[GoogleML] Neural Style Transfer

2023. 10. 11. 01:37ใ†ArtificialIntelligence/2023GoogleMLBootcamp

 

 

 

What is Neural Style Transfer?

C + S -> G

 

 

 

What are deep ConvNets learning?

input์— ๊ฐ€๊นŒ์šด ์–•์€ layer๋“ค์€ ์ถ”์ƒ์ ์ธ edge 

output์— ๊ฐ€๊นŒ์šด deeper layer -> ๋” ๋””ํ…Œ์ผํ•œ ๋‹จ์œ„๋กœ detect 

 

 

 

layer 2 -> ์„ , ๋„ํ˜• ๋‹จ์œ„๋ฅผ ํ™•์ธํ•  ์ˆ˜ ์žˆ๋‹ค

 

 

 

๊ตฌ์ฒด์ ์ธ ๊ฐ์ฒด๊นŒ์ง€ ์žก์•„๋‚ด๋Š” ๋งˆ์ง€๋ง‰ ๋‹จ๊ณ„ layer๋“ค 

 

 

 

Cost Function

๋‘ cost func์„ ํ•ฉํ•˜์—ฌ total 

 

 

 

 

 

 

Content Cost Function

 

 

 

Style Cost Function

์ฑ„๋„์„ 5๊ฐœ๋กœ (์ƒ‰์œผ๋กœ) ๊ตฌ๋ถ„

 

 

 

์˜ค๋ Œ์ง€์™€ ๋…ธ๋ž€์ƒ‰์˜ ์Šคํƒ€์ผ์ด ์œ ์‚ฌํ•˜๋‹ค - correlated

์ฆ‰, vertical๊ณผ ์ฃผํ™ฉ๋น›์ด ์„œ๋กœ correlated (์ฆ๊ฐ€ํ•˜๋ฉด ์ฆ๊ฐ)

 

 

 

๋‘ activation ๊ฐ’์ด ์ปค์ง€๋ฉด (correlated) G๊ฐ’๋„ ์ฆ๊ฐ€ํ•จ

๋‘˜์ด uncorrelated -> G๊ฐ’๋„ ์ž‘์•„์ง

 

 

 

 

 

 

 

 

 

1D and 3D Generalizations

1D data์—๋„ ์“ฐ์ผ ์ˆ˜ ์žˆ๋‹ค

 

 

 

depth๋ผ๋Š” ์ถ”๊ฐ€์ ์ธ ์ฐจ์›!

 

 

 

3์ฐจ์› conv ์—ฐ์‚ฐ -> filter ๋˜ํ•œ 3์ฐจ์›์ด๋‹ค! 

๋’ค์— ๋ถ™๋Š” 4๋ฒˆ์งธ ์š”์†Œ๊ฐ€ ํ•„ํ„ฐ ์ˆ˜

 

 

 

 

 

'ArtificialIntelligence > 2023GoogleMLBootcamp' ์นดํ…Œ๊ณ ๋ฆฌ์˜ ๋‹ค๋ฅธ ๊ธ€

[GoogleML] Recurrent Neural Networks  (1) 2023.10.17
[GoogleML] Convolutional Neural Networks ์ˆ˜๋ฃŒ  (0) 2023.10.11
[GoogleML] Face Recognition  (0) 2023.10.09
[GoogleML] Object Detection  (1) 2023.10.09
[GoogleML] Convolutional Neural Networks Case Studies  (1) 2023.10.06