ÂÛÎĵص㣺

GreedyNAS: Towards Fast One-Shot NAS with Greedy Supernet

https://arxiv.org/abs/2003.11236

µ¼¶Á£º

ÔÚCVPR 2020ÉÏ£¬£¬£¬£¬£¬£¬ÉÌÌÀÒƶ¯ÖÇÄÜÊÂҵȺ-3DAR-Éí·ÝÈÏÖ¤ÓëÊÓƵ¸ÐÖª×éÌá³öÁË»ùÓÚÌ°Ðij¬ÍøÂçµÄOne-Shot NASÒªÁì£¬£¬£¬£¬£¬£¬ÏÔÖøÌáÉýÁ˳¬ÍøÂçÖ±½ÓÔÚ´ó¹æÄ£Êý¾Ý¼¯ÉϵÄËÑË÷ѵÁ·Ð§ÂÊ£¬£¬£¬£¬£¬£¬²¢ÔÚ±ê×¼ImageNetÊý¾Ý¼¯ÉÏÈ¡µÃÁË300M FLOPsÁ¿¼¶µÄSOTA¡£¡£¡£¡£¡£¡£¡£¡£GreedyNASÂÛÎÄͨ¹ýÌá³öÒ»ÖÖÌ°Ðĵij¬ÍøÂç½á¹¹²ÉÑùѵÁ·ÒªÁì£¬£¬£¬£¬£¬£¬¸ÄÉÆÁËѵÁ·»ñµÃµÄ³¬ÍøÂç¶Ô½á¹¹µÄÆÀ¹ÀÄÜÁ¦£¬£¬£¬£¬£¬£¬½ø¶ø×ÊÖúËÑË÷Ëã·¨»ñµÃ¾«¶È¸ü¸ßµÄ½á¹¹¡£¡£¡£¡£¡£¡£¡£¡£

ÄîÍ·ÓëÅä¾°

ÔÚÏÖÔÚµÄÉñ¾­½á¹¹ËÑË÷ÁìÓòÖÐ£¬£¬£¬£¬£¬£¬One-Shot NASÒªÁìÓÉÓÚÆäËÑË÷¿ªÏúС±»ÆÕ±éÓ¦ÓÃ£¬£¬£¬£¬£¬£¬ÕâЩҪÁìʹÓÃÒ»¸öȨÖع²ÏíµÄ³¬ÍøÂ磨supernet£©×÷Ϊ²î±ðÍøÂç½á¹¹µÄÐÔÄÜÆÀ¹ÀÆ÷£¬£¬£¬£¬£¬£¬Òò´Ë£¬£¬£¬£¬£¬£¬supernetµÄѵÁ·¶ÔËÑË÷Ч¹ûµÄÓÅÁÓÖÁ¹ØÖ÷Òª¡£¡£¡£¡£¡£¡£¡£¡£È»¶ø£¬£¬£¬£¬£¬£¬ÏÖÔÚµÄÒªÁìÒ»Ñùƽ³£½ÓÄÉÁËÒ»¸ö»ù±¾µÄ¼ÙÉè£¬£¬£¬£¬£¬£¬¼´supernetÖÐÿһ¸ö½á¹¹ÊÇÒ»ÂÉÖ÷ÒªµÄ£¬£¬£¬£¬£¬£¬supernetÓ¦¸Ã¶Ôÿ¸ö½á¹¹¾ÙÐÐ׼ȷÆÀ¹À»òÏà¶ÔÅÅÐò¡£¡£¡£¡£¡£¡£¡£¡£È»¶ø£¬£¬£¬£¬£¬£¬supernetÖÐËù°üÀ¨µÄ½á¹¹Á¿¼¶£¨ËÑË÷¿Õ¼äµÄsize£©ÊǺÜÊÇÖØ´óµÄ£¨Èç 7^{21} £©£¬£¬£¬£¬£¬£¬Òò´Ë׼ȷµÄÆÀ¹À¹ØÓÚsupernetÀ´ËµÊǺÜÊÇÄÑÌâµÄ£¬£¬£¬£¬£¬£¬µ¼ÖÂsupernetÖнṹµÄÌåÏÖÓëÆäÕæʵÌåÏÖÏà¹ØÐԺܲî [1]¡£¡£¡£¡£¡£¡£¡£¡£

