µ¼¶Á£ºÔÚCVPR 2020ÉÏ£¬£¬£¬£¬£¬ÉÌÌÀÑо¿ÔºÁ´½ÓÓë±àÒë×éºÍ±±¾©º½¿Õº½Ìì´óѧÁõÏéÁúÏÈÉúÍŶÓÌá³öÁËÒ»ÖÖÖ¼ÔÚÓÅ»¯Ç°ºóÏòÈö²¥ÖÐÐÅÏ¢Á÷µÄÊÊÓᢸßЧµÄÍøÂç¶þÖµ»¯ÐÂËã·¨IR-Net¡£¡£¡£¡£¡£¡£²î±ðÓÚÒÔÍù¶þÖµÉñ¾­ÍøÂç´ó¶à¹Ø×¢Á¿»¯Îó²î·½Ãæ£¬£¬£¬£¬£¬±¾ÎÄÊ״δÓͳһÐÅÏ¢µÄ½Ç¶ÈÑо¿Á˶þÖµÍøÂçµÄÇ°ÏòºÍºóÏòÈö²¥Àú³Ì£¬£¬£¬£¬£¬ÎªÍøÂç¶þÖµ»¯»úÖƵÄÑо¿ÌṩÁËȫеÄÊӽǡ£¡£¡£¡£¡£¡£Í¬Ê±£¬£¬£¬£¬£¬¸ÃÊÂÇéÊ×´ÎÔÚARM×°±¸ÉϾÙÐÐÁËÏȽø¶þÖµ»¯Ë㷨ЧÂÊÑéÖ¤£¬£¬£¬£¬£¬ÏÔʾÁËIR-Net°²ÅÅʱµÄÓÅÒìÐÔÄܺͼ«¸ßµÄÊÊÓÃÐÔ£¬£¬£¬£¬£¬ÓÐÖúÓÚ½â¾ö¹¤Òµ½ç¹Ø×¢µÄÉñ¾­ÍøÂç¶þÖµ»¯ÂäµØµÄ½¹µãÎÊÌâ¡£¡£¡£¡£¡£¡£

ÄîÍ·

¶þÖµÉñ¾­ÍøÂçÒòÆä´æ´¢Á¿Ð¡¡¢ÍÆÀíЧÂʸ߶øÊܵ½Éç»áµÄÆÕ±é¹Ø×¢ [1]¡£¡£¡£¡£¡£¡£È»¶øÓëÈ«¾«¶ÈµÄ¶ÔÓ¦ÒªÁìÏà±È£¬£¬£¬£¬£¬ÏÖÓеÄÁ¿»¯ÒªÁìµÄ¾«¶ÈÈÔÈ»±£´æÏÔÖøµÄϽµ¡£¡£¡£¡£¡£¡£

