è«æ
[1703.05192] Learning to Discover Cross-Domain Relations with Generative Adversarial Networks
èè
Taeksoo Kim, Moonsu Cha, Hyunsoo Kim, Jung Kwon Lee, Jiwon Kim
èæ¯
æã 人éã¯è±èªã®æã¨ãããã翻訳ãããã©ã³ã¹èªã®æã®é¢ä¿æ§ã容æã«èªèããäºãå¯è½ã§ããã ã¾ããã¹ã¼ãã®ä¸çãè¦ããããã¨å¯¾ã«ãªãåããã¡ã¤ã³ã®æãé´ãé¸ã¶ãã¨ãã§ããã
ãã®ããã«ãï¼ã¤ã®ãã¡ã¤ã³éã®é¢ä¿æ§ãèªèãããã¨ã¯äººéã«ã¨ã£ã¦ã¯ããããããã³ã³ãã¥ã¼ã¿ã«ã¨ã£ã¦ã¯é常ã«é£ããåé¡ã¨ãªã£ã¦ããã ï¼ã¤ã®ãã¡ã¤ã³éã®é¢ä¿æ§ãèªèãããã¨ã¯ãä¸æ¹ã®ãã¡ã¤ã³ã«åå¨ããç»åããããä¸æ¹ã®ãã¡ã¤ã³ã«é©åããç»åãçæããã¨ããåé¡ã«ç½®ãæãããã¨ãã§ããã
ãã®åé¡ã解ãããã«ã¯ãé¢ä¿æ§ã示ããç»åã®ãã¢ãè¨å¤§ã«ç¨æããå¿ è¦ãããã ä¸æ¹ã§ããã¡ã¤ã³éã§å¯¾ã«ãªã£ã¦ããç»åã¯é常æã«å ¥ããªãã ä»®ã«æã«å ¥ã£ãã¨ãã¦ããï¼å¯¾å¤ã®çµã«ãªã£ã¦ãã¾ããã¨ãèµ·ããããã
ç®çã¨ã¢ããã¼ã
ç®ç
- ãã¡ã¤ã³éã®é¢ä¿æ§ã®èªå¾çãªå¦ç¿
ã¢ããã¼ã
- Discovers cross-domain relations with Generative Adversarial Networks (DiscoGAN)
- 2ã¤ã®ã©ãã«ç¡ãç»åã®éåãç¨ããå¦ç¿
- äºåå¦ç¿ãä¸è¦ãªå¦ç¿ã¢ãã«
ææ¡ææ³
DiscoGAN
å¾æ¥ã®GANã®ã¢ãã«ã®åé¡ã(a)ã«ç¤ºãã å¾æ¥ã®GANã®ææ³ã§ã¯ãå ¥åç»åããGeneratorãéããã¨ã§ãå¥ã®éåã«ä¼¼ãããã§ã¤ã¯ã®ç»åãçæããã ãã®çµæãæ師ãã¼ã¿ã¨çµã¿åããã¦Discriminatorã®å ¥åã¨ãããããã®æåããã¡ã¤ã³ã«å±ããç»åã®éåã§ãã£ã¦ãå¤æåã¨å¤æå¾ã®ç»åã対å¿ä»ããå¿ è¦ãããã
次ã«ãææ¡ææ³ãå°å ¥ããå ´åã(b)ã«ç¤ºãã ä¸æ¹ã®ãã¡ã¤ã³ã«å±ããç»åãGeneratorã«å ¥åããã¨ãä»æ¹ã®ãã¡ã¤ã³ã«å±ããç»åã«å¤æããã(Handbag -> Shoe, Shoe->Handbag) ãã®ããã«2ã¤ã®ãããã³ã°ï¼A->B, B->Aï¼ãå©ç¨ãããã¨ã§ãåæ¹åã®å¤æãå¯è½ã«ãªãã
ãããã¯ã¼ã¯è¨è¨
DiscoGANã®ãããã¯ã¼ã¯ã®æ¦è¦ãä¸è¨ã«ç¤ºãã
