¾àÀëÈ«ÇòÅÌËã»úÊÓ¾õ¶¥¼¶¾Û»áIEEE CVPR 2018 (IEEE¹ú¼ÊÅÌËã»úÊÓ¾õÓëģʽʶ±ð¾Û»á) ÕÙ¿ªµÄ½Å²½Ô½À´Ô½½ü¡£¡£¡£¡£¡£¡£¡£ÉÏÖÜ£¬£¬£¬£¬£¬ÉÌÌÀ¿Æ¼¼Ðû²¼ÒÔ44ƪÂÛÎÄÈëÑ¡CVPR 2018£¬£¬£¬£¬£¬ÎüÒýÁËÖî¶àÅóÙ­µãÔÞ¡£¡£¡£¡£¡£¡£¡£½ÓÏÂÀ´¼¸ÖÜ£¬£¬£¬£¬£¬ÎÒÃǽ«Â½ÐøÐû²¼ÉÌÌÀ¿Æ¼¼ÈëÑ¡CVPR 2018µÄÖصãÂÛÎĽâ¶Á±¨µÀ£¬£¬£¬£¬£¬½Ó´ý¸÷È˹Ø×¢¡£¡£¡£¡£¡£¡£¡£

ÒÔÏÂÊÇÔÚ´ó¹æÄ£ÂþÑÜʽѵÁ·ÁìÓò£¬£¬£¬£¬£¬ÉÌÌÀ¿Æ¼¼½ÒÏþµÄһƪÖØ°õ´ó»á¿ÚÍ·±¨¸æ(Oral)ÂÛÎÄ£¬£¬£¬£¬£¬Ìá³ö»ùÓÚÂþÑÜʽѵÁ·µÄÉî¶ÈÔöǿѧϰBlockQNNËã·¨£¬£¬£¬£¬£¬×Ô¶¯Éè¼ÆÉñ¾­ÍøÂç½á¹¹¡£¡£¡£¡£¡£¡£¡£±¾ÎÄÓÉÉÌÌÀ¿Æ¼¼ÊµÏ°Ñо¿Ô±ÖÓîÈÔÚÉÌÌÀ¿Æ¼¼Ñо¿Ôº¸±Ôº³¤ãÆ¿¡½ÜºÍÑо¿Ë¾ÀíÎäΰָµ¼ÏÂÍê³É¡£¡£¡£¡£¡£¡£¡£

ÂÛÎÄ£ºPractical Block-wise Neural Network Architecture Generation

×÷ÕߣºZhao Zhong, Junjie Yan, Wei Wu, Jing Shao, Cheng-Lin Liu

ÂÛÎÄÈ«ÎÄÁ´½Ó£º

https://arxiv.org/abs/1708.05552

¼ò½é

Éñ¾­ÍøÂç½á¹¹Éè¼ÆÒ»Ö±ÊÇÉî¶ÈѧϰÀïµÄ½¹µãÎÊÌâ¡£¡£¡£¡£¡£¡£¡£ÔÚ»ùÓÚÉî¶ÈѧϰµÄ·ÖÀà¡¢¼ì²â¡¢Ö§½â¡¢¸ú×ÙµÈʹÃüÖÐ£¬£¬£¬£¬£¬»ù´¡Éñ¾­ÍøÂçµÄ½á¹¹¶ÔÕûÌåËã·¨µÄÐÔÄÜÓÅÁÓÓÐמöÒéÐÔµÄÓ°Ïì¡£¡£¡£¡£¡£¡£¡£¹Å°åµÄÉñ¾­ÍøÂç½á¹¹Éè¼ÆÐèÒª´ó×Úר¼ÒµÄ֪ʶºÍÊÔ´í±¾Ç®£¬£¬£¬£¬£¬ÉõÖÁ»¹ÐèҪһЩÁé¸ÐºÍ¡°ÉñÀ´Ö®±Ê¡±£¬£¬£¬£¬£¬Ã¿Äê½öÓм¸¸öÓнϴóÓ°ÏìµÄÐÂÍøÂç½á¹¹±»Éè¼Æ³öÀ´£¬£¬£¬£¬£¬Òò´Ë£¬£¬£¬£¬£¬È˹¤Éè¼ÆÍøÂç½á¹¹ÊÇÒ»¼þ¼«ÆäÄÑÌâµÄÊÂÇé¡£¡£¡£¡£¡£¡£¡£ÍøÂç½á¹¹µÄ×Ô¶¯Éè¼Æ/ËÑË÷µÄÒªÁì×î½üÊܵ½ÁËÆÕ±éµÄÖØÊÓ£¬£¬£¬£¬£¬ÆäÊÔͼ³¯×Å×Ô¶¯»¯Éî¶ÈѧϰÕâ¸öÄ¿µÄ£¬£¬£¬£¬£¬°ÑÈË´ÓÖØ´ó·±ËöµÄÍøÂçÉè¼Æµ÷²ÎÖнâ·Å³öÀ´(ÈçͼһËùʾ)¡£¡£¡£¡£¡£¡£¡£

