StackGANã«ãããã©ã³ãã®é¬éè¡
è¿æ³
ããã¾ãã«ãä»ã¾ã§ã®ä¸å¯è½ãæéãç¶ããDeep Learningãèªåã§è¨ãã¨çã ç§å¦ã¨ãã¦ã®æ©æ¢°å¦ç¿ãé¶è½ãããã§ãããã¾ã Deep Learningã¯ä½ç³»åãããç¥æµã®éåä½ã¨ãã¦ã®æ£ããç§å¦ã®æ®µéã«ã¯ãã©ãçãã¦ããªãããã«æãã¾ããã©ã¡ããã¨è¨ãã¨é¬éè¡ã«è¿ãæããã
ãDeep Learningã¯ããã¤ãã¾ã è¦ã¬ççã¸ã¨äººé¡ãå°ãã¦ããããã§ãããããå人ãæå¾ ãã¦ããã¾ãã
ã¢ããã¼ã·ã§ã³
- æ¥æ¬èªã®ãã©ã³ããä½æããéã«ããã¶ã¤ãã¯è¨å¤§ãªæéçã»èä½çã»ç²¾ç¥çãªå´åãæããã¨ã«ãªã[1]
- ãã©ã³ãã®ãã¶ã¤ã³ãè¦ã¦ããã¨ããä¸å®ã®æ³åããããã¨ãããããããã¯ã大ã¾ããªãã©ã³ãã®ãã¶ã¤ã³ã®åå°ããã£ã¦ãã¶ã¤ã³ãè£ é£¾ããããã«ããªãããã®åºæãªè¡¨ç¾ãä»å ããã¦ãã
- ããã¯ãã£ã¼ãã©ã¼ãã³ã°ã§æ å ±ãä»å ããç¹å¾´ãè¦ããããå¤æãããã®ã«é©ãã好ä¾ã§ãã
- åç´ãªæ å ±ä»å ã§ã¯ãpix2pixã§ãå¯è½ã§ãããããStackGANã¨å¼ã°ããGANãè¤æ°åéãããã¨ã§ããé«åº¦ã«è¿ã¥ãããç»åã«è¿ã¥ããææ³ãé©å¿ãã[2]
å è¡ç 究
- Rewrite: Neural Style Transfer For Chinese Fonts
- End2Endã£ã½ãã®ã§StackGANã¯ï¼åè¿ããã£ã¼ãã®ãããã¯ã¼ã¯ã使ã£ã¦ããã®è¶ ããããå£ã§ã¯ãããã
- Analyzing 50k fonts using deep neural networks
- Fontæ å ±ã®ã¨ã³ãããã£ã³ã°åããããillustã§ããã¨illustration2vecã«ãªã
- StackGAN: Text to Photo-realistic Image Synthesis with Stacked Generative Adversarial Networks
- GANãè¤æ°åéãããã¨ã§ããé«ç»è³ªåããï¼ãã£ãã)
StackGANã«ã¤ãã¦
ãStackGANã¯å°ãªãæ
å ±ãããããã¨ã®çµµã«è¿ãæ
å ±ã復å·ããã®ã«é©ãã¦ãããããã¯ã¼ã¯ã§ããã
ããã¨ã®è«æã§ã¯Skip Thought Vectorsãç¨ãã¦æç« ããã®ç»åã®çæã§ããããSkip Thought Vectorsã®ä»£ããã«ãç»åã®ãã¯ãã«ãå
¥åãããæ¡ä»¶ãåãããã«ãããããDCGANã§ç¨ãããããã¤ãºãä»ä¸ããã
ããããã¯ã¼ã¯æ§æå³ãæ¸ãç´ããã
å®é¨ææ³
- è«æãèªãã§Chainerã§å®è£ ãã
- ä½åº¦ãå®é¨ããã¨ãããäºæ®µç®ã®GANã«ã¦å¾é æ¶å¤±ãèµ·ããã¦ãã¾ãã®ããä¸æ®µç®ã®GANã®åºåçµæã«å¼·ãå¼ã£å¼µããã¦ãã¾ã£ãã®ã§ãå¦ç¿ã¬ã¼ããå¤§å¹ ã«å¼ãä¸ããã¨ããå¦çãå ¥ãããæ¬å½ã¯ä¸å®ã®iterationãåãã¾ã§ãå¦ç¿ãé 延ããããã®ã ããä»ã®ã¨ããå®è£ ã«è³ã£ã¦ããªãã
- ãã©ã³ããï¼ç¨®é¡è¦ããã
- æ¸éã®ãããªãã©ã³ãã§ããéæ³é·æ¸ãã©ã³ããwindowsã§æ¨æºã§ã¤ã³ã¹ãã¼ã«ããã¦ããHGPåµè±è§ãããä½ã®ï¼ã¤ãæ¤è¨¼å¯¾è±¡ã¨ãã
- ãã©ã³ãã®é¸ææ¡ä»¶ã¯ãå¤ç®æ¼±ç³ã®ãåã£ã¡ãããããããããããå¾è¼©ã¯ç«ã§ãããã§ä½¿ç¨ããã¦ãã3313èªãç¨ããããã®ãã¡ã800èªãå¥éæ¤è¨¼ç¨ãã¼ã¿ã¨ãã¦å¦ç¿ç¨ãã¼ã¿ã¨ã¯åºå¥ããã[3]
