๋ณธ๋ฌธ ๋ฐ”๋กœ๊ฐ€๊ธฐ
1๏ธโƒฃ AI•DS/๐Ÿ“’ Deep learning

[์ธ๊ณต์ง€๋Šฅ] GAN

by isdawell 2022. 6. 13.
728x90

๐Ÿ“Œ ๊ต๋‚ด '์ธ๊ณต์ง€๋Šฅ' ์ˆ˜์—…์„ ํ†ตํ•ด ๊ณต๋ถ€ํ•œ ๋‚ด์šฉ์„ ์ •๋ฆฌํ•œ ๊ฒƒ์ž…๋‹ˆ๋‹ค. 

 

 

1๏ธโƒฃ  Generative model  


 

๐Ÿ‘€ CNN, RNN 

 

  • ์ด๋ฏธ ์žˆ๋Š” ๋ฐ์ดํ„ฐ ๋ถ„ํฌ๋ฅผ ์ž˜ ์ถ”์ถœํ•˜๋Š” ๋„คํŠธ์›Œํฌ 
  • P(Y|X) ๐Ÿ‘‰ ๋ฐ์ดํ„ฐ ๊ฐ๊ฐ์„ ์ž˜ ๊ตฌ๋ถ„ํ•˜๋Š” ํ™•๋ฅ  ๋ถ„ํฌ๋ฅผ ์ถ”์ถœ 
  • discriminative model 

 

 

 

๐Ÿ‘€ ์ƒ์‚ฐ์  ์ ๋Œ€ ์‹ ๊ฒฝ๋ง GAN 

 

  • data instance ๋ฅผ ์ƒˆ๋กœ ์ƒ์„ฑํ•˜๋Š” ๋„คํŠธ์›Œํฌ
  • ๋ฐ์ดํ„ฐ๋ฅผ ๋งŒ๋“œ๋Š” ์ž‘์—… 
  • ๋‹ค์–‘ํ•œ ์‘์šฉ๋ถ„์•ผ์— ์ ์šฉ๋  ์ˆ˜ ์žˆ์Œ

 

 

โ‘   So far we've learn 

 

 

๐Ÿ”˜ Discriminative model 

 

  • ์ง€๊ธˆ๊นŒ์ง€ DNN, CNN, RNN ๋“ฑ ์ฃผ์–ด์ง„ ๋ฐ์ดํ„ฐ ๋ถ„ํฌ๋ฅผ ์ž˜ ๊ตฌ๋ถ„ํ•˜๊ธฐ ์œ„ํ•ด ์„ค๊ณ„ํ•œ ๋ชจ๋ธ๋“ค์„ ๋ฐฐ์› ์Œ

 

 

โ‘ก Generative model  VS Discriminative model 

 

๐Ÿ”˜ Discriminative model 

 

  • ๋‹ค๋ฅธ ์ข…๋ฅ˜์˜ Data instance ์‚ฌ์ด๋ฅผ ๊ตฌ๋ถ„ํ•˜๋Š” ๊ฒƒ
  • P(Y|X) : ์–ด๋–ค input X ๋ฅผ ์ž…๋ ฅํ–ˆ์„ ๋•Œ, Y ์˜ ๊ฒฐ๊ณผ๋ฅผ ์ถœ๋ ฅํ•˜๋Š” ์กฐ๊ฑด๋ถ€ ํ™•๋ฅ ์— ๋Œ€ํ•ด ํ•™์Šตํ•จ 
  • ex. ์ด๋ฏธ์ง€ ๋ฐ์ดํ„ฐ๋ฅผ ๋„ฃ์—ˆ์„ ๋•Œ, ๊ณ ์–‘์ด์ธ์ง€ ๊ฐ•์•„์ง€์ธ์ง€ ๊ตฌ๋ถ„ํ•˜๋Š” task 
  • ๋ฐ์ดํ„ฐ ๊ณต๊ฐ„์ƒ์—์„œ ๊ฒฝ๊ณ„์„ ์„ (boundaries) ๊ทธ๋ฆฌ๋Š” ๋ชจ๋ธ์ด๋‹ค. 

