GAN Ò²¿ÉÒÔ´ó·ùѹËõ£¬MIT º«ËÉÍŶӵÄ×îÐÂÑо¿ÈÃÖÚ¶àÑо¿ÕßÃÇΪ֮һÕñ¡£
Éú³ÉÄ£ÐÍ GAN ÊÇ»úÆ÷ѧϰÁìÓòÀï×îΪÖØÒªµÄ·¢Õ¹·½ÏòÖ®Ò»¡£µ«ÕâÀàËã·¨ÐèÒªÏûºÄ¾ÞÁ¿ËãÁ¦£¬´ó¶àÊýÑо¿ÕßÒѾºÜÄѵóöгɹû¡£½üÄêÀ´£¬ÕâÒ»·½ÏòÆÄÓб»´óÐÍ»ú¹¹Â¢¶ÏµÄÇ÷ÊÆ¡£
µ«½üÈÕ£¬À´×ÔÂéÊ¡Àí¹¤Ñ§Ôº£¨MIT£©¡¢Adobe¡¢ÉϺ£½»Í¨´óѧµÄÑо¿ÕßÌá³öÁËÒ»ÖÖÓÃÓÚѹËõÌõ¼þ GAN µÄͨÓ÷½·¨¡£Õâһм¼ÊõÔÚ±£³ÖÊÓ¾õ±£Õæ¶ÈµÄͬʱ£¬½« pix2pix£¬CycleGAN ºÍ GauGAN µÈ¹ã·ºÊ¹ÓõÄÌõ¼þ GAN Ä£Ð͵ļÆËãÁ¿¼õÉÙµ½ 1/9~1/21¡£¸Ã·½·¨ÊÊÓÃÓÚ¶àÖÖÉú³ÉÆ÷¼Ü¹¹¡¢Ñ§Ï°Ä¿±ê£¬Åä¶Ô»ò·ÇÅä¶ÔÉèÖá£
Ä¿Ç°¸ÃÑо¿µÄÂÛÎÄÒѱ» CVPR 2020 ´ó»áÊÕ¼£¬ÓÐ¹Ø GAN ѹËõ¿ò¼ÜµÄ PyTorch °æʵÏÖÒ²ÒÑ¿ªÔ´¡£
ÏîÄ¿Á´½Ó£ºhttps://github.com/mit-han-lab/gan-compression
GAN ѹËõºóµÄ¾ßÌåÐÔÄÜÈçºÎ£¿ÔÚÑо¿ÕßÃÇչʾµÄ Demo ÖУ¬Ê¹Óà CycleGAN ΪÊÓƵÖеÄÂíÌí¼Ó°ßÂíÌõÎÆËùÐèµÄËãÁ¦²»µ½ 1/16£¬Ö¡ÊýÌá¸ßÁËÈý±¶£¬¶øÇÒЧ¹û»¹ÓÐËùÌáÉý£º
ÖµµÃÒ»ÌáµÄÊÇ£¬¸ÃÑо¿ËùʹÓõÄÓ²¼þƽ̨ÊÇӢΰ´ïµÄ±ßÔµ AI ¼ÆËãоƬ Jetson Xavier GPU¡£¸ù¾Ý¹Ù·½¹«²¼µÄÊý¾Ý£¬Jetson Xavier µÄ INT8 ËãÁ¦Îª 22+10TOPS£¬æçÁú 865 ÔòÊÇ 15TOPS¡£Ñ¹ËõºóµÄ GAN ÏÖÔÚ¿´À´ÒѾ¿ÉÒÔÅÜÔÚ»úÆ÷ÈË¡¢ÎÞÈË»úµÈСÐÍÉ豸ÉÏÁË£¬Î´À´Èû½øÊÖ»úÖ¸ÈÕ¿É´ý¡£
ÂÛÎÄÁ´½Ó£ºhttps://arxiv.org/pdf/2003.08936v1.pdf