ͼһ£º(ÓÉ×óÖÁÓÒ)

´ÓÈ˹¤Éè¼ÆµÄÍøÂç½á¹¹µ½

Ëã·¨×Ô¶¯Éè¼ÆµÄÍøÂç½á¹¹

½üÆÚµÄÍøÂç½á¹¹×Ô¶¯Éè¼Æ/ËÑË÷Ë㷨ͨ³£ÐèÒªãýÃðÖØ´óµÄÅÌËã×ÊÔ´(ÀýÈç£¬£¬£¬£¬£¬GoogleµÄNASËã·¨ÐèҪʹÓÃÊý°Ù¿éGPUÒÔ¼°½üÒ»¸öÔµÄѵÁ·Ê±¼ä)£¬£¬£¬£¬£¬²¢ÇÒÌìÉúµÄÄ£×Ó¿ÉǨáãÐÔ²»Ç¿£¬£¬£¬£¬£¬ÄÑÒÔ×öµ½ÕæÕýµÄÊÊÓû¯¡£¡£¡£¡£¡£¡£¡£±¾ÎÄÌá³öµÄBlockQNNËã·¨Äܹ»½â¾öÏÖÓÐÍøÂç½á¹¹×Ô¶¯Éè¼Æ/ËÑË÷ÒªÁìЧÂʺͷº»¯ÐÔµÄÎÊÌâ¡£¡£¡£¡£¡£¡£¡£

»ùÓÚ¡°Block¡±µÄÍøÂç½á¹¹ÌåÏÖ

ÏÖÔÚÖ÷Á÷µÄÉî¶ÈÉñ¾­ÍøÂç½á¹¹¼«ÆäÖØ´ó£¬£¬£¬£¬£¬ÆäÉî¶ÈÍùÍùµÖ´ïÊý°Ù²ã£¬£¬£¬£¬£¬Ö±½ÓËÑË÷Õû¸öÍøÂç½á¹¹µÄËÑË÷¿Õ¼äºÜÊÇÖØ´ó£¬£¬£¬£¬£¬ÕâÒ²ÊÇ֮ǰ×Ô¶¯ÍøÂç½á¹¹Éè¼Æ/ËÑË÷Ëã·¨½ÏÁ¿µÍЧµÄÔµ¹ÊÔ­ÓÉÖ®Ò»¡£¡£¡£¡£¡£¡£¡£Í¬Ê±£¬£¬£¬£¬£¬Ö±½ÓÉè¼ÆÕû¸öÍøÂçµÄÕ½ÂԻᵼÖÂÍøÂç½á¹¹²»¾ß±¸·º»¯ÐÔ¡£¡£¡£¡£¡£¡£¡£ÀýÈç£¬£¬£¬£¬£¬»ùÓÚCIFARÊý¾Ý¼¯ËÑË÷³öÀ´µÄÍøÂç½á¹¹£¬£¬£¬£¬£¬ÓÉÓÚÆäÊý¾ÝÊäÈëµÄ¾ÞϸÏÞÖÆ£¬£¬£¬£¬£¬Ö»Êʺϴ¦Öóͷ£ÊäÈëΪ32x32Çø·ÖÂʵÄÊý¾Ý£¬£¬£¬£¬£¬´Ó¶øʹµÃËÑË÷³öµÄÍøÂç½á¹¹·º»¯ÐÔÄܽÏÈõ¡£¡£¡£¡£¡£¡£¡£

