µ¼¶Á

AutoML£¨×ÔÄîͷеѧϰ£©ÒÔ¼°NAS£¨Éñ¾­¼Ü¹¹ËÑË÷£©ÊÇÉî¶ÈѧϰÁìÓòÏÖÔÚ×î¾ßDZÁ¦ºÍÓ°ÏìÁ¦µÄÆ«ÏòÖ®Ò»£¬£¬£¬£¬£¬ÆäÄ¿µÄÊÇÔ½·¢×Ô¶¯»¯µØʵÏÖÉñ¾­ÍøÂçµÄ½á¹¹Éè¼ÆÒÔ¼°³¬²ÎÓÅ»¯µÈ·½Ãæ¡£¡£¡£¡£¡£

2020Äê5ÔÂ7ÈÕ£¬£¬£¬£¬£¬ÉÌÌÀ̩̹¹ûÕæ¿ÎµÚ010ÆÚ¡¢CVPR2020ÂÛÎĽâ¶ÁµÚ¶þÆÚµÄÏßÉϿγ̣¬£¬£¬£¬£¬ÓÉÉÌÌÀËÄλÑо¿Ô±·ÖÏíÁËÍŶÓÔÚAutoML·½Ãæ×îеÄÑо¿ÊÂÇ飬£¬£¬£¬£¬Ð¡ÖúÊÖÒÑΪ¸÷ÈË×öÁËͼÎÄÕûÀí£¬£¬£¬£¬£¬¿ìÀ´Ò»¶ÃΪ¿ì°É~

Part 1  AutoMLÁìÓò×ÛÊö

Part 2  ÉÌÌÀAutoMLÍŶÓÏà¹ØÑо¿Ï£Íû

 2.1 Augmentation

 2.2 NAS

 2.3 LossFunction

Part 3  Q&AÎÊ´ð

Part 4  ´ðÒɽ»Á÷

¿Î³ÌPPT¼°ÊÓƵ»Ø·Å»ñÈ¡·½·¨¼ûÎÄÄ©ÌáÐÑ~

Part 1 AutoML×ÛÊö

Éî¶ÈѧϰÊÇÒ»¸öÔÚ2013Äê±»Òý±¬µÄ¿´·¨£¬£¬£¬£¬£¬×Ô´ÓAlexNetÍƳöÒÔÀ´£¬£¬£¬£¬£¬ÔÚÐí¶àCVʹÃüµÄbenchmarkÉϻῴµ½Éî¶ÈѧϰΪ½¹µãµÄËã·¨µÄÉíÓ°£¬£¬£¬£¬£¬ÏÖÔÚ¿ÉÒÔ³Æ×÷ÊÇ¡°Éî¶ÈѧϰµÄʱ´ú¡±£¬£¬£¬£¬£¬ÓÉÓÚһЩ¹Å°åµÄËã·¨ÒѾ­µ­³öÁËÊÓÒ°£¬£¬£¬£¬£¬³£ËµµÄ¡°Éî¶ÈѧϰÈý¾ÞÍ·¡±ÔÚ2018ÄêÒÀ¸½ÔÚÉî¶ÈѧϰƫÏòÉϵÄТ˳»ñµÃÁËͼÁé½±£¬£¬£¬£¬£¬×ãÒÔ¼ûµÃÉî¶ÈѧϰÊÇÏÖÔڷdz£»£»£»£»£»£»ðÈȵÄÆ«Ïò¡£¡£¡£¡£¡£

Éî¶ÈѧϰÊÇÒ»¸öÔÚ2013Äê±»Òý±¬µÄ¿´·¨£¬£¬£¬£¬£¬²¢ÔÚÕû¸ö»úеѧϰÁìÓòÕ¼ÓÐÁËÖ÷µ¼Ö°Î»¡£¡£¡£¡£¡£Éî¶ÈѧϰºÍ·ÇÉî¶ÈѧϰµÄÖ÷ÒªÇø±ðÔÚÓÚÓÃÓÚ×öÎïÌåʶ±ðµÄfeatureÊÇÊÖ¹¤Éè¼ÆÕÕ¾É×Ô¶¯Éè¼Æ£¬£¬£¬£¬£¬Éî¶ÈѧϰÌṩÁËÒ»¸ö´ÓdataÖлñÈ¡featureµÄ·½·¨£¬£¬£¬£¬£¬¹ØÓÚÉî¶ÈѧϰÀ´Ëµ£¬£¬£¬£¬£¬ÏÂÒ»²½µÄ×ßÏò»áÊÇÄÇÀ£¿£¿£¿£¿

ÏÖÔÚÒÔΪÉî¶ÈѧϰµÄÏÂÒ»²½ÊÇ×Ô¶¯»¯µÄÉî¶Èѧϰ¡£¡£¡£¡£¡£Éî¶Èѧϰ½â¾öÁËÒ»¸ö´ÓdataÖÐÍÚ¾òºÏÊʵÄfeatureµÄÎÊÌ⣬£¬£¬£¬£¬Ñо¿Ö°Ô±ÔÚÕû¸öѵÁ·Àú³ÌÖлáÓöµ½Ðí¶àÎÊÌ⣺ÊÇ·ñҪʹÓÃÊý¾ÝÔöÇ¿£¿£¿£¿£¿£¿ÒªÊ¹ÓÃʲôÊý¾ÝÔöÇ¿£¿£¿£¿£¿£¿ÊÇ·ñÓÃÁ˺ÏÊʵÄloss function£¿£¿£¿£¿£¿ÓÐûÓв½·¥Ñ°ÕÒ¸üºÃµÄloss function£¿£¿£¿£¿£¿Éî¶ÈѧϰµÄ¼Ü¹¹ÒªÓü¸¶à²ã£¿£¿£¿£¿£¿²ãÓë²ãµÄÍØÆ˹ØϵҪÔõÑùÈ·¶¨£¿£¿£¿£¿£¿ÔõÑù·ÖÅÉÅÌËãÁ¿£¿£¿£¿£¿£¿

