è«æ
https://arxiv.org/abs/1710.10196
èè
Tero Karras, Timo Aila, Samuli Laine, Jaakko Lehtinen
NVIDIA
èæ¯
çæçææ³ã®ä¸ã§ãç¾å¨ç¹ã«åªãã¦ãããã®ã«ã¯ãautoregressive modelsãvariational autoencoders (VAE) ãã㦠generative adversarial networks (GAN)ãªã©ãæããããã ãããã«ã¯ããããé·æã»çæããããç¾å¨ãç 究ãé²ãããã¦ããã
ãã®ãã¡ãGANã¯çæããåå¸ï¼ç»åãªã©ï¼ã®è§£å度ãé«ããªãã«é£ãã¦ã©ã³ãã è¦ç´ ã®å½±é¿ãè²æ¿ããªããDiscriminatorã¯Generatorã®çæåå¸ã¨æ師ãã¼ã¿ãåºå¥ãããã¨ã容æã«ãªã£ã¦ãã¾ãããããã¯ã¼ã¯å ¨ä½ã®å¦ç¿ãä¸å®å®ã«ãªãã ãã®ããã«ãGANã¯é«è§£å度ã®ç»åãåºåãããã¨ãé£ããã¨ããåé¡ãããã
ç®çã¨ã¢ããã¼ã
ç®ç
- GANãç¨ããé«è§£å度åå¸ã®çæ
ã¢ããã¼ã
PGGANï¼Progressive Growing of GANsï¼
- 段éçã«ãããã¯ã¼ã¯å±¤ãå¢å ããã追å ãã層ã®å½±é¿ãå¦ç¿ã«åº¦åã«å¿ãã¦æ¯éãå¤å
- ãããããæ¨æºåå·®ãæ´»ç¨ããå¤æ§æ§ã®åä¸
- ãã«ãã¹ã±ã¼ã«ã«ããã¯ããµã¼ã¹ã¿ã¤ã³è¨éãç¨ããçµ±è¨çé¡ä¼¼åº¦ã®ä½¿ç¨
ææ¡ææ³
PGGANï¼Progressive Growing of GANsï¼
PGGANï¼Progressive Growing of GANsï¼ã®æ¦å¿µå³ãä¸è¨ã«ç¤ºãã
PGGANã§ã¯ãä½è§£å度ã®ç»åï¼ä¸ã§ã¯4Ã4ï¼ããå¦ç¿ãéå§ããå¾ã ã«å±¤ã追å ãããã¨ã§å¯¾è±¡ã®ç»åã®è§£å度ãåä¸ãããã 層ãæ°ãã追å ããéã«ã¯ãå¦ç¿ããªããã¦ããªãããã©ã³ãã è¦ç´ ãå¼·ããGeneratorã®åºåã«ã©ã³ãã è¦ç´ ãå¼·ãåæ ããã¦ãã¾ãã ããã§ãPGGANã§ã¯ä¸å³ã®ããã«ãåºåã«å¯¾ããæ°ãã追å ãã層ã®å½±é¿åº¦(α)ã段éçã«å¢ããã
ããã«ããããããã¯ã¼ã¯å ¨ä½ã®åºåã¨ãã¦å¦ç¿ãå®äºãã¦ããªã層ã®ã©ã³ãã è¦ç´ ã軽æ¸ãããã¨ãå¯è½ã«ãªããå¦ç¿ãå®å®åããã
ãããããæ¨æºåå·®ãæ´»ç¨ããå¤æ§æ§ã®åä¸
GANã§ã¯ãå¦ç¿ãã¼ã¿ã®ç¯å²ã§ããç¹å¾´ãå¦ç¿ã§ãããå¦ç¿ãã¼ã¿ã®éãããç¹å¾´ã«å¤§ããå½±é¿ãããå¦ç¿ããã¦ãã¾ãæããããã ãã®åé¡ã解決ããããã«ããããããã®æ¨æºåå·®ãç¨ããæ£è¦åã«ãã£ã¦ãããããã®ã¾ã¨ã¾ãããå¾ãããç¹å¾´ãå¦ç¿ããæ¹æ³ãå°å ¥ããã
