ͼÉñ¾ÍøÂç·¢Õ¹µ½Ê²Ã´³Ì¶ÈÁË£¿ÏÖÔÚÎÒÃÇÓÐÁËרÓÃµÄ Benchmark ¹¤¾ßÀ´½øÐÐÆÀ²â¡£
½üÆڵĴóÁ¿Ñо¿ÒѾÈÃÎÒÃÇ¿´µ½ÁËͼÉñ¾ÍøÂçÄ£ÐÍ£¨GNN£©µÄÇ¿´óDZÁ¦£¬ºÜ¶àÑо¿ÍŶӶ¼ÔÚ²»¶Ï¸Ä½øºÍ¹¹½¨»ù´¡Ä£¿é¡£µ«´ó¶àÊýÑо¿Ê¹ÓõÄÊý¾Ý¼¯¶¼ºÜС£¬Èç Cora ºÍ TU¡£ÔÚÕâÖÖÇé¿öÏ£¬¼´Ê¹ÊÇ·ÇͼÉñ¾ÍøÂçµÄÐÔÄÜÒ²Êǿɹ۵ġ£Èç¹û½øÐнøÒ»²½µÄ±È½Ï£¬Ê¹ÓÃÖеȴóСµÄÊý¾Ý¼¯£¬Í¼Éñ¾ÍøÂçµÄÓÅÊƲÅÄÜÏÔÏÖ³öÀ´¡£
ÔÚ˹̹¸£Í¼Éñ¾ÍøÂç´óÅ£ Jure µÈÈË·¢²¼¡¶Open Graph Benchmark¡·Ö®ºó£¬ÓÖÒ»¸öÖ¼ÔÚ¹¹½¨¡¸Í¼Éñ¾ÍøÂçµÄ ImageNet¡¹µÄÑо¿³öÏÖÁË¡£½üÈÕ£¬À´×ÔÄÏÑóÀí¹¤´óѧ¡¢ÂåÔ¼ÀÂíÀûÃÉÌØ´óѧ¡¢ÃÉÌØÀû¶û´óѧºÍ MILA µÈ»ú¹¹µÄÂÛÎı»Ìá½»µ½ÁËÂÛÎÄÔ¤Ó¡°æƽ̨ÉÏ£¬¶øÇÒÕâһеÄÑо¿ÓÐÉî¶ÈѧϰÏÈÇý Yoshua Bengio µÄ²ÎÓ룬ҲµÃµ½ÁË Yann LeCun µÄ¹Ø×¢¡£
ÂÛÎÄÁ´½Ó£ºhttps://arxiv.org/abs/2003.00982
ÔÚ¸ÃÑо¿ÖУ¬×÷ÕßÒ»´ÎÒýÈëÁËÁù¸öÖеȴóСµÄ»ù×¼Êý¾Ý¼¯£¨12k-70k ͼ£¬8-500 ½Úµã£©£¬²¢¶ÔһЩÓдú±íÐÔµÄͼÉñ¾ÍøÂç½øÐÐÁ˲âÊÔ¡£³ýÁËÖ»ÓýڵãÌØÕ÷µÄ»ù×¼ÏßÄ£ÐÍÖ®Í⣬ͼÉñ¾ÍøÂç·Ö³É´ø»ò²»´ø¶Ô±ß¶Ô×¢ÒâÁ¦Á½´óÀà¡£GNN Ñо¿ÉçÇøÒ»Ö±ÔÚÑ°ÇóÒ»¸ö¹²Í¬µÄ»ù×¼ÒÔ¶ÔÐÂÄ£Ð͵ÄÄÜÁ¦½øÐÐÆÀ²â£¬ÕâÒ»¹¤¾ß»òÐí¿ÉÒÔÈÃÎÒÃÇʵÏÖÄ¿±ê¡£
ÏÖÔÚ£¬Èç¹ûÄãÏë²âÊÔÒ»ÏÂ×Ô¼ºµÄͼÉñ¾ÍøÂçÄ£ÐÍ£¬¿ÉÒÔʹÓÃËüµÄ¿ªÔ´ÏîÄ¿½øÐвâÊÔÁË¡£