ÔÚ±¾ÆªÂÛÎÄÖÐ£¬£¬£¬£¬£¬£¬ÎÒÃÇÌá³öÒ»ÖÖÌ°Ðij¬ÍøÂçÀ´¼õÇásupernetµÄÆÀ¹ÀѹÁ¦£¬£¬£¬£¬£¬£¬Ê¹µÃsupernetÔ½·¢Ì°ÐĵØ×¢ÖØÓÚÓÐDZÁ¦µÄºÃ½á¹¹£¬£¬£¬£¬£¬£¬¶ø²»ÊÇÈ«Ìå¡£¡£¡£¡£¡£¡£¡£¡£Ïêϸ¶øÑÔ£¬£¬£¬£¬£¬£¬ÔÚsupernetѵÁ·Àú³ÌÖÐ£¬£¬£¬£¬£¬£¬ÎÒÃÇÌá³öÁËÒ»Öֶ෾¶¾Ü¾øʽ²ÉÑùÒªÁì(multi-path sampling with rejection)À´¾ÙÐз¾¶Â˲¨ (path filtering)£¬£¬£¬£¬£¬£¬Ê¹µÃÓÐDZÁ¦µÄºÃ½á¹¹»ñµÃѵÁ·¡£¡£¡£¡£¡£¡£¡£¡£Í¨¹ýÕâÖÖÒªÁì£¬£¬£¬£¬£¬£¬supernetµÄѵÁ·´ÓÕû¸öËÑË÷¿Õ¼äÌ°ÐĵØËõСµ½ÁËÓÐDZÁ¦µÄ½á¹¹×é³ÉµÄ¿Õ¼äÖÐ£¬£¬£¬£¬£¬£¬Òò´ËѵÁ·µÄЧÂÊ»ñµÃÁËÌáÉý¡£¡£¡£¡£¡£¡£¡£¡£Í¬Ê±£¬£¬£¬£¬£¬£¬ÎªÁ˽øÒ»²½Ôö´óÓÐDZÁ¦½á¹¹µÄ²ÉÑù¸ÅÂÊÓëÌá¸ßѵÁ·Ð§ÂÊ£¬£¬£¬£¬£¬£¬ÎÒÃÇ»ùÓÚexploration and exploitation×¼Ôò£¬£¬£¬£¬£¬£¬Ê¹ÓÃÒ»¸öÂÄÀú³Ø´æ´¢ÆÀ¹À¹ýµÄ¡°ºÃ¡±½á¹¹£¬£¬£¬£¬£¬£¬ÓÃÀ´ÔöÇ¿Ì°ÐĶȲ¢ÎªºóÐøµÄËÑË÷ÌṩºÃµÄ³õʼµã¡£¡£¡£¡£¡£¡£¡£¡£±¾ÂÛÎÄËÑË÷³öµÄ½á¹¹ÔÚImageNet (mobile setting) ÏÂÈ¡µÃÁË SOTA(state-of-the-art) µÄЧ¹û¡£¡£¡£¡£¡£¡£¡£¡£

Ö÷Ҫ˼Ð÷

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¸üÐÂ£¬£¬£¬£¬£¬£¬½µµÍsupernetµÄѵÁ·Ð§ÂÊ¡£¡£¡£¡£¡£¡£¡£¡£

