[DL輪読会]Large Scale GAN Training for High Fidelity Natural Image Synthesis

DEEP LEARNING JP [DL Papers] Large Scale GAN Training for High Fidelity Natural Image Synthesis Kohei Nishimura, DeepX, Inc. http://deeplearning.jp/ 1

• http://deeplearning.jp/

• • ) • ( 2 ) 1 ,1 09 1 B DA CI 2

• https://openreview.net/forum?id=B1xsqj09Fm

• – – • C A O A 3

• ,32- 32 –, – ,3- 32 I 2,. n iG – r C • N I 32 ,35. N d G N N Cc aA Go > S = o 53 ., 32 I r t l nlG e G N j I N ps C N = 4

• dtHas – H • S – r O • – L O H R h l E kHc T b :- e oGmE E r - (, - • n gHas • • • , • ) – L - pD as iu H e - pD : as - 5

• . – ( ) !, # , – ( , 6

Shared Embedding • ( –7 ! – % ) ) E 3 B 7

Hierarchical Latent Spaces • 1 1 – 4 , 8 L • • – % H 8 1 G LzL 4 c 8 8 1 S S i • oheAR!B – S NL s ? S p t r u t S – ka • • u nl t 8

( • )z shared embedding ,/ – + 2= – ,/ • 0 + 2 = 8 8 ( * ()* +( . / . / 1 / 1 / 18 18 / , =10 1 .100 . =10,1 .100 + 1 .100 , 1 1, /8 / 18 , + . , = 1/ , / / =10,1 .100 ,/ , , +(# , = 1/ , / / =10,1 .100 ,/ , !"#$"% 9

Truncation Trick • t – – G i r u i s f n . ) . G u ep • . ) . f a – • – i o r u ep : ) • • . c .( ) u .( . r 10

Truncation Trick • 0 – – . 0 . ( 0 . ) 0 ) 11

Truncation Trick • D F . – . • • I . I 横軸がIS(Inception Score) 縦軸がFID 12

Truncation Trick • O ] . • ts o iO > [ u iO o dml tsl < a R> o >OG TO c n hr k ? f [ • z dml ts . Oeg = 13

Orthogonal Regularization • 0 • – G : : • – – 0: 0 % % :16 :16 G G 14

Generator • i g h – – . h 3 – h . h e 3 15

( ) • • 16

Discriminator • ( – ) Z Z D – - C ed a ( ) 17

Discriminator • 8, - ,0 ,0 ov n • 9 d –n – • : : : tg zyg n : eD tr D 5 5 % a : : i lc u g u g zy G p 9 18

• dtHas – H • S – r O • – L O H R h l E kHc T b :- e oGmE E r - (, - • n gHas • • • , • ) – L - pD as iu H e - pD : as - 19

• P – – – – • B SDU D 5 a D 18 N e D Gdc 38 Gdc 1 2 8 ( 15 D)D N A)D 31 G c e T – 83 2 1 3 20

1 • 5 68 – )2 10 0 ) – ,( 21

1 • 22

1 • 23

1 • – , , • – • • 24

2 • T – 0.4 • • • 2 2 , 2 F F 1 J F M n )1 F e FI F – ) • ) F F TS K 856 98 ( • 3 . -F – 0.4 2Dci J o 25

2 • 26

ImageNet • S – 5 1 1 – . • – 5 1 1 – . • . o M • . - : 0 3 : 0 3 S JFT-300M 30 30 mJ AT M- N k I ut 3 1 M 3 F iF G c F k G a T g r N M e c nG 27

• T i N . x T pi S.O A e G ( ) G d n 28

• 6 D : H 2/7 :TJ NFN – • 06 2 J – • LN. D :L FHF T N - 3FD 1FA HF T 7 P H 4I D - H 7 I 9 DPH F F J 4IL FJD 0 L 6 JFJD :L H 7 I HF F J LN. RRR NHFA N J 0 DPH F F J FIL FJD J I HF F J D JN 06 – HF SF FJFJD : H / L6 JFJD58 D J HF FHF T JF J2 J LN. RRR NHFA N J 0 L6 D J F A N F H J R GN F /A JFJD58 () AH NL A L H H J I JFJDNL 2/7N H N F H 7 R GN AHN H JF J 29

(128x128) • 30