(a)ã«ä¸è¬çãªGANã®ã¢ãã«ã示ãã¦ããã ãã¨ãã¨ã¯ãä¹±æ°ãå ¥åã¨ãã¦MNISTã®ãããªç»åãåºåãã¦ããã
ãã¡ã¤ã³Aã¨ãã¡ã¤ã³Bã対å¿ä»ãããã®ã(b)ã«ç¤ºãã ãã®ã¢ãã«ã§ã¯A->Bã¸ã®å¤æã¨B->Aã¨ãçµã¿åããããã¨ã§ãç»åããã¨ã®ãã¡ã¤ã³Aã«æ»ãã ãããåãã®å ¥åç»åã¨æ¯è¼ãããã¨ã§Construction Lossãåãã
ææ¡ææ³ã§ã¯ã(c)ã®ããã«ä¸è¨ã®ã¢ãã«ã対ã«ãããã¨ã§2ã¤ã®ãã¡ã¤ã³ã®åæ¹åã®å¤æãå¯è½ã«ããã
Loss
DiscoGANã§ã®Lossã¯ä¸è¨ã®ããã«è¡¨ç¾ãããã
ããã§ã
ã¨ãªã£ã¦ããã ãGenerative Adversarial Lossã ã Reconstruction Lossã¨å¼ãã§ããã
ãããã¯ãé常ã®GANã®Lossã«å¯¾ãã¦Reconstruction Lossãä»å ããå½¢ã¨ãªã£ã¦ããäºããããã
Generative Adversarial Loss
é常ã®GANã§ã¯ãGeneratorãåºåããç»åãDiscriminatorã«å ¥åãã¦ç»åãæ師ãã¼ã¿ã§ãããè¤è£½ãã¼ã¿ã§ããããå¤æããããã¨ã§å¦ç¿ãè¡ãã
æ°å¦çã«ã¯ãåãã¡ã¤ã³A, Bã«å±ããç»åãã¡ã¤ã³éã®é¢ä¿æ§ã¯ã®ããã«è¡¨ãããã
ãã®ã¨ã Generative Adversarial Lossã¯ä¸è¨ã®ããã«ç¤ºãããã
]
]
ã¾ããDiscriminatorã®Lossã¯
Reconstruction Loss
ä¸è¨ã®ããã«å¤æãè¡ãå ´åãå ¥åç»åã¯å¤æãçµã¦ãã¨ã®ãã¡ã¤ã³ã«æ»ããã¨ã«ãªãã
ããã¨å ¥åç»åã®å·®åãè¨ç®ãããã¨ã§Lossã¨ãã¦åãæ±ããã¨ãã§ããã å³å¯ã«åãç»åã«æ»ãããã§ã¯ãªããã¨ãèæ ®ããã¨ãä¸è¨ã®ããã«ãªãã
ã¾ããMode Collapseåé¡ã«ãææ¡ææ³ã¯æå¹ã¨ãªããä¸è¨ã®å³ã®ãããªå¤æãè¡ãããã
(a)ã®ããã«åããã¡ã¤ã³ã®ç»åããå¥ã®ãã¡ã¤ã³ã«å±ããç»åãçæããã ä¸æ¹ã§ãMode Collapseã®ç¾è±¡ãçºçããå ´åã¯(b)ã®ããã«åä¸ã®ç»åããçæã§ããªããªãã
ããã§Reconstruction Lossãå«ã¾ããã°(c)ã®ããã«ãåæ¡ä»¶ã®ã¤ã³ãããç»åããã§ããç»åã®å¾©å ã®çµè·¯ã«ãã£ã¦Reconstruction Lossã大ãããªããããä¸æ¹åã®ç»åçæããmode Collapseã«å¯¾ãã¦æå¹ã¨ãªãã
è©ä¾¡
åç´è©ä¾¡
ã¾ããåç´ãªGANã¨reconstruction lossã使ç¨ããGANãããã¦ææ¡ææ³ã使ã£ãGANã«é¢ããåç´è©ä¾¡ãè¡ãã 説æã®ããã«ãã¡ã¤ã³ã¯A, Bã®ï¼ã¤ã¨ããGaussian mixture modelsã«ãã£ã¦ãµã³ãã«ãæç»ããã