ÔÚÉî¶Èѧϰ¿ò¼ÜÏÂÕâЩÎÊÌâÈÔÈ»ÐèÒªÊÖµ÷£¬£¬£¬£¬£¬ÎÒÃÇÏ£Íû´ÓÊý¾ÝÖлñÈ¡ÃÕµ×£¬£¬£¬£¬£¬ÄÇôAutoML¾ÍÊÇÆäÖÐÒ»ÖÖÊÖÒÕõ辶¡£¡£¡£¡£¡£ÏÖÔÚËüÓöµ½µÄÖ÷ÒªÎÊÌâÊÇÅÌËãÁ¿ÐèÇóºÜÊÇ´ó£¬£¬£¬£¬£¬´ÓһЩÊÂÇé¿ÉÒÔ¿´³ö£¬£¬£¬£¬£¬ÔÚNASNetÖÐÌáµ½ÆäÐèÒªµÄGPU DaysÊÇ1800Ì죬£¬£¬£¬£¬ËäÈ»¿ÉÒÔ»ñµÃ²»´íµÄЧ¹û£¬£¬£¬£¬£¬µ«¿ªÏú¹ØÓÚһЩС¹«Ë¾À´½²ÊÇÎÞ·¨½ÓÊܵÄ¡£¡£¡£¡£¡£ØʺóÐøÊÂÇéAmoebaNetʹÓÃÁËEvolutionµÄ·½·¨×öÍøÂç½á¹¹ËÑË÷£¬£¬£¬£¬£¬ÆäGPU DaysÉõÖÁ±È1800»¹Òª¸ß¡£¡£¡£¡£¡£ÔõÑù½â¾öcomputational budgetÊǸ÷ÈËÔÚ×öAutoMLʱºÜÊÇÌåÌùµÄÎÊÌâ¡£¡£¡£¡£¡£

ÏÖÔÚÒÔΪAutoMLÍêÕûµÄÁ÷³Ì£¬£¬£¬£¬£¬Ö÷Òª·ÖΪdata¡¢modelºÍoptimizationÈý¸ö²¿·Ö£¬£¬£¬£¬£¬Æä¹ØϵÈçÏÂͼËùʾ£º

´ÓData setµ½×îÖÕµÄÄ£×Ó»áÂÄÀúºÜ³¤µÄÀú³Ì£¬£¬£¬£¬£¬Ê×ÏÈÒª¶Ôdata¾ÙÐÐϴ媣¬£¬£¬£¬£¬È»ºóÈ·¶¨Ä£×ӽṹ£¬£¬£¬£¬£¬×îºóÊǶÔÄ£×Ó¾ÙÐÐoptimization¡£¡£¡£¡£¡£ÏÖÔÚÍŶӲ»µ«ÔÚNAS·½ÃæÓнϸ»ºñµÄЧ¹û£¬£¬£¬£¬£¬Í¬Ê±Ò²ÔÚAugmentationºÍLoss Function·½ÃæÓÐÒ»¶¨ÊµÑ飬£¬£¬£¬£¬²¢È¡µÃ²»´íµÄÏ£Íû¡£¡£¡£¡£¡£

ÏÂÃ潫ƾ֤ÒÔÉÏÏÈÈݵÄAutoMLµÄÈý·½Ã棬£¬£¬£¬£¬»®·ÖÏÈÈÝÉÌÌÀÑо¿ÍŶӽüÆÚÈ¡µÃµÄЧ¹û¡£¡£¡£¡£¡£

Part 2 ÂÛÎĽâ¶Á

1 Augmentation

ÕâƪµÄÊÂÇéÓÉgoogleÊ×´ÎʵÏÖ£¬£¬£¬£¬£¬pipeline»áºÍËûÃǵÄNASÒªÁìºÜÊÇ¿¿½ü£¬£¬£¬£¬£¬Ê×ÏÈÓÐÒ»¸öRNN Controller generated strategy£¬£¬£¬£¬£¬ÆäÖаüÀ¨ÁË×öAugmentationµÄ¸ÅÂʺͷù¶È£¬£¬£¬£¬£¬ÔÚ»ñµÃÁËAugmentation policyºó»áѵÁ·Ò»¸öÍøÂ磬£¬£¬£¬£¬´Ó³õʼ»¯ÑµÁ·µ½ÊÕÁ²£¬£¬£¬£¬£¬ÊÕÁ²ºó»áÔÚvalidation setÉÏ»ñµÃreward¡£¡£¡£¡£¡£Í¨¹ýrewardÀ´¼àÊÓRNN Controller£¬£¬£¬£¬£¬Ê¹µÃRNN Controller²ÉÑùreward¸ü¸ßµÄstrategy¡£¡£¡£¡£¡£Õâ¸öÁ÷³ÌÖеĵÚÒ»¸öÎÊÌâ¾ÍÊÇevaluationºÜÂý£¬£¬£¬£¬£¬Àú³ÌºÜÊǺÄʱ£¬£¬£¬£¬£¬Æä´ÎÈôÊÇReinforce Learning´îÅäRNN Controller¾ÙÐвÎÊýµÄѵÁ·£¬£¬£¬£¬£¬µü´ú´ÎÊýºÜÊǶ࣬£¬£¬£¬£¬×ÛºÏÆðÀ´»áÈÃʱ¼ä¿ªÏúºÜÊÇ´ó¡£¡£¡£¡£¡£ÎªÏàʶ¾öºÄʱµÄÎÊÌ⣬£¬£¬£¬£¬ËûÃDzÉÑùÁËÊý¾Ý¼¯£¬£¬£¬£¬£¬¶ÔCIFAR-10½ö²ÉÑù8%£¬£¬£¬£¬£¬¶ÔIMAGENET½ö²ÉÑù0.5%¡£¡£¡£¡£¡£²¢ÇÒ£¬£¬£¬£¬£¬Ö»ÄܶԼòµ¥½á¹¹¾ÙÐÐËÑË÷£¬£¬£¬£¬£¬ÆÚ´ýËÑË÷Ч¹ûÔÚ±ðµÄÍøÂçÉÏ»ñµÃºÃµÄЧ¹û¡£¡£¡£¡£¡£ÕâЩ¶¼ÊÇÒÔÉÏÊÂÇéµÄһЩÎó²î¡£¡£¡£¡£¡£ÎÒÃǽÓÄÉÁ˲î±ðµÄÊӽǣ¬£¬£¬£¬£¬°Ñaugmentation policy search¿´×÷Ò»¸ö³¬²ÎÓÅ»¯¡£¡£¡£¡£¡£¹Å°å³¬²ÎÓÅ»¯µÄÎÊÌâÒ²·ºÆðÔÚµü´úÉÏ£¬£¬£¬£¬£¬ÎÒÃÇÌá³öÁËOnline Hyper-parameter LearningµÄÒªÁìÀ´½â¾ö´ËÎÊÌâ¡£¡£¡£¡£¡£OHLÓë¹Å°å·½·¨µÄÇø±ðÊÇ£¬£¬£¬£¬£¬Ö»¾ÙÐе¥¸öÄ£×ÓµÄѵÁ·£¬£¬£¬£¬£¬single runÊÇÎÒÃÇÖ÷ÒªµÄÌص㣬£¬£¬£¬£¬ÇÒ²¢²»ÐèҪģ×Ó×öretrain£¬£¬£¬£¬£¬µ±ËÑË÷Íê³Éºó¿É»ñµÃ×îÖÕÏëÒªµÄÄ£×Ó¡£¡£¡£¡£¡£½«µ¥¸öÄ£×ÓµÄѵÁ·Àú³ÌÇгÉËéƬ£¬£¬£¬£¬£¬ÔËÐжà¸öÄ£×Ó£¬£¬£¬£¬£¬Ã¿¸öGPUÔÚÅÜÒ»¸öСtrunkʱµÄ³õʼ²ÎÊýÊÇÏàͬµÄ£¬£¬£¬£¬£¬²ÉÑù²î±ðµÄaugmentation policy£¬£¬£¬£¬£¬ÔËÓÃÔÚ²î±ðµÄÄ£×ÓÉÏ£¬£¬£¬£¬£¬»ñµÃ²î±ðµÄreward£¬£¬£¬£¬£¬¼àÊÓP0×ßÏòºÃµÄaugmentation policyÂþÑÜ¡£¡£¡£¡£¡£