 

 

 

๐Ÿ”˜ Generative model 

 

  • ์ƒˆ๋กœ์šด data instance ๋ฅผ ๋งŒ๋“œ๋Š” ์—ญํ• 
  • input X์™€ output Y ๋ฅผ ๋…๋ฆฝ์ ์œผ๋กœ ๋ณด๊ณ  ๊ทธ ์‚ฌ์ด์˜ ๊ด€๊ณ„๋ฅผ ํฌ์ฐฉํ•˜๊ฑฐ๋‚˜ input ์ž์ฒด์˜ ํ™•๋ฅ ๋งŒ์„ ํฌ์ฐฉํ•˜๋Š” ๋ชจ๋ธ
  • GAN ์€ generative model ์˜ ์—ฌ๋Ÿฌ ์ข…๋ฅ˜ ์ค‘์— ํ•˜๋‚˜์ž„ 
  • GAN : ์ ๋Œ€์  ๊ฒฝ์Ÿ๊ด€๊ณ„์— ๋†“์ธ ๋„คํŠธ์›Œํฌ๊ฐ€ ์ƒˆ๋กœ์šด ๋ฐ์ดํ„ฐ๋ฅผ ํ˜•์„ฑํ•œ๋‹ค.
  • ๋ฐ์ดํ„ฐ๊ฐ€ ๊ณต๊ฐ„์ƒ์—์„œ ์–ด๋–ป๊ฒŒ ๋ถ„ํฌ๋˜๋Š”์ง€ ๋ฐ์ดํ„ฐ ๋ถ„ํฌ ์ž์ฒด๋ฅผ ๋ชจ๋ธ๋งํ•œ๋‹ค.

 

 

 

 

โ— ์ฃผ์˜ : ๊ธฐ๊ณ„๋ฒˆ์—ญ ๋ชจ๋ธ์€ output ์œผ๋กœ ๋ฒˆ์—ญ๋œ ๋ฌธ์žฅ์ด ์ƒ์„ฑ๋˜์–ด์„œ generative model ๋กœ ํ—ท๊ฐˆ๋ฆด ์ˆ˜ ์žˆ๋Š”๋ฐ, ์ด ๋˜ํ•œ ์ฃผ์–ด์ง„ ์˜์–ด ์–ดํœ˜ ์†์—์„œ ํ”„๋ž‘์Šค์–ด ์–ดํœ˜์— ๋งž๋Š” ์ ์ ˆํ•œ ๋‹จ์–ด๋ฅผ ์„ ํƒํ•˜๋Š” ๊ฒƒ์ด๋ฏ€๋กœ discriminative model ์— ํ•ด๋‹นํ•จ 

 

 

โ‘ข Application of Generative models 

 

๐Ÿ”˜ ๋ฏธ์ˆ /์ด๋ฏธ์ง€์˜์—ญ 

 

 

  • ๋ชจ๋„ค์˜ ๊ทธ๋ฆผ์œผ๋กœ๋ถ€ํ„ฐ ์‚ฌ์ง„์ฒ˜๋Ÿผ ๋ณด์ด๋Š” ์ด๋ฏธ์ง€๋ฅผ ์ƒ์„ฑํ•˜๊ฑฐ๋‚˜ ์‹ค์ œ ์ดฌ์˜ํ•œ ์‚ฌ์ง„์œผ๋กœ๋ถ€ํ„ฐ ๋ชจ๋„ค์˜ ๊ทธ๋ฆผ์ฒ˜๋Ÿผ ๋ฌ˜์‚ฌํ•œ ์ด๋ฏธ์ง€๋ฅผ ์ƒ์„ฑํ•˜๋Š” ๊ฒƒ ๐Ÿ‘‰ ๊ธฐ์กด์— ์กด์žฌํ•˜์ง€ ์•Š์•˜๋˜ data instance ๋ฅผ ์ƒ์„ฑํ•ด์ค€๋‹ค.

 

 

  • ์‚ฌ๋žŒ ์–ผ๊ตด ์ด๋ฏธ์ง€์—์„œ ๋จธ๋ฆฌ ์ƒ‰๊น”์„ ๋ฐ”๊พธ๊ฑฐ๋‚˜, ์„ฑ๋ณ„/๋‚˜์ด/ํ”ผ๋ถ€์ƒ‰/ํ‘œ์ •์„ ๋ฐ”๊พธ๋Š” ๊ฒƒ ๐Ÿ‘‰ ๊ธฐ์กด์— ์กด์žฌํ•˜์ง€ ์•Š์•˜๋˜ data instance ๋ฅผ ์ƒ์„ฑ

 

 

๐Ÿ”˜ ๋ฐ˜๋„์ฒด ์˜์—ญ 

 

 