åã¢ãã«ã«ã¤ãã¦50000iteration å¦ç¿ããçµæã«ã¤ãã¦ä¸è¨ã«ç¤ºãã
ä¸å³ã«é¢ããè²ä»ãã®èæ¯ã¯Discriminatorã«ããåºå (0~1) ã表ãã¦ãããçé«ç·ã¯Discriminatorã®åºåå¤ãåãé¨åã示ãã¦ããã 'x'å°ã¯ãã¡ã¤ã³Bã®ç°ãªããµã³ãã«ã表ãã¦ããã è²ä»ãã®â¯ã¯ãã¡ã¤ã³Aãããã¡ã¤ã³Bã«ãããã³ã°ããçµæã表ãã¦ããããã®è²ã¯å ¥åã®ãµã³ãã«ã表ãã¦ããã
ä¸å³ã®(a)ã¯åæç¶æ ã表ãããã¡ã¤ã³Aã®ã¢ãã«ããã¹ã¦åä¸å°ç¹ã«å¯¾ãã¦ãããã³ã°ããã¦ããäºããããã
åç´å®è£ ã®GANã®å ´å(B)ã¯ãè¤æ°ã®ãµã³ãã«ãå ¥åã¨ããGenerator ã®åºåããã¡ã¤ã³Bã§åããµã³ãã«ã®è¿ãã«ãããã³ã°ããã¦ããã ã¤ã¾ããå¤å¯¾ï¼ã®ãããã³ã°ã«ãªã£ã¦ãã¾ã£ã¦ããäºããããã
reconstruction lossã使ç¨ããGANã®å ´å(C)ã¯ãmode collapseãé¡èã«ç¾ãã¦ãããããã¤ãã¼ãç·ãæ°´è²ã«ã¤ãã¦ã¯2 ~ 3 ã®ã¢ãã«ã«åæãã¦ããã çé«ç·ã確èªããã¨ãåæç¶æ ã«æ¯ã¹ã¦Bã®ç°ãªãã¢ãã«ã®å¨å²ã«çé«ç·ã確èªã§ããã
ã¾ãã(B)ã¨(C)ã®ä¸¡æ¹ã«å ±éãã¦ãããã³ã°ãåå°ã«ãªã£ã¦ãããã¨ã確èªãããã
ææ¡ææ³(D)ã§ã¯ãmode collapseã¯è¦åãããããããã«ãããã³ã°ãå ¨åå°ã«ãªã£ã¦ããã
ãã®çµæãããå¾æ¥ã®ææ³ããé¢ä¿æ§ãæè»ã«å¦ç¿ãããããã³ã°ãå¯è½ã«ãªã£ã¦ããäºããããã
å®ç»åã®è©ä¾¡
è©ä¾¡ã¯ã®æ¡ä»¶ã¯ä¸è¨ã®éãã¨ãªã£ã¦ããã
- ç»å : 64Ã64Ã3
- å¦ç¿ç : 0.0002
- æé©å : Adam Optimizer
- ,
- ããããã¼ãã©ã¤ã¼ã¼ã·ã§ã³
- åãã¨çµãããé¤ãããã¹ã¦ã®Conv, deconv 層ã«é©ç¨
- decay regularization coefficient :
- ããããã : 200
- GPU : Nvidia Titan X Pascal GPU
- CPU : Intel(R) Xeon(R) E5-1620 CPU
CAR to CAR
3Dã®è»ã®ã¢ãã«ã«å¯¾ãã¦ã«ã¡ã©ã®ä½ç½®ã15°ãã¤ããããç»åã«é¢ãã¦è©ä¾¡ãè¡ã£ãçµæãä¸è¨ã«ç¤ºãã
â»å ¥åã®è§åº¦ã«å¯¾ãã¦ãæ師ãã¼ã¿ã¯ä¸å¿ã§å転ãããç»åï¼é¡ã¤ã¡ã¼ã¸ï¼ã¨ãªã£ã¦ãã模æ§ã
