pauldubois98/musicocr-models

MusicOCR models (copista and DocAligner, ONNX)

ONNX exports of KobiMusic's copista networks, and a mirror of DocsaidLab's DocAligner page-corner model, used by MusicOCR, a fully client-side sheet-music recognizer that runs in the browser with ONNX Runtime Web.

File	What it is
copista-2m-trunk.onnx, copista-2m-verifier.onnx	copista-2m symbol detector (two graphs: convolutional trunk, verifier over the top-K peaks)
copista-28m-trunk.onnx, copista-28m-verifier.onnx	copista-28m symbol detector, same split
copista-evidence.onnx	copisteria's 2M-parameter evidence model
copista-meta.json	class list, attribute vocabularies, decoding constants
docaligner-lcnet100.onnx	DocAligner page-corner detector (lcnet100_h_e_bifpn_256_fp32): image [1,3,256,256] → 4 corner heatmaps [1,4,128,128]

Source: https://github.com/kobimusic/copisteria @ c685f280914d91a694494588d9f6a376d7917eba. Converted with scripts/models/export_copista.py; every graph was checked against PyTorch (max abs difference ≤ 1e-4).

License and attribution

Apache License 2.0, like the original weights (kobimusic/copista-2m, kobimusic/copista-28m) and code (kobimusic/copisteria). These files are converted (modified) copies of KobiMusic's weights; see NOTICE.

docaligner-lcnet100.onnx is an unmodified copy of DocsaidLab's DocAligner model (Apache License 2.0, LICENSE-DocAligner), mirrored here because the original is only published on Google Drive, which browsers cannot load from.