Éú²ú¶Ô¿¹ÍøÂ磨GANS£©Éó¤ºÏ³ÉÊ®·Ö±ÆÕæµÄͼÏñ¡£GAN µÄÒ»ÖÖ±äÌ塪¡ªÌõ¼þʽÉú³É¶Ô¿¹ÍøÂ磨conditional generative adversarial network£¬cGAN£©Ôںܶà¼ÆËã»úÊÓ¾õºÍͼÏñѧӦÓÃÖж¼Äܹ»ÊµÏÖ¿É¿ØÖƵÄͼÏñºÏ³É¡£µ«ÕâЩӦÓôó¶¼ÐèҪģÐÍÓëÈË»¥¶¯£¬Òò´ËÐèÒªµÍÑÓ³ÙµÄÉ豸ÉϲÅÄÜ»ñµÃ¸üºÃµÄÓû§ÌåÑé¡£
È»¶ø£¬½üÀ´ÍƳöµÄһЩ cGAN ÔÚ¼ÆËãÇ¿¶ÈÉϱȵ±Ç°Ê¶±ð¾í»ýÉñ¾ÍøÂ磨CNN£©´ó 1 ÖÁ 2 ¸öÁ¿¼¶¡£¾ÙÀý¶øÑÔ£¬GanGAN ÿÕÅͼÏñÏûºÄ 281G Macs£¬¶ø MobileNet-v3 ½öÐè 0.44G Macs£¬Õâ¾Íµ¼ÖÂÇ°ÕߺÜÄÑÓÃÓÚ½»»¥Ê½²¿Êð¡£
¶øÇÒ£¬¾ÍÄ¿Ç°À´¿´£¬±ßÔµÉ豸ÓÖ¶àÊýÊܵ½ÄÚÈÝÒÔ¼°µç³ØÖ®ÀàÓ²¼þ²ãÃæÉϵÄÏÞÖÆ£¬Ò²×è°ÁË GAN ÔÚ±ßÔµÉ豸ÉϵIJ¿Êð¡£
Òò´Ë£¬»ùÓÚ GAN ºÍ cGAN ÔÚͼÏñºÏ³ÉÁìÓò´æÔÚµÄÕâЩÎÊÌ⣬º«ËÉÍŶÓÌá³öÁË GAN ѹËõ£¬ÕâÊÇÒ»ÖÖͨÓÃѹËõ·½·¨£¬ÓÃÓÚ¼õÉÙ GAN µÄÍÆÀíʱ¼äÒÔ¼°¼ÆËã³É±¾¡£Í¬Ê±£¬Ñ¹ËõÉú³ÉÄ£ÐÍÃæÁÙÁ½¸ö»ù±¾À§ÄÑ£ºGAN ѵÁ·²»Îȶ¨£¬ÓÈÆäÊÇÔÚδÅä¶ÔµÄÇé¿öÏ£»Éú³ÉÆ÷Óë CNN ²»Í¬£¬Òò´ËºÜÄÑʹÓÃÏÖÓÐµÄ CNN Éè¼Æ¡£ÎªÁ˽â¾ö´ËÎÊÌ⣬ÍŶӽ«ÖªÊ¶´ÓÔʼµÄ½ÌʦÉú³ÉÆ÷Öмä±íʾ²ã´«µÝµ½ÆäÏàÓ¦µÄѧÉúÉú³ÉÆ÷²ãÖС£
ΪÁ˽µµÍѵÁ·³É±¾£¬ÍŶӻ¹Í¨¹ýѵÁ·°üº¬ËùÓпÉÄÜͨµÀÊýµÄ¡¸once- for-all network¡¹£¬½«Ä£ÐÍѵÁ·Óë¼Ü¹¹ËÑË÷·ÖÀë¡£Õâ¸ö¡¸once-for-all network¡¹¿ÉÒÔͨ¹ýȨÖع²Ïí²úÉúÐí¶à×ÓÍøÂ磬ÎÞÐèѵÁ·¾Í¿ÉÒÔÆÀ¹Àÿ¸ö×ÓÍøÂçµÄÐÔÄÜ¡£¸ÃÄ£ÐÍ¿ÉÓ¦ÓÃÖÁ¸÷ÖÖÌõ¼þÏ嵀 GAN Ä£ÐÍ£¬²»¹ÜÆäÊôÓÚÄÄÖÖÄ£Ðͼܹ¹¡¢Ñ§Ï°Ëã·¨»ò¼à¶½ÉèÖã¨Åä¶Ô»òδÅä¶Ô£©¡£