Ч¹û£º¿ÉÒÔ¿´µ½³ýÁËAMOEBANET-B±ÈÕÕGoogle»áÉÔÑ·Ò»³ïÍ⣬£¬£¬£¬£¬ÆäËûµÄ¼¸¸öÄ£×Ó¶¼¿ÉÒÔÎȹÌÁè¼ÝgoogleµÄЧ¹û¡£¡£¡£¡£¡£

ÔÚImageNetÉϵIJâÊÔЧ¹û£º

ÂÛÎĵص㣺https://arxiv.org/abs/1905.07373

2 Network Architecture

1£©EcoNAS: Ϊ¹Å°åÍøÂçËÑË÷ÒªÁìÑ°ÕÒ×î¼ÑµÄÊðÀí

ÔÚÍøÂçËÑË÷ÖкÜÊÇÖ÷ÒªµÄ²¿·Ö¾ÍÊÇÆÀ¹ÀºòÑ¡×ӽṹµÄÓÅÁÓ¡£¡£¡£¡£¡£ÎÒÃÇͨ³£ÒªÊ¹ÓÃÊðÀí£¬£¬£¬£¬£¬ÔõÑùÆÀ¹ÀÊðÀíµÄÓÅÁÓÄØ£¿£¿£¿£¿£¿ÎÒÃdz£ÓõÄÒªÁìÊÇȥȨºâÊðÀíÌõ¼þÏ»ñµÃÄ£×ӵľ«¶ÈºÍÄ£×ÓÕæʵµÄ¾«¶ÈÅÅÐòµÄÒ»ÖÂÐÔ¡£¡£¡£¡£¡£¹þ¹þ(haha)ÌåÓýʵÑé²ÉÑùÁË50¸öÍøÂ磬£¬£¬£¬£¬°ÑËûÃÇѵÁ·µ½ÍêÈ«ÊÕÁ²£¬£¬£¬£¬£¬ÓûñµÃµÄ×îÖÕ¾«¶ÈµÄÅÅÐòºÍÔÚÊðÀíÇéÐÎÏÂÅÅÐòµÄ˹Ƥ¶ûÂüÅÅÐòÏà¹ØϵÊý£¨Spearman ranking coeffficient£©À´ÆÀ¹ÀËûÃÇÖ®¼äµÄ¹ØÁªÐÔ£¬£¬£¬£¬£¬Ïà¹ØϵÊýÔ½¸ßÔòÊðÀíÇéÐμá³ÖÍøÂçÅÅÐòµÄÄÜÁ¦£¨Ò²³Æ±£ÐòÐÔ£©Ô½¸ß£¬£¬£¬£¬£¬¹ØÓÚNASËã·¨Ô½ÓÐÀû¡£¡£¡£¡£¡£¿£¿£¿£¿£¿ÉÒÔ¸ø³öÏÂͼ£¬£¬£¬£¬£¬ºáÖáÊÇÊðÀíµÄ¼ÓËÙ±ÈÀý£º

¶ÔÉÏͼµÄʵÑéЧ¹û¾ÙÐÐÆÊÎöºóµÃÖª£¬£¬£¬£¬£¬¸ü³¤µÄѵÁ·µü´úºÍʹÓÃËùÓÐÊý¾Ý¾ÙÐÐѵÁ·¶Ô±£ÐòÊÇÓÐÀûµÄ£¬£¬£¬£¬£¬½µµÍÇø·ÖÂʶԱ£ÐòµÄÓ°Ïì²¢²»¼ÍÂÉ£¬£¬£¬£¬£¬¿ÉÒÔÊʵ±½µµÍÇø·ÖÂÊ£¬£¬£¬£¬£¬Í¬Ê±Í¨µÀ½µµÍ¶Ô±£ÐòÓÐÀû¡£¡£¡£¡£¡£»£»£»£»£»£»ùÓÚÎÒÃÇÉÏÊöµÄÊӲ죬£¬£¬£¬£¬ÎÒÃÇʹÓÃÁËʹµÃѵÁ·¿ìËÙµ«Ïà¹ØϵÊý´óµÄÉèÖã¬£¬£¬£¬£¬²¢Éè¼ÆÁËEcoNAS¡£¡£¡£¡£¡£EcoNAS½ÓÄÉÁËËÑË÷»ñµÃµÄ¸ßЧÉèÖã¬£¬£¬£¬£¬ÔÚÕâ¸öÉ趨Ï£¬£¬£¬£¬£¬µ¥Ä£×ÓѵÁ·20 epochÖ»ÐèÒªµ¥¿¨Êý·ÖÖÓ¡£¡£¡£¡£¡£¹ØÓÚѵÁ·µü´ú£¬£¬£¬£¬£¬ÎÒÃÇÉè¼ÆÁËÒ»Ì×·Ö²ãËã·¨£¬£¬£¬£¬£¬¶ÔÔçÆÚÌåÏÖÇ·ºÃµÄ±¸Ñ¡ÍøÂç¾ÙÐÐÔçÍ££¬£¬£¬£¬£¬´Ó¶ø½øÒ»²½ïÔÌ­ÅÌËã×ÊÔ´£º

ÕâÈý¸öÜöÝÍËæ×ÅËÑË÷µÄ¾ÙÐоÞϸÓɴ˽µµÍ£¬£¬£¬£¬£¬Ò²¾ÍÊÇ˵ֻÓкÜÉÙµÄÄ£×ӻᱻѵÁ·µ½3N´Îµü´ú£¬£¬£¬£¬£¬ÕâÊÇ»ùÓÚÎÒÃÇ֮ǰ·¢Ã÷µÄepochÔ½¶à£¬£¬£¬£¬£¬¶ÔÄ£×ÓµÄÆÀ¹ÀÔ½ºÃµÄÊÓ²ìÉè¼ÆµÄËã·¨¡£¡£¡£¡£¡£