Õë¶ÔÕâЩÎÊÌâ£¬£¬£¬£¬£¬±¾ÎĽè¼øÁËÏÖ´úÖ÷Á÷Éî¶ÈÉñ¾­ÍøÂçµÄÉè¼ÆÍ·ÄÔ£¬£¬£¬£¬£¬ºÃ±ÈResNet¡¢InceptionµÈÍøÂç¡£¡£¡£¡£¡£¡£¡£ÕâЩÍøÂçÊÇÓÉͬÑù½á¹¹µÄ×ÓÍøÂçÖظ´×éºÏÔÚÒ»ÆðÐγÉ£¬£¬£¬£¬£¬±¾ÎÄ°ÑÕâÖÖÄÜÖظ´×éºÏµÄ×ӽṹ³ÆΪblock¡£¡£¡£¡£¡£¡£¡£Í¨¹ýÉè¼Æblock½á¹¹£¬£¬£¬£¬£¬¿ÉÒÔÈÃÍøÂç½á¹¹µÄËÑË÷¿Õ¼ä´ó´ó¼õС£¬£¬£¬£¬£¬²¢ÇÒblock½á¹¹×Ô¼º¾ßÓÐÇ¿Ê¢µÄ·º»¯ÐÔ£¬£¬£¬£¬£¬Õë¶Ô²î±ðµÄÊý¾Ý¼¯»òÕßʹÃü£¬£¬£¬£¬£¬Ö»ÐèÒªµþ¼Ó²î±ð¸öÊýµÄblock¼´¿ÉÍê³É(Èçͼ¶þËùʾ)¡£¡£¡£¡£¡£¡£¡£

ͼ¶þ£º

»ùÓÚBlockÉè¼ÆµÄÍøÂçÕûÌå¿ò¼Ü

×ó±ßΪÕë¶ÔCIFARµÄ¿ò¼Ü£¬£¬£¬£¬£¬ÓÒ±ßÕë¶ÔImageNet

±íÒ»£º

ÍøÂç½á¹¹±àÂë±í

ΪÁËÌåÏÖÍøÂçblock½á¹¹£¬£¬£¬£¬£¬±¾ÎÄÉè¼ÆÁËÒ»Ì×ÍøÂç½á¹¹±àÂë£¬£¬£¬£¬£¬°ÑÉñ¾­ÍøÂç¿´×öÒ»¸öÓÐÏòÎÞ»·Í¼£¬£¬£¬£¬£¬Ã¿¸ö½ÚµãÌåÏÖÍøÂçÖеÄÿһ²ã£¬£¬£¬£¬£¬¶ø±ß¾ÍÌåÏÖÊý¾ÝÁ÷¶¯µÄÆ«Ïò¡£¡£¡£¡£¡£¡£¡£Õû¸ö±àÂë°üÀ¨Éñ¾­ÍøÂçµÄ²ãÊýÐòºÅ£¬£¬£¬£¬£¬ÀàÐÍ£¬£¬£¬£¬£¬ºËµÄ¾Þϸ£¬£¬£¬£¬£¬ÒÔ¼°Á½¸öÇ°Ðò½ÚµãµÄÐòºÅ¡£¡£¡£¡£¡£¡£¡£Ê¹ÓÃÕâÖÖ±àÂë·½·¨¾Í¿ÉÒÔÌåÏÖí§ÒâµÄÉñ¾­ÍøÂç½á¹¹£¬£¬£¬£¬£¬ÀýÈçResNetºÍInceptionµÄblock½á¹¹¾ÍÄÜʹÓÃͼÈýÖеıàÂë¾ÙÐÐÌåÏÖ¡£¡£¡£¡£¡£¡£¡£

