{"id":330,"date":"2025-07-29T11:14:27","date_gmt":"2025-07-29T03:14:27","guid":{"rendered":"http:\/\/anyone.anuoblog.cn\/?p=330"},"modified":"2025-08-12T11:37:25","modified_gmt":"2025-08-12T03:37:25","slug":"an-image-is-worth-16x16-words-transformers-for-image-recognition-at-scale","status":"publish","type":"post","link":"http:\/\/anyone.anuoblog.cn\/index.php\/2025\/07\/29\/an-image-is-worth-16x16-words-transformers-for-image-recognition-at-scale\/","title":{"rendered":"An Image is Worth 16×16 Words: Transformers for Image Recognition at Scale"},"content":{"rendered":"\n

arXiv\u8bba\u6587\u5730\u5740<\/a><\/p>\n\n\n\n

\u300a\u5c06 Transformer \u67b6\u6784\u5e94\u7528\u4e8e\u56fe\u50cf\u8bc6\u522b\u4efb\u52a1\u300b<\/p>\n\n\n\n

Submitted on 22 Oct 2020 (<\/em>v1<\/a>), last revised 3 Jun 2021 (this version, v2<\/em>\uff09\uff0cICLR 2021<\/p>\n\n\n\n

\u7b80\u4ecb<\/h1>\n\n\n\n

\u8fd9\u7bc7\u8bba\u6587\u4e3b\u8981\u4ecb\u7ecdICLR 2021\u4e0a\u7684\u3001\u4e00\u4e2a\u57fa\u4e8eTransformer\u6a21\u578b\u5e94\u7528\u4e8e\u5904\u7406\u56fe\u50cf\u8bc6\u522b\u4efb\u52a1\u7684ViT\u6a21\u578b\uff0c\u8bc1\u660e\u4e86\u7eaf Transformer \u6a21\u578b\u76f4\u63a5\u5904\u7406\u56fe\u50cf\u5757\u5e8f\u5217\u5728\u56fe\u50cf\u5206\u7c7b\u4efb\u52a1\u4e2d\u80fd\u6709\u51fa\u8272\u8868\u73b0\u3002<\/p>\n\n\n\n

0 \u6458\u8981<\/h1>\n\n\n\n
\u7ffb\u8bd1<\/summary>\n

\u5c3d\u7ba1 Transformer \u67b6\u6784\u5df2\u6210\u4e3a\u81ea\u7136\u8bed\u8a00\u5904\u7406\u4efb\u52a1\u7684\u4e8b\u5b9e\u6807\u51c6\uff0c\u4f46\u5176\u5728\u8ba1\u7b97\u673a\u89c6\u89c9\u9886\u57df\u7684\u5e94\u7528\u4ecd\u5341\u5206\u6709\u9650\u3002\u5728\u89c6\u89c9\u4efb\u52a1\u4e2d\uff0c\u6ce8\u610f\u529b\u673a\u5236\u901a\u5e38\u8981\u4e48\u4e0e\u5377\u79ef\u7f51\u7edc\u7ed3\u5408\u4f7f\u7528\uff0c\u8981\u4e48\u7528\u4e8e\u66ff\u6362\u5377\u79ef\u7f51\u7edc\u7684\u67d0\u4e9b\u7ec4\u4ef6\uff0c\u540c\u65f6\u4fdd\u7559\u5377\u79ef\u7f51\u7edc\u7684\u6574\u4f53\u7ed3\u6784\u3002<\/p>\n\n\n\n

\u672c\u6587\u8868\u660e\uff0c\u8fd9\u79cd\u5bf9\u5377\u79ef\u795e\u7ecf\u7f51\u7edc\uff08CNNs\uff09\u7684\u4f9d\u8d56\u5e76\u975e\u5fc5\u9700 \u2014\u2014\u5c06\u7eaf Transformer \u76f4\u63a5\u5e94\u7528\u4e8e\u56fe\u50cf\u5757\u5e8f\u5217<\/strong>\uff0c\u5728\u56fe\u50cf\u5206\u7c7b\u4efb\u52a1\u4e2d\u5c31\u80fd\u8868\u73b0\u51fa\u8272\u3002\u5f53\u5728\u5927\u91cf\u6570\u636e\u4e0a\u8fdb\u884c\u9884\u8bad\u7ec3\uff0c\u5e76\u8fc1\u79fb\u5230\u591a\u4e2a\u4e2d\u578b\u6216\u5c0f\u578b\u56fe\u50cf\u8bc6\u522b\u57fa\u51c6\u4efb\u52a1\uff08\u5982 ImageNet\u3001CIFAR-100\u3001VTAB \u7b49\uff09\u65f6\uff0cVision Transformer\uff08ViT\uff09\u4e0e\u6700\u5148\u8fdb\u7684\u5377\u79ef\u7f51\u7edc\u76f8\u6bd4\uff0c\u80fd\u53d6\u5f97\u4f18\u5f02\u7684\u7ed3\u679c\uff0c\u540c\u65f6\u8bad\u7ec3\u6240\u9700\u7684\u8ba1\u7b97\u8d44\u6e90\u663e\u8457\u51cf\u5c11\u3002<\/p>\n<\/details>\n\n\n\n

\u52a8\u673a\uff1a<\/strong>Transformer\u5e7f\u6cdb\u5e94\u7528\u4e8e\u81ea\u7136\u8bed\u8a00\u5904\u7406\uff0c\u4f46\u5728\u8ba1\u7b97\u673a\u89c6\u89c9\u9886\u57df\u7684\u5e94\u7528\u4ecd\u5341\u5206\u6709\u9650\uff1b<\/p>\n\n\n\n