¶à·¾¶¾Ü¾øʽ²ÉÑù

Õë¶ÔÉÏÊöÎÊÌâ£¬£¬£¬£¬£¬£¬Ò»¸öÖ±½ÓµÄÏë·¨¾ÍÊÇ»ùÓÚÌ°ÐÄÕ½ÂÔ£¬£¬£¬£¬£¬£¬ÔÚѵÁ·Àú³ÌÖÐֻѵÁ·ºÃµÄ½á¹¹¡£¡£¡£¡£¡£¡£¡£¡£µ«ºÜÏÔÈ»£¬£¬£¬£¬£¬£¬ÎÒÃDz¢²»ÖªµÀÒ»¸öËæ»ú³õʼ»¯µÄËÑË÷¿Õ¼äÖÐÄÄЩ½á¹¹ÊǺõġ£¡£¡£¡£¡£¡£¡£¡£¼ÙÉè¹ØÓÚÒ»¸ösupernet£¬£¬£¬£¬£¬£¬ÎÒÃÇ˼Á¿ÆäËÑË÷¿Õ¼äAµÄÒ»¸öÍêÕû»®·Ö£¬£¬£¬£¬£¬£¬¼´£º

ËÑË÷¿Õ¼ä¿ÉÒÔÈçÉÏ»®·ÖΪºÃµÄ¿Õ¼äÓë²îµÄ¿Õ¼ä£¬£¬£¬£¬£¬£¬ÇҺÿռäÖÐÿһ¸ö½á¹¹µÄACC¾ù´óÓÚ²î¿Õ¼ä£¬£¬£¬£¬£¬£¬¼´

ÓÚÊÇ£¬£¬£¬£¬£¬£¬Ò»¸öÀíÏëµÄ²ÉÑùÕ½ÂÔÊÇÖ±½ÓÔںÿռäÖоÙÐвÉÑù¼´¿É¡£¡£¡£¡£¡£¡£¡£¡£È»¶øƾ֤ÉÏÃæµÄ²»µÈʽ£¬£¬£¬£¬£¬£¬È·¶¨ËùÓнṹÖÐÄÄЩÊÇÀ´×ԺÿռäÐèÒª±éÀúÕû¸öËÑË÷¿Õ¼ä£¬£¬£¬£¬£¬£¬ÅÌË㿪ÏúÊÇÎÞ·¨½ÓÊܵġ£¡£¡£¡£¡£¡£¡£¡£ÎªÏàʶ¾öÕâ¸öÎÊÌâ£¬£¬£¬£¬£¬£¬ÎÒÃÇÊ×ÏÈ˼Á¿´ÓÈ«¿Õ¼äÖоÙÐеÄÒ»¸öÔȳƲÉÑù£¬£¬£¬£¬£¬£¬ÄÇôÿ¸öpathÀ´×ԺÿռäµÄ¸ÅÂÊΪ£º

