[Paper Study] AutoAugment : Learning Augmentation Strategies from Data

field : CV,RL
understanding : ๐Ÿ˜ƒ๐Ÿ˜ƒ

Paper study
Author

hoyeon

Published

May 17, 2023

ํ•œ ์ค„ ์š”์•ฝ

  • image Augmentation์— ๊ฐ•ํ™”ํ•™์Šต \(\to\) sota!

Intro & Abstract

  • Data augmentation์€ image classifier์˜ ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ์‰ฌ์šด ๋ฐฉ๋ฒ•.
  • ์™œ image augmentatio์ด ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ? \(\to\) ๋ฐ์ดํ„ฐ์˜ invariance๋“ค์„ ์ถฉ๋ถ„ํžˆ ํ•™์Šต
    • invariance(๋ถˆ๋ณ€์„ฑ)๋ž€? : ์ฐจ์ด๊ฐ€ ์žˆ๊ฑฐ๋‚˜ ๋ณ€ํ™˜์ด ์ ์šฉ๋œ ํ›„์—๋„ ๋ฌด์–ธ๊ฐ€๊ฐ€ ๋™์ผํ•˜๊ฒŒ ์œ ์ง€๋˜๋Š” ์†์„ฑ,์ƒํƒœ๋ฅผ ์˜๋ฏธํ•จ
    • ์˜ˆ๋ฅผ ๋“ค์–ด ์ž๋™์ฐจ,์‚ฌ๊ณผ
  • ๊ทธ๋Ÿฌ๋‚˜ ๋ฐ์ดํ„ฐ์…‹๋งˆ๋‹ค ์ด๋ฏธ์ง€์˜ ๋ถ„ํฌ๊ฐ€ ๋‹ค๋ฅด๊ธฐ ๋•Œ๋ฌธ์— ์ˆ˜๋™์ ์œผ๋กœ augmentation์ „๋žต์„ ๋‹ค๋ค„์ค˜์•ผ ํ–ˆ์Œ.
    • ๊ณผ์ผ์ด ๋งŽ์€ ๋ฐ์ดํ„ฐ์— ์ƒ‰์— ๋Œ€ํ•œ transform์„ ๋งŽ์ด ์ ์šฉํ•˜๋ฉด? \(\to\) ์‚ฌ๊ณผ๊ฐ€ ์‚ฌ๊ณผ๊ฐ€ ์•„๋‹ˆ๊ฒŒ ๋˜๊ฒ ์ฃ ?
  • ์ด ๋…ผ๋ฌธ์—์„œ๋Š” ๋ฐ์ดํ„ฐ๋งˆ๋‹ค ์ ์ ˆํ•œ augmentation policy์„ ์ž๋™์ ์œผ๋กœ ์ฐพ๊ธฐ์œ„ํ•ด ๊ฐ•ํ™”ํ•™์Šต์„ ์‚ฌ์šฉํ•จ.
  • Imagenet๊ณผ CIFAR-10์—์„œ Sota๋ฅผ ๋‹ฌ์„ฑํ–ˆ์Œ. ํ›„์— ๋ํŒ์™• ๋Š๋‚Œ์˜ efficient net์ด ๋‚˜์˜ค๋Š”๋ฐ ๊ทธ ๋•Œ์˜ augmentation์—๋„ ํ™œ์šฉ.

Method

Fig1

  • ํฌ๊ฒŒ ๋‘ ๊ฐ€์ง€์˜ ํŒŒํŠธ๋กœ ๊ตฌ๋ถ„
    1. Controller(RNN) : Polict๊ฐ€ ์žˆ๋Š” Search Space์—์„œ ํ•˜๋‚˜์˜ augmentation policy๋ฅผ sampling.
    2. Child network : ์ด๋ฏธ์ง€ ๋ถ„๋ฅ˜๊ธฐ sampling๋œ policy๋กœ ํ•™์Šต ํ›„ validation accuracy R(reward)์„ ๊ณ„์‚ฐ.
  • ๊ณผ์ •(์ถฉ๋ถ„ํžˆ ๋ฐ˜๋ณต)
    policy sampling(by controller)
    \(\to\) fitting classifier,calculating reward
    \(\to\) contoller update,policy sampling
    \(\to\) fitting classifier,calculating reward
    \(\to\) contoller update,policy sampling
    \(\to\) fitting classifier,calculating reward
    \(\quad\quad\quad\quad\quad\quad\quad \vdots\)
    \(\to\) optimal policy(converge to best augmentation strategy)