\u672c\u6587\u7684\u8d21\u732e\u3001\u5b8c\u6210\u7684\u4efb\u52a1\uff1a\u8bc1\u660e\u8ba1\u7b97\u673a\u89c6\u89c9\u4efb\u52a1\u5bf9CNNs\u7684\u4f9d\u8d56\u4e0d\u662f\u5fc5\u8981\u7684\uff0c\u5c06\u7eaf Transformer \u76f4\u63a5\u5e94\u7528\u4e8e\u56fe\u50cf\u5757\u5e8f\u5217<\/strong>\uff0c\u5728\u56fe\u50cf\u5206\u7c7b\u4efb\u52a1\u4e2d\u5c31\u80fd\u8868\u73b0\u51fa\u8272\u3002\u672c\u6587\u7684ViT\uff08Vision Transformer\uff09\u6a21\u578b\u4e0e\u6700\u5148\u8fdb\u7684CNN\u76f8\u6bd4\uff0c\u5f97\u5230\u4e86\u66f4\u52a0\u4f18\u5f02\u7684\u7ed3\u679c\uff0c\u5e76\u4e14\u51cf\u5c11\u4e86\u8bad\u7ec3\u6240\u9700\u8981\u7684\u8ba1\u7b97\u8d44\u6e90\uff1b<\/p>\n\n\n\n

1 \u5f15\u8a00<\/h1>\n\n\n\n
\u7ffb\u8bd1<\/summary>\n

\u57fa\u4e8e\u81ea\u6ce8\u610f\u529b\u673a\u5236\u7684\u67b6\u6784\uff0c\u5c24\u5176\u662f Transformer\uff08Vaswani \u7b49\u4eba\uff0c2017\uff09\uff0c\u5df2\u6210\u4e3a\u81ea\u7136\u8bed\u8a00\u5904\u7406\uff08NLP\uff09\u4efb\u52a1\u7684\u9996\u9009\u6a21\u578b\u3002\u5176\u4e3b\u6d41\u65b9\u6cd5\u662f\u5728\u5927\u578b\u6587\u672c\u8bed\u6599\u5e93\u4e0a\u9884\u8bad\u7ec3\uff0c\u7136\u540e\u5728\u8f83\u5c0f\u7684\u7279\u5b9a\u4efb\u52a1\u6570\u636e\u96c6\u4e0a\u5fae\u8c03\uff08Devlin \u7b49\u4eba\uff0c1989\uff09\u3002\u5f97\u76ca\u4e8e Transformer \u7684\u8ba1\u7b97\u6548\u7387\u548c\u53ef\u6269\u5c55\u6027\uff0c\u8bad\u7ec3\u89c4\u6a21\u7a7a\u524d\u7684\u6a21\u578b\u6210\u4e3a\u53ef\u80fd\uff0c\u53c2\u6570\u6570\u91cf\u5df2\u8d85\u8fc7 1000 \u4ebf\uff08Brown \u7b49\u4eba\uff0c2020\uff1bLepikhin \u7b49\u4eba\uff0c2020\uff09\u3002\u968f\u7740\u6a21\u578b\u548c\u6570\u636e\u96c6\u7684\u4e0d\u65ad\u6269\u5927\uff0c\u6027\u80fd\u5c1a\u672a\u663e\u73b0\u51fa\u9971\u548c\u7684\u8ff9\u8c61\u3002<\/p>\n\n\n\n

\u7136\u800c\uff0c\u5728\u8ba1\u7b97\u673a\u89c6\u89c9\u9886\u57df\uff0c\u5377\u79ef\u67b6\u6784\u4ecd\u7136\u5360\u636e\u4e3b\u5bfc\u5730\u4f4d\uff08LeCun \u7b49\u4eba\uff0c1989\uff1bKrizhevsky \u7b49\u4eba\uff0c2012\uff1bHe \u7b49\u4eba\uff0c2016\uff09\u3002\u53d7 NLP \u9886\u57df\u6210\u529f\u6848\u4f8b\u7684\u542f\u53d1\uff0c\u591a\u9879\u7814\u7a76\u5c1d\u8bd5\u5c06\u7c7b CNN \u67b6\u6784\u4e0e\u81ea\u6ce8\u610f\u529b\u673a\u5236\u7ed3\u5408\uff08Wang \u7b49\u4eba\uff0c2018\uff1bCarion \u7b49\u4eba\uff0c2020\uff09\uff0c\u4e5f\u6709\u90e8\u5206\u7814\u7a76\u5b8c\u5168\u66ff\u6362\u5377\u79ef\u64cd\u4f5c\uff08Ramachandran \u7b49\u4eba\uff0c2019\uff1bWang \u7b49\u4eba\uff0c2020a\uff09\u3002\u5c3d\u7ba1\u540e\u8005\u5728\u7406\u8bba\u4e0a\u5177\u6709\u6548\u7387\u4f18\u52bf\uff0c\u4f46\u7531\u4e8e\u4f7f\u7528\u4e86\u4e13\u95e8\u7684\u6ce8\u610f\u529b\u6a21\u5f0f\uff0c\u5c1a\u672a\u80fd\u5728\u73b0\u4ee3\u786c\u4ef6\u52a0\u901f\u5668\u4e0a\u6709\u6548\u6269\u5c55\u3002\u56e0\u6b64\uff0c\u5728\u5927\u89c4\u6a21\u56fe\u50cf\u8bc6\u522b\u4e2d\uff0c\u7ecf\u5178\u7684\u7c7b ResNet \u67b6\u6784\u4ecd\u662f\u5f53\u524d\u7684\u6700\u4f18\u9009\u62e9\uff08Mahajan \u7b49\u4eba\uff0c2018\uff1bXie \u7b49\u4eba\uff0c2020\uff1bKolesnikov \u7b49\u4eba\uff0c2020\uff09\u3002<\/p>\n\n\n\n

