Views
No views yet
smt_ocr inference engine (pure C/C++ + ggml, no Python/PyTorch at runtime).bekern token sequence, in one end-to-end pass (no staff/system segmentation).| File | Size | Notes |
|---|---|---|
smt-fp-grandstaff-f32.gguf | 46 MB | full precision |
smt-fp-grandstaff-q8_0.gguf | 16 MB | recommended; greedy decode token-identical to f32 |
smt-fp-grandstaff-q5_k.gguf | 13 MB | smallest usable; 0.04% token-CER vs f32 (one accidental on a ~2300-token page) |
q4_k is not provided — even with the encoder + LM head kept at Q8_0 it only
reaches ~0.5% token-CER while ending up larger than q5_k (the ConvNext encoder
is most of the weights), and plain q4_k degenerates into repetition. Use q5_k.[64,128,256], 16× reduction) + 8-layer
cross-attention Transformer decoder (d_model 256, 4 heads, 1× ReLU FFN),
sinusoidal 1-D/2-D positional encodings, greedy AR decode to a 181-token
bekern vocabulary (maxlen 4353). This is the antoniorv6/SMT forward:
scaled dot-product attention (d_head**-0.5), no ReLU before the linear LM
head; input is used at native resolution (reduce_ratio=1.0) and grayscaled
(no inversion — matching SMT/data_augmentation.convert_img_to_tensor).1crispembed -m smt-fp-grandstaff-q8_0.gguf --image page.png
2# or let the model manager fetch it:
3crispembed --model smt-fp --image page.pngA. Ríos-Vila, D. Rizo, J. M. Iñesta, J. Calvo-Zaragoza. End-to-End Full-Page Optical Music Recognition for Pianoform Sheet Music. International Journal of Computer Vision (IJCV), 2026. arXiv:2405.12105.
models/convert-smt-to-gguf.py from CrispEmbed.PRAIG.mit. This repository redistributes under the same terms; it grants no rights the upstream licence does not.