ͨ¹ý´óÁ¿µÄʵÑ飬ÍŶÓÒÑÖ¤Ã÷ÁË´Ë·½·¨¿ÉÒÔ½« pix2pix£¬CycleGAN ÒÔ¼° GauGAN ÈýÖֹ㷺ʹÓÃµÄ GAN Ä£ÐͼÆËãÁ¿¼õÉÙÖÁ 1/9 µ½ 1/21£¬Í¬Ê±»¹²»»áËðʧÉú³ÉͼÏñµÄ±£Õæ¶È¡£
ÎÒÃǶ¼ÖªµÀ£¬¶ÔÓÃÓÚ½»»¥Ê½Ó¦ÓõÄÌõ¼þʽÉú³ÉÄ£ÐͽøÐÐѹËõ¾ßÓÐÌôÕ½ÐÔ£¬ÕâÖ÷ÒªÊÇÓÉÒÔÏÂÁ½·½ÃæÔÒòÔì³ÉµÄ¡£ÆäÒ»£¬´Ó±¾ÖÊÉϽ²£¬GAN µÄ¶¯Ì¬ÑµÁ··Ç³£²»Îȶ¨£»Æä¶þ£¬Ê¶±ðºÍÉú³ÉÄ£ÐÍÖ®¼ä´æÔڵľ޴ó¼Ü¹¹²îÒìµ¼ÖºÜÄÑÖ±½ÓʹÓÃÏÖÓÐµÄ CNN ѹËõËã·¨¡£
»ùÓÚÕâЩÔÒò£¬Ñо¿ÕßÌá³öÁËרÃÅÕë¶Ô¸ßЧÉú³ÉÄ£Ð͵ÄѵÁ··½°¸£¬²¢ÀûÓÃÉñ¾¼Ü¹¹ËÑË÷£¨NAS£©½øÒ»²½Ôö¼ÓѹËõ±È¡£GAN ѹËõ¿ò¼ÜÕûÌå¼Ü¹¹ÈçÏÂͼ 3 Ëùʾ£¬ÆäÖÐËûÃÇÀûÓà ResNet Éú³ÉÆ÷×÷ΪʾÀý¡£ÐèҪǿµ÷µÄÊÇ£¬Í¬Ò»¿ò¼Ü¿ÉÒÔÓ¦ÓÃÓÚ²»Í¬µÄÉú³ÉÆ÷¼Ü¹¹ºÍѧϰĿ±ê¡£
ͼ 3£ºÎÄÖÐ GAN ѹËõ¿ò¼ÜµÄÕûÌå¼Ü¹¹Í¼¡£
1. ͳһÅä¶ÔѧϰºÍ·ÇÅä¶Ôѧϰ
¹ã·ºµÄѵÁ·Ä¿±êʹµÃ¹¹½¨Í¨ÓÃѹËõ¿ò¼Ü·Ç³£À§ÄÑ¡£ÎªÁ˽â¾öÕâÒ»ÎÊÌ⣬Ñо¿ÕßÔÚÄ£ÐÍѹËõÉèÖÃÖÐͳһÁËÅä¶ÔºÍ·ÇÅä¶Ôѧϰ£¬²»¹Ü½ÌʦģÐÍ×î³õÊÇÈçºÎѵÁ·µÄ¡£¸ø¶¨Ôʼ½ÌʦÉú³ÉÆ÷ G¡ä£¬Ñо¿Õß½«·ÇÅä¶ÔѵÁ·ÉèÖÃת»»ÎªÅä¶ÔÉèÖ᣶ÔÓÚ·ÇÅä¶ÔÉèÖ㬿ÉÒÔ½«ÔʼÉú³ÉÆ÷Êä³öÊÓΪÕæÖµ²¢Ê¹ÓÃÅä¶ÔѵÁ·Ä¿±êѵÁ·Ñ¹ËõºóµÄÉú³ÉÆ÷ G¡£
ѧϰĿ±ê×ܽáÈçÏ£º
»ùÓÚÕâЩÐ޸ģ¬ÏÖÔÚ¿ÉÒÔ½«Í¬Ò»¸öѹËõ¿ò¼ÜÓ¦ÓÃÓÚ²»Í¬ÀàÐ굀 cGAN¡£¶øÇÒ£¬ÓëÔʼµÄ·ÇÅä¶ÔѵÁ·ÉèÖÃÏà±È£¬Ê¹ÓÃÉÏÊöα¶Ô£¨pseudo pair£©½øÐÐѧϰ¿ÉÒÔʹѵÁ·¸ü¼ÓÎȶ¨£¬²¢²úÉú¸üºÃµÄ½á¹û¡£