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

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

ÓÚÊÇ£¬£¬£¬£¬£¬£¬ÎÒÃÇ¿ÉÒÔ¶Ô²ÉÑùµ½µÄ¶à¸ö½á¹¹¾ÙÐÐÆÀ¹ÀɸѡµÄÒªÁìÌáÉý²ÉÑùµ½¡°ºÃ¡±½á¹¹µÄ¸ÅÂÊ£¬£¬£¬£¬£¬£¬¼´Ã¿´Î²ÉÑùm¸ö½á¹¹£¬£¬£¬£¬£¬£¬´ÓÖÐÑ¡È¡ÆÀ¹ÀÖ¸±ê×î¸ßµÄk¸ö½á¹¹¾ÙÐÐѵÁ·¡£¡£¡£¡£¡£¡£¡£¡£¿ÉÊÇ£¬£¬£¬£¬£¬£¬¶Ô½á¹¹¾ÙÐÐÆÀ¹ÀÐèÒªÔÚÑéÖ¤¼¯ÉÏÅÌËãÆäACC£¬£¬£¬£¬£¬£¬ÕâÑù»áÔöÌíºÜÊǶàÅÌËãÁ¿£¨¹þ¹þ(haha)ÌåÓýImageNetÑéÖ¤¼¯¾ÞϸΪ50k£©¡£¡£¡£¡£¡£¡£¡£¡£ÎªÁËïÔÌ­ÆÀ¹ÀÍøÂçµÄÏûºÄ£¬£¬£¬£¬£¬£¬ÎÒÃÇ´ÓÑéÖ¤¼¯ÖÐËæ»úÑ¡È¡ÁËһС²¿·Ö£¨Èç1000ÕÅͼ£©×é³ÉСÑéÖ¤¼¯£¬£¬£¬£¬£¬£¬²¢Ê¹ÓÃÔÚСÑéÖ¤¼¯ÉϵÄloss×÷Ϊ½á¹¹µÄÅÅÐòÖ¸±ê¡£¡£¡£¡£¡£¡£¡£¡£Ê¹ÓÃСÑéÖ¤¼¯¾ÙÐÐÆÀ¹À£¬£¬£¬£¬£¬£¬ÔÚ°ü¹ÜÆÀ¹À׼ȷÐÔµÄÌõ¼þÏÂ£¬£¬£¬£¬£¬£¬Ïà½Ïuniform samplingÒªÁìÖ»ÔöÌíÁ˺ÜÉÙµÄÅÌËã¼ÛÇ®£¬£¬£¬£¬£¬£¬Ïê¼ûÂÛÎÄʵÑ鲿·Ö¡£¡£¡£¡£¡£¡£¡£¡£

»ùÓÚexploration and exploitationÕ½ÂԵķ¾¶ºòÑ¡³Ø

ÔÚÇ°ÃæÌáµ½µÄ·¾¶Â˲¨ÖÐ£¬£¬£¬£¬£¬£¬ÎÒÃÇͨ¹ýÆÀ¹À¿ÉÒÔÇø·Ö³ö½ÏºÃµÄ½á¹¹£¬£¬£¬£¬£¬£¬ÎªÁ˽øÒ»²½ÌáÉýѵÁ·Ð§ÂÊ£¬£¬£¬£¬£¬£¬ÊÜÃÉÌØ¿¨ÂåÊ÷ËÑË÷(Monte Carlo tree search)[4] ºÍ deep Q-learning[5] Öг£ÓõÄexploration and exploitationÕ½ÂÔÆô·¢£¬£¬£¬£¬£¬£¬ÎÒÃÇÌá³öʹÓÃÒ»¸ö·¾¶ºòÑ¡³ØÓÃÓÚ´æ·ÅѵÁ·Àú³ÌÖÐÆÀ¹À¹ýµÄ¡°ºÃ¡±½á¹¹£¬£¬£¬£¬£¬£¬²¢¾ÙÐÐÖظ´Ê¹Óᣡ£¡£¡£¡£¡£¡£¡£Ïêϸ¶øÑÔ£¬£¬£¬£¬£¬£¬ºòÑ¡³Ø¿ÉÒÔ¿´×÷ÊÇÒ»¸öÀο¿¾ÞϸµÄÓÐÐòÐÐÁÐ£¬£¬£¬£¬£¬£¬ÆäÖ»»á´æ´¢ËùÓÐÆÀ¹À¹ý½á¹¹Öе÷ÖÇ°n(ºòÑ¡³Ø¾Þϸ) µÄ½á¹¹¡£¡£¡£¡£¡£¡£¡£¡£

ÓÐÁ˺òÑ¡³ØµÄ×ÊÖú£¬£¬£¬£¬£¬£¬ÎÒÃÇ¿ÉÒÔÑ¡Ôñ´ÓËÑË÷¿Õ¼äÖлòºòÑ¡³ØÖвÉÑù½á¹¹¡£¡£¡£¡£¡£¡£¡£¡£´ÓºòÑ¡³ØÖвÉÑùµÄ½á¹¹ÊǺýṹµÄ¸ÅÂʸü¸ß£¬£¬£¬£¬£¬£¬µ«¿ÉÄÜ»áÎþÉü½á¹¹µÄ¶àÑùÐÔ¡£¡£¡£¡£¡£¡£¡£¡£ÎªÁËƽºâexplorationÓëexploitation£¬£¬£¬£¬£¬£¬ÎÒÃǽÓÄÉ?-²ÉÑùÕ½ÂÔ£¬£¬£¬£¬£¬£¬¼´ÒÔÒ»¶¨µÄ¸ÅÂÊ´ÓÕû¸öËÑË÷¿Õ¼äA»òºòÑ¡³ØPÖвÉÑù½á¹¹¦Á£º

