[DL輪読会]A Style-Based Generator Architecture for Generative Adversarial Networks

>100 Views

January 25, 19

スライド概要

2019/01/25
Deep Learning JP:
http://deeplearning.jp/seminar-2/

シェア

またはPlayer版

埋め込む »CMSなどでJSが使えない場合

(ダウンロード不可)

関連スライド

各ページのテキスト
1.

DEEP LEARNING JP [DL Papers] A Style-Based Generator Architecture for Generative Adversarial Networks http://deeplearning.jp/ 1

2.

• • • • • • • ( ) ( 2

3.

• deb – gN G IG • de S Q • A :8 :1 1 :8 82 28: :8 10 , 0 : 8 aD P • V f i 3

4.

• G • e , l – l • p g e – l r do n n g s t o • • i a , , , 4

5.

• Ø 5

6.

• • 6 2 ) ) 0 12 ( L. Gatys et al. “Image Style Transfer Using Convolutional Neural Networks Leon”, CVPR, 2016 6 6

7.

Content loss Style loss • G • – • – 7

8.

• , – • , , 8

9.

AdaIN • gaN H • X • ) , H X N I I ]I I ., 1 , Nc A ie ( 3, 1 , 1 ] ., ( d [ 9

10.

AdaIN • – I B • – I B 10

11.

AdaIN • – I B どちらも, γとβは学習パラ メータ • – I B 11

12.

AdaIN • , – – – N Adc A I a X. Huang et al. “Arbitrary Style Transfer in Real-time with Adaptive Instance Normalization”, 2017. N e 12

13.

• • N • , A AN A I 13

14.

Style based generator • N 2 I • , 2 41 2 I A 5 • 5 5 2 14

15.

• N – – – – – – ( L • • F ow – ( S x mn A ) g iFdl A ) -S D - B Se v pH ty C E a , h kQ P S S u r I pH (( F a z G F a ( ) - F s w 15

16.

style-based generator • Ø - , , A I 16

17.

mixing regularization • A 1 I • , 2 1 例) 異なる潜在変数 から生成したパラ メータを用いる 17

18.

mixing regularization • 18

19.

mixing regularization • 648 , - 6 03 043 • 6 :1 20 ,4 6 6 :1 Styleをmix ) – • 20--1 6 :1 ( – • 03 6 :1 ( – Ø 6 :1 19

20.

• , – 20

21.

disentanglement • • , , G , G Ø , Ø A - N 21

22.

Perceptual path length • e e o a p S. Laine. Feature-based metrics for exploring the latent space of generative models. ICLR workshop poster, 2018. • • • n , ] e a o l [ , e o 4 i 22

23.

Perceptual path length • b e a • ( h d l • 2-) 2 ), -, ) - , ) , ) - ) i • 2 -) 2)( ) ) e • W -) -, ) c h , d W g b n , ) - , 23

24.

Perceptual path length • • - 24

25.

Linear separability • , 2 • 2 , 25

26.

Linear separability c 23 1 34350 H c t S mt H ( V Y :. – • • .) , S S r )s do H , do n ) i Va V Va X V n M | 26

27.

The FFHQ datase • ( • -.0., a • E ( /1 20C U Ua )hg Lc e d QR A:7 4 H AB U 7 27

28.

• g a • t s i • • / o xd z –p –e – r • y r v h s h : nz v l : // ./ ) z ( 28

29.

7 . 2 3 1 K L S 6L K . G LH BCL L H G LCN N K C 3 L H T . L K L S0F A 6L 7 GK 8KCGA ,HGNH LCHG 3 L H K 2 HGT , 4 GA L S CL 6L 7 GK CG LCF CLB 0GKL G 3H F CR LCHGT 6 2 CG S- L K F L C K H PI H CGA LB L GL KI A G LCN FH KT 0,2 H KBHI IHKL ILCN H 29