ÏîÄ¿µØÖ·£ºhttps://github.com/graphdeeplearning/benchmarking-gnns
²»Í¬ÈÎÎñÉϵIJâÊԽű¾£¬Ã¿Ò»¸ö Notebook ¶¼»áÊÖ°ÑÊÖ½ÌÄãÈçºÎ²âÊÔ²»Í¬µÄͼÉñ¾ÍøÂç¡£
¸Ã¿ª·Å»ù×¼¼Ü¹¹»ùÓÚ DGL ¿â£¬DGL ÓÉ AWS ÉϺ£ AI Ñо¿Ôº¡¢Å¦Ô¼´óѧ¡¢ÉϺ£Å¦Ô¼´óѧ¿ª·ÅºÍά»¤£¬ÊÇÒµ½çÁìÏȵÄͼÉñ¾ÍøÂçѵÁ·Æ½Ì¨£¬²¢ÎÞ·ìÖ§³ÖÖ÷Á÷Éî¶ÈÍøÂçƽ̨¡£Benchmarking gnn ½¨Á¢ÔÚ DGL µÄ PyTorch °æ±¾Ö®ÉÏ¡£
AWS ÉϺ£ AI Ñо¿ÔºÊ×ÈÎÔº³¤¡¢ÉϺ£Å¦Ô¼´óѧÕÅ´ÌÊÚÆÀÂÛ˵£º¡¸ÕâƪÂÛÎÄÀ´µÃºÜ¼°Ê±£¬Ò²ÓÐÒâÒå¡£µÚÒ»£¬ËµÃ÷ÏÖÓеÄÊý¾Ý¼¯Ì«Ð¡¡¢ÒÔÖ³ÉΪǰ½øµÄÕÏ°£¬ÒѾ³ÉΪѧ½çµÄ¹²Ê¶¡£ÖµµÃÔÞÑïµÄÊÇÕâƪÎÄÕµÄ×÷Õß²¢Ã»ÓÐÒòΪ OGB µÄ·¢²¼¾Í¸éϲ»ÅªÁË¡£ÔÚÎÒ¿´À´£¬ËûÃǵÄÊý¾Ý¼¯ºÍ OGB ÓкÜÇ¿µÄ»¥²¹ÐÔ£¬³ÊÏÖÁËͼÉñ¾ÍøÂç¸ü·á¸»µÄÓ¦Óó¡¾°£¬±ÈÈç°ÑͼÏñÊý¾Ýת»»³ÉͼÊý¾Ý£¬ËäÈ»ÊÇ´Ó MNIST ºÍ CIFAR ¿ªÊ¼£¬Ò²Òþº¬Á˵߸²»ò¸Ä±ä»ùÓÚ¾í»ýÍøÂç CNN µÄ½â¾ö·½°¸£¬ÔÙ±ÈÈçÂÃÐÐÍÆÏúÔ±ÎÊÌâÊÇÒ»¸ö¾µäµÄÓÅ»¯ÎÊÌ⣬µÈµÈ¡£¡¹
¡¸ÁíÍ⣬»ùÓÚÕâһϵÁеÄÊý¾ÝµÃµ½µÄ½áÂÛÓбȽϸߵĿÉÐŶȣ¬±ÈÈçÊý¾Ý¶àÆðÀ´Í¼Éñ¾ÍøÂç¸üÄÜ·¢»ÓÓÅÊÆ£¬±ÈÈç´ø×¢ÒâÁ¦µÄͼÉñ¾ÍøÂçËäÈ»²ÎÊý¸ü¶à£¬µ«ÐÔÄÜÒ²¸üºÃ¡£×ÜÖ®£¬ÕâЩ½á¹û¶Ô¼¤Àø¸ü¶àµÄÄ£ÐÍÑо¿ºÍÍØÕ¹Ó¦Óó¡¾°·Ç³£ÓÐÒâÒå¡£¡¹ÕŽÌÊÚ˵
