

License

Apache-2.0