\u53d7 Transformer \u5728 NLP \u9886\u57df\u89c4\u6a21\u5316\u6210\u529f\u7684\u542f\u53d1\uff0c\u6211\u4eec\u5c1d\u8bd5\u5c06\u6807\u51c6 Transformer \u76f4\u63a5\u5e94\u7528\u4e8e\u56fe\u50cf\uff0c\u4e14\u5c3d\u53ef\u80fd\u5c11\u5730\u8fdb\u884c\u4fee\u6539\u3002\u5177\u4f53\u800c\u8a00\uff0c\u6211\u4eec\u5c06\u56fe\u50cf\u5206\u5272\u4e3a\u591a\u4e2a\u8865\u4e01\uff08patches\uff09\uff0c\u5e76\u5c06\u8fd9\u4e9b\u8865\u4e01\u7684\u7ebf\u6027\u5d4c\u5165\u5e8f\u5217\u4f5c\u4e3a Transformer \u7684\u8f93\u5165\u3002\u56fe\u50cf\u8865\u4e01\u7684\u5904\u7406\u65b9\u5f0f\u4e0e NLP \u5e94\u7528\u4e2d\u7684\u6807\u8bb0\uff08tokens\uff0c\u5982\u5355\u8bcd\uff09\u5b8c\u5168\u76f8\u540c\u3002\u6211\u4eec\u4ee5\u6709\u76d1\u7763\u7684\u65b9\u5f0f\u5728\u56fe\u50cf\u5206\u7c7b\u4efb\u52a1\u4e0a\u8bad\u7ec3\u8be5\u6a21\u578b\u3002<\/p>\n\n\n\n

\u5f53\u5728\u4e2d\u7b49\u89c4\u6a21\u6570\u636e\u96c6\uff08\u5982 ImageNet\uff09\u4e0a\u8bad\u7ec3\u4e14\u672a\u4f7f\u7528\u5f3a\u6b63\u5219\u5316\u65f6\uff0c\u8fd9\u4e9b\u6a21\u578b\u7684\u7cbe\u5ea6\u8868\u73b0\u5e73\u5e73\uff0c\u6bd4\u540c\u7b49\u89c4\u6a21\u7684 ResNet \u4f4e\u51e0\u4e2a\u767e\u5206\u70b9\u3002\u8fd9\u4e00\u7ed3\u679c\u770b\u4f3c\u4ee4\u4eba\u6cae\u4e27\uff0c\u4f46\u5176\u5b9e\u5728\u610f\u6599\u4e4b\u4e2d\uff1aTransformer \u7f3a\u4e4f CNN \u56fa\u6709\u7684\u4e00\u4e9b\u5f52\u7eb3\u504f\u7f6e\uff08\u5982\u5e73\u79fb\u7b49\u53d8\u6027\u548c\u5c40\u90e8\u6027\uff09\uff0c\u56e0\u6b64\u5728\u8bad\u7ec3\u6570\u636e\u4e0d\u8db3\u65f6\u6cdb\u5316\u80fd\u529b\u8f83\u5dee\u3002<\/p>\n\n\n\n

\u7136\u800c\uff0c\u82e5\u5728\u66f4\u5927\u7684\u6570\u636e\u96c6\uff081400 \u4e07 – 3 \u4ebf\u5f20\u56fe\u50cf\uff09\u4e0a\u8bad\u7ec3\uff0c\u60c5\u51b5\u4f1a\u6709\u6240\u6539\u89c2\u3002\u6211\u4eec\u53d1\u73b0\uff0c\u5927\u89c4\u6a21\u8bad\u7ec3\u7684\u6548\u679c\u4f18\u4e8e\u5f52\u7eb3\u504f\u7f6e\u3002\u5f53\u5728\u8db3\u591f\u5927\u7684\u89c4\u6a21\u4e0a\u9884\u8bad\u7ec3\u540e\uff0c\u6211\u4eec\u63d0\u51fa\u7684 Vision Transformer\uff08ViT\uff09\u5728\u8fc1\u79fb\u5230\u6570\u636e\u91cf\u8f83\u5c11\u7684\u4efb\u52a1\u65f6\u80fd\u53d6\u5f97\u4f18\u5f02\u7ed3\u679c\u3002\u5728\u516c\u5f00\u7684 ImageNet-21k \u6570\u636e\u96c6\u6216\u5185\u90e8\u7684 JFT-300M \u6570\u636e\u96c6\u4e0a\u9884\u8bad\u7ec3\u540e\uff0cViT \u5728\u591a\u4e2a\u56fe\u50cf\u8bc6\u522b\u57fa\u51c6\u6d4b\u8bd5\u4e2d\u63a5\u8fd1\u6216\u8d85\u8fc7\u4e86\u5f53\u524d\u6700\u4f18\u6c34\u5e73\u3002\u5177\u4f53\u800c\u8a00\uff0c\u6700\u4f18\u6a21\u578b\u5728 ImageNet \u4e0a\u8fbe\u5230 88.55% \u7684\u51c6\u786e\u7387\uff0c\u5728 ImageNet-ReaL \u4e0a\u8fbe\u5230 90.72%\uff0c\u5728 CIFAR-100 \u4e0a\u8fbe\u5230 94.55%\uff0c\u5728\u5305\u542b 19 \u4e2a\u4efb\u52a1\u7684 VTAB \u5957\u4ef6\u4e0a\u8fbe\u5230 77.63%\u3002<\/p>\n<\/details>\n\n\n\n

1\u3001Transformer\u5728\u8ba1\u7b97\u6548\u7387\u3001\u53ef\u6269\u5c55\u6027\u65b9\u9762\u5177\u6709\u4f18\u52bf\uff1b\u5f97\u76ca\u4e8eTransformer\u7684\u4f18\u52bf\uff0c\u968f\u7740\u6a21\u578b\u548c\u6570\u636e\u96c6\u7684\u4e0d\u65ad\u6269\u5927\uff0c\u6027\u80fd\u4e5f\u53d8\u5f97\u8d8a\u6765\u8d8a\u597d\uff1b<\/p>\n\n\n\n