2. ´Ó½ÌʦÅбðÆ÷ÖÐѧϰ
¾¡¹Ü´ËÑо¿ÖÂÁ¦ÓÚѹËõÉú³ÉÆ÷£¬µ«ÅбðÆ÷ D Öд¢´æ×Å GAN µÄÓÐÓÃÐÅÏ¢¡£Òò´Ë£¬Ñо¿ÕßÀûÓÃÁËͬһ¸öÅбðÆ÷¼Ü¹¹£¬Ê¹ÓÃÀ´×Ô½ÌʦÅбðÆ÷µÄԤѵÁ·È¨ÖØ£¬ÓëѹËõÉú³ÉÆ÷Ò»Æð΢µ÷ÅбðÆ÷¡£
ÔÚʵÑéÖУ¬Ñо¿Õ߹۲쵽£¬Ô¤ÑµÁ·ÅбðÆ÷¿ÉÒÔÖ¸µ¼Ñ§ÉúÉú³ÉÆ÷µÄѵÁ·¡£Ê¹ÓÃÒ»¸öËæ»ú³õʼ»¯µÄÅбðÆ÷ͨ³£»áµ¼ÖÂѵÁ·²»Îȶ¨¼°Í¼ÏñÖÊÁ¿Ï½µ¡£Õâ¸ö GAN µÄÄ¿±ê¿ÉÒÔд³ÉÒÔÏÂÐÎʽ£º
ÔÚÉÏʽÖУ¬Ñо¿ÕßʹÓÃÀ´×Ô½ÌʦÅбðÆ÷ D¡äµÄȨÖØÀ´³õʼ»¯Ñ§ÉúÅбðÆ÷ D¡£ËûÃÇʹÓÃÒ»¸ö±ê×¼µÄ¼«Ð¡»¯¼«´óÓÅ»¯Æ÷ѵÁ· G ºÍ D¡£
3. ÖмäÌØÕ÷ÕôÁó
֪ʶÕôÁóÊÇ CNN Ä£ÐÍѹËõÓõ½µÄÒ»ÖÖÆձ鷽·¨¡£Í¨¹ýÆ¥ÅäÊä³ö²ã logit µÄ·Ö²¼£¬¿ÉÒÔ½«À´×Ô½ÌʦģÐ͵ݵ֪ʶ£¨dark knowledge£©Ç¨ÒƵ½Ñ§ÉúÄ£ÐÍÖУ¬Ìá¸ßѧÉúÄ£Ð͵ÄÐÔÄÜ¡£È»¶ø£¬Ìõ¼þ GAN ͨ³£»áÊä³öÒ»¸öÈ·¶¨µÄͼÏñ£¬¶ø²»ÊǸÅÂÊ·Ö²¼¡£
ΪÁ˽â¾öÉÏÊöÎÊÌ⣬Ñо¿ÕßÆ¥ÅäÁ˽ÌʦÉú³ÉÆ÷µÄÖмä±íʾ¡£Öмä²ã°üº¬µÄͨµÀÔ½¶à£¬ËüÃÇËùÄÜÌṩµÄÐÅÏ¢Ô½·á¸»£¬Ñ§ÉúÄ£ÐÍËù»ñÈ¡µÄÊä³öÖ®ÍâµÄÐÅÏ¢Ò²Ô½¶à¡£ÕôÁóÄ¿±êÈçÏ£º
ÆäÖУ¬G_t(x) ºÍ G¡ä_t(x) ÊÇѧÉúºÍ½ÌʦģÐÍÖÐµÚ t ¸ö±»Ñ¡²ãµÄÖмäÌØÕ÷¼¤»î£¬T ±íʾ²ãÊý¡£
4. ÍêÕûÓÅ»¯Ä¿±ê
×îÖÕÄ¿±ê¿ÉÒÔд³ÉÈçÏÂÐÎʽ£º
ÆäÖУ¬³¬²ÎÊý¦Ë_recon ºÍ ¦Ë_distill ¿ØÖÆÿ¸öÏîµÄÖØÒªÐÔ¡£
Ñ¡ÔñÒ»¸öÉè¼ÆÁ¼ºÃµÄѧÉú¼Ü¹¹¶Ô×îÖÕ֪ʶÕôÁóµÄЧ¹ûÊÇÖÁ¹ØÖØÒªµÄ£¬Ñо¿Õß·¢ÏÖ£¬¹âÊÇËõ¼õ½ÌʦģÐ͵ÄͨµÀÊýÁ¿²¢²»ÄÜʹѧÉúÄ£Ð͸ü½ô´Õ£ºµ±¼ÆËãËõ¼õÁ¿³¬¹ý 4 ±¶Ê±£¬ÐÔÄܾͻáÏÔÖøϽµ¡£