ÔÚ8 GPU DaysÄÚEcoNAS¼´¿ÉËÑË÷£¬£¬£¬£¬£¬ÎÒÃÇËÑË÷µ½µÄ²ÎÊýÔÚ½«AMOEBANET¼ÓËÙ½ü260±¶µÄͬʱ¶ÔÐÔÄÜÒ²ÓÐÒ»¶¨ÌáÉý£¬£¬£¬£¬£¬¼ÓÈëÌõÀí»¯ÊðÀíºó»¹»á´øÀ´Á½±¶µÄËÙÂÊÌáÉýºÍ¸ü¸ßµÄÐÔÄÜ£¬£¬£¬£¬£¬¹þ¹þ(haha)ÌåÓýÑо¿Åú×¢ÎúÊðÀí±£ÐòÐÔ¶ÔNASËã·¨µÄÖ÷ÒªÐÔ¡£¡£¡£¡£¡£

ÂÛÎĵص㣺https://arxiv.org/abs/2001.01233

2£©PC-NAS: »ùÓÚ²ÎÊýºóÑéÂþÑÜÌØÕ÷Ìá¸ßÍøÂç½á¹¹ËÑË÷ÖÐÊðÀíµÄÓÐÓÃÐÔ

Ä¿½ñÍøÂç½á¹¹ËÑË÷Ëã·¨´ó×ÚµÄÅÌËãÁ¿ÐèÇó³ÉΪÁËÆäÓ¦ÓÃÖÐ×î´óµÄÎÊÌâ¡£¡£¡£¡£¡£È»¶ø½â¾öÅÌËãÁ¿ÎÊÌâµÄÒªÁìÏÖÔÚ»¹²»ÍêÉÆ£¬£¬£¬£¬£¬Ê¹Óù²Ïí²ÎÊýµÄ³¬Íø×öÊðÀí³ÉΪÁËÏÖÔÚ½ÏÁ¿ÆÕ±éµÄÑ¡Ôñ¡£¡£¡£¡£¡£µ«ÕâÀàÒªÁìÊÜÏÞÓÚÊðÀíºÍÕæÕæÏàÐεÄÎó²î¡£¡£¡£¡£¡£Í¨Ì«¹ýÎöËæ»úÌݶÈÓÅ»¯Çó½âµÄ²ÎÊýºóÑéÂþÑܵÄKLÉ¢¶È£¬£¬£¬£¬£¬ÎÒÃǵóöÁËïÔÌ­ËÑË÷¿Õ¼äÄÚÍøÂç½á¹¹ÊýÄ¿¿ÉÒÔÌá¸ßone-shotÏà¹ØÐԵĽáÂÛ¡£¡£¡£¡£¡£

ΪÁËÔÚѵÁ·È¨Öع²ÏíµÄͬʱ£¬£¬£¬£¬£¬¼õСȨÖع²ÏíµÄÍøÂç½á¹¹ÊýÄ¿£¬£¬£¬£¬£¬ÎÒÃÇÌá³öÒ»ÖÖѵÁ·Ê±µÄ¿Õ¼äÐÞ¼ôÒªÁì¡£¡£¡£¡£¡£¼ÙÉèÎÒÃÇÏÖÔÚÓÐÒ»¸öÈçÉÏͼËùʾµÄ³¬Íø£¬£¬£¬£¬£¬Ã¿Ò»²ãÓµÓÐÒ»¸ömixop£¨mixopÊÇn¸öºòÑ¡opµÄºÏ³Æ£©ÓÚÊǸó¬ÍøÖÐ×ܹ²°üÀ¨n^²ãÊý¸ö²î±ðµÄ½á¹¹¡£¡£¡£¡£¡£ÎªÁ˾ÙÐÐËÑË÷¿Õ¼äµÄÐÞ¼ô£¬£¬£¬£¬£¬ÎÒÃÇÌá³öÁ˲¿·Ö½á¹¹³ØµÄ¿´·¨(partial model pool)¡£¡£¡£¡£¡£¼ÙÉèÍøÂç½á¹¹ÓÐL²ã£¬£¬£¬£¬£¬²¿·Ö½á¹¹³ØÔÚÕû¸öËÑË÷Àú³ÌÖн«»áÓÐL¸ö״̬£¨Stage£©¡£¡£¡£¡£¡£Ã¿´ÓL-1״̬µ½L״̬£¬£¬£¬£¬£¬²¿·Ö½á¹¹³ØÖеÄ×ӽṹµÄ²ãÊý½«ÔöÌí1¡£¡£¡£¡£¡£

ˢкóµÄproxyºÍÔ­À´µÄone-shot proxyµÄ±ÈÕÕ¡£¡£¡£¡£¡£PC-NAS×îÖÕµÄËÑË÷¿Õ¼äºÜС£¡£¡£¡£¡£¬£¬£¬£¬£¬ÒÔÊÇ»ñµÃµÄ×îÖÕ¾«¶ÈºÜÊÇ¿¿½üµ¥¶ÀѵÁ·Ê±µÄÅÅÐò¡£¡£¡£¡£¡£

ÂÛÎĵص㣺https://arxiv.org/abs/1910.02543

3£©GreedyNAS: »ùÓÚÌ°?³¬?ÂçµÄOne-Shot NAS?·¨

ÎÒÃÇÌá³öÒ»ÖÖÌ°Ðij¬ÍøÂçÀ´¼õÇásupernetµÄÆÀ¹ÀѹÁ¦£¬£¬£¬£¬£¬Ê¹µÃsupernetÔ½·¢Ì°ÐĵØ×¢ÖØÓÚÓÐDZÁ¦µÄºÃ½á¹¹£¬£¬£¬£¬£¬Ìá¸ßËÑË÷µÄ¾«¶È¡£¡£¡£¡£¡£