ÔÚÍøÂç¸Õ×îÏÈѵÁ·Ê±ÆÀ¹À¹ýµÄ½á¹¹½ÏÉÙ£¬£¬£¬£¬£¬£¬ºòÑ¡³ØÖд洢µÄ½á¹¹ÊǺýṹµÄ¿ÉÐŶȲ»¸ß£¬£¬£¬£¬£¬£¬Òò´Ë´ÓºòÑ¡³ØÖвÉÑùµÄ¸ÅÂÊ ? ÔÚ×îÏÈʱÉèΪ0£¬£¬£¬£¬£¬£¬²¢ÏßÐÔÔöÌíÖÁÒ»¸ö½Ï¸ßµÄÖµ£¨ÔÚʵÑéÖÐ£¬£¬£¬£¬£¬£¬ÎÒÃÇ·¢Ã÷0.8ÊÇÒ»¸ö½ÏÓŵÄÖµ£©¡£¡£¡£¡£¡£¡£¡£¡£ÈôºòÑ¡³ØÖеĽṹ¶¼À´×Ժÿռä£¬£¬£¬£¬£¬£¬Í¨¹ýʹÓúòÑ¡³Ø£¬£¬£¬£¬£¬£¬¶¨Àí1ÖкÃÍøÂçµÄ²ÉÑù¸ÅÂÊqÌáÉýΪ£º

Òò´Ë£¬£¬£¬£¬£¬£¬²ÉÑù10¸ö½á¹¹£¬£¬£¬£¬£¬£¬ÖÁÉÙÓÐ5¸öºÃ½á¹¹µÄ¸ÅÂÊÓÉ88.38%ÌáÉýÖÁ99.36% ¡£¡£¡£¡£¡£¡£¡£¡£

»ùÓÚºòÑ¡³ØµÄearly stoppingÕ½ÂÔ

ºòÑ¡³ØµÄ¸üÐÂΪsupernetµÄѵÁ·ÇéÐÎÌṩÁËÒ»¸öºÜºÃµÄ²Î¿¼¡£¡£¡£¡£¡£¡£¡£¡£ÈôºòÑ¡³Ø±¬·¢µÄ¸üУ¨Ë³Ðò±ä»»¡¢ÊÕÖ§£©½ÏÉÙ£¬£¬£¬£¬£¬£¬¿ÉÒÔÒÔΪ³¬ÍøÂçÖнϺõÄÍøÂçά³Ö×ÅÒ»¸öÏà¶ÔÎȹ̵ÄÅÅÐò£¬£¬£¬£¬£¬£¬ÕâҲ˵Ã÷´ËʱµÄsupernetÒѾ­ÊÇÒ»¸ö½ÏºÃµÄÐÔÄÜÆÀ¹ÀÆ÷£¬£¬£¬£¬£¬£¬Òò´ËѵÁ·Àú³Ì¿ÉÒÔÌáÇ°¿¢ÊÂ£¬£¬£¬£¬£¬£¬¶ø²»ÐèҪѵÁ·ÖÁÍøÂçÍêÈ«ÊÕÁ²¡£¡£¡£¡£¡£¡£¡£¡£