1. ¾í»ý·Ö½âºÍ²ãÁéÃô¶È
ÏÖÓеÄÉú³ÉÆ÷ͨ³£²ÉÓô«Í³¾í»ýÀ´Æ¥Åä CNN ·ÖÀàºÍ·Ö¶ÎÉè¼Æ¡£½üÀ´Ò»Ð©¸ßЧµÄ CNN Éè¼Æ¹ã·º²ÉÓþí»ýµÄ·Ö½âÐÎʽ£¨depthwise + pointwise£©£¬ÔÚÐÔÄܺͼÆËã¶þÕßÖ®¼äµÄ¾ùºâÐÔ¸üºÃ¡£Ñо¿Õß·¢ÏÖ£¬·Ö½âµÄ¾í»ýÒ²¿ÉÒÔÓÃÔÚ cGAN µÄÉú³ÉÆ÷Éè¼ÆÉÏ¡£
2. ʹÓà NAS ʵÏÖ×Ô¶¯²Ã¼ôͨµÀ
ÏÖÓеÄÉú³ÉÆ÷ÔÚËùÓвãÉÏʹÓÃÊÖ¶¯Éè¼Æ£¨²¢ÇÒ¼¸ºõͳһ£©µÄͨµÀÊý£¬Èç´ËÒ»À´¾Í»á²úÉúÈßÓ࣬Զ·Ç×îÓÅ·½·¨¡£ÎªÁ˽øÒ»²½ÌáÉýѹËõЧÂÊ£¬Ñо¿ÕßʹÓÃͨµÀ¼ôÖ¦£¨channel pruning£©À´×Ô¶¯Ñ¡ÔñÉú³ÉÆ÷ÖеÄͨµÀ¿í¶È£¬´Ó¶ø¼õÉÙÈßÓ࣬¶þ´Î¼õÉÙ¼ÆËãÁ¿¡£ÕâÒ»·½·¨Ö§³ÖÓйØͨµÀÊýµÄϸÁ£¶ÈÑ¡Ôñ£¬Õë¶Ôÿ¸ö¾í»ý²ã£¬¿ÉÒÔ´Ó 8 µÄ±¶ÊýÖÐÑ¡Ôñ¾í»ý²ã£¬´Ó¶øƽºâ MAC ºÍÓ²¼þ²¢ÐÐÐÔ¡£
½âñîѵÁ·Óë½á¹¹ËÑË÷
Ñо¿ÕßÒÀÕÕ×î½ü one-shot µÄ NAS ·½·¨µÄÑо¿£¬½«Ä£ÐÍѵÁ·Óë¼Ü¹¹ËÑË÷Íѹ³¡£Ê×ÏÈ£¬ÑµÁ·Ò»¸öÖ§³Ö²»Í¬Í¨µÀÊýÁ¿µÄ¡¸once-for-all¡¹ÍøÂ磬ÆäÖеÄÿ¸ö×ÓÍøÂ綼¾¹ýÁËͬµÈѵÁ·¡£Í¼ 3 ²ûÊÍÁËÕû¸ö¿ò¼Ü¡£Ñо¿Õß¼ÙÉèÔʼµÄ½ÌʦÉú³ÉÆ÷ÓÐ
¸öͨµÀ£¬¶ÔÓÚ¸ø¶¨µÄͨµÀÊý
£¬´Ó¡¸once-for-all¡¹µÄȨÖØÕÅÁ¿ÌáÈ¡µÚÒ»¸öͨµÀ£¬»ñµÃ×ÓÍøÂçµÄȨÖØÍøÂç
¡£
ÔÚÿһ¸öѵÁ·²½ÖèÖУ¬Ê¹ÓÃѧϰĿ±ê¶Ô¾ß±¸Ä³Ò»Í¨µÀÊýµÄ×ÓÍøÂç½øÐÐËæ»úÈ¡Ñù£¬¼ÆËãÊä³öºÍÌݶȣ¬¸üÐÂÌáÈ¡µÄȨÖØ£¨¹«Ê½ 4£©¡£ÓÉÓÚ×îÏȱ»³éÈ¡µÄ¼¸¸öͨµÀ¸üÐÂƵÂʸü¸ß£¬ËüÃÇÔÚËùÓÐȨÖØÖ®ÖаçÑÝןüΪ¹Ø¼üµÄ½ÇÉ«¡£
