Apppublic

johnlockejrr/regnetx-8gf-polyline-baseline-stage0

sourceHugging Faceapache-2.0updated 1mo agoView on Hugging Face

1likes

App README

RegNetX Polyline Baselines — Stage-0 demo

Gradio Space (ZeroGPU / CUDA, CPU fallback) for `johnlockejrr/regnetx-8gf-polyline-baseline-stage0`.

Upload a historical manuscript page → detect text-line baselines (polylines) → inspect an overlay with optional BLLA polygonal environments. Layer toggles (baselines / polygons / scores) re-draw without re-running the model.

This is the multiscript Stage-0 foundation checkpoint, not the Hebrew/Samaritan domain fine-tune. For the specialized Hebrew + Samaritan release, see `johnlockejrr/regnetx-8gf-polyline-hebrew-samaritan`.

What is included

This folder is an inference-only package (regnetx_infer/). It does not contain training loops, Lightning modules, Arrow compilers, matchers, losses, or experiment CLIs.

Local run

bash

cd gradio
python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
# optional local weights to skip Hub download:
# export REGNETX_WEIGHTS=/path/to/PRETRAIN/regnetx_stage0/best_cbad_f1.safetensors
python app.py

Override Hub repo without editing code:

bash

export REGNETX_REPO=johnlockejrr/regnetx-8gf-polyline-baseline-stage0
python app.py

Space tips

—Set Space hardware to ZeroGPU (Settings → Hardware).
—Upload the Stage-0 model repo first (public best_cbad_f1.safetensors).
—Canvas is fixed at 1280×1280 (training resolution).
—Detection runs under @spaces.GPU (falls back to CPU locally).
—Disable polygonize for a faster baselines-only pass.
—Default confidence 0.40 (Stage-0 val cbad_f1_max); 0.50 is a close alternative. Do not use 0.1.
—Default reading order ltr — switch to rtl for Hebrew/Samaritan pages or ttb for vertical scripts.
—Override device with REGNETX_DEVICE=cpu or cuda if needed.

Dependencies

Stay on ZeroGPU hardware. Set Gradio only via README sdk_version: 6.1.0 (do not also pin gradio==… in requirements.txt — Spaces installs gradio[oauth,mcp]==<sdk_version> and pip will conflict).

Model weights load on CPU in process memory at startup; CUDA runs only inside @spaces.GPU on Detect — that is still ZeroGPU, not a CPU Space.

app.py injects a small fetch patch so if HF reports APP_STARTING while the app is already up, the browser treats it as RUNNING and mounts the UI.

After changing sdk_version / requirements.txt, do a Factory reboot.

D-FINE HybridEncoder: https://arxiv.org/abs/2410.13842 Weights: https://huggingface.co/johnlockejrr/regnetx-8gf-polyline-baseline-stage0