¶ÔÉñ¾­ÍøÂçµÄÑо¿Åú×¢£¬£¬£¬£¬£¬ÍøÂçµÄ¶àÑùÐÔÊÇÄ£×ÓµÖ´ï¸ßÐÔÄܵÄÒªº¦[2]£¬£¬£¬£¬£¬¼á³ÖÕâÖÖ¶àÑùÐÔµÄÒªº¦ÊÇ£º(1) ÍøÂçÔÚÇ°ÏòÈö²¥Àú³ÌÖÐÄܹ»Ð¯´ø×ã¹»µÄÐÅÏ¢£»£»£»£»£»£»£»£»(2) ·´ÏòÈö²¥Àú³ÌÖÐ£¬£¬£¬£¬£¬×¼È·µÄÌݶÈΪÍøÂçÓÅ»¯ÌṩÁË׼ȷµÄÐÅÏ¢¡£¡£¡£¡£¡£¡£¶þÖµÉñ¾­ÍøÂçµÄÐÔÄÜϽµÖ÷ÒªÊÇÓɶþÖµ»¯µÄÓÐÏÞÌåÏÖÄÜÁ¦ºÍÀëÉ¢ÐÔÔì³ÉµÄ£¬£¬£¬£¬£¬Õâµ¼ÖÂÁËÇ°ÏòºÍ·´ÏòÈö²¥µÄÑÏÖØÐÅÏ¢Ëðʧ£¬£¬£¬£¬£¬Ä£×ӵĶàÑùÐÔ¼±¾çϽµ¡£¡£¡£¡£¡£¡£Í¬Ê±£¬£¬£¬£¬£¬ÔÚ¶þÖµÉñ¾­ÍøÂçµÄѵÁ·Àú³ÌÖÐ£¬£¬£¬£¬£¬ÀëÉ¢¶þÖµ»¯ÍùÍùµ¼ÖÂÌݶȽû¾øÈ·ºÍÓÅ»¯Æ«Ïò¹ýʧ¡£¡£¡£¡£¡£¡£ÔõÑù½â¾öÒÔÉÏÎÊÌâ£¬£¬£¬£¬£¬»ñµÃ¸ü¸ß¾«¶ÈµÄ¶þÖµÉñ¾­ÍøÂ磿£¿£¿£¿£¿£¿£¿ÕâÒ»ÎÊÌâ±»Ñо¿ÕßÃÇÆÕ±é¹Ø×¢£¬£¬£¬£¬£¬±¾ÎĵÄÄîÍ·ÔÚÓÚ£ºÍ¨¹ýÐÅÏ¢±£´æµÄ˼Ð÷£¬£¬£¬£¬£¬Éè¼Æ¸ü¸ßÐÔÄܵĶþÖµÉñ¾­ÍøÂç¡£¡£¡£¡£¡£¡£

»ùÓÚÒÔÉÏÄîÍ·£¬£¬£¬£¬£¬±¾ÎÄÊ״δÓÐÅÏ¢Á÷µÄ½Ç¶ÈÑо¿ÁËÍøÂç¶þÖµ»¯£¬£¬£¬£¬£¬Ìá³öÁËÒ»ÖÖеÄÐÅÏ¢¼á³ÖÍøÂ磨IR-Net£©£º(1)ÔÚÇ°ÏòÈö²¥ÖÐÒýÈëÁËÒ»ÖÖ³ÆΪLibra²ÎÊý¶þÖµ»¯£¨Libra-PB£©µÄƽºâ±ê×¼»¯Á¿»¯ÒªÁì£¬£¬£¬£¬£¬×î´ó»¯Á¿»¯²ÎÊýµÄÐÅÏ¢ìغÍ×îС»¯Á¿»¯Îó²î£»£»£»£»£»£»£»£» (2) ÔÚ·´ÏòÈö²¥ÖнÓÄÉÎó²îË¥¼õÔ¤¼ÆÆ÷£¨EDE£©À´ÅÌËãÌݶÈ£¬£¬£¬£¬£¬°ü¹ÜѵÁ·×îÏÈʱµÄ³ä·Ö¸üкÍѵÁ·¿¢ÊÂʱµÄ׼ȷÌݶȡ£¡£¡£¡£¡£¡£

IR-NetÌṩÁËÒ»¸öȫеĽǶÈÀ´Ã÷È·¶þÖµÉñ¾­ÍøÂçÊÇÔõÑùÔËÐеÄ£¬£¬£¬£¬£¬²¢ÇÒ¾ßÓкܺõÄͨÓÃÐÔ£¬£¬£¬£¬£¬¿ÉÒÔÔÚ±ê×¼µÄÍøÂçѵÁ·Á÷³ÌÖоÙÐÐÓÅ»¯¡£¡£¡£¡£¡£¡£×÷ÕßʹÓÃCIFAR-10ºÍImageNetÊý¾Ý¼¯ÉϵÄͼÏñ·ÖÀàʹÃüÀ´ÆÀ¹ÀÌá³öµÄIR-Net£¬£¬£¬£¬£¬Í¬Ê±½èÖú¿ªÔ´¶þÖµ»¯ÍÆÀí¿âdaBNN¾ÙÐÐÁË°²ÅÅЧÂÊÑéÖ¤¡£¡£¡£¡£¡£¡£