ÏÈÇ°µÄÒªÁìÊ×ÏȽṹһ¸ö³¬ÍøÂ磬£¬£¬£¬£¬°üÀ¨ËÑË÷¿Õ¼äÀïµÄËùÓÐOPºÍ²ÎÊý£¬£¬£¬£¬£¬Í¬Ê±ÔÚÿһ´Î³¬ÍøÂçѵÁ·Ê±sample³öÒ»¸öСµÄ×ÓÍøÂç¾ÙÐÐѵÁ·¡£¡£¡£¡£¡£ÖØ´óµÄËÑË÷¿Õ¼ä´øÀ´µÄÆÀ¹ÀѹÁ¦Ê¹supernetÄÑÒÔ׼ȷµØÇø·Ö½á¹¹µÄÓÅÁÓ£¬£¬£¬£¬£¬ÓÉÓÚËùÓнṹµÄȨÖض¼ÊÇÔÚsupernetÖи߶ȹ²ÏíµÄ£¬£¬£¬£¬£¬ÈôÊÇÒ»¸ö²îµÄ½á¹¹±»ÑµÁ·µ½ÁË£¬£¬£¬£¬£¬ºÃµÄ½á¹¹µÄȨÖØÒ²»áÊܵ½×ÌÈÅ¡£¡£¡£¡£¡£ÕâÑùµÄ×ÌÈÅ»áÏ÷ÈõºÃÍøÂçµÄ×îÖÕÆÀ¹À¾«¶È£¬£¬£¬£¬£¬Ó°ÏìÍøÂçµÄËÑË÷Ч¹û¡£¡£¡£¡£¡£Í¬Ê±£¬£¬£¬£¬£¬¶Ô²îµÄ½á¹¹¾ÙÐÐѵÁ·Ï൱ÓÚ¶ÔȨÖؾÙÐÐÁËûÓÐÐëÒªµÄ¸üУ¬£¬£¬£¬£¬½µµÍsupernetµÄѵÁ·Ð§ÂÊ¡£¡£¡£¡£¡£

Òò´ËÎÒÃÇÏ£Íû½µµÍËÑË÷¿Õ¼äµÄ´óС£¡£¡£¡£¡£¬£¬£¬£¬£¬ÌáÉýËüµÄÆÀ¹ÀÐÔÄÜ¡£¡£¡£¡£¡£

¼ÙÉèËÑË÷¿Õ¼ä¿ÉÒÔÈçÉÏ»®·ÖΪºÃµÄ¿Õ¼äÓë²îµÄ¿Õ¼ä£¬£¬£¬£¬£¬ÇҺÿռäÖÐÿһ¸ö½á¹¹µÄACC¾ù´óÓÚ²î¿Õ¼ä£¬£¬£¬£¬£¬È»¶øÈ·¶¨ËùÓнṹÖÐÄÄЩÊÇÀ´×ԺÿռäÐèÒª±éÀúÕû¸öËÑË÷¿Õ¼ä£¬£¬£¬£¬£¬ÅÌË㿪ÏúÊÇÎÞ·¨½ÓÊܵÄ¡£¡£¡£¡£¡£

ΪÏàʶ¾öÕâÒ»ÎÊÌ⣬£¬£¬£¬£¬ÎÒÃÇÌá³öÁËÒ»ÖÖmulti-path sampling with rejectionµÄ¶à·ÖÖ§¹ýÂ˲ÉÑùÒªÁì¡£¡£¡£¡£¡£ÔÚÏÈÈÝÕâ¸öÒªÁìÇ°£¬£¬£¬£¬£¬ÎÒÃÇÊ×ÏȶÔÏÖÔÚµÄÔȳƲÉÑùÒªÁì¾ÙÐÐÆÊÎö¡£¡£¡£¡£¡£

ÓÉÓÚ·ÖÖ§µÄ²ÉÑùÊÇÔȳÆÇÒ×ÔÁ¦µÄ£¬£¬£¬£¬£¬ÄÇôÿ¸öpathÀ´×ԺÿռäµÄ¸ÅÂÊΪ£º

ΪÁË»ñµÃÐèÒªµÄÀ´×ÔÓںÿռäµÄ½á¹¹£¬£¬£¬£¬£¬ÎÒÃǽøÒ»²½Ë¼Á¿Ò»¸ö¶àάµÄBernoulliʵÑ飬£¬£¬£¬£¬ÄÇôÓÐÈçϵĽáÂÛ£º

ÎÒÃÇÈ¡m = 10ºÍm = 20£¬£¬£¬£¬£¬¶Ô¶¨Àí1µÄ¸ÅÂʾÙÐлæÖÆ£¬£¬£¬£¬£¬ÈçFigure 2Ëùʾ£¬£¬£¬£¬£¬¿É¼ûÕâÖÖ²ÉÑùÏ»ñµÃÀ´×ԺÿռäÖеÄpathµÄ¸ÅÂÊÊǺܸߵÄ¡£¡£¡£¡£¡£

ÓÚÊÇ£¬£¬£¬£¬£¬ÎÒÃÇ¿ÉÒÔ¶Ô²ÉÑùµ½µÄ¶à¸ö½á¹¹¾ÙÐÐÆÀ¹ÀɸѡµÄÒªÁìÌáÉý²ÉÑùµ½¡°ºÃ¡±½á¹¹µÄ¸ÅÂÊ£¬£¬£¬£¬£¬¼´Ã¿´Î²ÉÑùm¸ö½á¹¹£¬£¬£¬£¬£¬´ÓÖÐÑ¡È¡ÆÀ¹ÀÖ¸±ê×î¸ßµÄk¸ö½á¹¹¾ÙÐÐѵÁ·¡£¡£¡£¡£¡£ÎªÁËïÔÌ­ÆÀ¹ÀÍøÂçµÄÏûºÄ£¬£¬£¬£¬£¬ÎÒÃÇ´ÓÑéÖ¤¼¯ÖÐËæ»úÑ¡È¡ÁËһС²¿·Ö£¨Èç1000ÕÅͼ£©×é³ÉСÑéÖ¤¼¯£¬£¬£¬£¬£¬²¢Ê¹ÓÃÔÚСÑéÖ¤¼¯ÉϵÄloss×÷Ϊ½á¹¹µÄÅÅÐòÖ¸±ê¡£¡£¡£¡£¡£Ê¹ÓÃСÑéÖ¤¼¯¾ÙÐÐÆÀ¹À£¬£¬£¬£¬£¬ÔÚ°ü¹ÜÆÀ¹À׼ȷÐÔµÄÌõ¼þÏ£¬£¬£¬£¬£¬Ïà½Ïuniform samplingÒªÁìÖ»ÔöÌíÁ˺ÜÉÙµÄÅÌËã¼ÛÇ®¡£¡£¡£¡£¡£

ÔÚÇ°ÃæÌáµ½µÄ·¾¶Â˲¨ÖУ¬£¬£¬£¬£¬ÎÒÃÇͨ¹ýÆÀ¹À¿ÉÒÔÇø·Ö³ö½ÏºÃµÄ½á¹¹£¬£¬£¬£¬£¬ÎªÁ˽øÒ»²½ÌáÉýѵÁ·Ð§ÂÊ£¬£¬£¬£¬£¬ÎÒÃÇÌá³öʹÓÃÒ»¸ö·¾¶ºòÑ¡³ØÓÃÓÚ´æ·ÅѵÁ·Àú³ÌÖÐÆÀ¹À¹ýµÄ¡°ºÃ¡±½á¹¹£¬£¬£¬£¬£¬²¢¾ÙÐÐÖظ´Ê¹Óá£¡£¡£¡£¡£Ïêϸ¶øÑÔ£¬£¬£¬£¬£¬ºòÑ¡³Ø¿ÉÒÔ¿´×÷ÊÇÒ»¸öÀο¿¾ÞϸµÄÓÐÐòÐÐÁУ¬£¬£¬£¬£¬ÆäÖ»»á´æ´¢ËùÓÐÆÀ¹À¹ý½á¹¹Öе÷ÖÇ°n(ºòÑ¡³Ø¾Þϸ) µÄ½á¹¹¡£¡£¡£¡£¡£