ͼÈý£º

InceptionºÍResnetÖÐblock½á¹¹±àÂëÌåÏÖ

»ùÓÚÇ¿»¯Ñ§Ï°µÄÍøÂç½á¹¹×Ô¶¯Éè¼Æ

½ÓÏÂÀ´µÄ½¹µãÎÊÌâ¼´ÊÇÔõÑù»ñµÃ×îÓŵÄÍøÂç½á¹¹¡£¡£¡£¡£¡£¡£¡£Ö»¹ÜÍøÂç½á¹¹µÄËÑË÷¿Õ¼äÒѾ­Í¨¹ýÉè¼Æblock´ó´ó¼õС£¬£¬£¬£¬£¬¿ÉÊÇÖ±½Ó±©Á¦ËÑË÷ËùÓпÉÄܽṹ£¬£¬£¬£¬£¬ÒÀȻʮ·ÖãýÃðÅÌËã×ÊÔ´¡£¡£¡£¡£¡£¡£¡£±¾ÎÄÒò´ËÌá³öÒ»ÖÖ»ùÓÚÇ¿»¯Ñ§Ï°µÄÍøÂçÉè¼ÆÒªÁì£¬£¬£¬£¬£¬×Ô¶¯Ñ§Ï°»ñµÃÍøÂç½á¹¹¡£¡£¡£¡£¡£¡£¡£

ÔÚÍøÂçÉè¼ÆÇ¿»¯Ñ§Ï°ÖÐ£¬£¬£¬£¬£¬±¾ÎÄ°ÑÄ¿½ñÉñ¾­ÍøÂç²ã½ç˵ΪÔöǿѧϰÖеÄÏÖÔÚ״̬(current state)£¬£¬£¬£¬£¬¶øÏÂÒ»²ã½á¹¹µÄ¾öÒé½ç˵ΪÔöǿѧϰÖеÄÐж¯(action)¡£¡£¡£¡£¡£¡£¡£ÕâÀïʹÓÃ֮ǰ½ç˵µÄÉñ¾­ÍøÂç½á¹¹±àÂëÀ´ÌåÏÖÿһ²ãÍøÂç¡£¡£¡£¡£¡£¡£¡£ÕâÑù£¬£¬£¬£¬£¬Í¨¹ýһϵÁеÄÐж¯¾öÒé£¬£¬£¬£¬£¬¾ÍÄÜ»ñµÃÒ»ÌõÌåÏÖblock½á¹¹µÄ±àÂë(ÈçͼËÄËùʾ)£¬£¬£¬£¬£¬¶øÌá³öµÄÇ¿»¯Ñ§Ï°Ë㷨ͨ¹ýÓÅ»¯Ñ°»ñ×îÓŵÄÐж¯¾öÒéÐòÁС£¡£¡£¡£¡£¡£¡£±¾ÎÄʹÓÃQ-learningËã·¨À´¾ÙÐÐѧϰ£¬£¬£¬£¬£¬ÏêϸµÄ¹«Ê½²»ÔÙÕö¿ª¡£¡£¡£¡£¡£¡£¡£