- å種ãã©ã¡ã¼ã¿ã¯ãä¸åç®ã®GANã¯Facadeã¨åæ§ã§ãäºéç®ã®GANã¯ããã1/5ã«ãããã®ã§ããã詳細ãªãã©ã¡ã¼ã¿ãµã¼ãã¯å¥éãæ¬æ ¼çãªçæã¿ã¹ã¯ã§è¡ããã¨ã«ããã
çµæ
- Inputãå ¥åãã¯ãã«ä½æã«ç¨ããç»å
- Predictãåºåçµæ
- Ground Truthã人éã®è·äººãä½æãããã©ã³ãã§ãã
èå¯
- Ground Truthã«æ å ±éçã«è¿ãç¿åã®åã§ã¯ãªãããã©ãç¿åã£ã½ãåã表ç¾å¯è½ãªã®ã¯GANæ ã«ã ãããå ¥åºåã®èª¤å·®ã®æå°åãç®çã§ãªãã®ã§ã人éãæ¸ããæåã¯æ½å¨çã«å¤§ããªãã¤ãºã誤差ãå«ãããã®ç¹ãå¸åãã¦ããã¦ããããã«æãã
- ãã¨ãã¨ã¯å人ãã©ã³ãä½å®¶ã®ä½æã³ã¹ããä¸ãããããããªä»çµã¿ãããã°ã¨æã£ã¦ãã³ã¼ãã£ã³ã°ãéå§ãããã®ã§ãããä»ç¾å¨ãéçºãæ¢ã¾ã£ã¦ãã¾ã£ã¦ããæ±äºéå·¥ãã©ã³ãã®è¶³ãã¦ãªããã©ã³ãã®ç©´åããã人éããã©ã³ããä½æããéã®ä¸æ¸ãã«ç¨ãããããè¯ãã¨æãã[3]
Future Work
- StackGANã®ãã©ã¡ã¼ã¿ã調æ´ãã¦ãä¸æ¸ããã漫ç»ãä¸æ°ã«ä½æãããªã©ãããï¼ãã¼ã¿ãªãï¼
- é³å£°ã®åé¡ã«ãé©å¿ã§ããæ°ããã¦ã¦ãå¨æ³¢æ°æåãã«ã¼ã«ã«ãããã£ã¦ä¸ãããä¸ãããããä½æ¥ããGANã«ããããã°ãã人éã£ã½ããªãã£ããã£ã¦ææããã£ã¦ãç·æ§ã§ã女æ§ã®ãã£ã©ã¯ã¿ãæ¼ãããããã®éãã§ããããã«æããã
- ãã©ã¡ã¼ã¿ãµã¼ããããããã¯ã¼ã¯æ§é ã®åé¡ã ãã§ãªããä»åã¯å¨ãã®å¦ç¿ã®å®å®åº¦ã«å¿ãã¦å¦ç¿ãéå§ãã¦ã»ããã¨ãããã¼ãºãçºçããã®ã§ããã®è¾ºãå®è£ ãã¦ãããã
- ä¸å¿ConditionalStackGANã«ç°¡åã«ããæ¹æ³ããã£ã¦ããã®æ¹æ³ãå³ç¤ºããã¨ããããªãµãã«ãªãï¼ã¨æãï¼
åèæç®
[1] æåæ°14000åï¼ ææ¸ãæ¯çãã©ã³ããã§ããã¾ã§ http://portal.nifty.com/kiji/151118195086_1.htm
[2] StackGAN: Text to Photo-realistic Image Synthesis
with Stacked Generative Adversarial Networks https://arxiv.org/pdf/1612.03242v1.pdf
[3] ãã·ããã¢ã®é¨å£«ãã«åºã¦ãããã©ã³ããåç¾ãããæ±äºéå·¥é»åæ¸ä½åè¨ç»ããå
¬éä¸æ¢ã« http://gigazine.net/news/20140702-toa-heavy-industries-font/
è¬è¾
ãã¤ã³ãã«ã¨ã³ã¶ã風éªã ãããããªããç±ãä¸ãããªããé£ã®å¸ã®ä¸å¸ãã¤ã³ãã«ã§ä¼ãã§ããã®ã§ãå¤åã¤ã³ãã«ãªãã ãããã¤ããã
ããããªä¸ãä»é±ã¯æ¥åã§ã¯ä¸»ã«AWSã¨Ansibleã¨ã®æ¦ãã«æéãå²ãã¦ããã®ã ãããã¯ãæ©æ¢°å¦ç¿ãã£ã¦ãæ¹ãæ°ã楽...(ã¨ã³ã¸ãã¢ã«åãã¦ãªãçæ)
é²æãªããã¨ãããããããã£ã±ãé³æ¥½ã¨MMDã®ã¡ããã§ãªãã¨ãä¹ãåã£ããMMDã§ã¯è¦ããã§å±±é¢¨ã®ã¢ãã«ãå¤æ°ã®ã¢ã¼ãã£ã¹ãã«ãã£ã¦é »ç¹ãªã¢ããã°ã¬ã¼ããè¡ããã¦ãããé常ã«ã¯ãªãªãã£ãé«ãã尊いã
ä½åä¸ã§ãã山風ã¯æ±é¢¨ã¨æµ·é¢¨ã¨ä¸ç·ã«ããããããã¨ãå¤ãã®ã ãããã®ä¸äººãçµã¿åãããã¨ã¤ãããããããã®æ¹åæ§ãç°ãªãæ§æ ¼ã®ãã¯ãã«ãè¦äºã«èª¿åãã¦ããæãã«ã¾ã¨ã¾ã£ã¦ãããèå³ã®ãã人ã¯YouTubeã§ã山風 MMDãã§æ¤ç´¢ãããã
MMDè·äººã«æè¬ã§ãããGPL v3ã©ã¤ã»ã³ã¹ã¨ãã«ããã»ããããã®ã§ã¯ã¨æãã