ÔÚÍøÂç¸Õ×îÏÈѵÁ·Ê±ÆÀ¹À¹ýµÄ½á¹¹½ÏÉÙ£¬£¬£¬£¬£¬ºòÑ¡³ØÖд洢µÄ½á¹¹ÊǺýṹµÄ¿ÉÐŶȲ»¸ß£¬£¬£¬£¬£¬Òò´Ë´ÓºòÑ¡³ØÖвÉÑùµÄ¸ÅÂÊ ? ÔÚ×îÏÈʱÉèΪ0£¬£¬£¬£¬£¬²¢ÏßÐÔÔöÌíÖÁÒ»¸ö½Ï¸ßµÄÖµ£¬£¬£¬£¬£¬ÈôºòÑ¡³ØÖеĽṹ¶¼À´×Ժÿռ䣬£¬£¬£¬£¬Í¨¹ýʹÓúòÑ¡³Ø£¬£¬£¬£¬£¬²ÉÑù10¸ö½á¹¹£¬£¬£¬£¬£¬ÖÁÉÙÓÐ5¸öºÃ½á¹¹µÄ¸ÅÂÊÓÉ88.38%ÌáÉýÖÁ99.36% ¡£¡£¡£¡£¡£

ºòÑ¡³ØÖ÷ÒªÓÐÒÔÏÂÓŵ㣺

¡ñ ÌáÉýѵÁ·Ð§ÂÊ

¡ñ ÌáÉý²ÉÑù¾«¶È

¡ñ ¿ÉÒÔͨ¹ýºòÑ¡³Ø¸üÐÂƵÂÊÅжÏsupernetÊÇ·ñ¾ß±¸×ã¹»µÄÆÀ¹ÀÐÔÄÜ£¬£¬£¬£¬£¬Òò¶ø¿ÉÒÔÓÃ×÷×èֹѵÁ·µÄ²Î¿¼

¡ñ ʹ½ø»¯Ëã·¨ÓиüºÃµÄ³õʼ£¬£¬£¬£¬£¬Ìá¸ßËÑË÷µ½½á¹¹µÄÕûÌ徫¶È

ÂÛÎĵص㣺https://arxiv.org/abs/2003.11236

4£©CRNAS: ¼ì²âÖеÄËãÁ¦ÖØ·ÖÅÉ

ÎÒÃÇ»»ÁËÒ»¸öÑо¿µÄÊӽǣ¬£¬£¬£¬£¬´Ó¼ì²âÕâ¸öÎÊÌâ×Ô¼ºÈ¥Ë¼Á¿NAS¡£¡£¡£¡£¡£Ö®Ç°ÔÚNASÉÏ×ö¼ì²âµÄÎÄÕ¸ü¶àµÄÊÇÖ±½Ó°Ñ·ÖÀàµÄÂÄÀúÄõ½¼ì²âÉÏ×öËÑË÷£¬£¬£¬£¬£¬¹þ¹þ(haha)ÌåÓýÊÂÇéÊÇ»ùÓÚ¼ì²âµÄʹÃüÈ¥×öNAS£¬£¬£¬£¬£¬¶ø²»Êǽ«NASÖ±½ÓÓ¦Óõ½¼ì²âµÄʹÃüÉÏÀ´¡£¡£¡£¡£¡£

¶ÔRes50·ÖÀà¶øÑÔ£¬£¬£¬£¬£¬ImageNetÊäÈë¾ÞϸÊÇ224*224£¬£¬£¬£¬£¬¿ÉÊǹØÓÚ¼ì²â¶øÑÔ£¬£¬£¬£¬£¬ÄÃCOCOΪÀý£¬£¬£¬£¬£¬ÊäÈë¿ÉÄܵ½800*1333£¬£¬£¬£¬£¬²¢ÇÒ»¹ÐèҪȥhandle¸÷¸ö±ê×¼µÄÎïÌå¡£¡£¡£¡£¡£

Õâ¸öʱ¼ä£¬£¬£¬£¬£¬ÎÒÃǽ«ÍøÂçµÄÓÐÓøÐÊÜÒ°ERFÓëÍøÂçËãÁ¦·ÖÅÉģʽÁªÏµÁËÆðÀ´¡£¡£¡£¡£¡£ÔÚFaster RCNN + FPNÕâÒ»¾­µäbaselineÖУ¬£¬£¬£¬£¬ÎÒÃÇÏ£ÍûFPNµÄ¸÷²ãÓкÏÊÊÄÜÁ¦È¥¼ì²â¶ÔÓ¦¾ÞϸµÄÎïÌå(Anchor Size)¡£¡£¡£¡£¡£µ±ÎÒÃÇ»­³öR50-FPNµÄ¸÷²ãµÄERFʱ£¬£¬£¬£¬£¬·¢Ã÷¹ØÓÚP3¶øÑÔ(СÎïÌå)£¬£¬£¬£¬£¬ERFͦ´óÓÐÈßÓ࣬£¬£¬£¬£¬µ«¹ØÓÚP5¶øÑÔ(´óÎïÌå)£¬£¬£¬£¬£¬ERFÓÖ²»·óÓеãÏ£º±£¬£¬£¬£¬£¬¼ûÏÂͼFigure 1ÓÒÉϽÇ¡£¡£¡£¡£¡£ÕâÒ²¾ÍÊÇÖ±½ÓÓ¦ÓÃRes50×÷ΪbackboneµÄÎÊÌâ¡£¡£¡£¡£¡£

¾­Óɹþ¹þ(haha)ÌåÓýstageµÄÖØ·ÖÅÉÖ®ºó£¬£¬£¬£¬£¬¿ÉÒÔÏÔ×Å¿´µ½¸÷²ãFPNÓÐÔ½·¢Æ½ºâµÄERF£¬£¬£¬£¬£¬°ü¹ÜÁË×îºóµÄÐÔÄÜ¡£¡£¡£¡£¡£