ÔÚÏÖʵʹÓÃÖÐ£¬£¬£¬£¬£¬£¬ÎÒÃÇ»á½ÏÁ¿Ä¿½ñºòÑ¡³ØPÓëtÂÖµü´úÇ°µÄºòÑ¡³ØP_tµÄ²î±ð¶È£¬£¬£¬£¬£¬£¬Èô²î±ð¶ÈµÍÓÚij¸öÊýÖµ£¨¹þ¹þ(haha)ÌåÓýʵÑéʹÓÃ0.08£©£¬£¬£¬£¬£¬£¬ÑµÁ·×èÖ¹¡£¡£¡£¡£¡£¡£¡£¡£²î±ð¶ÈµÄ½ç˵ÈçÏ£º

»ùÓÚºòÑ¡³ØµÄËÑË÷Õ½ÂÔ

SupernetѵÁ·¿¢Êºó£¬£¬£¬£¬£¬£¬ÎÒÃÇ¿ÉÒÔʹÓÃÑéÖ¤¼¯µÄACCÆÀ¹À½á¹¹µÄÓÅÁÓ¡£¡£¡£¡£¡£¡£¡£¡£±¾ÎÄʹÓà NSGA-II ½ø»¯Ëã·¨[3] ¾ÙÐнṹËÑË÷¡£¡£¡£¡£¡£¡£¡£¡£ÎÒÃÇÔÚ½ø»¯Ëã·¨ÖÐʹÓúòÑ¡³ØÖеĽṹ¾ÙÐÐpopulationµÄ³õʼ»¯£¬£¬£¬£¬£¬£¬Ïà½ÏÓÚËæ»ú³õʼ»¯£¬£¬£¬£¬£¬£¬½èÖúÓÚºòÑ¡³ØÄܹ»Ê¹½ø»¯Ëã·¨ÓÐÒ»¸ö¸üºÃµÄ³õʼ£¬£¬£¬£¬£¬£¬ÌáÉýËÑË÷ЧÂʼ°×îÖյľ«¶È¡£¡£¡£¡£¡£¡£¡£¡£

Èç Figure 3 Ëùʾ£¬£¬£¬£¬£¬£¬ÎÒÃÇÔÚͳһ¸öѵÁ·ºÃµÄsupernetÉÏʹÓÃÁËËæ»ú³õʼ»¯ÓëºòÑ¡³Ø³õʼ»¯Á½ÖÖ·½·¨¾ÙÐÐËÑË÷£¬£¬£¬£¬£¬£¬Ê¹ÓúòÑ¡³Ø³õʼ»¯ËÑË÷µ½µÄ½á¹¹µÄ׼ȷÂÊƽ¾ù»á±ÈËæ»ú³õʼ»¯Òª¸ß¡£¡£¡£¡£¡£¡£¡£¡£

ʵÑéЧ¹û

ΪÁËÓëÏÖÔÚµÄone-shotÒªÁì¾ÙÐбÈÕÕ£¬£¬£¬£¬£¬£¬ÎÒÃÇÊ×ÏÈÔÚÓë ProxylessNAS[6] Ò»ÑùµÄ MobileNetV2 ËÑË÷¿Õ¼äÉϾÙÐнṹËÑË÷£¬£¬£¬£¬£¬£¬Ð§¹û¼û Table 1 ¡£¡£¡£¡£¡£¡£¡£¡£

ͬʱΪÁ˽øÒ»²½ÌáÉýÍøÂçÐÔÄÜ£¬£¬£¬£¬£¬£¬ÎÒÃÇÔÚ¼ÓÈëÁËSEµÄ¸ü´óËÑË÷¿Õ¼äÉϾÙÐÐËÑË÷£¬£¬£¬£¬£¬£¬Ð§¹û¼û Table 2 ¡£¡£¡£¡£¡£¡£¡£¡£