  • Photomask ๋ฅผ ์„ค๊ณ„ํ•  ๋•Œ, Wafer ์— ์˜๋„ํ•œ๋Œ€๋กœ ํšŒ๋กœ๊ฐ€ ์ž…ํ˜€์ง€๋„๋ก ํ•˜๋Š” GAN ๋ชจ๋ธ์„ ๋งŒ๋“ฆ 

 

 

 

 

2๏ธโƒฃ GAN 


 

โ‘   Key Idea : Police-Criminal Analogy

 

๐Ÿ”˜ Generator VS Discriminator 

 

  • Generator network ์™€ discriminator network ๋ฅผ ์ ๋Œ€์  ๊ฒฝ์Ÿ๊ด€๊ณ„๋กœ ๋งŒ๋“ค์–ด ์›ํ•˜๋Š” (์ƒˆ๋กœ์šด) ๋ฐ์ดํ„ฐ๋ฅผ ์ž˜ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•จ

 

 

  • Criminal : ์œ„์กฐ์ง€ํ๋ฒ”์€ ์ง„์งœ์ฒ˜๋Ÿผ ๋ณด์ด๋Š” ๊ฐ€์งœ ์ง€ํ๋ฅผ ๋งŒ๋“ค์–ด ๊ฒฝ์ฐฐ์„ ์†์ด๋ ค๊ณ  ํ•จ ๐Ÿ‘‰ Generator 
  • Police : ์ง„์งœ์™€ ๊ฐ€์งœ์ง€ํ๋ฅผ ๊ตฌ๋ถ„ํ•˜๊ณ ์ž ํ•จ ๐Ÿ‘‰ Discriminator 

 

๐Ÿ’จ Generator ๋Š” ๋” ์ •๊ตํ•˜๊ฒŒ ์ง€ํ๋ฅผ ๋งŒ๋“ค๊ณ ์ž ํ•  ๊ฒƒ์ด๊ณ , Discriminator ๋Š” ์ง„ํ’ˆ๊ณผ ๊ฐ€ํ’ˆ์„ ๊ตฌ๋ถ„ํ•˜๊ธฐ ์œ„ํ•ด ๋”์šฑ ๋…ธ๋ ฅํ•˜๋ฉฐ ์„œ๋กœ ์ ๋Œ€์  ๊ฒฝ์Ÿ๊ด€๊ณ„๋ฅผ ๊ฐ€์ง„๋‹ค. ํ•™์Šต์„ ์ง„ํ–‰ํ• ์ˆ˜๋ก fake data ๋Š” real data ์™€ ๊ตฌ๋ถ„ํ•  ์ˆ˜ ์—†์„ ์ •๋„๋กœ ๋น„์Šทํ•ด์ง€๊ฒŒ ๋œ๋‹ค. 

 

โ‘ก GAN overview

 

๐Ÿ”˜ ์‚ฌ๋žŒ ์–ผ๊ตด ์ด๋ฏธ์ง€๋ฅผ ์ƒ์„ฑํ•˜๋Š” GAN model 

 

 

  • Input : random noise → ๋žœ๋คํ•œ ๋…ธ์ด์ฆˆ๋ฅผ input ์œผ๋กœ ๋„ฃ์Œ 
  • Generator network ๋Š” ์‚ฌ๋žŒ์ฒ˜๋Ÿผ ๋ณด์ด๋Š” ๊ฐ€์งœ ์–ผ๊ตด ์ด๋ฏธ์ง€๋ฅผ ๋งŒ๋“ค์–ด๋‚ด๊ณ , Discriminator ๋Š” ์ง„์งœ ์‚ฌ๋žŒ ์–ผ๊ตด ์ด๋ฏธ์ง€์™€ ๊ฐ€์งœ ์–ผ๊ตด ์ด๋ฏธ์ง€๋ฅผ ๊ตฌ๋ถ„ํ•˜๊ณ ์ž ํ•จ 
  • ๋‘ network ๋Š” ์ ๋Œ€์  ๊ฒฝ์Ÿ๊ด€๊ณ„์— ๋†“์—ฌ์ ธ ์žˆ๊ธฐ ๋•Œ๋ฌธ์— fake image ๊ฐ€ real image ์— ๊ฐ€๊นŒ์›Œ์ง€๊ฒŒ ๋œ๋‹ค. 
  • ๊ฐ network ๋Š” DNN, CNN, RNN ๋“ฑ ์–ด๋Š ์ข…๋ฅ˜์˜ ๋„คํŠธ์›Œํฌ๋ผ๋„ ์˜ฌ ์ˆ˜ ์žˆ๋‹ค. ์ด ์˜ˆ์ œ์—์„œ๋Š” ์ด๋ฏธ์ง€ ๋ฐ์ดํ„ฐ์ด๋ฏ€๋กœ CNN ์„ ์‚ฌ์šฉํ•˜๋Š” ๊ฒƒ์ด ์ข‹๋‹ค. 

