<content>

  • open-source implemention์„ ํ™œ์šฉํ•˜๋ผ
  • Transfer Learning ์ด์šฉํ•˜๊ธฐ
  • Data Augmentation
  • ๊ทธ ์™ธ

 


convNet ์‚ฌ์šฉ์— ๋„์›€์ด ๋  ์ง€์‹

 

1. open-source implemention์„ ํ™œ์šฉํ•˜๋ผ

  • ๋งŽ์€ neural network๋Š” ๋ณต์ œํ•˜๊ธฐ๊ฐ€ ์–ด๋ ต๋‹ค.
    • learning decay๋‚˜ ๋‹ค๋ฅธ hyperparameter์— ๋”ฐ๋ผ ์„ฑ๋Šฅ ์ฐจ์ด๊ฐ€ ๋‚˜๊ธฐ ๋•Œ๋ฌธ์ด๋‹ค.
  • ์‚ฌ์šฉํ•ด๋ณด๊ณ  ์‹ถ์€ ๋…ผ๋ฌธ์ด ์žˆ๋‹ค๋ฉด ๋จผ์ € ์˜คํ”ˆ์†Œ์Šค๋ฅผ ์ฐพ์•„๋ณด๋ผ.
    • ๋”ฅ๋Ÿฌ๋‹ ์—ฐ๊ตฌ์ž๊ฐ€ github์— ๊ณต๊ฐœํ•œ ์ฝ”๋“œ๋ฅผ ํ™œ์šฉํ•˜์ž.
    • git cloneํ•ด์„œ ์‚ฌ์šฉํ•˜๊ธฐ

2. Transfer Learning ์ด์šฉํ•˜๊ธฐ

  • ๋‹ค๋ฅธ ์‚ฌ๋žŒ๋“ค์ด ํ›ˆ๋ จ์‹œ์ผœ ๋†“์€ ๊ฐ€์ค‘์น˜๋ฅผ ์ด์šฉํ•˜์—ฌ pre-training model๋กœ ํ™œ์šฉํ•˜๋ผ
    • ์ž‘์€ ๋ฐ์ดํ„ฐ๋กœ๋„ ์ข‹์€ ์„ฑ๋Šฅ์„ ๋‚ผ ์ˆ˜ ์žˆ๋Š” ๋ฐฉ๋ฒ•์ด๋‹ค 
    • ์ดํ›„ fine-tuning ์ง„ํ–‰
    • ๋ฐ์ดํ„ฐ์˜ ์ˆ˜๊ฐ€ ๋งŽ์„์ˆ˜๋ก ๊ฐ€์ค‘์น˜ ๊ณ ์ •ํ•˜๋Š” layer์˜ ์ˆ˜ ์ค„์ผ ์ˆ˜ ์žˆ๋‹ค.

3. Data Augmentation

 

  • 1. mirroring
    • ์ขŒ์šฐ ๋ฐ˜์ „
    • ์ƒํ•˜ ๋ฐ˜์ „
  • 2. random cropping
    • ์ด๋ฏธ์ง€์˜ ์ผ๋ถ€๋ถ„๋งŒ ์ž˜๋ผ์„œ ํ•™์Šตํ•˜๋Š” ๊ฒƒ
  • 3. rotation
    • ์ขŒ์šฐ ํšŒ์ „
    • ์ƒํ•˜ ํšŒ์ „
  • 4. shearing
    • ์‚ฌ์ง„ ์ฐŒ๊ทธ๋Ÿฌ๋œจ๋ฆฌ๊ธฐ (์™œ๊ณก)
  • 5. local warping
    • ์ง€์—ญ์ ์ธ ๋’คํ‹€๋ฆผ ์ ์šฉ
  • 6. color shifting
    • model์ด ์ƒ‰์ƒ์˜ ๋ณ€ํ™”์— robustํ•ด์ง€๋„๋ก ์ƒ‰์ƒ ๋ณ€ํ™”
      • ex1) R +20, G -20, B +20
      • ex2) R -20, G +20, B -20
    • PCA๋ฅผ ์ด์šฉํ•œ color augmentation
      • PCA๋ฅผ ํ†ตํ•ด ์ฐพ์€ ์ฃผ์š” ์„ฑ๋ถ„์ด R, B๋ผ๋ฉด R, B ์œ„์ฃผ๋กœ color shifting์„ ์ˆ˜ํ–‰ํ•˜๋Š” ๊ฒƒ