Ablation Study

¶à·¾¶¾Ü¾øʽ²ÉÑùÖеÄ·¾¶ÆÀ¹ÀÏà¹ØÐÔ

ÎÒÃǶÔËæ»ú³õʼ»¯¡¢uniform sampling¡¢greedyÒªÁìѵÁ·»ñµÃµÄsupernetϵÄСÑéÖ¤¼¯Ö¸±êÓëÍêÕûÑéÖ¤¼¯ACCÏà¹ØÐÔ¾ÙÐÐÁËÆÀ¹À£¬£¬£¬£¬£¬£¬Èç Table 3 Ëùʾ¡£¡£¡£¡£¡£¡£¡£¡£¿£¿£¿£¿£¿£¿É¼ûÔÚСÑéÖ¤¼¯ÉÏʹÓÃlossÏà±ÈACC»á»ñµÃ¸ü¸ßµÄÏà¹ØÐÔ£¬£¬£¬£¬£¬£¬¹þ¹þ(haha)ÌåÓýÌ°ÐÄÒªÁìѵÁ·³öµÄsupernetÓÐןüºÃµÄÏà¹ØÐÔ¡£¡£¡£¡£¡£¡£¡£¡£

ÎÒÃǶԲî±ð¾ÞϸСÑéÖ¤¼¯ÓëÍêÕûÑéÖ¤¼¯µÄÏà¹ØÐÔ¾ÙÐÐÁËÆÀ¹À£¨Figure 4 ×óͼ£©£¬£¬£¬£¬£¬£¬Í¬Ê±¶Ôuniform-samplingËã·¨ÔÚ²î±ðµü´úÂÖÊýϵÄÏà¹ØÐÔ×÷ÁËÆÀ¹À£¨Figure 4 ÓÒͼ£©¡£¡£¡£¡£¡£¡£¡£¡£¿£¿£¿£¿£¿£¿ÉÒÔ¿´³ö¹þ¹þ(haha)ÌåÓýËã·¨ÔÚ½ÏСµÄÑéÖ¤¼¯¾ÞϸÉÏÈÔÄܼá³Ö½Ï¸ßµÄÏà¹ØÐÔ£¬£¬£¬£¬£¬£¬ÔÚ¾«¶ÈÓëЧÂʵÄȨºâÏÂ£¬£¬£¬£¬£¬£¬ÎÒÃÇ×îÖÕÑ¡È¡ 1000 ×÷ΪСÑéÖ¤¼¯¾Þϸ¡£¡£¡£¡£¡£¡£¡£¡£

¶Ô¶à·¾¶Ì°ÐIJÉÑù¼°ºòÑ¡³ØЧ¹ûµÄÆÀ¹À

ÎÒÃÇÔÚMobileNetV2µÄsearch spaceÏÂÆÀ¹ÀÁ˶à·¾¶Ì°ÐIJÉÑù¼°ºòÑ¡³ØµÄЧ¹û£¬£¬£¬£¬£¬£¬Èç Table 4 Ëùʾ¡£¡£¡£¡£¡£¡£¡£¡£

×ܽá

³¬ÍøÂçѵÁ·Êǵ¥·ÖÖ§One-Shot NAS ÒªÁìµÄÒªº¦¡£¡£¡£¡£¡£¡£¡£¡£ÓëÏÖÔÚÒªÁìµÄ¶ÔËùÓзÖÖ§Ò»ÊÓͬÈʲî±ð£¬£¬£¬£¬£¬£¬¹þ¹þ(haha)ÌåÓýÒªÁìÌ°ÐĵØ×¢ÖØÓÚÓÐDZÁ¦µÄºÃ·ÖÖ§µÄѵÁ·¡£¡£¡£¡£¡£¡£¡£¡£ÕâÖÖÌ°ÐĵطÖÖ§Â˲¨¿ÉÒÔͨ¹ýÎÒÃÇÌá³öµÄ¶à·ÖÖ§²ÉÑùÕ½ÂÔ±»¸ßЧµØʵÏÖ¡£¡£¡£¡£¡£¡£¡£¡£ÎÒÃÇÌá³öµÄ GreedyNAS ÔÚ׼ȷÂʺÍѵÁ·Ð§ÂÊÉϾùÕ¹ÏÖ³öÁËÏÔÖøµÄÓÅÊÆ