 

 

โ‘ข Training GAN : Two-Player Minimax Game

 

 

 

๐Ÿ”˜ GAN ์„ ์ˆ˜ํ•™์ ์œผ๋กœ ๊ธฐ์ˆ ํ•˜๊ธฐ 

 

๐Ÿ“Œ D(x) : Discriminator output ์œผ๋กœ input ์ด real image ์— ํ•ด๋‹นํ•˜๋Š”์ง€ ์—ฌ๋ถ€๋ฅผ 0๊ณผ 1 ์‚ฌ์ด์˜ ๊ฐ’์œผ๋กœ ํ‘œํ˜„ํ•œ๋‹ค. 1์— ๊ฐ€๊นŒ์šธ์ˆ˜๋ก real, 0์— ๊ฐ€๊นŒ์šธ์ˆ˜๋ก fake image ์— ํ•ด๋‹นํ•˜๋Š” ๊ฒƒ์ด๋‹ค. 

 

 

๐Ÿ’จ D (discriminator) ์ž…์žฅ์—์„œ๋Š” Value function ์„ ์ตœ๋Œ€ํ™” 

 

 

โญ real image ๋Š” real ๋กœ, fake image ๋Š” fake ๋กœ ์ž˜ ๊ตฌ๋ถ„ํ•˜๋„๋ก ํ›ˆ๋ จํ•œ๋‹ค. 

 

  • X~Pdata(x) : training data ๋กœ๋ถ€ํ„ฐ ์ถ”์ถœ๋œ x ๋Š” real image ์— ํ•ด๋‹นํ•˜๋Š”๋ฐ, D ์ž…์žฅ์—์„œ real ์ด๋ผ๊ณ  ์ž˜ ํŒ๋‹จํ•ด์•ผ ํ•˜๋ฏ€๋กœ D(x) ๊ฐ€ 1์— ๊ฐ€๊นŒ์›Œ์ง€๋„๋ก ํ›ˆ๋ จํ•œ๋‹ค ๐Ÿ‘‰ log(D(x)) ์ตœ๋Œ€ํ™” 
  • Z~Pz(z) : noise input ์— ๋Œ€ํ•˜์—ฌ G(z) , ์ฆ‰ generator network ๋ฅผ ๊ฑฐ์ณ์„œ, D(G(z)) discriminator ์— ๋„์ฐฉํ–ˆ์„ ๋•Œ, D ์ž…์žฅ์—์„œ๋Š” fake image ๋ฅผ fake ๋ผ๊ณ  ์ž˜ ๋ถ„๋ฅ˜ํ•ด์•ผ ํ•˜๋ฏ€๋กœ D(G(z)) ๋ฅผ 0์— ๊ฐ€๊นŒ์›Œ์ง€๋„๋ก ํ›ˆ๋ จํ•œ๋‹ค ๐Ÿ‘‰  log(1-D(G(z))) ์ตœ๋Œ€ํ™” 

 

 

 

๐Ÿ’จ G (generator) ์ž…์žฅ์—์„œ๋Š” Value function ์„ ์ตœ์†Œํ™” 

 

 

โญ fake image ๋ฅผ discriminator ๊ฐ€ real ๋กœ ์ž˜๋ชป ํŒ๋‹จํ•˜๊ฒŒ๋” ๋งŒ๋“ค๋„๋ก ํ›ˆ๋ จํ•œ๋‹ค. 