ÒªÁìÉè¼Æ

¸ß¾«¶È¶þÖµÉñ¾­ÍøÂçѵÁ·µÄÆ¿¾±Ö÷ÒªÔÚÓÚѵÁ·Àú³ÌÖÐÑÏÖصÄÐÅÏ¢Ëðʧ¡£¡£¡£¡£¡£¡£Ç°Ïòsignº¯ÊýºÍºóÏòÌݶÈÆȽüËùÔì³ÉµÄÐÅÏ¢ËðʧÑÏÖØÓ°ÏìÁ˶þÖµÉñ¾­ÍøÂçµÄ¾«¶È¡£¡£¡£¡£¡£¡£ÎªÏàʶ¾öÒÔÉÏÎÊÌâ£¬£¬£¬£¬£¬±¾ÎÄÌá³öÁËÒ»ÖÖеÄÐÅÏ¢¼á³ÖÍøÂ磨IR-Net£©Ä£×Ó£¬£¬£¬£¬£¬Ëü±£´æÁËѵÁ·Àú³ÌÖеÄÐÅÏ¢£¬£¬£¬£¬£¬ÊµÏÖÁ˶þÖµ»¯Ä£×ӵĸ߾«¶È¡£¡£¡£¡£¡£¡£

Ç°ÏòÈö²¥ÖеÄLibra Parameter Binarization£¨Libra-PB£©

ÔÚ´Ë֮ǰ£¬£¬£¬£¬£¬¾ø´ó´ó¶¼ÍøÂç¶þÖµ»¯ÒªÁìÊÔͼ¼õС¶þÖµ»¯²Ù×÷µÄÁ¿»¯Îó²î¡£¡£¡£¡£¡£¡£È»¶ø£¬£¬£¬£¬£¬½öͨ¹ý×îС»¯Á¿»¯Îó²îÀ´»ñµÃÒ»¸öÓÅÒìµÄ¶þÖµÍøÂçÊDz»·óµÄ¡£¡£¡£¡£¡£¡£Òò´Ë£¬£¬£¬£¬£¬Libra-PBÉè¼ÆµÄÒªº¦ÔÚÓÚ£ºÊ¹ÓÃÐÅÏ¢ìØÖ¸±ê£¬£¬£¬£¬£¬×î´ó»¯¶þÖµÍøÂçÇ°ÏòÈö²¥Àú³ÌÖеÄÐÅÏ¢Á÷¡£¡£¡£¡£¡£¡£

ƾ֤ÐÅÏ¢ìصĽç˵£¬£¬£¬£¬£¬ÔÚ¶þÖµÍøÂçÖÐ£¬£¬£¬£¬£¬¶þÖµ²ÎÊýQx(x)µÄìØ¿ÉÒÔͨ¹ýÒÔϹ«Ê½ÅÌË㣺

ÈôÊÇ´¿´âµØ×·ÇóÁ¿»¯Îó²î×îС»¯£¬£¬£¬£¬£¬ÔÚ¼«¶ËÇéÐÎÏÂ£¬£¬£¬£¬£¬Á¿»¯²ÎÊýµÄÐÅÏ¢ìØÉõÖÁ¿ÉÒÔ¿¿½üÓÚÁã¡£¡£¡£¡£¡£¡£Òò´Ë£¬£¬£¬£¬£¬Libra-PB½«Á¿»¯ÖµµÄÁ¿»¯Îó²îºÍ¶þÖµ²ÎÊýµÄÐÅÏ¢ìØͬʱ×÷ΪÓÅ»¯Ä¿µÄ£¬£¬£¬£¬£¬½ç˵Ϊ:

ÔÚ²®Å¬ÀûÂþÑܼÙÉèÏÂ£¬£¬£¬£¬£¬µ±p=0.5ʱ£¬£¬£¬£¬£¬Á¿»¯ÖµµÄÐÅÏ¢ìØÈ¡×î´óÖµ¡£¡£¡£¡£¡£¡£

Òò´Ë£¬£¬£¬£¬£¬ÔÚLibra-PBͨ¹ý±ê×¼»¯Çå¾²ºâ²Ù×÷»ñµÃ±ê×¼»¯Æ½ºâȨÖØ£¬£¬£¬£¬£¬Èçͼ2Ëùʾ£¬£¬£¬£¬£¬ÔÚBernoulliÂþÑÜÏÂ£¬£¬£¬£¬£¬ÓÉLibra-PBÁ¿»¯µÄ²ÎÊý¾ßÓÐ×î´óµÄÐÅÏ¢ìØ¡£¡£¡£¡£¡£¡£ÓÐȤµÄÊÇ£¬£¬£¬£¬£¬¶ÔȨÖصļòÆӱ任Ҳ¿ÉÒÔ¼«´ó¸ÄÉÆÇ°ÏòÀú³ÌÖ줻îµÄÐÅÏ¢Á÷¡£¡£¡£¡£¡£¡£ÓÉÓÚ´Ëʱ£¬£¬£¬£¬£¬¸÷²ãµÄ¶þÖµ¼¤»îÖµÐÅÏ¢ìØͬÑù¿ÉÒÔ×î´ó»¯£¬£¬£¬£¬£¬ÕâÒâζ×ÅÌØÕ÷ͼÖÐÐÅÏ¢¿ÉÒÔ±»±£´æ¡£¡£¡£¡£¡£¡£

ÔÚÒÔÍùµÄ¶þÖµ»¯ÒªÁìÖÐ£¬£¬£¬£¬£¬ÎªÁËʹÁ¿»¯Îó²î¼õС£¬£¬£¬£¬£¬ÏÕЩËùÓÐÒªÁ춼»áÒýÈ븡µã±ê×¼Òò×ÓÀ´´ÓÊýÖµÉÏÆȽüԭʼ²ÎÊý£¬£¬£¬£¬£¬ÕâÎÞÒɽ«¸ß°ºµÄ¸¡µãÔËËãÒýÈëÆäÖС£¡£¡£¡£¡£¡£ÔÚLibra-PBÖÐ£¬£¬£¬£¬£¬ÎªÁ˽øÒ»²½¼õСÁ¿»¯Îó²î£¬£¬£¬£¬£¬Í¬Ê±×èÖ¹ÒÔÍù¶þÖµ»¯ÒªÁìÖмÛÇ®¸ß°ºµÄ¸¡µãÔËËã£¬£¬£¬£¬£¬Libra-PBÒýÈëÁËÕûÊýÒÆλ±êÁ¿s£¬£¬£¬£¬£¬À©Õ¹Á˶þֵȨÖصÄÌåÏÖÄÜÁ¦¡£¡£¡£¡£¡£¡£