ΪÁ˾ÙÐÐËãÁ¦ÖØ·ÖÅÉ£¬£¬£¬£¬£¬ÎÒÃǹ¹½¨ÁËÒ»¸öÁ½¼¶ËÑË÷À´ÅäºÏÕâ¸öÖØ´óµÄËÑË÷¿Õ¼ä¡£¡£¡£¡£¡£ÎÒÃÇÏȾÙÐÐstageµÄËÑË÷£¬£¬£¬£¬£¬È»ºó¾ÙÐÐspatialµÄËÑË÷¡£¡£¡£¡£¡£¼òÆÓÀ´Ëµ£¬£¬£¬£¬£¬ÏÈËÑË÷¸÷¸östageµÄblock number£¬£¬£¬£¬£¬ÔÙËÑË÷¸÷¸öλÖõÄdilated convµÄrate£¬£¬£¬£¬£¬ÓÉÓÚdilated convͨ¹ý¸Ä±ä²ÉÑùµãµÄÏ£º±Ë®Æ½£¬£¬£¬£¬£¬¿ÉÒԸıäÍøÂçµÄ¸ÐÊÜÒ°¡£¡£¡£¡£¡£

ÎÒÃÇʹÓÃOne-shot NASµÄ·½·¨¹¹½¨³¬Í¼£¬£¬£¬£¬£¬µ±³¬Í¼ÑµÁ·Íê³Éºó£¬£¬£¬£¬£¬ÎÒÃÇÈ¥²âÊÔÖÖÖÖ×éºÏµÄÐÔÄÜ£¬£¬£¬£¬£¬´ÓÆäÖÐÌôÑ¡ÖÐ×îºÃµÄ×éºÏ¡£¡£¡£¡£¡£

ÎÒÃÇÒ²Éè¼ÆÁË¿ÉÒÔÖظ´Ê¹Óõij¬Í¼£¬£¬£¬£¬£¬Ò²¾ÍÊÇÄãѵÁ·ÁËÒ»¸ö³¬Í¼Ö®ºó£¬£¬£¬£¬£¬ÎÒÃÇ¿ÉÒÔ´ÓÖÐÌôÑ¡³öÕâ¸öϵÁÐÖÐÈκÎÅÌËãÁ¿ÏÂ×îºÃµÄÄ£×Ó£¬£¬£¬£¬£¬ÈçRes18µ½Res101¡£¡£¡£¡£¡£

ÔÚ»ñµÃstageµÄЧ¹ûÖ®ºó£¬£¬£¬£¬£¬ÎÒÃÇʹÓÃdilated convÀ´Íê³É²î±ðλÖÃÏ£º±µÄ²ÉÑù¡£¡£¡£¡£¡£ÓëµÚ1²½Ò»Ñù£¬£¬£¬£¬£¬ÎÒÃÇÒ²ÊÇʹÓÃOne-shot NASµÄ·½·¨¹¹½¨³¬Í¼¡£¡£¡£¡£¡£

´ÓЧ¹ûÉÏÃæÀ´¿´£¬£¬£¬£¬£¬¾­Óɹþ¹þ(haha)ÌåÓýCRNAS£¬£¬£¬£¬£¬ËùÓеÄÄ£×ÓÔÚCOCOÉϾùÓÐ1.5-1.9 mAP×óÓÒµÄÌáÉý£¬£¬£¬£¬£¬²¢ÇÒÕâЩÌáÉý¶¼Ã»ÓдøÀ´ÈκεÄÅÌËãÁ¿ÔöÌí¡£¡£¡£¡£¡£ÇÒÄ£×ÓÓнϺõÄǨáãÐÔÄÜ¡£¡£¡£¡£¡£

ÂÛÎĵص㣺https://arxiv.org/abs/1912.11234

3 Network Architecture

1£©AM-LFS£ºAutoML for Loss Function Search

Loss functionÊÇÔÚÐí¶àͼÏñʹÃüÉϽÏÁ¿Ö÷ÒªµÄµ÷²ÎµÄ¹¤¾ß£¬£¬£¬£¬£¬ÏÖÔÚ±£´æµÄLoss functionÓÐÐí¶àÎÊÌ⣬£¬£¬£¬£¬ÏàÒËÓÃAutoML½â¾ö£¬£¬£¬£¬£¬¿ÉÒÔͨ¹ýÊý¾ÝÈ·¶¨Loss functionµÄÐÎʽ¡£¡£¡£¡£¡£»£»£»£»£»£»úеÊÇ·ñÄÜѧϰLoss function£¿£¿£¿£¿£¿½«ÒÑÍùµÄLoss function·ÅÔÚloss¿Õ¼äÖУ¬£¬£¬£¬£¬·¢Ã÷ÁËÒ»¸öͳһµÄ±í´ï£º

ÔÚÑ¡Óòî±ðµÄÏêϸ±í´ïʱ£¬£¬£¬£¬£¬»áÓ¦¶ÔÔ­¿Õ¼äÖвî±ðµÄloss function¡£¡£¡£¡£¡£ÔÚÓÐÁËÕâÒ»±í´ïºó£¬£¬£¬£¬£¬¶ÔÕâÒ»º¯ÊýµÄËÑË÷ÎÒÃǽÓÄÉÁ˼òÆӵķֶκ¯Êý£¬£¬£¬£¬£¬ÔÚ·Ö¶ÎÄÚ²¿ÊǼòÆÓµÄÏ߶Ρ£¡£¡£¡£¡£

ÎÒÃÇÌá³öÁ˽«loss function searchת»¯ÎªÁíÍâÁ½¸öº¯ÊýµÄËÑË÷£¬£¬£¬£¬£¬Ê¹ÓÃÁËÇ°ÃæÌáµ½µÄOHL framework£¬£¬£¬£¬£¬Óòî±ðµÄtrunk¸üÐÂÁ½¸öº¯ÊýµÄ³¬²Î£¬£¬£¬£¬£¬Ê¹·Ö¶Îº¯Êý¿ÉÒÔÄâºÏÉÏÄ¿µÄº¯Êý¡£¡£¡£¡£¡£×îºóÔÚCifar10ÉÏʵÑé±ÈÕÕbaselineÔÚnoisy ratioµÄÇéÐÎÏ¿ÉÒÔÌáÉý2¸÷µã×óÓÒ£¬£¬£¬£¬£¬´øÓÐnoiseµÄÇéÐÎÏ»áÌáÉý¸ü¶à¡£¡£¡£¡£¡£

ÂÛÎĵص㣺https://arxiv.org/abs/1905.07375

Part 3 Q&A

Q£ºÍ¨µÀÊý½µµÍ¿ÉÒÔ»ñµÃ¸üºÃµÄÐÔÄÜÆÀ¹À£¬£¬£¬£¬£¬ÕâÒ»µã¹ØÓÚ²î±ðµÄÍøÂç½á¹¹ºÍloss function¶¼ÊÊÓÃÂ𣿣¿£¿£¿£¿