2\u3001\u672c\u6587\u5c06Transformer\u5e94\u7528\u4e8e\u56fe\u50cf\u65b9\u7684\u65b9\u6cd5\uff1a\u5c06\u56fe\u50cf\u5206\u5272\u4e3a\u591a\u4e2apatches\uff08\u8865\u4e01\uff09\uff0c\u5e76\u5c06\u8fd9\u4e9b\u8865\u4e01\u7684\u7ebf\u6027\u5d4c\u5165\u5e8f\u5217\u4f5c\u4e3aTransformer\u7684\u8f93\u5165\uff0c\u56fe\u50cf\u8865\u4e01\u7684\u5904\u7406\u65b9\u5f0f\u4e0e\u81ea\u7136\u8bed\u8a00\u5904\u7406\u5e94\u7528\u4e2d\u7684tokens\uff08\u5355\u8bcd\u7b49\uff09\u5b8c\u5168\u76f8\u540c\uff1b\u4ee5\u6709\u76d1\u7763\u7684\u65b9\u5f0f\uff08\u6307\u7684\u662f\u57fa\u4e8e\u6807\u7b7e\u7684\u56fe\u50cf\u5206\u7c7b\u8bad\u7ec3\u6a21\u5f0f<\/strong>\uff0c\u6bcf\u4e2a\u56fe\u7247\u90fd\u6709real target\uff09\u5728\u56fe\u50cf\u5206\u7c7b\u4efb\u52a1\u4e0a\u8bad\u7ec3ViT\u6a21\u578b\uff1b<\/p>\n\n\n\n

\u4f46\u662f\u8fd9\u79cd\u65b9\u6cd5\u5728\u5904\u7406\u89c4\u6a21\u4e0d\u5927\u7684\u6570\u636e\u96c6\u4e2d\uff0c\u4f1a\u56e0\u6570\u636e\u4e0d\u8db3\u800c\u5bfc\u81f4\u6a21\u578b\u6cdb\u5316\u80fd\u529b\u8f83\u5dee\uff1b<\/p>\n\n\n\n

3\u3001\u672c\u6587\u6210\u679c\uff1a\u5728ImageNet\u4e0a\u53d6\u5f9788.55%\u51c6\u786e\u7387\uff0cImageNet-ReaL \u4e0a\u8fbe\u5230 90.72%\uff0c\u5728 CIFAR-100 \u4e0a\u8fbe\u5230 94.55%\uff0c\u5728\u5305\u542b 19 \u4e2a\u4efb\u52a1\u7684 VTAB \u5957\u4ef6\u4e0a\u8fbe\u5230 77.63%\uff1b<\/p>\n\n\n\n

2 \u76f8\u5173\u5de5\u4f5c<\/h1>\n\n\n\n
\u7ffb\u8bd1<\/summary>\n

Transformer \u7531 Vaswani \u7b49\u4eba\uff082017\uff09\u63d0\u51fa\uff0c\u6700\u521d\u7528\u4e8e\u673a\u5668\u7ffb\u8bd1\u4efb\u52a1\uff0c\u6b64\u540e\u6210\u4e3a\u4f17\u591a\u81ea\u7136\u8bed\u8a00\u5904\u7406\uff08NLP\uff09\u4efb\u52a1\u7684\u6700\u5148\u8fdb\u65b9\u6cd5\u3002\u5927\u578b\u57fa\u4e8e Transformer \u7684\u6a21\u578b\u901a\u5e38\u5728\u5927\u578b\u8bed\u6599\u5e93\u4e0a\u9884\u8bad\u7ec3\uff0c\u518d\u9488\u5bf9\u5177\u4f53\u4efb\u52a1\u8fdb\u884c\u5fae\u8c03\uff1aBERT\uff08Devlin \u7b49\u4eba\uff0c2019\uff09\u91c7\u7528\u53bb\u566a\u81ea\u76d1\u7763\u9884\u8bad\u7ec3\u4efb\u52a1\uff0c\u800c GPT \u7cfb\u5217\u5219\u4ee5\u8bed\u8a00\u5efa\u6a21\u4f5c\u4e3a\u9884\u8bad\u7ec3\u4efb\u52a1\uff08Radford \u7b49\u4eba\uff0c2018\uff1b2019\uff1bBrown \u7b49\u4eba\uff0c2020\uff09\u3002<\/p>\n\n\n\n