ͼÉñ¾ÍøÂçÒѳÉΪ·ÖÎöºÍѧϰͼÐÎÊý¾ÝµÄ±ê×¼¹¤¾ß£¬²¢Òѳɹ¦µØÓ¦ÓÃÔںܶàÁìÓòÖУ¬°üÀ¨»¯Ñ§¡¢ÎïÀí¡¢Éç»á¿Æѧ¡¢ÖªÊ¶Í¼Æס¢ÍƼöϵͳÒÔ¼°Éñ¾¿ÆѧµÈ¡£Ëæן÷ÁìÓòµÄ·¢Õ¹£¬È·¶¨¼Ü¹¹ÀàÐÍÒÔ¼°¹Ø¼üµÄ»úÖÆÏÔµÃÓÈΪÖØÒª£¬ÕâЩ¼Ü¹¹Óë»úÖÆ¿ÉÒÔÔÚ¿çͼÐδóСµÄÇé¿öϽøÐзº»¯£¬Ê¹µÃÎÒÃÇÄܹ»´¦Àí¸ü¶à¸ü´ó¸ü¸´ÔÓµÄÊý¾Ý¼¯ÒÔ¼°ÁìÓò¡£
µ«ÊÇ£¬ÔÚȱ·¦¾ßÓÐÒ»ÖÂÐÔµÄʵÑéÉèÖúʹóÁ¿Êý¾Ý¼¯Ã»Óбê×¼»¯»ù×¼µÄÇé¿öÏ£¬ºâÁ¿Ð嵀 GNN ÓÐЧÐÔÒÔ¼°¶Ô±ÈÄ£ÐͱäµÃÔ½À´Ô½À§ÄÑ¡£ÔÚ±¾ÂÛÎÄÖУ¬×÷ÕßÌá³öÁËÒ»¸ö¿É¸´ÖÆ»¯µÄ GNN »ù×¼²âÊÔ¿ò¼Ü£¬¿ÉÒÔÈÃÑо¿ÈËÔ±·½±ãµØÌí¼ÓеÄÊý¾Ý¼¯ÒÔ¼°Ä£ÐÍ¡£´ÓÊýѧ½¨Ä£¡¢¼ÆËã»úÊÓ¾õ¡¢»¯Ñ§ºÍ×éºÏÎÊÌâµÈ¶à·½Ã潫ÕâÒ»»ù×¼¿ò¼ÜÓ¦ÓÃÖÁ×îеÄÖг߶ÈͼÐÎÊý¾Ý¼¯ÀÒÔ±ãÓÚÔÚÉè¼ÆÓÐЧµÄ GNN ʱ½¨Á¢Æð¹Ø¼üµÄ²Ù×÷¡£¸ü׼ȷµÄÀ´Ëµ£¬Í¼¾í»ý¡¢¸÷ÏîÒìÐÔÀ©É¢¡¢²Ð²îÁ¬½Ó¡¢¹éÒ»»¯²ãÊÇ¿ª·¢Â³°ôÐÔÒÔ¼°¿ÉÀ©Õ¹ÐÔ GNN µÄͨÓù¹¼þ¡£
ÕâÏ×÷µÄÄ¿±êÖ®Ò»ÊÇÌṩһ¸öÒ×ÓÚʹÓõÄÖеȹæÄ£Êý¾Ý¼¯£¬ÔÚÕâЩÊý¾Ý¼¯ÉÏ£¬ÃæÏò¹ýÈ¥¼¸ÄêÖÐËùÌá³öµÄ²»Í¬ GNN ¼Ü¹¹ÔÚÐÔÄܱíÏÖÉÏÓÐÃ÷ÏԵIJîÒ졣ͬʱ£¬ÕâЩ²îÒì´Óͳ¼ÆµÄ½Ç¶ÈÉÏÀ´ËµÊǾßÓÐÏ൱µÄÒâÒ壬¸Ã»ù×¼°üº¬ 6 ¸öÊý¾Ý¼¯£¬Èç±í 1£º
ÌáÒé»ù×¼Êý¾Ý¼¯µÄ»ã×Üͳ¼ÆÐÅÏ¢¡£
¶ÔÓÚÕâÁ½¸ö¼ÆËã»úÊÓ¾õÊý¾Ý¼¯£¬À´×Ô¾µäµÄ MNIST (LeCun et al., 1998) ÒÔ¼° CIFAR10 (Krizhevsky et al., 2009) Êý¾Ý¼¯ÖеÄÿ¸öͼÏñ¶¼Ê¹ÓÃÁËËùνµÄ³¬ÏñËØת»»³Éͼ¡£