Õâ¸ö¡¸once-for-all¡¹ÍøÂçѵÁ·ºÃÖ®ºó£¬Ñо¿ÕßÖ±½ÓÔÚÑéÖ¤¼¯ÉÏÆÀ¹ÀÁËÿ¸ö×ÓÍøÂçµÄÐÔÄÜ£¬ÕÒ³öÁË×î¼ÑµÄ×ÓÍøÂç¡£¡¸once-for-all¡¹ÍøÂçÒѾ¾¹ýÁËȨֵ¹²ÏíµÄ³¹µ×ѵÁ·£¬ÎÞÐèÔÙ½øÐÐ΢µ÷£¨fine-tuning£©¡£Õâ¸ö½á¹û½üËÆÓÚ´ÓÍ·¿ªÊ¼ÑµÁ·Ä£Ð͵ÄÐÔÄÜ¡£
ͨ¹ýÕâÖÖ·½Ê½£¬Ö»ÐèҪѵÁ·Ò»´Î£¬ÇÒÎÞÐèÔÙ½øÒ»²½ÑµÁ·£¬¾ÍÄÜÆÀ¹ÀËùÓÐÆÀ¹ÀËùÓÐͨµÀµÄÅäÖ㬲¢ÒÀ¾ÝËÑË÷½á¹ûÕÒµ½ÆäÖÐ×î¼Ñ¡£µ±È»£¬Ò²¿ÉÒÔ¶ÔÑ¡Öеļܹ¹½øÐÐ΢µ÷À´½øÒ»²½ÌáÉýÆäÐÔÄÜ¡£
Ñо¿ÕßÔÚÒÔÏÂÈýÖÖÌõ¼þʽ GAN Ä£ÐÍÉϽøÐÐʵÑéÒÔÑéÖ¤ÎÄÖÐ GAN ѹËõ¿ò¼ÜµÄ·º»¯ÐÔ£¬ËüÃÇ·Ö±ðÊÇ CycleGAN¡¢Pix2Pix ºÍ GauGAN¡£ËùʹÓõÄËĸöÊý¾Ý¼¯Îª Horse?zebra¡¢Edges¡úshoes¡¢Cityscapes ºÍ Map?aerial photo¡£
ϱí 1 չʾÁËÔÚÉÏÊöËĸöÊý¾Ý¼¯ÉÏѹËõ CycleGAN¡¢Pix2Pix ºÍ GauGAN Ä£Ð͵ÄÁ¿»¯½á¹û¡£
±í 1£ºÈýÖÖÌõ¼þʽ GAN Ä£ÐÍѹËõµÄÁ¿»¯ÆÀ¹À£¬ÆäÖÐ Cityscapes Êý¾Ý¼¯ÉÏʹÓà mAP ¶ÈÁ¿£¨Ô½¸ßÔ½ºÃ£©£¬ÆäËûÊý¾Ý¼¯ÉÏʹÓà FID ¶ÈÁ¿¡£½á¹ûÏÔʾ£¬ÔÚÐÔÄÜÇá΢ϽµµÄÇé¿öÏ£¬ÎÄÖÐ GAN ѹËõ·½·¨¿ÉÒÔÔÚ MACs ÉϽ«µ±Ç° SOTA Ìõ¼þʽ GAN ѹËõ 7 ÖÁ 21 ±¶£¬ÔÚÄ£ÐÍ´óСÉÏѹËõ 5 ÖÁ 33 ±¶¡£¶ÔÓÚ CycleGAN Ä£ÐÍѹËõ£¬ÎÄÖÐ GAN ѹËõ·½·¨Ô¶ÓÅÓÚÒÔÍùµÄ CycleGAN-specific Co-evolution ·½·¨¡£
ÐÔÄÜÓë¼ÆËãÖ®¼äµÄȨºâ