ͼËÄ£º

»ùÓÚÇ¿»¯Ñ§Ï°µÄÍøÂç½á¹¹×Ô¶¯Éè¼ÆÁ÷³Ì

ÖµµÃ×¢ÖصÄÒ»µãÊÇ£¬£¬£¬£¬£¬ÓëÒ»Ñùƽ³£µÄÇ¿»¯Ñ§Ï°ÎÊÌâ²î±ð£¬£¬£¬£¬£¬¸ÃʹÃüÖ»ÔÚ¿¢ÊÂÕû¸öÐòÁеľöÒéºó(¼´ÌìÉúÍêÕûÍøÂç½á¹¹ºó)²Å»á»ñµÃÒ»¸öreward£¬£¬£¬£¬£¬¶ø֮ǰµÄÿ¸ö¾öÒéÊǶÔÓ¦reward¡£¡£¡£¡£¡£¡£¡£ÓÉÓÚ»ñµÃ×îÖÕrewardµÄ±¾Ç®ºÜÊǸß(ÐèÒªÔÚÊý¾ÝÉÏÖØÐÂѵÁ·Ð»ñµÃµÄÍøÂç½á¹¹)£¬£¬£¬£¬£¬ÎªÁ˼ÓËÙËüµÄÊÕÁ²£¬£¬£¬£¬£¬×÷ÕßʹÓÃÁËreward shapingµÄ¼¼ÇÉ(ÈçͼÎåËùʾ)£¬£¬£¬£¬£¬Òò¶øѵÁ·³õʼ½×¶ÎÖÕÖ¹²ãµÄQÖµ²»»á¹ý¸ß£¬£¬£¬£¬£¬ÈÃËã·¨²»»áÔÚѵÁ·³õʼ½×¶ÎÇãÏòÓÚÌìÉú²ãÊý¹ýdzµÄÍøÂç½á¹¹¡£¡£¡£¡£¡£¡£¡£

ͼÎ壺

Reward shaping¶ÔÊÕÁ²Ð§ÂʵÄÓ°Ïì

ÌáÇ°×èÖ¹Õ½ÂÔ

ËäÈ»Äܹ»Ê¹ÓöàÖÖ¼¼ÇÉÀ´Ê¹×Ô¶¯»¯ÍøÂç½á¹¹Éè¼Æ±äµÄÔ½·¢¸ßЧ¡£¡£¡£¡£¡£¡£¡£¿ÉÊÇ×Ô¶¯ÍøÂçÉè¼ÆÖÐãýÃðʱ¼äµÄÒªº¦ÕÕ¾ÉÔÚÓÚÿ´Î»ñµÃrewardµÄʱ¼ä±¾Ç®ºÜÊǸß£¬£¬£¬£¬£¬ÐèÒª½«ÌìÉúµÄÍøÂç½á¹¹ÔÚ¶ÔÓ¦µÄÊý¾Ý¼¯ÉÏѵÁ·ÖÁÊÕÁ²£¬£¬£¬£¬£¬È»ºó»ñµÃÏìÓ¦µÄ׼ȷ¶ÈÀ´ÌåÏֽṹµÄÓÅÁÓ²¢ÇÒÓÃ×÷reward¡£¡£¡£¡£¡£¡£¡£±¾ÎÄ×÷Õß·¢Ã÷£¬£¬£¬£¬£¬Í¨¹ýµ÷½âѧϰÂÊ£¬£¬£¬£¬£¬Ö»ÐèÒªÕý³£ÑµÁ·30·ÖÖ®Ò»µÄÀú³Ì(ÀýÈç£¬£¬£¬£¬£¬CIFAR-100Êý¾Ý¼¯ÉÏѵÁ·12¸öepoch)£¬£¬£¬£¬£¬¾Í¿ÉÒÔ»ñµÃÍøÂçµÄ´óÖÂ×îÖÕ¾«¶È£¬£¬£¬£¬£¬ÕâÑù¿ÉÒÔ´ó´ó½µµÍʱ¼ä±¾Ç®¡£¡£¡£¡£¡£¡£¡£¿ÉÊÇ£¬£¬£¬£¬£¬ÕâÑùµÄÍøÂç½á¹¹¾«¶È¼°Æä¹ØÁªµÄreward»áÓÐÎó²î£¬£¬£¬£¬£¬µ¼ÖÂÎÞ·¨Ï¸ÄåÇø·ÖÍøÂç½á¹¹µÄÓÅÁÓ£¬£¬£¬£¬£¬±¾ÎÄÌá³öÒ»¸öƾÂÄÀúµÄ½â¾ö¹«Ê½£º

¼´ÕæʵµÄrewardºÍÌáÇ°×èÖ¹µÄ׼ȷ¶È³ÉÕý±È£¬£¬£¬£¬£¬¿ÉÊǺÍÍøÂç½á¹¹µÄÅÌËãÖØƯºóºÍ½á¹¹ÅþÁ¬ÖØƯºó(blockÖбßÊý³ýÒÔµãÊý)³É·´±È¡£¡£¡£¡£¡£¡£¡£Í¨¹ýÕâÑùµÄ¹«Ê½½ÃÕý£¬£¬£¬£¬£¬»ñµÃµÄreward¶ÔÍøÂç½á¹¹µÄÓÅÁÓÔ½·¢¾ß±¸¿ÉÅбðÐÔ(ÈçͼÁùËùʾ)¡£¡£¡£¡£¡£¡£¡£