Òò´Ë×îÖÕ£¬£¬£¬£¬£¬Õë¶ÔÕýÏòÈö²¥µÄLibra²ÎÊý¶þÖµ»¯¿ÉÒÔÌåÏÖÈçÏ£º

IR-NetµÄÖ÷ÒªÔËËã²Ù×÷¿ÉÒÔÌåÏÖΪ£º

·´ÏòÈö²¥ÖеÄError Decay Estimator£¨EDE£©

ÓÉÓÚ¶þÖµ»¯µÄ²»Ò»Á¬ÐÔ£¬£¬£¬£¬£¬ÌݶȵĽüËƹØÓÚ·´ÏòÈö²¥ÊDz»¿É×èÖ¹µÄ£¬£¬£¬£¬£¬ÕâÖÖ¶Ôsignº¯ÊýµÄ½üËÆ´øÀ´ÁËÁ½ÖÖÌݶȵÄÐÅÏ¢Ëðʧ£¬£¬£¬£¬£¬°üÀ¨½Ø¶Ï¹æÄ£Íâ²ÎÊý¸üÐÂÄÜÁ¦Ï½µÔì³ÉµÄÐÅÏ¢Ëðʧ£¬£¬£¬£¬£¬ºÍ½Ø¶Ï¹æÄ£ÄÚ½üËÆÎó²îÔì³ÉµÄÐÅÏ¢Ëðʧ¡£¡£¡£¡£¡£¡£ÎªÁ˸üºÃµÄ±£´æ·´ÏòÈö²¥ÖÐÓÉËðʧº¯Êýµ¼³öµÄÐÅÏ¢£¬£¬£¬£¬£¬Æ½ºâ¸÷ѵÁ·½×¶Î¹ØÓÚÌݶȵÄÒªÇó£¬£¬£¬£¬£¬EDEÒýÈëÁËÒ»ÖÖ½¥½øµÄÁ½½×¶Î½üËÆÌݶÈÒªÁì¡£¡£¡£¡£¡£¡£

µÚÒ»½×¶Î£º±£´æ·´ÏòÈö²¥Ëã·¨µÄ¸üÐÂÄÜÁ¦¡£¡£¡£¡£¡£¡£½«ÌݶÈÔ¤¼Æº¯ÊýµÄµ¼ÊýÖµ¼á³ÖÔÚ¿¿½ü1µÄˮƽ£¬£¬£¬£¬£¬È»ºóÖð²½½«½Ø¶ÏÖµ´ÓÒ»¸ö´óµÄÊý×Ö½µµ½1¡£¡£¡£¡£¡£¡£Ê¹ÓÃÕâÒ»¹æÔò£¬£¬£¬£¬£¬½üËƺ¯Êý´Ó¿¿½üIdentityº¯ÊýÑÝ»¯µ½Clipº¯Êý£¬£¬£¬£¬£¬´Ó¶ø°ü¹ÜÁËѵÁ·ÔçÆڵĸüÐÂÄÜÁ¦¡£¡£¡£¡£¡£¡£

µÚ¶þ½×¶Î£ºÊ¹0ÖÜΧµÄ²ÎÊý±»¸ü׼ȷµØ¸üС£¡£¡£¡£¡£¡£½«½Ø¶Ï¼á³ÖΪ1£¬£¬£¬£¬£¬²¢Öð½¥½«µ¼ÊýÇúÏßÑݱ䵽õ辶º¯ÊýµÄÐÎ×´¡£¡£¡£¡£¡£¡£Ê¹ÓÃÕâÒ»¹æÔò£¬£¬£¬£¬£¬½üËƺ¯Êý´ÓClipº¯ÊýÑݱ䵽signº¯Êý£¬£¬£¬£¬£¬´Ó¶ø°ü¹ÜÁËÇ°ÏòºÍ·´ÏòÈö²¥µÄÒ»ÖÂÐÔ¡£¡£¡£¡£¡£¡£

¸÷½×¶ÎEDEµÄÐÎ״ת±äÈçͼ3(c)Ëùʾ¡£¡£¡£¡£¡£¡£Í¨¹ý¸ÃÉè¼Æ£¬£¬£¬£¬£¬EDE¼õСÁËÇ°Ïò¶þÖµ»¯º¯ÊýºÍºóÏò½üËƺ¯ÊýÖ®¼äµÄ²î±ð£¬£¬£¬£¬£¬Í¬Ê±ËùÓвÎÊý¶¼ÄÜ»ñµÃºÏÀíµÄ¸üС£¡£¡£¡£¡£¡£

ʵÑéЧ¹û

×÷ÕßʹÓÃÁËÁ½¸ö»ù×¼Êý¾Ý¼¯£ºCIFAR-10ºÍImageNet£¨ILSVRC12£©¾ÙÐÐÁËʵÑé¡£¡£¡£¡£¡£¡£ÔÚÁ½¸öÊý¾Ý¼¯ÉϵÄʵÑéЧ¹ûÅú×¢£¬£¬£¬£¬£¬IR-Net±ÈÏÖÓеÄ×îÏȽøÒªÁì¸ü¾ß¾ºÕùÁ¦¡£¡£¡£¡£¡£¡£

