pj-ocr-jaxjs-medium / docs /pp-doclayout-v3.md
hf-dev-01's picture
Promote the all-JAXJS medium release
9c59d5c verified
|
Raw
History Blame Contribute Delete
2.81 kB
# Canonical medium JAXJS release
This composed release lane assembles canonical FP16 PP-DocLayoutV3, its explicit FP32 rollback, the streaming OCR recognizer, and the three all-JAXJS MinerU table artifacts before it installs `model-sets/medium.json`.
The historical FP32 startup evidence remains relevant to the rollback artifact, while the FP16 browser result and CLI/WASM memory boundary are recorded separately.
- Upstream: `PaddlePaddle/PP-DocLayoutV3`
- Upstream revision: `7b48a7566925fa464281f930c58eee04fe2c862a`
- Canonical FP16 weight SHA-256: `da8613530ea128d8e85d00be530fe02a29830e59308f6dc6c67e12fdb03e6a8f`
- Canonical FP16 weight dtype: `float16`
- FP32 rollback weight SHA-256: `13492ac1f96d5340caa0f75e8c894a3b7d8e95d7c2e485fd5b671bee9a3df6f3`
- FP32 rollback weight dtype: `float32`
- Published layout artifact revision: `3ebf4493fe544793ad72ece74cf4777b281460d6`
- Streaming OCR artifact state: `published` (`3ebf4493fe544793ad72ece74cf4777b281460d6`)
- JAXJS table artifact state: `published` (`3ebf4493fe544793ad72ece74cf4777b281460d6`)
- Layout release descriptor: [`model-sets/releases/medium-v3.release.json`](../model-sets/releases/medium-v3.release.json)
- Streaming OCR release descriptor: [`model-sets/releases/medium-ppocr-v6-medium-rec-streaming.release.json`](../model-sets/releases/medium-ppocr-v6-medium-rec-streaming.release.json)
- Table release descriptor: [`model-sets/releases/medium-jaxjs-table.release.json`](../model-sets/releases/medium-jaxjs-table.release.json)
- Layout license: Apache-2.0
- Streaming OCR license: Apache-2.0
- SLANet-plus license: Apache-2.0
- Table classifier and wired-table U-Net licenses: AGPL-3.0
- Precision report: [`reports/pp-doclayout-v3/precision-experiment.json`](../reports/pp-doclayout-v3/precision-experiment.json)
- Promotion decision: [`reports/pp-doclayout-v3/promotion-decision.json`](../reports/pp-doclayout-v3/promotion-decision.json)
- Browser WebGPU experiment: [`reports/pp-doclayout-v3/webgpu-fp16-experiment.json`](../reports/pp-doclayout-v3/webgpu-fp16-experiment.json)
The FP32 model remains an explicit non-default verification and rollback profile. Canonical `medium` selects FP16 PP-DocLayoutV3, the streaming OCR recognizer, and the all-JAXJS table trio; it never falls back to FP32 or ONNX implicitly. The layout, streaming OCR, and table descriptors all pin immutable artifact revision `3ebf4493fe544793ad72ece74cf4777b281460d6`.
## Reproduce
From the OCR JS workspace with the pinned Paddle source available:
```bash
pnpm jaxjs:convert-pp-doclayout-v3
pnpm jaxjs:convert-pp-doclayout-v3-fp16
pnpm --filter @ocr-js/pipeline-runner verify:pp-doclayout-v3
pnpm jaxjs:stage-medium-release -- /path/to/pj-ocr-jaxjs-medium
```
No WebGPU execution is required for conversion or the frozen WASM oracle checks.