ͼÁù£º

ÌáÇ°×èÖ¹µÄ׼ȷ¶ÈÓëÍøÂçÕæʵ׼ȷ¶È¡¢

ÅÌËãÖØƯºó¡¢ÅþÁ¬ÖØƯºóµÄ¹Øϵ

ʵÑéЧ¹û

±¾ÎÄʹÓÃÁË32¸öGPU£¬£¬£¬£¬£¬¾­ÓÉ3ÌìµÄËÑË÷£¬£¬£¬£¬£¬¿ÉÒÔÔÚCIFARÊý¾Ý¼¯ÉÏÕÒµ½ÐÔÄִܵïÏÖÔÚÏȽøˮƽµÄÍøÂç½á¹¹(ÈçͼÆßËùʾ)¡£¡£¡£¡£¡£¡£¡£

ͼÆߣº

ËÑË÷Àú³ÌÒÔ¼°Ñ§Ï°»ñµÃµÄ×îÓÅÍøÂç½á¹¹

Ïà±È֮ǰµÄ×Ô¶¯ÍøÂçËÑË÷ÒªÁì(ÈçGoogle NASËã·¨µÄÊý°Ù¿éGPUÒÔ¼°Ò»¸öÔÂʱ¼ä)£¬£¬£¬£¬£¬BlockQNNËã·¨¿Éνʮ·Ö¸ßЧ(Èç±í¶þ¡¢±íÈýËùʾ)¡£¡£¡£¡£¡£¡£¡£

±í¶þ£º

²î±ð×Ô¶¯ÍøÂç½á¹¹ÉèÅÌËã·¨µÄ

ÅÌËã×ÊÔ´±ÈÕÕ

±íÈý£º

ÔÚCIFARÊý¾Ý¼¯Éϲî±ðÍøÂç½á¹¹µÄ¹ýʧÂÊ

ÁíÒ»·½Ãæ£¬£¬£¬£¬£¬Ñ§Ï°»ñµÃµÄÍøÂç½á¹¹Ò²¿ÉÒÔ¸üÈÝÒ×µÄǨáãµ½ImageNetʹÃüÉÏ£¬£¬£¬£¬£¬È¡µÃÁ˲»´íµÄ¾«¶È(Èç±íËÄËùʾ)¡£¡£¡£¡£¡£¡£¡£

±íËÄ£º

ÔÚImagenetÊý¾Ý¼¯Éϲî±ðÍøÂç½á¹¹µÄ¹ýʧÂÊ

ΪÁËÑéÖ¤BlockQNNËã·¨ÊÇ·ñÕæµÄѧϰµ½ÁËÉè¼ÆÍøÂçµÄ·½·¨£¬£¬£¬£¬£¬ÔÚͳһËÑË÷¿Õ¼äϺͱ©Á¦ËÑË÷¾ÙÐбÈÕÕ£¬£¬£¬£¬£¬¿ÉÒÔ·¢Ã÷±¾ÎÄÌá³öµÄBlockQNNËã·¨ÄÜ»ñµÃ±È±©Á¦ËÑË÷¸üºÃµÄÄ£×Ó(Èçͼ°ËËùʾ)¡£¡£¡£¡£¡£¡£¡£