ã¯ããã«ããããããå ã®ãã³ã½ã«ï¼HÃWÃCï¼ã§åãä½ç½®ã«å¯¾å¿ããç»ç´ ã«ã¤ãã¦æ¨æºåå·®ãè¨ç®ããHÃWÃCã®ãã³ã½ã«ãä¸ã¤æ±ããã 次ã«ãã®ãã³ã½ã«å ã®ãã¹ã¦ã®ãã¯ã»ã«ã®å¹³åãè¨ç®ããã¹ã«ã©å¤ãä¸ã¤å¾ãã æå¾ã«ãã®ã¹ã«ã©å¤ãè¤è£½ããä¸ã¤ã®ãã³ã½ã«ãä½æããã
ãã®å¦çã¯ãçè«ä¸ãããã¯ã¼ã¯ã®ã©ãã«ã§ãæ¿å ¥ãããã¨ãã§ãããã諸ã ã®äºæ ããDiscriminatorã®æçµå±¤ã«æ¿å ¥ããã ï¼è©³ç´°ã¯è«æã®Appendixåç §ï¼
è£è¶³
ã¤ã¡ã¼ã¸ã¨ãã¦ã¯ãããªæããããã§ãã
Generatorã¨Discriminatorã®æ£è¦å
å¦ç¿ãã¼ã¿ã«ç¹å¾´çãªå¦ç¿ãã¼ã¿ãå«ã¾ãã¦ããéã«ããã®ãã¼ã¿ã«å¦ç¿çµæãå¼·ãå½±é¿ãåãã¦ãã¾ãåé¡ãããã ä¸è¬ã«ãã®åé¡ã«å¯¾ãã¦ã¯ããããã¼ãã©ã¤ã¼ã¼ã·ã§ã³ã使ç¨ãã¦å¯¾å¦ãããã¨ãå¤ãã ããããæã ã¯GANã®ãããã¯ã¼ã¯èªä½ãåé¡ãªã®ã§ã¯ãªããå¦ç¿ä¿¡å·ã®å¤§ããã¨ãã®è©ä¾¡ã«å¶éããªããã¨ã«åé¡ãããã¨èããã
ããã§æã ã¯ä¸è¨ã®ï¼ã¤ã®æåã使ç¨ãããã¨ã§ãã®åé¡ã«å¯¾å¦ããã
Equalized learning rate
å¾æ¥ã®ãããã¯ã¼ã¯ã®éã¿ã®ç¹ç´°ãªåæåã¯è¡ããã0~1ã®ç¯å²ã§ã®åæåã«å¯¾ãã¦å®è¡æã«ãããã¯ã¼ã¯ã®éã¿ãå¤åãããã ãããã¯ã¼ã¯ã®éã¿ã«ä¸è¨ã®å¼ãå°å ¥ããã
ããã§ãã¯ãããã¯ã¼ã¯ã®éã¿ãã¯ã¬ã¤ã¤ãã¨ã®æ¨æºåå®æ°ã表ãã
é常Adamãªã©ã§ã¯ãå ¥åãã©ã¡ã¼ã¿ã¨ã¯ç¬ç«ã«æ¨æºåå·®ããã¨ã«å¾é ã決ããã ãã®ããããã¤ãããã¯ã¬ã³ã¸ã大ãããã©ã¡ã¼ã¿ã¯ããã§ãªããã©ã¡ã¼ã¿ãããå¦ç¿ã«æéãè¦ããã æã ã®ææ³ã§ã¯ããã¤ãããã¯ã¬ã³ã¸ãèæ ®ãããããããã¹ã¦ã®ãã©ã¡ã¼ã¿ã§å質ã®å¦ç¿ã¹ãã¼ãã確ä¿ãããã¨ãã§ããã
Pixelwise feature vector normalization in generator
Generatorã¨Discriminatorãã³ã³ããã¼ã«ã§ããªããªããªãããã«ãGeneratorã®åç³ã¿è¾¼ã¿å±¤ã®ãã¨ã«åãã¯ã»ã«ãã¨ã«feature vectorãæ£è¦åããå¦çãè¡ãã
ããã§ãã¯ãããã¯ã¼ã¯ã®éã¿ãã¯ç¹å¾´ãããã®æ°(ãã£ã³ãã«æ°)ããã¯ãããããªãªã¸ãã«ã¨æ£è¦åãããç¹å¾´ãã¯ãã«ã表ãã
ããã«ãããå¤åã®å°ããªå¦ç¿ã®éã«ã¯å½±é¿ã¯å°ãããå¤åã®å¤§ããªå¦ç¿ã®éã«ã¯å¿ è¦ã«å¿ãã¦ä¿¡å·ãæ¸è¡°ããããã¨ãã§ããã
