Grape 2.0

Grape 2.0, the FineNotes handwriting model.

• A 5x192 network that reads Bezier curves instead of points: smaller on disk than it sounds (5.7 MB), and about a third faster to run than Grape 1.0.
• Trained on label-checked data, and with formulas written on one line.
• On the developer's own iPad notes, 13.4 % of characters wrong against 21.2 % for Grape 1.0; on a copy sheet 9.7 % against 14.0 %. Public test sets are level (UNIPEN, IAM-OnDB) or better (ImadocSen, BRUSH); DeepWriting is 1 point worse. Full table in eval.md.
• Model file format 2 (see FORMAT.md). reader.py reads format 1 only.
• Apache 2.0, same attributions as before plus HWRT (see NOTICE).

SHA-256: f553c71367e8881f2918c9a34a44b3f215a54e86b39d6547ff885aede1fa6d04

Grape 1.0

Grape 1.0: the first public version of the FineNotes handwriting recogniser. Work in progress: real stylus handwriting data is the main constraint, and fine-tuning to a writer works very well (14.2 % to 3.4 % character error rate on one writer). See the README, MODEL-CARD.md and eval.md. Apache 2.0; keep NOTICE with the model. SHA-256 of grape-1.0.bin: 87a9ddd56b6bcf01e66a027ff74f1b8ba9a58fe0d4b0077e6abb6455710d21f9