ͼ°Ë£º

Ç¿»¯Ñ§Ï°Ó뱩Á¦ËÑË÷»ñµÃÍøÂç½á¹¹

ÔÚCIFARÊý¾Ý¼¯µÄ׼ȷÂÊ

¶ÔÕû¸öËÑË÷Àú³ÌºÍЧ¹ûÍøÂç½á¹¹¾ÙÐÐÆÊÎö(Èçͼ¾ÅËùʾ)£¬£¬£¬£¬£¬±¾ÎÄ×÷Õß·¢Ã÷ѧϰ»ñµÃµÄÓÅÒì½á¹¹ÓµÓÐһЩ¹²ÐÔ¡£¡£¡£¡£¡£¡£¡£ºÃ±Èmulti-branch½á¹¹¡¢short-cutÅþÁ¬·½·¨µÈÕâЩÏÖÔÚ³£ÓõÄÉè¼ÆÍ·ÄÔ¡£¡£¡£¡£¡£¡£¡£Í¬Ê±£¬£¬£¬£¬£¬×÷ÕßÒ²·¢Ã÷ÁËһЩ²»Ì«³£¼ûµÄ½á¹¹¹²ÐÔ£¬£¬£¬£¬£¬ºÃ±È¾í»ý²ãÖ®¼äµÄaddition²Ù×÷·ºÆðµÄÊ®·ÖƵÈÔ£¬£¬£¬£¬£¬ÕâЩѧϰ»ñµÃµÄÍøÂç½á¹¹ÉÐÓдý½øÒ»²½µÄÆÊÎöºÍÑо¿¡£¡£¡£¡£¡£¡£¡£

ͼ¾Å£º

BlockQNNËã·¨Éè¼ÆµÄÍøÂç½á¹¹Ê¾Àý

½áÂÛ

±¾ÎÄÌá³öÁËÒ»ÖÖ¸ßЧµÄ»ùÓÚÇ¿»¯Ñ§Ï°µÄÍø·½á¹¹×Ô¶¯ÉèÅÌËã·¨¡ª¡ªBlockQNN£¬£¬£¬£¬£¬Æäͨ¹ý¡°Block¡±µÄÉè¼ÆÍ·ÄÔ£¬£¬£¬£¬£¬ÈÃÍøÂç½á¹¹ËÑË÷¿Õ¼ä´ó´ó½µµÍ£¬£¬£¬£¬£¬²¢ÇÒʹµÃѧϰ»ñµÃµÄÍøÂç¾ß±¸ºÜÊÇÇ¿µÄ¿ÉǨáãÐÔ¡£¡£¡£¡£¡£¡£¡£Í¬Ê±£¬£¬£¬£¬£¬±¾ÎÄʹÓá°ÌáÇ°×èÖ¹¡±ºÍÂþÑÜʽ¼Ü¹¹À´¼ÓËÙÕû¸öѧϰÀú³Ì£¬£¬£¬£¬£¬µÖ´ïÁË°Ù±¶ÓÚ֮ǰGoogle NASËã·¨µÄËÙÂÊ(½öÐèÒª32¿éGPUÒÔ¼°3ÌìµÄѵÁ·)¡£¡£¡£¡£¡£¡£¡£ÊµÑéÅú×¢£¬£¬£¬£¬£¬Æäѧϰ»ñµÃµÄÍø·½á¹¹ÔÚCIFARÊý¾Ý¼¯ÉϵִïÈËÀàÉè¼ÆÍøÂç½á¹¹µÄ¾«¶È£¬£¬£¬£¬£¬²¢ÇÒÆä½á¹¹¿ÉÒÔǨáãµ½´ó¹æÄ£µÄImageNetÊý¾Ý¼¯ÉÏ£¬£¬£¬£¬£¬Ò²ÄÜÈ¡µÃ²»´íµÄÐÔÄÜ¡£¡£¡£¡£¡£¡£¡£×÷ÕßÏ£Íû³¯×Å×Ô¶¯»¯Éî¶ÈѧϰÕâ¸öÄ¿µÄ¼ÌÐøÉú³¤Ëã·¨£¬£¬£¬£¬£¬×îÖÕ°ÑÈË´ÓÖØ´ó·±ËöµÄÍøÂçÉè¼ÆºÍµ÷²ÎÖнâ·Å³öÀ´£¬£¬£¬£¬£¬ÊµÏÖÉî¶Èѧϰ¸ü½øÒ»²½µÄ¡°ÖÇÄÜ»¯¡±¡£¡£¡£¡£¡£¡£¡£