\u82e5\u5c06\u81ea\u6ce8\u610f\u529b\u673a\u5236\u76f4\u63a5\u5e94\u7528\u4e8e\u56fe\u50cf\uff0c\u4f1a\u8981\u6c42\u6bcf\u4e2a\u50cf\u7d20\u4e0e\u5176\u4ed6\u6240\u6709\u50cf\u7d20\u8fdb\u884c\u6ce8\u610f\u529b\u4ea4\u4e92\u3002\u7531\u4e8e\u8ba1\u7b97\u6210\u672c\u4e0e\u50cf\u7d20\u6570\u91cf\u6210\u4e8c\u6b21\u65b9\u5173\u7cfb\uff0c\u8fd9\u79cd\u65b9\u5f0f\u65e0\u6cd5\u6269\u5c55\u5230\u5b9e\u9645\u7684\u8f93\u5165\u5c3a\u5bf8\u3002\u56e0\u6b64\uff0c\u8fc7\u53bb\u4e3a\u4e86\u5728\u56fe\u50cf\u5904\u7406\u4e2d\u5e94\u7528 Transformer\uff0c\u7814\u7a76\u4eba\u5458\u5c1d\u8bd5\u4e86\u591a\u79cd\u8fd1\u4f3c\u65b9\u6cd5\u3002Parmar \u7b49\u4eba\uff082018\uff09\u4ec5\u5728\u6bcf\u4e2a\u67e5\u8be2\u50cf\u7d20\u7684\u5c40\u90e8\u90bb\u57df\u5185\u5e94\u7528\u81ea\u6ce8\u610f\u529b\uff0c\u800c\u975e\u5168\u5c40\u8303\u56f4\u5185\u3002\u6b64\u7c7b\u5c40\u90e8\u591a\u5934\u70b9\u79ef\u81ea\u6ce8\u610f\u529b\u5757\u53ef\u5b8c\u5168\u66ff\u4ee3\u5377\u79ef\u64cd\u4f5c\uff08Hu \u7b49\u4eba\uff0c2019\uff1bRamachandran \u7b49\u4eba\uff0c2019\uff1bZhao \u7b49\u4eba\uff0c2020\uff09\u3002\u53e6\u4e00\u7c7b\u7814\u7a76\u4e2d\uff0c\u7a00\u758f Transformer\uff08Child \u7b49\u4eba\uff0c2019\uff09\u91c7\u7528\u53ef\u6269\u5c55\u7684\u5168\u5c40\u81ea\u6ce8\u610f\u529b\u8fd1\u4f3c\u65b9\u6cd5\uff0c\u4ee5\u9002\u7528\u4e8e\u56fe\u50cf\u3002\u53e6\u4e00\u79cd\u6269\u5c55\u6ce8\u610f\u529b\u7684\u65b9\u5f0f\u662f\u5728\u4e0d\u540c\u5927\u5c0f\u7684\u5757\u4e0a\u5e94\u7528\u6ce8\u610f\u529b\uff08Weissenborn \u7b49\u4eba\uff0c2019\uff09\uff0c\u6781\u7aef\u60c5\u51b5\u4e0b\u4ec5\u6cbf\u5355\u4e2a\u8f74\u5e94\u7528\uff08Ho \u7b49\u4eba\uff0c2019\uff1bWang \u7b49\u4eba\uff0c2020a\uff09\u3002\u8fd9\u4e9b\u4e13\u95e8\u8bbe\u8ba1\u7684\u6ce8\u610f\u529b\u67b6\u6784\u5728\u8ba1\u7b97\u673a\u89c6\u89c9\u4efb\u52a1\u4e0a\u5c55\u73b0\u51fa\u826f\u597d\u524d\u666f\uff0c\u4f46\u9700\u8981\u590d\u6742\u7684\u5de5\u7a0b\u5b9e\u73b0\u624d\u80fd\u5728\u786c\u4ef6\u52a0\u901f\u5668\u4e0a\u9ad8\u6548\u8fd0\u884c\u3002<\/p>\n\n\n\n

\u4e0e\u6211\u4eec\u7684\u5de5\u4f5c\u6700\u76f8\u5173\u7684\u662f Cordonnier \u7b49\u4eba\uff082020\uff09\u7684\u6a21\u578b\uff0c\u8be5\u6a21\u578b\u4ece\u8f93\u5165\u56fe\u50cf\u4e2d\u63d0\u53d6 2\u00d72 \u5927\u5c0f\u7684\u8865\u4e01\uff0c\u5e76\u5728\u5176\u4e4b\u4e0a\u5e94\u7528\u5b8c\u6574\u7684\u81ea\u6ce8\u610f\u529b\u3002\u8be5\u6a21\u578b\u4e0e ViT \u975e\u5e38\u76f8\u4f3c\uff0c\u4f46\u6211\u4eec\u7684\u5de5\u4f5c\u8fdb\u4e00\u6b65\u8bc1\u660e\uff0c\u5927\u89c4\u6a21\u9884\u8bad\u7ec3\u80fd\u4f7f vanilla Transformer\uff08\u57fa\u7840 Transformer\uff09\u4e0e\u6700\u5148\u8fdb\u7684 CNN \u76f8\u7ade\u4e89\uff08\u751a\u81f3\u66f4\u4f18\uff09\u3002\u6b64\u5916\uff0cCordonnier \u7b49\u4eba\uff082020\uff09\u4f7f\u7528 2\u00d72 \u50cf\u7d20\u7684\u5c0f\u8865\u4e01\u5c3a\u5bf8\uff0c\u8fd9\u4f7f\u5f97\u6a21\u578b\u4ec5\u9002\u7528\u4e8e\u4f4e\u5206\u8fa8\u7387\u56fe\u50cf\uff0c\u800c\u6211\u4eec\u7684\u6a21\u578b\u4e5f\u80fd\u5904\u7406\u4e2d\u7b49\u5206\u8fa8\u7387\u56fe\u50cf\u3002<\/p>\n\n\n\n

\u5c06\u5377\u79ef\u795e\u7ecf\u7f51\u7edc\uff08CNN\uff09\u4e0e\u81ea\u6ce8\u610f\u529b\u673a\u5236\u76f8\u7ed3\u5408\u7684\u7814\u7a76\u4e5f\u5907\u53d7\u5173\u6ce8\uff0c\u4f8b\u5982\u901a\u8fc7\u589e\u5f3a\u7279\u5f81\u56fe\u7528\u4e8e\u56fe\u50cf\u5206\u7c7b\uff08Bello \u7b49\u4eba\uff0c2019\uff09\uff0c\u6216\u4f7f\u7528\u81ea\u6ce8\u610f\u529b\u8fdb\u4e00\u6b65\u5904\u7406 CNN \u7684\u8f93\u51fa\uff0c\u5982\u7528\u4e8e\u76ee\u6807\u68c0\u6d4b\uff08Hu \u7b49\u4eba\uff0c2018\uff1bCarion \u7b49\u4eba\uff0c2020\uff09\u3001\u89c6\u9891\u5904\u7406\uff08Wang \u7b49\u4eba\uff0c2018\uff1bSun \u7b49\u4eba\uff0c2019\uff09\u3001\u56fe\u50cf\u5206\u7c7b\uff08Wu \u7b49\u4eba\uff0c2020\uff09\u3001\u65e0\u76d1\u7763\u76ee\u6807\u53d1\u73b0\uff08Locatello \u7b49\u4eba\uff0c2020\uff09\uff0c\u6216\u7edf\u4e00\u7684\u6587\u672c – \u89c6\u89c9\u4efb\u52a1\uff08Chen \u7b49\u4eba\uff0c2020c\uff1bLu \u7b49\u4eba\uff0c2019\uff1bLi \u7b49\u4eba\uff0c2019\uff09\u3002<\/p>\n\n\n\n