ãã«ãã¹ã±ã¼ã«ã«ãããçµ±è¨çé¡ä¼¼æ§ã®ä½¿ç¨
ç°ãªãGANã®ã¢ãã«ãè©ä¾¡ããã«ã¯ãè¨å¤§ãªæ°ã®çæç»åãè©ä¾¡ããå¿ è¦ãããã主観çã«ãªã£ã¦ãã¾ãã ãããã£ã¦ããããã®è©ä¾¡ã«ã¯ç»åããå¾ãããææ¨ãç¨ãã¦èªåçã«è©ä¾¡ããå¿ è¦ãããã ããã¾ã§ããmulti-scale structural similarity (MS- SSIM)ã®ãããªææ³ããã£ãããè²ãå½¢ç¶ã®å¤æ§æ§ã«ããç´°ããªå¹æã«ã¯å½±é¿ãã«ããã
æã ã¯ãåªããGeneratorã¯ã©ããªã¹ã±ã¼ã«ã«ããã¦ãå¦ç¿ãã¼ã¿ã¨é¡ä¼¼ããå±æçç»åæ§é ãçæãããã®ã ã¨èãã¦ããã ããã§ãæ師ãã¼ã¿ã¨ã©ãã©ã·ã¢ã³ãã©ãããã«ãã£ã¦å¾ãããå±æã¤ã¡ã¼ã¸ãçµ±è¨çã«æ¯è¼ããææ³ãç¨ããã é£ç¶ãããã©ãããé層ã«ããã¦ãåã®ã¬ãã«ãã¢ãããµã³ããªã³ã°ãããã®ã¨æ¬¡ã®ã¬ãã«ã¨ã®å·®åãè¦ããã¨ã§ããããå®ç¾ããã
å層ã¯ï¼Ãï¼ã®ç»ç´ è¿åã«ãã£ã¦è¡¨ç¾ãããè¨è¿°å(å層128å)ã§è¡¨ç¾ãããã ããããã®è¨è¿°åãæ£è¦åããæ¨æºåå·®ãè¨ç®ããã®å¾ã¯ããµã¼ã¹ã¿ã¤ã³è¨éã«ãã£ã¦é¡ä¼¼åº¦ãå¤å®ããã ããã«ãããå°ãã解å度é¨åã§è¦ãã¨ãã®å·®åã¯ç»åå ¨ä½ã®å¤§ã¾ããªæ§é ãæããé«è§£å度é¨åã§è¦ãã¨ãã®å·®åã¯è¼ªéããã¤ãºã¨ãã£ãé¨åã¾ã§è¡¨ç¾ããã
è£è¶³
ã¤ã¡ã¼ã¸ã¨ãã¦ã¯ãããªæãã¿ããã§ãã
è©ä¾¡
å®éã«çæãããç»åã¯ãã¡ãã
çµ±è¨çé¡ä¼¼åº¦ã«é¢ããè©ä¾¡
sliced Wasserstein distance (SWD) ã¨multi-scale structural similarity (MS- SSIM)ã«é¢ããè©ä¾¡ãè¡ãã
æ¡ä»¶ã¯ä¸è¨ã®ã¨ããã
- ææ°ã®loss function (WGAN-GP)ã¨å¦ç¿ã®è¨å®ã使ç¨
- å¦ç¿ãã¼ã¿ : CelebA, LSUN BEDROOM
- 解å度 : 128Ã128
次ã®è¡¨ã¯æ°å¤çã«SWDã¨MS-SSIMã«é¢ãã¦ã¯ããµã¼ã¹ã¿ã¤ã³è¨éã¨MS-SSIMãè¨ç®ããçµæã§ããã
ã¾ããææ¡ææ³ã«ãã£ã¦ææ¡ãããç»åãä¸è¨ã«ç¤ºãã
ç´æçã«ãè¯ãè©ä¾¡ææ¨ã§ã¯æ§ã ãªè²ãå½¢ç¶ãè¦ç¹ãåæ ããããã£ã¨ããããç»åãè¯ãè©ä¾¡ã¨ãããã¹ãã§ããã ã¤ã¾ããï¼hï¼ã®è¡ãæãå¤ãå°ãããªãã¯ãã§ããã ããããä»åãããMS-SSIMã«ã¯è¦ãããªãã ãã®ãããMS-SSIMããææ¡ææ³ã®SWDã®ã»ããè¯ãææ¨ã¨èããããã