 

  • random noise 'z' ๊ฐ€ Generator ๋กœ๋ถ€ํ„ฐ ์ž…๋ ฅ๋˜์–ด fake image ๊ฐ€ ์ƒ์„ฑ๋˜์—ˆ์„ ๋•Œ, G ์ž…์žฅ์—์„œ๋Š” D(G(z)) ๊ฐ€ 1์ด ๋˜๋„๋ก ์ฆ‰, fake image ๋ฅผ real ๋กœ ํŒ๋‹จํ•˜๋„๋ก ํ›ˆ๋ จํ•œ๋‹ค ๐Ÿ‘‰ log(1-D(G(z)))  ์ตœ์†Œํ™” 

 

 

 

โœ” ์ฐธ๊ณ  : log x , log(1-x)  ๊ทธ๋ž˜ํ”„ ๋ชจ์–‘ 

 

์ดˆ๋ก : log(x) , ๋นจ๊ฐ• : log(1-x)

 

โ‘ฃ Training Steps 

 

๐Ÿ”˜ ์•„๋ž˜์˜ ๋‘ ๊ณผ์ •์„ ๋ฐ˜๋ณตํ•˜๋ฉด์„œ ํ›ˆ๋ จ์„ ์ง„ํ–‰ํ•œ๋‹ค. 

 

 

1.  Gradient ascent on disriminator ๊ฒฝ์‚ฌ ์ƒ์Šน๋ฒ• (maximize์ด๋ฏ€๋กœ)

 

2.  Gradient descent on disriminator ๊ฒฝ์‚ฌ ํ•˜๊ฐ•๋ฒ• (minimize ์ด๋ฏ€๋กœ) 
 - discriminator ๊ฐ€ ๋งž์ถœ ๊ฐ€๋Šฅ์„ฑ likelihood ๋ฅผ ๋‚ฎ์ถ˜๋‹ค. 

 

 

๐Ÿ‘€ ๊ฐœ๋…์ ์œผ๋กœ๋Š” ๋‘ ์ˆ˜์‹์ด ์ ๋Œ€์  ๊ด€๊ณ„๋ฅผ ์ž˜ ๋ฌ˜์‚ฌํ•˜๋Š” ์‹์ด ๋˜๋Š”๋ฐ, ํ›ˆ๋ จ ์ค‘์—๋Š” ์ด๋Ÿฌํ•œ ๊ด€๊ณ„๊ฐ€ ์ž˜ ์ž‘๋™๋˜์ง€ ์•Š๋Š”๋‹ค. 

 

 

  • ํ›ˆ๋ จ ์ดˆ๋ฐ˜๋ถ€์— discriminator (๊ฒฝ์ฐฐ ์—ญํ• ) ์€ ์ž˜ํ•˜์ง€๋งŒ, generator (์œ„์กฐ์ง€ํ์กฐ์ž‘๋ฒ”) ๋Š” ์ง€ํ๋ฅผ ์ž˜ ๋งŒ๋“ค์ง€ ๋ชปํ•œ๋‹ค. 
  • Gradient descent ๋ฐฉ๋ฒ•์œผ๋กœ generator ์— ๋Œ€ํ•ด ํ›ˆ๋ จ์„ ์ง„ํ–‰ํ•˜๋ฉด, ํ›ˆ๋ จ ์ดˆ๋ฐ˜์—๋Š” gradient ๊ฐ€ ์ž‘์•„์„œ ํ›ˆ๋ จ์ด ์ดˆ๋ฐ˜์— ์ž˜ ์ด๋ฃจ์–ด์ง€์ง€ ์•Š๊ฒŒ ๋œ๋‹ค ๐Ÿ‘‰ log(1-x) ๊ทธ๋ž˜ํ”„์—์„œ ๋ณด๋ฉด x๋ฅผ 1๋กœ ๊ฐ€๊นŒ์›Œ์ง€๊ฒŒ ๋งŒ๋“œ๋Š” ๋ฐฉํ–ฅ (์—ฌ๊ธฐ์„œ๋Š” D(G(z)) ๊ฐ€ 1์— ๊ฐ€๊นŒ์›Œ์ง€๋„๋ก ๋งŒ๋“œ๋Š” ์ƒํ™ฉ์ž„ → fake ๋ฅผ real ๋กœ ์˜คํŒํ•˜๊ฒŒ ๋งŒ๋“œ๋Š”) ์œผ๋กœ ๊ทธ๋ž˜ํ”„๋ฅผ ์‚ดํŽด๋ณผ ๋•Œ, ํ›ˆ๋ จ ์ดˆ๋ฐ˜ ์˜์—ญ์ธ ๊ทธ๋ž˜ํ”„ ์™ผ์ชฝ ์˜์—ญ์—์„œ ๊ธฐ์šธ๊ธฐ๊ฐ€ ๊ฐ€ํŒŒ๋ฅด์ง€ ์•Š์Œ

 

 

 