Deployment Efficiency

ΪÁ˽øÒ»²½ÑéÖ¤IR-NetÔÚÏÖʵÒƶ¯×°±¸ÖеݲÅÅЧÂÊ£¬£¬£¬£¬£¬×÷ÕßÔÚ1.2GHz 64λËĺËARM Cortex-A53µÄRaspberry Pi 3BÉϽøÒ»²½ÊµÏÖÁËIR-Net£¬£¬£¬£¬£¬²¢ÔÚÏÖʵӦÓÃÖвâÊÔÁËÆäÕæʵËÙÂÊ¡£¡£¡£¡£¡£¡£±í5ÏÔʾ£¬£¬£¬£¬£¬IR-NetµÄÍÆÀíËÙÂÊÒª¿ìµÃ¶à£¬£¬£¬£¬£¬Ä£×ӳߴçÒ²´ó´ó¼õС£¬£¬£¬£¬£¬²¢ÇÒIR-NetÖеÄλÒƲÙ×÷ÏÕЩ²»»á´øÀ´ÌØÁíÍâÍÆÀíʱ¼äºÍ´æ´¢ÏûºÄ¡£¡£¡£¡£¡£¡£

´«ËÍÃÅ

IR-Net´úÂëÏÖÔÚÒѾ­¿ªÔ´£¬£¬£¬£¬£¬½Ó´ýÁÐλͬÑâʹÓúͽ»Á÷¡£¡£¡£¡£¡£¡£

ÂÛÎĵص㣺https://arxiv.org/abs/1909.10788

ÏîÄ¿µØµã£ºhttps://github.com/htqin/IR-Net

ÂÛÎĽâ˵£º½üÆÚ½«ÓÐÏßÉÏÖ±²¥µÄÂÛÎĽâ˵¼°×¨Ìâ½»Á÷¿Î³Ì£¬£¬£¬£¬£¬½Ó´ýµã»÷Ï·½¡°ÉÌÌÀ̩̹¹ûÕæ¿Î¡±¿Î³ÌרÌâÁ´½Ó¾ÙÐб¨Ãû

±¨ÃûÁ´½Ó£ºhttps://jinshuju.net/f/RKOLf9

ÕÐƸ

ÉÌÌÀÑо¿Ôº-Á´½ÓÓë±àÒëÍŶÓÖÂÁ¦ÓÚͨ¹ýSystem+AIÊÖÒÕ´òÔ춥¼âµÄÉî¶Èѧϰ½¹µãÒýÇæ¡£¡£¡£¡£¡£¡£¿£¿£¿£¿£¿£¿£¿ª·¢µÄÄ£×ÓѵÁ·ºÍÄ£×Ó°²ÅŹ¤¾ßÁ´ÒÑЧÀÍÓÚ¹«Ë¾¶à¸ö½¹µãÓªÒµ¡£¡£¡£¡£¡£¡£ÍŶÓÔÚÁ¿»¯Ä£×ÓµÄÔÚÏß/ÀëÏßÉú²ú¡¢°²ÅŶÔÆë¡¢¼ÓËÙѵÁ·µÈÆ«ÏòÓÐ×ÅÃ÷È·µÄÊÖÒÕÍýÏë¡£¡£¡£¡£¡£¡£

ÍŶÓÏÖÔÚÕýÔÚÕÐƸģ×ÓÁ¿»¯Æ«ÏòµÄϵͳËã·¨Ñо¿Ô±/ʵϰÉú£¬£¬£¬£¬£¬½Ó´ý¸ÐÐËȤµÄСͬ°éͶµÝ¼òÀúµ½yufengwei@sensetime.com¡£¡£¡£¡£¡£¡£

Reference

[1] Rastegari M, Ordonez V, Redmon J, et al. Xnor-net: Imagenet classification using binary convolutional neural networks[C]//ECCV. Springer, Cham, 2016: 525-542.

[2] Xie B, Liang Y, Song L. Diverse neural network learns true target functions[J]. arXiv preprint arXiv:1611.03131, 2016.