4. ์ถ”๊ฐ€ ๋‚ด์šฉ

  • two source of knowledge
    • labeled data
    • hand engineering (feature ๋””์ž์ธ) - labeled data๊ฐ€ ์ ๋‹ค๋ฉด hand engineering์— ๋Œ€ํ•œ ๋…ธ๋ ฅ์ด ๋” ํ•„์š”ํ•˜๋‹ค
  • Tips for doing well on winning competitions
    • ensemble
      • ์‹ค์ œ ์ œ๊ณตํ•˜๋Š” ์„œ๋น„์Šค ๋งŒ๋“ค ๊ฒฝ์šฐ์— ๊ฑฐ์˜ ์‚ฌ์šฉํ•˜์ง€ ์•Š๋Š”๋‹ค. (๋ฉ”๋ชจ๋ฆฌ ๊ณต๊ฐ„์ด ๋งŽ์ด ํ•„์š”ํ•˜๊ณ , ๊ณ„์‚ฐ๋น„์šฉ์ด ํฌ๊ธฐ ๋•Œ๋ฌธ์—)
      • ๋Œ€ํšŒ์—์„œ ์„ฑ๋Šฅ์„ ์กฐ๊ธˆ์ด๋ผ๋„ ๋” ์˜ฌ๋ฆฌ๊ธฐ ์œ„ํ•ด ์ฃผ๋กœ ์‚ฌ์šฉํ•œ๋‹ค.
      • ๊ฐ๊ฐ์˜ ๋ชจ๋ธ(3~15)์„ ๋…๋ฆฝ์ ์œผ๋กœ ํ•™์Šตํ•œ ํ›„, output์„ average ์ทจํ•œ๋‹ค.
    • multi-crop at test time
      • test image์˜ ์—ฌ๋Ÿฌ๊ฐœ์˜ crop์— ๋Œ€ํ•ด์„œ ๋ชจ๋ธ ์˜ˆ์ธก์น˜๋ฅผ ๊ตฌํ•˜๊ณ , ํ‰๊ท  ์ทจํ•˜์—ฌ ์ตœ์ข… ์˜ˆ์ธก์น˜๋กœ ์„ ์ •
  • ํŠน์ • ์ด๋ฏธ์ง€์—์„œ ์ž˜ ๋™์ž‘ํ•˜๋Š” ๋ชจ๋ธ์€ ๋‹ค๋ฅธ ์ด๋ฏธ์ง€์—์„œ๋„ ์ž˜ ๋™์ž‘ํ•  ํ™•๋ฅ ์ด ๋†’๋‹ค.
    • ๋”ฐ๋ผ์„œ ์ƒˆ๋กœ์šด ๋ชจ๋ธ์„ ๊ตฌ์ถ•ํ•  ๋•Œ ๊ธฐ์กด์˜ ์‹ ๊ฒฝ๋ง ๊ตฌ์กฐ์—์„œ๋ถ€ํ„ฐ ์‹œ์ž‘ํ•˜๋Š” ๊ฒƒ์ด ์ข‹๋‹ค.
  • ๊ฐ€๋Šฅํ•˜๋‹ค๋ฉด ์˜คํ”ˆ์†Œ์Šค๋ฅผ ํ™œ์šฉํ•˜๋ผ (์—ฌ๋Ÿฌ๊ฐ€์ง€ hyperparameter๋ฅผ ์ด๋ฏธ ํ…Œ์ŠคํŠธํ•ด๋ณธ ๊ฒฐ๊ณผ์ด๊ธฐ ๋•Œ๋ฌธ์—)
  • pretrained model์„ ์‚ฌ์šฉํ•˜๊ณ , ์šฐ๋ฆฌ์˜ ๋ฐ์ดํ„ฐ๋กœ fine-tuningํ•˜์ž

'๐Ÿ™‚ > Coursera_DL' ์นดํ…Œ๊ณ ๋ฆฌ์˜ ๋‹ค๋ฅธ ๊ธ€

WEEK6 : Object Detection (2)  (0) 2020.12.23
WEEK6 : Object Detection (1)  (0) 2020.12.23
WEEK6 : Inception (googLeNet)  (0) 2020.12.23
WEEK6 : ResNet  (0) 2020.12.23
WEEK5 : CNN (convolutional neural network)  (0) 2020.12.21

+ Recent posts