\u53e6\u4e00\u9879\u8fd1\u671f\u76f8\u5173\u7684\u6a21\u578b\u662f image GPT\uff08iGPT\uff09\uff08Chen \u7b49\u4eba\uff0c2020a\uff09\uff0c\u5b83\u5728\u964d\u4f4e\u56fe\u50cf\u5206\u8fa8\u7387\u548c\u538b\u7f29\u8272\u5f69\u7a7a\u95f4\u540e\uff0c\u5c06 Transformer \u5e94\u7528\u4e8e\u56fe\u50cf\u50cf\u7d20\u3002\u8be5\u6a21\u578b\u4ee5\u65e0\u76d1\u7763\u65b9\u5f0f\u4f5c\u4e3a\u751f\u6210\u6a21\u578b\u8bad\u7ec3\uff0c\u5176\u5f97\u5230\u7684\u7279\u5f81\u8868\u793a\u53ef\u901a\u8fc7\u5fae\u8c03\u6216\u7ebf\u6027\u63a2\u6d4b\u7528\u4e8e\u5206\u7c7b\u4efb\u52a1\uff0c\u5728 ImageNet \u4e0a\u6700\u9ad8\u8fbe\u5230 72% \u7684\u51c6\u786e\u7387\u3002<\/p>\n\n\n\n

\u6211\u4eec\u7684\u5de5\u4f5c\u52a0\u5165\u4e86\u8d8a\u6765\u8d8a\u591a\u63a2\u7d22\u8d85\u5927\u89c4\u6a21\u56fe\u50cf\u8bc6\u522b\u7684\u7814\u7a76\u884c\u5217\uff0c\u8fd9\u4e9b\u7814\u7a76\u4f7f\u7528\u7684\u6570\u636e\u96c6\u89c4\u6a21\u8d85\u8fc7\u6807\u51c6\u7684 ImageNet\u3002\u5229\u7528\u989d\u5916\u6570\u636e\u6e90\u6709\u52a9\u4e8e\u5728\u6807\u51c6\u57fa\u51c6\u6d4b\u8bd5\u4e0a\u53d6\u5f97\u6700\u5148\u8fdb\u7ed3\u679c\uff08Mahajan \u7b49\u4eba\uff0c2018\uff1bTouvron \u7b49\u4eba\uff0c2019\uff1bXie \u7b49\u4eba\uff0c2020\uff09\u3002\u6b64\u5916\uff0cSun \u7b49\u4eba\uff082017\uff09\u7814\u7a76\u4e86 CNN \u6027\u80fd\u5982\u4f55\u968f\u6570\u636e\u96c6\u5927\u5c0f\u6269\u5c55\uff0cKolesnikov \u7b49\u4eba\uff082020\uff09\u548c Djolonga \u7b49\u4eba\uff082020\uff09\u5219\u5b9e\u8bc1\u63a2\u7d22\u4e86\u4ece ImageNet-21k \u548c JFT-300M \u7b49\u5927\u89c4\u6a21\u6570\u636e\u96c6\u8fc1\u79fb\u5b66\u4e60\u7684 CNN \u6a21\u578b\u3002\u6211\u4eec\u540c\u6837\u805a\u7126\u4e8e\u8fd9\u4e24\u4e2a\u6570\u636e\u96c6\uff0c\u4f46\u8bad\u7ec3\u7684\u662f Transformer \u800c\u975e\u5148\u524d\u5de5\u4f5c\u4e2d\u57fa\u4e8e ResNet \u7684\u6a21\u578b\u3002<\/p>\n<\/details>\n\n\n\n

<\/p>\n\n\n\n

\u672f\u8bed\u8bf4\u660e<\/summary>\n

Transformer\uff1a<\/strong> \u7528\u4e8e\u673a\u5668\u7ffb\u8bd1\u7684\u65b9\u6cd5\uff0c\u88ab\u5e7f\u6cdb\u7528\u4e8eNLP\u9886\u57df<\/p>\n\n\n\n

BERT\uff1a<\/strong> \u4f7f\u7528\u53bb\u566a\u81ea\u6211\u76d1\u7763\u7684\u8bad\u7ec3\u524d\u4efb\u52a1<\/p>\n\n\n\n

\u5c40\u90e8\u591a\u5934\u70b9\u79ef\u81ea\u6211\u6ce8\u610f\u5757\uff1a<\/strong> \u53ea\u5728\u6bcf\u4e2a\u67e5\u8be2\u50cf\u7d20\u7684\u5c40\u90e8\u793e\u533a\u4e2d\u5e94\u7528\u81ea\u6ce8\u610f\u529b\uff0c\u53ef\u4ee5\u5b8c\u5168\u53d6\u4ee3\u5377\u79ef<\/p>\n\n\n\n

\u7a00\u758fTransformer\uff1a<\/strong> \u91c7\u7528\u4e86\u5bf9\u5168\u5c40\u81ea\u5173\u6ce8\u7684\u53ef\u6269\u5c55\u8fd1\u4f3c\uff0c\u4ee5\u4fbf\u9002\u7528\u4e8e\u56fe\u50cf<\/p>\n\n\n\n

\u5728\u4e0d\u540c\u5927\u5c0f\u7684\u5757\u4e2d\u5e94\u7528\uff1a<\/strong> \u5728\u6781\u7aef\u60c5\u51b5\u4e0b\uff0c\u53ea\u6cbf\u7740\u4e2a\u522b\u8f74\u7ebf\u5e94\u7528<\/p>\n\n\n\n

iGPT\uff1a<\/strong> \u65e0\u76d1\u7763\u7684\u65b9\u5f0f\uff0c\u5728\u964d\u4f4e\u56fe\u50cf\u5206\u8fa8\u7387\u548c\u8272\u5f69\u7a7a\u95f4\u540e\u5c06Transformers\u5e94\u7528\u4e8e\u56fe\u50cf\u50cf\u7d20<\/p>\n<\/details>\n\n\n\n