¸Ã·½·¨³ýÁËÄܹ»ÊµÏֽϴóµÄѹËõÂÊÒÔÍ⣬ͬʱÆäÒ²¿ÉÒÔÌáÉý²»Í¬Ä£ÐÍ´óСµÄÐÔÄÜ¡£ÏÂͼ 6 ÏÔʾÁËÔÚ pix2pix Ä£ÐÍÖУ¬ÔÚ²»Í¬Êý¾Ý¼¯ÉϵÄÐÔÄÜÓë¼ÆËãµÄȨºâ¡£
ͼ 6£ºÔÚ Cityscapes Óë Edges¡úShoes Êý¾Ý¼¯ÉÏ pix2pix µÄȨºâÇúÏß¡£¼ôÖ¦ÓëÕôÁó·½·¨ÔÚ´óÐÍÄ£ÐÍÖг¬³öÁË´ÓÍ·¿ªÊ¼ÑµÁ·£¬È»¶øÔÚÄ£Ðͱ»¼±¾çѹËõʱ±íÏֽϲ
Ч¹ûչʾ
ÏÂͼ 4 ΪʹÓø÷½·¨µÃµ½µÄЧ¹ûչʾ¡£Í¼Öзֱð¸ø³öÁËÊäÈëÊý¾Ý¡¢±ê×¼Êä³ö¡¢ÔʼģÐÍÊä³öÒÔ¼°Ñ¹ËõºóÄ£Ð͵ÄÊä³ö¡£´ÓͼÖпÉÒÔ¿´³ö£¬¼´Ê¹ÔڽϴóµÄѹËõÂÊÏ£¬Ñо¿ÕßËùÌá·½·¨ÈÔÈ»Äܱ£³ÖÊä³öͼÏñµÄÊÓ¾õ¿ÉÐŶȡ£
ͼ 4£ºCityscapes¡¢Edges¡úShoes ÒÔ¼° Horse¡úZebra Êý¾Ý¼¯ÉÏЧ¹û±È½Ï¡£
Ó²¼þÍÆÀí¼ÓËÙ
¶ÔÓÚÕæʵ³¡¾°½»»¥Ó¦ÓöøÑÔ£¬ÔÚÓ²¼þÉ豸ÉÏÍÆÀí¼ÓËÙµÄÖØÒªÐÔÒªÔ¶´óÓÚ¼õÉÙ¼ÆËãÏûºÄ¡£Èçϱí 2 Ëùʾ£¬ÎªÑéÖ¤ËùÌá·½·¨ÔÚʵ¼ÊÓ¦ÓÃÖеÄÓÐЧÐÔ£¬Ñо¿ÕßÔÚ¾ßÓв»Í¬ÔËËãÐÔÄܵÄÉ豸ÉÏÃæ²âÊÔÁËѹËõÄ£Ð͵ÄÍÆÀíËٶȡ£
±í 2£ºÔÚ NVIDIA Jetson AGX Xavier¡¢NVIDIA Jetson Nano¡¢1080Ti GPU ºÍ Xeon CPU Éϲ⵽µÄÄÚ´æºÍÑÓ³ÙϽµ¡£
½áÂÛ
ÔÚ±¾ÎÄÖУ¬º«ËÉÍŶÓÌá³öµÄͨÓÃѹËõ¿ò¼ÜÄܹ»ÏÔÖø½µµÍÌõ¼þʽ GAN ÖÐÉú³ÉÆ÷µÄ¼ÆË㻨ÏúºÍÄ£ÐÍ´óС£¬²¢ÇÒͨ¹ý֪ʶÕôÁóºÍ×ÔÈ»¼Ü¹¹ËÑË÷À´ÌáÉýѵÁ·Îȶ¨ÐÔÒÔ¼°Ä£ÐÍЧÂÊ¡£ÊµÑé±íÃ÷£¬ÎÄÖÐÌá³öµÄ GAN ѹËõ·½·¨Äܹ»ÔÚ±£³ÖÊÓ¾õÖÊÁ¿µÄͬʱѹËõÊýÖÖÌõ¼þʽ GAN Ä£ÐÍ¡£ËûÃDZíʾ£¬Î´À´µÄÑо¿¹¤×÷½«²àÖØÓÚ½µµÍÄ£ÐÍÑÓ³ÙÒÔ¼°¹¹½¨Éú³ÉÊÓƵģÐ͵ĸßЧ¿ò¼Ü¡£
ÁìȡרÊô 10ÔªÎÞÃż÷ȯ
˽Ïí×îР¼¼Êõ¸É»õ