A£ºÕâÊÇÒ»¸ö½ÏÁ¿·´Ö±¾õµÄÕ÷Ïó£¬£¬£¬£¬£¬Óöµ½ÐµÄʹÃüʱ¿ÉÒÔÍùÕâ¸öÆ«ÏòʵÑ飬£¬£¬£¬£¬µ«Î´±Ø»áÕÒµ½Ò»¸öÄÇÀﶼÊÊÓõļÍÂÉ¡£¡£¡£¡£¡£

Q£ºlossµÄËÑË÷ʵÖÊÉÏÊǶÔÈ·¶¨ºÃµÄ¼¸ÖÖloss¾ÙÐÐϵÊýµÄ×Ô¶¯Ñ°ÓÅÂ𣿣¿£¿£¿£¿

A£º²»ÊǵÄ¡£¡£¡£¡£¡£ÎÒÃDzο¼Á˼¸×éloss functionµÄÌØÕ÷£¬£¬£¬£¬£¬Áýͳ³ÉÁËloss functionµÄ¿Õ¼ä£¬£¬£¬£¬£¬ÓÉÁ½¸öµ¥¶ÀµÄº¯ÊýÈ·¶¨×îÖյķ½·¨£¬£¬£¬£¬£¬ÎÒÃÇ°Ñloss function searchµÄº¯Êýת»¯µ½ÁËÇ°ÃæÌáµ½µÄÁ½¸öº¯ÊýÉÏ£¬£¬£¬£¬£¬È»ºóÓ÷ֶÎÏßÐÔº¯ÊýµÄ²ÎÊýÀ´ËÑË÷еĿռä¡£¡£¡£¡£¡£

Q£ºPaperÖÐÌáµ½£¬£¬£¬£¬£¬ÓÃ1000ÕÅͼƬÆÀ¼ÛÊDz»Ì«ÉÙÁË£¿£¿£¿£¿£¿

A£ºÔÚpaperÖУ¬£¬£¬£¬£¬ÎÒÃǶÔÖÖÖÖ¾ÞϸµÄÑéÖ¤¼¯ÓëÍêÕûÑéÖ¤¼¯µÄÅÅÐòÏà¹ØÐÔ×öÁËÏêϸµÄÆÊÎö£¬£¬£¬£¬£¬Ñ¡Ôñ1000ÕÅÊÇÓÉÓÚͼƬÊýÄ¿µ½1000ʱÅÅÐòÏà¹ØÐÔÒѾ­¿ÉÒÔÎȹ̵½90%ÒÔÉÏ£¬£¬£¬£¬£¬ÔÙÔöÌíÒ²²»»áÓиü´óµÄÌáÉý£¬£¬£¬£¬£¬Òò´ËΪÁ˼æ¹ËЧÂÊ£¬£¬£¬£¬£¬ÎÒÃÇÑ¡ÔñÁË1000ÕÅͼƬ¡£¡£¡£¡£¡£

Q£ººòÑ¡³ØÖеĽṹÓиü´ó¿ÉÄÜÐÔ±»¶à´ÎѵÁ·£¬£¬£¬£¬£¬Õâ¶Ôone shotÒªÁì×îÖÕÆÀ¹À»áÓнϴóÓ°Ï죬£¬£¬£¬£¬Ôõô°ü¹ÜÕÒµ½µÄ½á¹¹²»ÊÇÓÉÓÚÎÞÒâÒòËضø±»ÑµÁ·×î¶àµÄ½á¹¹£¿£¿£¿£¿£¿

A£ºÈôÊÇ´ÓºòÑ¡³ØÖвÉÑù£¬£¬£¬£¬£¬¿ÉÒÔÌáÉýЧÂÊ£¬£¬£¬£¬£¬µ«Ò²»áɥʧËÑË÷½á¹¹µÄ¶àÑùÐÔ£¬£¬£¬£¬£¬ÒÔÊÇÔÚʵÏÖÖÐΪÁË°ü¹Ü´Ó´ó¿Õ¼ä²ÉÑùµÄ¶àÑùÐÔÓë³ØÖÐËÑË÷ЧÂʵÄƽºâ£¬£¬£¬£¬£¬ÎÒÃÇÒ»×îÏÈÖ»»á´ÓÍêÕûµÄËÑË÷¿Õ¼ä¾ÙÐÐsample£¬£¬£¬£¬£¬¹«ÕýµØ²ÉÑùËùÓнṹ£¬£¬£¬£¬£¬µ±ÑµÁ·ÂÖÊýµÖ´ïÒ»¶¨Ë®Æ½ÔÙÒýÈëºòÑ¡³Ø¾ÙÐÐËÑË÷£¬£¬£¬£¬£¬ÇÒ´ÓºòÑ¡³ØÖвÉÑùµÄ¸ÅÂÊ»áÓÉ0ÏßÐÔÔöÌíÖÁ0.8¡£¡£¡£¡£¡£ÕâÑù´Ó¹«Õýµ½Ì°ÐÄѵÁ·Öð½¥ÑݱäµÄÀú³Ì£¬£¬£¬£¬£¬¿ÉÒÔïÔ̭ѵÁ·²»Æ½ºâ¡¢½á¹¹²»¹«ÕýµÄÇéÐΣ¬£¬£¬£¬£¬Í¬Ê±»ñµÃÌ°ÐÄ´øÀ´µÄЧÂÊÌáÉý¡£¡£¡£¡£¡£

Q£ºNAS»á²»»áÔÚËÑË÷µÄÊý¾Ý¼¯ÉϹýÄâºÏ£¬£¬£¬£¬£¬µ¼ÖÂÔÚÏÖʵ³¡¾°ÖÐÌåÏÖûÄÇôºÃ£¿£¿£¿£¿£¿

A£º¾ø´ó²¿·ÖCVʹÃü¶¼»ùÓÚImageNET£¬£¬£¬£¬£¬ÈôÊÇNASÄÜÔÚImageNET»ù´¡ÉÏËѳöÒ»¸öºÃµÄÄ£×Ó£¬£¬£¬£¬£¬»òÐíÂÊ¿ÉÒÔǨáãµ½ÆäËûÏÖʵʹÃüÉϵÄ¡£¡£¡£¡£¡£ÔÚһЩTask specific researchÖУ¬£¬£¬£¬£¬ÓпÉÄÜÕë¶Ô²î±ðµÄËÑË÷ʱ»áÄâºÏµ½ÄãËÑË÷µÄÊý¾Ý¼¯¡£¡£¡£¡£¡£

¡¸ÉÌÌÀ̩̹¹ûÕæ¿Î¡¹¹«Öںźǫ́»Ø¸´¡°010¡±£¬£¬£¬£¬£¬¼´¿É»ñÈ¡ÏìÓ¦µÄ¿Î³Ì×ÊÁÏŶ£¡