¶ø½ÓÏÂÀ´µÄÈÎÎñÊǽ«ÕâЩͼÐηÖÀà¡£ÔÚ PATTERN ºÍ CLUSTER Êý¾Ý¼¯ÖУ¬Í¼ÐÎÊǸù¾ÝËæ»ú¿éÄ£ÐÍÉú³ÉµÄ¡£ÕâЩÈÎÎñ°üÀ¨Ê¶±ðÌض¨µÄ×Óͼ½á¹¹ (PATTERN Êý¾Ý¼¯) »òÕßʶ±ð¼¯Èº ( CLUSTER Êý¾Ý¼¯)¡£ÕâЩ¶¼ÊôÓÚÊǽڵã·ÖÀàÈÎÎñ¡£
Tsp Êý¾Ý¼¯ÊÇ»ùÓÚÏúÊÛÈËÔ±ÂÃÐеÄÎÊÌâ (¼ÙÉè¸ø¶¨Ò»¸ö³ÇÊÐÁÐ±í£¬·ÃÎÊÿ¸ö³ÇÊв¢·µ»ØÔʼ³ÇÊеÄ×î¶Ì·¾¶ÊÇʲô?)
½«Ëæ»úÅ·ÊÏͼÉ쵀 TSP ÎÊÌâ×÷Ϊһ¸ö±ß½ç·ÖÀà»òÊÇÁ¬½ÓÔ¤²âµÄÈÎÎñ¿´´ý£¬ÆäÖÐ Concorde Solver ¸ø³öµÄ TSP ÂÃÐÐÖÐÿһ±ß½çµÄÕæʵÇé¿öÖµ¶¼ÊôÓÚÊÇÔÚÏÖʵÊÀ½çÖÐÒÑ´æÔڵķÖ×ÓÊý¾Ý¼¯¡£Ã¿¸ö·Ö×ӿɱ»×ª»»³ÉÒ»¸öͼÐÎ: ÆäÖÐÿ¸öÔ×ӿɳÉΪһ¸ö½Úµã£¬Ã¿¸ö¼ü¿É³ÉΪһ¸ö±ß¡£
GatedGCN-ÃÅ¿Øͼ¾í»ýÍøÂç (Bresson & Laurent£¬2017) ÊÇ¿¼ÂÇÖеÄ×îºóÒ»¸ö GNN¡£Èç¹ûÔÚÊý¾Ý¼¯ÖпÉÓõÄÇé¿öÏ£¬ÆäÖÐ GatedGCN-e ±íʾʹÓñßÔµÊôÐÔ/ÌØÕ÷µÄ°æ±¾¡£ÁíÍ⣬×÷ÕßҲʵÏÖÁËÒ»¸ö¼òµ¥µÄ²»Ê¹ÓÃͼ½á¹¹µÄ»ùÏßÄ£ÐÍ£¬Ëü´¦ÓÚ²¢ÐÐÇé¿ö϶Ôÿ¸ö½ÚµãµÄÌØÕ÷ÏòÁ¿Ê¹ÓÃÒ»¸ö MLP£¬ÇÒ¶ÀÁ¢ÓÚÆäËû½Úµã¡£
ÕâÊǺóÐø¿ÉÑ¡µÄÒ»¸öÃÅ¿Ø»úÖÆ£¬ÓÃÒÔÒÔ»ñµÃÃÅ¿Ø MLP »ùÏß (ÏêÇé¼û²¹³ä²ÄÁÏ)¡£×÷Õß¶Ô MNIST£¬CIFAR10£¬ZINC ÒÔ¼° TSP ÔÚ Nvidia 1080Ti GPU ÉϽøÐÐʵÑ飬¶Ô PATTERN ºÍ CLUSTER ÔÚ Nvidia 2080Ti GPU ÉϽøÐÐʵÑé¡£
ͼ·ÖÀàºÍ³¬ÏñËØÊý¾Ý¼¯
