aiflow-math-ink-0.9 / README.md
cwLeeDev's picture
Restore model-only AIFlow Math Ink 0.9 release
1227759 verified
|
Raw
History Blame Contribute Delete
2.61 kB
---
language:
- ko
- en
library_name: pytorch
license: other
pipeline_tag: feature-extraction
base_model: cwLeeDev/aiflow-math-ink-06-intermediate
tags:
- online-handwriting
- mathematical-expression-recognition
- trajectory
- research
---
# AIFlow Math Ink 0.9
AIFlow Math Ink 0.6 seed17 base/online adapter์— ๊ฒฐํ•ฉํ•˜๋Š” ๊ณต๊ฐœ ๋ฐ์ดํ„ฐ ์žฌํ•™์Šต additive calibrator ์—ฐ๊ตฌ ํ›„๋ณด์ž…๋‹ˆ๋‹ค.
## ์ฒดํฌํฌ์ธํŠธ ๊ณ„์•ฝ
- ํŒŒ์ผ: `normalized_balanced_calibrator.pt`
- SHA-256: `bf3bc80c07315b812e2419ffbddadb0ecb3bd5b73e00799fc271979a0124c2dd`
- ์ž…๋ ฅ: glyph bbox crop, aspect-preserving letterbox, local character context
- ์ถœ๋ ฅ: ๊ธฐ์กด 378 labels + `=`
- ํ•„์•• ์‚ฌ์šฉ: ์•„๋‹ˆ์˜ค
- ํ•„์ˆ˜ base: `cwLeeDev/aiflow-math-ink-06-intermediate` seed17์˜ `base_378.pt`, `online_adapter.pt`
์ด ํŒŒ์ผ๋งŒ์œผ๋กœ ์ถ”๋ก ํ•  ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค. base encoder embedding์— additive correction์„ ์ ์šฉํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.
## ํ•„๊ธฐ replay
| ์ง€ํ‘œ | baseline | seed31 | guarded seed31 |
|---|---:|---:|---:|
| ๋ฌธ์ž Top-1 | 50.29% | 87.72% | 87.72% |
| ๋ฌธ์ž Top-5 | 70.18% | 98.83% | 98.83% |
| ์ˆ˜์‹ exact | 26.53% | 67.35% | 67.35% |
| ์ˆ˜์‹ Top-5 oracle | 42.86% | 95.92% | 95.92% |
| ๊ฐœ์„ /ํšŒ๊ท€ ๋ฌธ์ž | - | 67/3 | 66/2 |
Guard๋Š” public-train support๊ฐ€ 3๊ฐœ ๋ฏธ๋งŒ์ธ label์—์„œ baseline Top-1 confidence๊ฐ€ 0.95 ์ด์ƒ์ด๋ฉด baseline์„ ๋ณด์กดํ•ฉ๋‹ˆ๋‹ค. ๊ฐ™์€ replay์—์„œ ํ™•์ธ๋œ ๊ทœ์น™์ด๋ฏ€๋กœ ๋…๋ฆฝ ์ œํ’ˆ ๊ฒ€์ฆ ๊ฒฐ๊ณผ๋Š” ์•„๋‹™๋‹ˆ๋‹ค.
## CROHME2019 ๋น„์ƒ์—… ๊ฒ€์ฆ
Kaggle `ntcuong2103/crohme2019`์˜ CC BY-NC 4.0 valid split์—์„œ parse ๊ฐ€๋Šฅํ•œ 985์‹์„ ํ‰๊ฐ€ํ–ˆ์œผ๋ฉฐ ํ•™์Šต์—๋Š” ์‚ฌ์šฉํ•˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค.
| ์ง€ํ‘œ | baseline | calibrator | guarded |
|---|---:|---:|---:|
| truth-group ๋ฌธ์ž Top-1 | 56.96% | 74.11% | 74.18% |
| truth-group ๋ฌธ์ž Top-5 | 79.97% | 92.45% | 92.28% |
| end-to-end group-token Top-1 recall | 48.20% | 62.17% | 62.22% |
| fully-covered ์ˆ˜์‹ exact | 7.95% | 11.98% | 12.10% |
0.6 geometry gridder์˜ partition exact๋Š” 40.81%, exact group recall์€ 82.48%, lattice group ceiling์€ 99.86%์˜€์Šต๋‹ˆ๋‹ค.
## ์ œํ•œ
- ์—ฐ๊ตฌ ํ›„๋ณด์ด๋ฉฐ `product_adopted=false`์ž…๋‹ˆ๋‹ค.
- ํ•„๊ธฐ replay์˜ contributor ID๊ฐ€ ์—†์–ด writer-disjoint๋ฅผ ์ฆ๋ช…ํ•˜์ง€ ๋ชปํ–ˆ์Šต๋‹ˆ๋‹ค.
- CROHME2019์€ CC BY-NC ์—ฐ๊ตฌ ํ‰๊ฐ€์ด๋ฉฐ ์ œํ’ˆ ์„ฑ๋Šฅ ๊ทผ๊ฑฐ๊ฐ€ ์•„๋‹™๋‹ˆ๋‹ค.
- ์ดํ›„ ๋น„์ƒ์—… gridder artifact๊ฐ€ ์ œ๊ฑฐ๋˜์–ด ๊ณต๊ฐœ ๊ฐ€๋Šฅํ•œ 0.6 geometry gridder๋ฅผ ํ‰๊ฐ€ํ–ˆ์Šต๋‹ˆ๋‹ค.
- ๊ฒฐ์ • ๊ณ„์ธต์˜ ์ˆ˜ํ•™์  ๋™์น˜ ์ •ํ™•๋„๋Š” ์ด๋ฒˆ ์‹คํ–‰์—์„œ ์ธก์ •ํ•˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค.
์ƒ์„ธ ์ˆ˜์น˜๋Š” `metrics.json`๊ณผ `crohme2019_valid.json`์„ ์ฐธ์กฐํ•˜์‹ญ์‹œ์˜ค.