PP-FormulaNet_plus-M ONNX, graph optimized
This repository contains a graph-rewritten ONNX inference model derived from the PP-FormulaNet_plus-M model distributed by RapidAI/RapidDoc. It is intended for formula recognition in OCR pipelines using ONNX Runtime.
What changed
This is not a fine-tuned model. The learned weights are unchanged. The
linear_292 encoder projection was moved from the decoder Loop body to the
outer graph because its input and weights are invariant across decoder steps.
The rewrite was generated by tools/hoist_formula_encoder_projection.py in
Rapid-OCR-RS. Dead copies of the moved initializers were removed from the Loop
body. The ONNX graph is therefore different, while its recognized token
sequences remain identical in the recorded regression sets.
Artifact
- File:
pp_formulanet_plus_m.onnx - Size: 593,916,167 bytes
- SHA-256:
6e509ec4b7cda604dc9bf80a876ae6494693deea4c7f4157a4dc0eba0ebda745 - Input:
FLOAT32 [N, 1, 384, 384] - Output:
INT64 [N, L] - Source model: RapidAI/RapidDoc PP-FormulaNet_plus-M
- Source model SHA-256:
71b6d389cf7b857e45252a4b98cfced1a3ffca7bf24d9497d02d052a41d9493b
Validation
The candidate was evaluated with the Rapid-OCR-RS Rust release evaluator, ONNX Runtime 1.28, CPUExecutionProvider, and pinned validation manifests.
- 64-image set: 64 scored; exact and normalized matches, mean CER, truncation, EOS positions, token IDs, and decoded LaTeX matched the source model.
- 501-image set: exact 190, normalized 194, mean CER 0.0796678406, one truncated sample, and zero pipeline failures; predictions matched the source model.
- Independent single-image latency: one candidate run measured 1,077.5 ms per image versus a three-run source-model mean of 1,119.7 ms per image. This is directional evidence, not a repeatable performance claim.
- Batch-8 over 501 images: two candidate runs measured 444.4 s and 458.6 s; the source-model rerun measured 501.8 s under the same conditions. The observed difference varied with machine load and is not a fixed speedup guarantee.
Full methodology and raw evidence are maintained in the
Rapid-OCR-RS repository,
under tests/baseline/formula-ocr-performance/optimized-model-adoption-report.md.
License and attribution
The source model is identified as Apache-2.0 by the RapidAI/RapidDoc ModelScope
model metadata and the upstream PaddleOCR model documentation. This graph
rewrite does not change the source weights or their license. The Apache-2.0
license text is included as LICENSE; attribution and derivation details are
provided here. Check the upstream model repository's current terms before
redistributing this artifact.
This repository distributes inference weights only. It does not include training data or claim ownership of the original model weights.