software-mansion/react-native-executorch-pp-doclayout-v3

Introduction

This repository hosts PaddleOCR PP-DocLayoutV3, an RT-DETR-based document layout detector (~33M params), for the React Native ExecuTorch library, exported to .pte for the ExecuTorch runtime (XNNPACK, CoreML, Vulkan). It finds and classifies document regions — titles, paragraphs, tables, figures, formulas, headers/footers, etc. — and is a companion to react-native-executorch-pp-ocrv6.

If you'd like to run these models in your own ExecuTorch runtime, refer to the official documentation for setup instructions.

The .pte is a pure tensor→tensor function; pre-processing (resize, normalize) and the final score threshold are the client's job.

Output contract

A single fixed-shape method forward (shape also declared in config.json; no shape-discovery companion methods on this model). The RT-DETR box decode is baked into the graph — outputs are ready-to-threshold:

in   [1, 3, 800, 800]     # RGB, ImageNet-normalized by the client: (x/255 - mean)/std
out  boxes   [300, 4]     # (x1, y1, x2, y2) in 800×800 model-input pixel space
     scores  [300]        # max-class sigmoid score per query
     classes [300]        # float class index per query (argmax)

PP-DocLayoutV3 is a DETR set-prediction model → no NMS. All 300 queries are returned; post-processing is just: keep rows with score ≥ threshold, scale boxes from the 800×800 input space to your image, and map classes[i] through labels.json (index → label).

Classes (25)

abstract, algorithm, aside_text, chart, content, formula, doc_title, figure_title, footer, footnote, formula_number, header, image, number, paragraph_title, reference, reference_content, seal, table, text, vision_footnote (some indices map to the same display label; use labels.json as the authoritative index→label map).

Backends, sizes & latency (warm)

backend	target	precision	size	latency
xnnpack	CPU	fp32	132 MB	~2.0 s (Galaxy S24)
coreml	Apple ANE	fp16	91 MB	~50 ms (Apple M-series ANE)
vulkan	Android GPU	fp16 (mixed-delegate)	66 MB	~0.86 s (Galaxy S24)

Vulkan is the recommended Android backend — ~2.4× faster than XNNPACK and half the size. It's mixed-delegate: most of RT-DETR runs fp16 on the GPU, while the box-head matmuls run on XNNPACK (they delegate as addmm→linear). XNNPACK stays fp32 because int8/int4 quantization loses whole boxes on this model.

Compatibility

If you intend to use these models outside of React Native ExecuTorch, make sure your runtime is compatible with the ExecuTorch version used to export the .pte files. For more details, see the compatibility note in the ExecuTorch GitHub repository. If you work with React Native ExecuTorch, the library constants guarantee compatibility with the runtime used behind the scenes.