  • SVHN dataset์— ์ ์šฉํ•œ ์˜ˆ์‹œ
  • ํ•˜๋‚˜์˜ Policy๋Š” 5๊ฐ€์ง€์˜ subpolicy๋กœ ๊ตฌ์„ฑ๋จ.
  • subpolicy๋Š” operation 2๊ฐœ์™€ probability,magnitude๋กœ ๊ตฌ์„ฑ๋จ
    • operation : ์ด๋ฏธ์ง€ ๋ณ€ํ˜•ํ•˜๋Š” ๋ฐฉ๋ฒ• (Rotate,Brightness,ShearX,Inver ๋“ฑ๋“ฑโ€ฆ ์ด 16๊ฐ€์ง€ ๋ฐฉ๋ฒ•์ด ์กด์žฌํ•จ)
    • probability : ์–ผ๋งˆ๋‚˜ ๋งŽ์ด ์ ์šฉํ• ๊ฑฐ๋ƒ
    • magnitude : ์–ด๋Š์ •๋„ ๊ฐ•๋„๋กœ ํ• ๊ฑฐ๋ƒ
  • ์œ„์˜ ์˜ˆ์‹œ์—์„œ ๋ฐฐ์น˜๋Š” ์ด 15๊ฐœ. ๊ฐ๊ฐ์˜ ๋ฐฐ์น˜๋งˆ๋‹ค ๊ท ์ผ๋ถ„ํฌ๋กœ ์–ด๋–ค subpolicy๊ฐ€ ํ• ๋‹น๋จ.(๊ทธ๋ฆผ์—์„œ๋Š” ์•„์˜ˆ 33333์”ฉ ํ• ๋‹น๋˜์—ˆ์ง€๋งŒ ์‹ค์ œ๋Š” ์•„๋‹ ์ˆ˜ ์žˆ์Œ.์ดํ•ด๋ฅผ ๋•๊ธฐ ์œ„ํ•จ)
  • Note
    • ๋™์ผํ•œ subpolicy๊ฐ€ ์ ์šฉ๋˜๋Š” ๋ฐฐ์น˜๋“ค์ด๋ผ๋„ ๊ฐ ๋ฐฐ์น˜๋งˆ๋‹ค ๋‹ค๋ฅธ operation์ด ์ ์šฉ๋  ์ˆ˜ ์žˆ์Œ. ์ด๋Š” probability ๋˜ํ•œ subpolicy์— ์žˆ๊ธฐ ๋–„๋ฌธ์ž„

  • ImageNet์— ์ ์šฉํ•œ ์˜ˆ์‹œ vs SVHN์—์„œ ์ ์šฉํ•œ์˜ˆ์‹œ
  • SVHN์—๋Š” ๊ธฐํ•˜ํ•™์  ๋ณ€ํ™˜์ด ๋งŽ์ด ๋†’์€ ํ™•๋ฅ ๋กœ ์ฒซ๋ฒˆ์งธ ๋ณ€ํ™˜์œผ๋กœ ์ ์šฉ๋˜๊ฒŒ ๋˜์–ด ์žˆ์Œ(Shear X,Shear Y)
    • SVHN์€ ์ˆซ์ž ์ด๋ฏธ์ง€ dataset $\(์ˆซ์ž๊ฐ€ ๋น„ํ‹€๋ ค์žˆ๊ฑฐ๋‚˜ ์™œ๊ณก๋˜์–ด ์žˆ๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์Œ.\)$ ๊ฐ•์ธํ•จ์„ policy๋Š” ๊ธฐํ•˜ํ•™์  ๋ณ€ํ™˜์„ ๋งŽ์ด ํฌํ•จ.
    • ๋˜ํ•œ ์ƒ‰๋ณ€ํ™˜๋„ ์–ด๋Š์ •๋„ ํฌํ•จ๋˜์–ด ์žˆ์Œ.(์ด๋Š” ๋ฐ์ดํ„ฐ์—์„œ ํ™•์ธ๊ฐ€๋Šฅ,๋ฐ์ดํ„ฐ ์ž์ฒด์— ๋ฐ˜์ „๋˜์–ด ์žˆ๋Š” ๊ฒฝ์šฐ๋„ ๋งŽ์Œ)
  • Imagenet์—๋Š” ์ƒ‰์ƒ ๋ณ€ํ™˜์ด ๋งŽ์ด ๋†’์€ ํ™•๋ฅ ๋กœ ์ฒซ๋ฒˆ์งธ ๋ณ€ํ™˜์œผ๋กœ ์ ์šฉ๋˜๊ฒŒ ๋˜์–ด ์žˆ์Œ(Shear X,Shear Y)
    • ImageNet์€ ๋‹ค์–‘ํ•œ ๋ฌผ์ฒด๋“ค์„ ํฌํ•จํ•˜๋Š” dataset \(\to\) ๋‹ค์–‘ํ•œ ์ƒ‰์ƒ๋“ค์ด ํฌํ•จ๋˜์–ด ์žˆ์Œ. \(\to\) ๊ฐ•์ธํ•จ์„ ๊ฐ€์ง€๊ธฐ ์œ„ํ•ด policy๋Š” ์ƒ‰์ƒ ๋ณ€ํ™˜์ด ๋งŽ์ด ํฌํ•จ

Result

Result on many datasets

Result on Imagenet dataset

Conclusion

  • ๋‹ค์–‘ํ•œ ๋ฐ์ดํ„ฐ์…‹์— ๊ฐ•ํ™”ํ•™์Šต์„ ์‚ฌ์šฉํ•œ autoaugmentation ์ „๋žต์„ ์ ์šฉ \(\to\) Sota!!
  • ์ถ”๊ฐ€์ ์œผ๋กœ, ๋ฐ์ด์ฝ˜,์บ๊ธ€ ๋“ฑ์—์„œ ์œ„์—์„œ ๋‚˜์˜จ augmentation์ „๋žต๋“ค์ด ๋งŽ์ด ์‚ฌ์šฉ๋จ. ์ ์šฉํ•˜๊ธฐ ์‰ฌ์›€
    ๋งํฌ ๋‚ด๊บผ ์ฝ”๋“œ