๐Ÿ”˜ ๋ณ€ํ˜•

 

  • G ์— ๋Œ€ํ•ด value function ์˜ training ์„ ์ง„ํ–‰ํ•  ๋•Œ, gradient descent ๋ฐฉ๋ฒ•์„ ์‚ฌ์šฉํ•  ๊ฒฝ์šฐ ์ดˆ๋ฐ˜์— ํ›ˆ๋ จ์ด ์ž˜ ์ด๋ฃจ์–ด์ง€์ง€ ์•Š๋Š”๋‹ค๋Š” ์ ์„ ๊ณ ๋ คํ•ด์„œ, ์ˆ˜์‹์„ ๋ณ€๊ฒฝํ•˜์—ฌ ํ›ˆ๋ จ ๋ฐฉ๋ฒ•์„ ๋ฐ”๊พธ์–ด์คŒ ๐Ÿ‘‰ Gradient ascent ๋ฐฉ๋ฒ•์œผ๋กœ ๋ณ€๊ฒฝ 

 

 

  • D(G(z)) ๋ฅผ 1๋กœ ๋งŒ๋“œ๋Š” ์ƒํ™ฉ์€ ๋˜‘๊ฐ™์€๋ฐ, log(D(G(z)) ๋กœ ์‹์„ ๋ณ€๊ฒฝํ•˜์—ฌ, ๊ฒฝ์ฐฐ์ด ์œ„์กฐ ์ง€ํ๋ฅผ ์ž˜๋ชป ๊ตฌ๋ถ„ํ•  ํ™•๋ฅ ์„ ๋†’์ด๋Š” ๋ฐฉ์‹์œผ๋กœ training ์„ ์ง„ํ–‰ํ•œ๋‹ค. 
  • log(1-x) ๋ฅผ ์ตœ์†Œํ™” = log(x) ๋ฅผ ์ตœ๋Œ€ํ™”, x~[0,1]
  • ์ฆ‰, ์œ„์กฐ์ง€ํ๋ฒ” ์ž…์žฅ์—์„œ๋Š” ๊ฒฝ์ฐฐ์ด ์ž˜๋ชป๋œ ํŒ๋‹จ์„ ํ–ˆ์„ ๋•Œ, ๊ทธ ์ ์„ ๋…ธ๋ ค์„œ ๋” ๋งŽ์€ ๊ฒƒ์„ ํ•™์Šตํ•˜๋Š” ๊ผด์ด ๋˜๋ฏ€๋กœ ์ˆ˜ํ•™์ ์œผ๋กœ ํ•ด์„ํ•˜๋ฉด gradient ๊ฐ€ ์ปค์ง„๋‹ค๊ณ  ํ•  ์ˆ˜ ์žˆ๋‹ค. 
  • log(x) ๊ทธ๋ž˜ํ”„๋ฅผ ๋ณผ ๋•Œ๋„, x๋ฅผ 1์— ๊ฐ€๊นŒ์›Œ์ง€๋„๋ก ํ›ˆ๋ จ์„ ์ง„ํ–‰ํ•  ๋•Œ, ํ›ˆ๋ จ ์ดˆ๋ฐ˜์˜ ๊ธฐ์šธ๊ธฐ (๊ทธ๋ž˜ํ”„์˜ ์™ผ์ชฝ๋ฒ”์œ„) ๊ฐ€ ๋งค์šฐ ๊ฐ€ํŒŒ๋ฆ„์„ ํ™•์ธํ•ด๋ณผ ์ˆ˜ ์žˆ๋‹ค. 

 

 

๐Ÿ”˜ GAN ์•Œ๊ณ ๋ฆฌ์ฆ˜์˜ ๋™์ž‘๋ฐฉ์‹ 

 

 

 

  • Discriminator ์— ๋Œ€ํ•ด gradient ascent ๋ฅผ k ๋ฒˆ ๋ฐ˜๋ณตํ•ด์ค€ ํ›„, generator ์— ๋Œ€ํ•ด gradient descent ๋ฅผ ํ•œ๋ฒˆ ๋ฐ˜๋ณตํ•˜๋Š” ํ˜•ํƒœ๋ฅผ ์ทจํ•œ๋‹ค. (์‹ค์ œ๋กœ generator ์— ๋Œ€ํ•ด์„œ๋Š” gradient ascent ๋ฅผ ์ ์šฉํ•œ๋‹ค) 

 

728x90

๋Œ“๊ธ€