ÕâÒ»²¿·ÖʹÓÃÁ˼ÆËã»úÊÓ¾õÁìÓòÀï×îÁ÷ÐÐµÄ MNIST ºÍ CIFAR10 ͼÏñ·ÖÀàÊý¾Ý¼¯¡£³¬·Ö±æÂʸñʽΪ SLIC£¨Knyazev et al., 2019£©¡£MNIST ÓµÓÐ 55000 ѵÁ·/5000 ÑéÖ¤/10000 ²âÊÔͼ£¬½ÚµãΪ 40-75 Ö®¼ä£¨¼´³¬ÏñËصÄÊýÁ¿£©£¬CI-FAR10 ÓÐ 45000 ѵÁ·/5000 ÑéÖ¤/10000 ²âÊÔͼ£¬½ÚµãÊýΪ 85-150¡£
ͼ 1. ʾÀýͼºÍ³¬ÏñËØͼ¡£SLIC µÄ³¬ÏñËØͼ£¨ÆäÖÐ MNIST ×î¶à 75 ½Úµã£¬CIFAR10 ×î¶à 150 ½Úµã£©ÊÇÅ·¼¸ÀïµÃ¿Õ¼äÖÐµÄ 8 ¸ö×î½üÁÚͼÐΣ¬½ÚµãÑÕÉ«±íʾƽ¾ùÏñËØÇ¿¶È¡£
±í 3. ²»Í¬·½·¨ÔÚ»ùÓÚ MNIST ºÍ CI-FAR10 µÄ±ê×¼²âÊÔ¼¯ÉϵIJâÊÔ½á¹û£¨ÊýÖµÔ½¸ßÔ½ºÃ£©¡£¸Ã½á¹ûÊÇʹÓà 4 ¸ö²»Í¬ÖÖ×ÓÔËÐÐËĴνá¹ûµÄƽ¾ùÖµ¡£ºìɫΪ×î¼Ñˮƽ£¬×ÏɫΪ¸ßˮƽ¡£´ÖÌåÔò±íʾ²Ð²îÁ´½ÓºÍ·Ç²Ð²îÁ¬½ÓÖ®¼äµÄ×î¼ÑÄ£ÐÍ£¨ÈçÁ½¸öÄ£ÐÍˮƽÏàͬÔò½ÔΪ´ÖÌåÏÔʾ£©¡£
ͼ»Ø¹éºÍ·Ö×ÓÊý¾Ý¼¯
ZINC ·Ö×ÓÊý¾Ý¼¯±»ÓÃÓÚ¶ÔÊÜÏÞÈܽâ¶È·Ö×ÓÌØÐÔ½øÐлعé·ÖÎö¡£ÔÚÕâÀï ZINC ÓÐ 10000 ѵÁ·/1000 ÑéÖ¤/1000 ²âÊÔͼ£¬½ÚµãÊý/Ô×ÓÊýΪ 9-37¡£¶ÔÓÚÿ¸ö·Ö×Óͼ£¬½ÚµãÌØÕ÷ÊÇÔ×ÓµÄÀàÐÍ£¬±ßÔµÌØÕ÷ÊDZßÔµµÄÀàÐÍ¡£
ÔÚ SBM?Êý¾Ý¼¯ÉϽøÐнڵã·ÖÀà
Ñо¿Õß¿¼ÂÇÁ˽ڵ㼶±ðµÄͼģʽʶ±ðÈÎÎñºÍ°ë¼à¶½Í¼¾ÛÀàÈÎÎñ¡£Í¼Ä£Ê½Ê¶±ðʱΪÁËÕÒµ½Ò»¸ö¹Ì¶¨µÄͼģʽ P£¬Ç¶ÈëÓÚ¸ü´óµÄͼ G ÖС£
¶ø°ë¼à¶½¾ÛÀàÈÎÎñÔòÊÇÍøÂç¿ÆѧÖеÄÁíÒ»¸öÖØÒªÈÎÎñ¡£Ñо¿ÕßÕë¶ÔÒÔÉÏÁ½¸öÈÎÎñ·Ö±ðÉú³ÉÁËÏàÓ¦µÄÊý¾Ý¼¯¡£
±í 5£ºÔÚ±ê×¼²âÊÔ¼¯ PATTERN ºÍ CLUSTER SBM ͼÉϵÄÐÔÄܱíÏÖ¡£
TSP Êý¾Ý¼¯Éϵı߷ÖÀà