3 \u65b9\u6cd5<\/h1>\n\n\n\n
\u7ffb\u8bd1<\/summary>\n

\u5728\u6a21\u578b\u8bbe\u8ba1\u4e2d\uff0c\u6211\u4eec\u5c3d\u53ef\u80fd\u8d34\u8fd1\u539f\u59cb Transformer\uff08Vaswani \u7b49\u4eba\uff0c2017\uff09\u7684\u7ed3\u6784\u3002\u8fd9\u79cd\u523b\u610f\u7b80\u5316\u7684\u8bbe\u8ba1\u6709\u4e00\u4e2a\u4f18\u52bf\uff1a\u53ef\u6269\u5c55\u7684\u81ea\u7136\u8bed\u8a00\u5904\u7406\uff08NLP\uff09Transformer \u67b6\u6784\u53ca\u5176\u9ad8\u6548\u5b9e\u73b0\u51e0\u4e4e\u53ef\u4ee5\u76f4\u63a5\u590d\u7528\u3002<\/p>\n<\/details>\n\n\n\n

<\/p>\n\n\n\n

3.1 Vision Transformer\uff08ViT\uff09<\/h2>\n\n\n\n

\u2460<\/p>\n\n\n\n

\"\"<\/div><\/figure>\n\n\n\n

\u2461<\/p>\n\n\n\n

\"\"<\/div><\/figure>\n\n\n\n

\u2462<\/p>\n\n\n\n

\"\"<\/div><\/figure>\n\n\n\n

\u2463<\/p>\n\n\n\n

\"\"<\/div><\/figure>\n\n\n\n

\u2464<\/p>\n\n\n\n

\"\"<\/div><\/figure>\n\n\n\n
\u7ffb\u8bd1<\/summary>\n

\u6807\u51c6 Transformer \u7684\u8f93\u5165\u662f\u4e00\u7ef4\u6807\u8bb0\u5d4c\u5165\u5e8f\u5217\uff08sequence of token embeddings\uff09\u3002 \u4e3a\u4e86\u5904\u7406 2D \u56fe\u50cf\uff0c\u6211\u4eec\u5c06\u56fe\u50cf\u2460 reshape\u4e3a\u4e00\u7cfb\u5217\u5c55\u5e73\u7684 2D patches\u2461 \uff0c\u5176\u4e2d (H, W ) \u662f\u539f\u59cb\u56fe\u50cf\u7684\u5206\u8fa8\u7387\uff0cC \u662f \u901a\u9053\u6570\uff0c(P, P ) \u662f\u6bcf\u4e2a\u56fe\u50cfpatch\u7684\u5206\u8fa8\u7387\uff0c\u2462\u662fpatches\u6570\u76ee\uff0c\u5b83\u4e5f\u7528\u4f5c Transformer \u7684\u6709\u6548\u8f93\u5165\u5e8f\u5217\u957f\u5ea6\u3002 Transformer \u5728\u5176\u6240\u6709\u5c42\u4e2d\u4f7f\u7528\u6052\u5b9a\u7684\u6f5c\u5728\u5411\u91cf\u5927\u5c0f D\uff0c\u56e0\u6b64\u6211\u4eec\u5c06patches\u5c55\u5e73\u5e76\u4f7f\u7528\u53ef\u8bad\u7ec3\u7684\u7ebf\u6027\u6295\u5f71\u6620\u5c04\u5230 D \u7ef4\uff08\u65b9\u7a0b 1\uff09\u3002 \u6211\u4eec\u5c06\u6b64\u6295\u5f71\u7684\u8f93\u51fa\u79f0\u4e3apatch embeddings\u3002<\/p>\n\n\n\n

\u4e0e BERT \u7684 [class] token\u7c7b\u4f3c\uff0c\u6211\u4eec\u5728\u5d4c\u5165patches\u5e8f\u5217 (sequence of embedded patches\u2463) \u4e2d\u6dfb\u52a0\u4e86\u4e00\u4e2a\u53ef\u5b66\u4e60\u7684embeddings\uff0c\u5176\u5728 Transformer encoder (\u2464) \u7684\u8f93\u51fa\u7aef\u7684\u72b6\u6001\u7528\u4f5c\u56fe\u50cf\u8868\u793a y\uff08\u65b9\u7a0b 4\uff09 . \u5728\u9884\u8bad\u7ec3\u548c\u5fae\u8c03\u671f\u95f4\uff0c\u5206\u7c7b\u5934\u90fd\u8fde\u63a5\u5230\u3002\u5206\u7c7b\u5934\u7531\u5e26\u6709\u4e00\u4e2a\u9690\u85cf\u5c42\u7684 MLP \u5728\u9884\u8bad\u7ec3\u65f6\u5b9e\u73b0\uff0c\u5728\u5fae\u8c03\u65f6\u7531\u5355\u4e2a\u7ebf\u6027\u5c42\u5b9e\u73b0\u3002<\/p>\n\n\n\n

position embeddings\u88ab\u6dfb\u52a0\u5230patch embeddings\u4e2d\u4ee5\u4fdd\u7559\u4f4d\u7f6e\u4fe1\u606f\u3002 \u6211\u4eec\u4f7f\u7528\u6807\u51c6\u7684\u53ef\u5b66\u4e60\u4e00\u7ef4position embeddings\uff0c\u56e0\u4e3a\u6211\u4eec\u6ca1\u6709\u89c2\u5bdf\u5230\u4f7f\u7528\u66f4\u9ad8\u7ea7\u7684\u4e8c\u7ef4\u611f\u77e5position embeddings\uff08\u9644\u5f55 D.4\uff09\u5e26\u6765\u7684\u663e\u7740\u6027\u80fd\u63d0\u5347\u3002 \u5f97\u5230\u7684embeddings\u5411\u91cf\u5e8f\u5217\u4f5c\u4e3aencoder\u7684\u8f93\u5165\u3002<\/p>\n\n\n\n