åæåã³å¦ç¿ã¹ãã¼ã
åæã¨å¦ç¿é度ã«é¢ããè©ä¾¡çµæãä¸è¨ã«ç¤ºãã
å³(a)ã¯progressive growingããªãå ´åã(b)ã¯progressive growingãããå ´åã示ãã¦ããã ããããæ¯è¼ããã¨ãprogressive growingã使ç¨ããã»ããåææã®SWDã®å¤ãå°ãããå¦ç¿æéãç縮ã§ãã¦ããäºããããã
progressive growingãªãã ã¨ãã¯ããªè¦ç¹ã¨ãã¯ããªè¦ç¹ã®ä¸¡æ¹ã®å¦ç¿ãåæã«ããªãã ä¸æ¹progressive growingããã®å ´åã¯ä½ã解å度ã®å¦ç¿ã¯çµãã£ã¦ããããã解å度ã大ããããã¨ãã«å®å®ãã¦å¦ç¿ã§ãã¦ããã
å³(b)ãããåã©ãã©ã·ã¢ã³ãã©ãããã®å±¤ã«ããã¦ãã¯ããã®ä½è§£å度ã®ã¨ãã¯ã¾ã£ããSWDãæ¸å°ãã¦ããã 解å度ãåä¸ãã¦ããä¸è²«ãã¦SWDã¯æ¸å°å¾åã示ãã¦ããã ä¸æ¹ãå³(a)ã§ã¯ãã¹ã¦ã®åã©ãã©ã·ã¢ã³ãã©ãããã®å±¤ã§ä¼¼ãã«ã¼ãã®åããè¦ãããã¼ã¿ã«ãã£ã¦å ±é³´ããå¦ç¿ããã©ã£ã¦ããã
å³(c)ãããprogressive growingã使ç¨ããå ´åã¯1024Ã1024ã®ç»åã«ã¤ãã¦åæããã¾ã§ã«96æéè¦ããã®ã«å¯¾ããprogressive growingã使ç¨ããªãã£ãå ´åã«ã¯520æéãè¦ããã ãã®ãããprogressive growingã使ç¨ããã»ããç´5.4åé«éåããäºããããã
CelebAãç¨ããé«è§£å度ç»åã®çæ
æ¬è«æã§1024Ã1024ã®é«è§£å度ã®ç»åãçæããçµæãä¸è¨ã«ç¤ºãã
ãããã®ç»åãçæããã®ã«ãTesla V100 GPUã8æ©ä½¿ç¨ãã¦ã4æ¥è¨ç®ããã
LSUNã®çµæ
LSUNã®BED-ROOMã«ã¤ãã¦ãçæç»åãæ¯è¼ãããã®ãä¸è¨ã«ç¤ºãã
ã¾ããä¸è¨ã«ã¯å¥ã®LSUNã«ãã´ãªã®å³ã示ãã
å ¨è¬çã«çæç»åã®è³ªã¯é«ãããå è¡ç 究ã¨å¤§å·®ãªãã
å®è£
èè ã®æ¹ãå®éã«æ¸ãã¦ããã¦ããã®ã§ããã¡ãããåç §ã
https://github.com/tkarras/progressive_growing_of_gans
ã³ã¼ããçºããã ããªããããã©ãå®éã«è¨ç®ã¯åå¼ãã¦ãã ããã è«æã®å®è£ ã§ã¯100ä¸ãããã®GPU8æ©ä½¿ã£ã¦ãããGitHubã®READMEã«ããã¨GPUã¯1000ä¸è¶ ããã¿ããã§ããã ããã¾ã§ããã¨ä¸è¬äººã§ã¯ãäºç®çã«å®è¡ãããã¨ããä¸å¯è½ã§ããã
ãã¾ã
æ¸ãã¦ãéä¸ã§è¶ ããããããããã°ã§è¦ã¤ãã¾ããã ãããåèã«ããã¦ããã ãã¾ããã ãããã¨ããããã¾ãã