TSP£¨Travelling Salesman Problem£©Ö¸µÄÊÇÂÃÐÐÍÆÏúÔ±ÎÊÌ⣺¸ø¶¨Ò»¸ö 2D µÄÅ·¼¸ÀïµÃͼ£¬Ëã·¨ÐèÒªÕÒµ½Ò»¸ö×îÓŵÄÐòÁнڵ㣬ÃûΪ Tour¡£ËüÓ¦µ±ÓÐ×Å×îÉٵıßȨÖØ¡£TSP µÄ´ó¹æÄ£ÌØÐÔʹµÃËü³ÉΪһ¸ö¾ßÓÐÌôÕ½ÐÔµÄͼÈÎÎñ£¬ÐèÒª¶Ô¾Ö²¿½ÚµãµÄ½üÁÚºÍÈ«¾Öͼ½á¹¹½øÐÐÍÆÀí¡£
¸üÖØÒªµÄÊÇ£¬×éºÏÓÅ»¯ÎÊÌâÒ²ÊÇ GNN ÖÐÓÐÑо¿ÒâÒåµÄÒ»¸öÓ¦Óó¡¾°¡£Ñо¿ÕâÀàÎÊÌ⣬²»½ö½öÔÚÏÖʵÖÐÓÐ׏㷺µÄÓ¦Ó㬻¹¶ÔÓÚÀí½âͼģÐ͵ÄÓÅ»¯ºÍѧϰ¹ý³Ì£¬Í¼ÍøÂç±¾ÉíµÄ¾ÖÏÞÐÔµÈÓÐÖØÒªÒâÒå¡£
ÔÚ»ù×¼²âÊÔÖУ¬Ñо¿Õß²ÉÓÃÁË»ùÓÚѧϰµÄ·½·¨£¬½¨Á¢ÁËÒ» GNN ×÷Ϊ¹Ç¼ÜÍøÂ磬À´¸øÿ¸ö±ßºÍÊÇ·ñËùÊôÔ¤²â½á¹û¼¯½øÐиÅÂÊÔ¤²â¡£ÕâÒ»¸ÅÂʾÓÉͼËÑË÷¼¼Êõ±»×ª»»ÎªÀëÉ¢¾ö²ß¡£Ñо¿Õ߷ֱ𴴽¨ÁË 10000 ¸öѵÁ·ÊµÀýºÍ 1000 ¸öÑéÖ¤¡¢1000 ¸ö²âÊÔʵÀý¡£
ͼ 2£ºTSP Êý¾Ý¼¯µÄÑù±¾Í¼¡£½ÚµãÒÔÀ¶É«±íʾ£¬ºìÉ«±íʾ groundtruth µÄ±ß¡£
±í 6£ºTSP ²âÊÔ¼¯µÄͼÐÔÄܱíÏÖ£¬·ÖΪÓÐ/Î޲вîÁ¬½ÓÁ¼ºÃ×ÜÇé¿ö¡£ºìÉ«±íʾ×îºÃµÄÄ£ÐÍÐÔÄÜ£¬×ÏÉ«±íʾģÐÍЧ¹û²»´í¡£
±í 7£ºÔÚ TSP ²âÊÔ¼¯Í¼ÉϵÄÐÔÄܱíÏÖ¡£Ä£ÐÍÊÇÉî¶È GNN£¬ÓÐ 32 ²ã¡£Ä£ÐÍ·ÖΪʹÓòвîÁ¬½ÓºÍûÓввîÁ¬½ÓÁ½ÖÖÇé¿ö¡£L ±íʾ²ãÊý£¬B ±íʾ×îºÃµÄ½á¹û£¨ÓввîÁ¬½ÓºÍÎ޲вîÁ¬½ÓµÄÇé¿ö£©¡£
±í 8£ºZINC¡¢CIFAR10?ºÍ CLUSTER ²âÊÔ¼¯Í¼ÔÚÓлòÕßûÓÐ BN¡¢GN µÄÇé¿öϵÄÐÔÄܱíÏÖ¡£
ÁìȡרÊô 10ÔªÎÞÃż÷ȯ
˽Ïí×îР¼¼Êõ¸É»õ