横軸ãå ¥åã®ç»åã®å転è§ã縦軸ãå¤æå¾ã®ç»åã®å転è§ã表ãã¦ããã å·¦ãããåç´ãªGAN(a)ã¨reconstruction lossã使ç¨ããGAN(b)ãããã¦ææ¡ææ³ã®DiscoGAN(c)ã«é¢ããçµæã示ãã¦ããã
(a)(b)ã«ã¤ãã¦ã¯èµ¤ã®ç¹ãããã¤ãã®ã¯ã©ã¹ã¿ãå½¢æãã¦ãã¾ã£ã¦ããäºããããã ãã®ãã¨ãããã»ã¨ãã©ã®å ¥åç»åã¯åãåºåç»åã«åæãã¦ãã¾ã£ã¦ããäºãããããmode collapseãçºçãã¦ããäºããããã
ä¸æ¹ææ¡ææ³ã§ã¯ãè§åº¦ã«å¯¾ãã¦åçã«çæç»åã®è§åº¦ãåå¸ãã¦ããããã®ãã¨ããè§åº¦ã®é¢ä¿æ§ãæ£ããèªèã§ãã¦ããã¨èããããã
FACE to FACE
次ã«ãä¸ã¨åæ§ã«äººéã®é¡ã§ãåæ§ã®è©ä¾¡ãè¡ã£ãã çµæãä¸è¨ã«ç¤ºãã
(a)ãå ¥åç»åã(b)ãåç´ãªGANã(c)ãreconstruction lossã使ç¨ããGANãããã¦(d)ãææ¡ææ³ã®DiscoGANã«é¢ããçµæã示ãã¦ããã ãã¡ããè»ã®ç»åã¨åæ§ã«ãææ¡ææ³ã«ããmode collapleãåé¿ã§ãã¦ããã
å¤æ°ã®ç¹å¾´ãå ±éããå ´åã®å¤æ
selebAãã¼ã¿ã»ããã使ç¨ãã¦ã人éã®é¡ãç¹å¾´ãç¶æããã¾ã¾å¤æã§ããããæ¤è¨¼ããçµæãä¸è¨ã«ç¤ºãã
æ¦ãç¹å¾´ã¯ç¶æã§ãã¦ãããã¨ã確èªã§ããã
å®å ¨ã«å¥ã®ãã®ãè§åº¦ã ãå ±éã«æã¤å ´åã®å¤æ
æ¤ åããè»ã¸ã®å¤æãè»ããé¡ã¸ã®å¤æã®çµæãä¸è¨ã«ç¤ºãã
ç»åã¯è§åº¦ãç¶æãã¦å¤æããã¦çæããã¦ãããé¢ä¿æ§ãååã«èªèã§ãã¦ããã
EDGES to PHOTOS
ç·ç»ããåçã¸ã®å¤æã®çµæãä¸è¨ã«ç¤ºãã
åçã®è²ãç¡æ°ã«åãããç¶æ³ã§ãã£ã¦ãå¤æã§ãã¦ããäºããããã
HANDBAG to SHOES, SHOES to HANDBAG
åé ã®ç»åãçµæã¨ãªã£ã¦ããã æ¦ãæ£ããå¤æã§ãã¦ããã
çµè«
ãã¡ã¤ã³éã®é¢ä¿æ§ãèªå¾çã«å¦ç¿ããDiscoGANãææ¡ãããããã«ããæ示çã«ç»åã«ã©ãã«ãä»ä¸ãããã¨ãªãå¦ç¿ãå¯è½ã«ãªãã DiscoGANã«ãã£ã¦æ§æ度ã®ç»åçæãå¯è½ã§ãããã¨ãè©ä¾¡ã«ãã£ã¦ç¤ºãããã ä»å¾ã®æ¹éã®ä¸ã¤ã¨ãã¦ã¯ãConditionalç³»ã®GANã¨ã®çµã¿åãããèããããã
ææ³
CycleGANã¨ãã£ããã
ï¼å人çã¡ã¢ï¼CycleGANã¨ã®éãï¼Lossã®ã¨ãæ¹
DiscoGANã¯ãããªæãã«å¥ã ã«èããã
CycleGANã¯ãããªæãã«ä¸æ¬ãã¦èããã