File size: 3,266 Bytes
434a065 fe5a411 ea25c16 fe5a411 ea25c16 fe5a411 ea25c16 fe5a411 ea25c16 fe5a411 ea25c16 fe5a411 ea25c16 fe5a411 ea25c16 fe5a411 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 | ---
license: apache-2.0
---
# Introduction
This repository hosts [PaddleOCR PP-DocLayoutV3](https://huggingface.co/PaddlePaddle/PP-DocLayoutV3_safetensors),
an RT-DETR-based **document layout detector** (~33M params), for the
[React Native ExecuTorch](https://www.npmjs.com/package/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`](https://huggingface.co/software-mansion/react-native-executorch-pp-ocrv6).
If you'd like to run these models in your own ExecuTorch runtime, refer to the
[official documentation](https://pytorch.org/executorch/stable/index.html) 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](https://github.com/pytorch/executorch/blob/main/runtime/COMPATIBILITY.md).
If you work with React Native ExecuTorch, the library constants guarantee compatibility with the
runtime used behind the scenes.
|