Transformer encoder\uff08Vaswani \u7b49\u4eba\uff0c2017 \u5e74\uff09\u7531multi-head self-attention\uff08MSA\uff0c\u53c2\u89c1\u9644\u5f55 A)\u548c MLP blocks\uff08\u65b9\u7a0b 2\u30013\uff09\u7684\u4ea4\u66ff\u5c42\u7ec4\u6210\u3002 \u5728\u6bcf\u4e2a\u5757\u4e4b\u524d\u5e94\u7528layerNorm (LN)\uff0c\u5728\u6bcf\u4e2a\u5757\u4e4b\u540e\u5e94\u7528\u6b8b\u5dee\u8fde\u63a5\uff08Wang \u7b49\u4eba\uff0c2019 \u5e74\uff1bBaevski \u548c Auli\uff0c2019 \u5e74\uff09\u3002<\/p>\n\n\n\n

\u5f52\u7eb3\u504f\u7f6e\uff1a\u6211\u4eec\u6ce8\u610f\u5230 Vision Transformer \u6bd4 CNN \u5177\u6709\u66f4\u5c11\u7684\u7279\u5b9a\u4e8e\u56fe\u50cf\u7684\u5f52\u7eb3\u504f\u5dee\u3002 \u5728 CNN \u4e2d\uff0c\u5c40\u90e8\u6027\u3001\u4e8c\u7ef4\u90bb\u57df\u7ed3\u6784\u548c\u5e73\u79fb\u7b49\u6548\u6027\u88ab\u4f20\u9012\u5230\u6574\u4e2a\u6a21\u578b\u7684\u6bcf\u4e00\u5c42\u4e2d\u3002 \u5728 ViT \u4e2d\uff0c\u53ea\u6709 MLP \u5c42\u662f\u5c40\u90e8\u548c\u5e73\u79fb\u4e0d\u53d8\u7684\uff0c\u800cself-attention\u5c42\u662f\u5168\u5c40\u7684\u3002 \u4e8c\u7ef4\u90bb\u57df\u7ed3\u6784\u7684\u4f7f\u7528\u975e\u5e38\u8c28\u614e\uff1a\u5728\u6a21\u578b\u5f00\u59cb\u65f6\u901a\u8fc7\u5c06\u56fe\u50cf\u5207\u5272\u6210patches\uff0c\u5e76\u5728\u5fae\u8c03\u65f6\u8c03\u6574\u4e0d\u540c\u5206\u8fa8\u7387\u56fe\u50cf\u7684position embeddings\uff08\u5982\u4e0b\u6240\u8ff0\uff09\u3002 \u9664\u6b64\u4e4b\u5916\uff0c\u521d\u59cb\u5316\u65f6\u7684position embeddings\u4e0d\u643a\u5e26\u6709\u5173patches\u7684 2D \u4f4d\u7f6e\u7684\u4fe1\u606f\uff0cpatches\u4e4b\u95f4\u7684\u6240\u6709\u7a7a\u95f4\u5173\u7cfb\u90fd\u5fc5\u987b\u4ece\u5934\u5f00\u59cb\u5b66\u4e60\u3002<\/p>\n\n\n\n

\u6df7\u5408\u67b6\u6784\uff1a\u4f5c\u4e3a\u539f\u59cb\u56fe\u50cfpatches\u7684\u66ff\u4ee3\u65b9\u6848\uff0c\u8f93\u5165\u5e8f\u5217\u53ef\u4ee5\u7531 CNN \u7684feature maps\u5f62\u6210\uff08LeCun \u7b49\uff0c1989\uff09\u3002 \u5728\u8fd9\u4e2a\u6df7\u5408\u6a21\u578b\u4e2d\uff0cpatch embedding\u6295\u5f71 E\uff08\u65b9\u7a0b 1\uff09\u5e94\u7528\u4e8e\u4ece CNN featuremap\u4e2d\u63d0\u53d6\u7684patches\u3002 \u4f5c\u4e3a\u4e00\u79cd\u7279\u6b8a\u60c5\u51b5\uff0cpatch \u7684\u7a7a\u95f4\u5927\u5c0f\u53ef\u4ee5\u4e3a 1×1\uff0c\u8fd9\u610f\u5473\u7740\u8f93\u5165\u5e8f\u5217\u662f\u901a\u8fc7\u7b80\u5355\u5730\u5c06feature maps\u7684\u7a7a\u95f4\u7ef4\u5ea6\u5c55\u5e73\u5e76\u6295\u5f71\u5230 Transformer \u7ef4\u5ea6\u6765\u83b7\u5f97\u7684\u3002 \u5982\u4e0a\u6240\u8ff0\u6dfb\u52a0\u4e86\u5206\u7c7b\u8f93\u5165embedding\u548cposition embeddings\u3002<\/p>\n<\/details>\n\n\n\n

ViT\u6a21\u578b\u7ed3\u6784<\/p>\n\n\n\n

\"\"<\/div><\/figure>\n\n\n\n

\u7b2c\u4e00\u90e8\u5206<\/strong>-\u8f93\u5165<\/h3>\n\n\n\n

\u5c06\u56fe\u50cf\u5206\u5272\u4e3a\u591a\u4e2apatches\uff0c\u5e76\u5c06\u8fd9\u4e9b\u8865\u4e01\u7684\u7ebf\u6027\u5d4c\u5165\u5e8f\u5217\u4f5c\u4e3aTransformer\u7684\u8f93\u5165<\/p>\n\n\n\n

\"\"<\/div><\/figure>\n\n\n\n

\u9996\u5148\u8f93\u5165\u4e00\u5f20\u56fe\u7247\uff0c\u5c06\u8fd9\u4e2a\u56fe\u7247\u5206\u4e3a\u4e869\u4e2apatch\uff0c\u7136\u540e\u5c06\u6bcf\u4e2apatch\u91cd\u7ec4\u6210\u4e00\u4e2a\u5411\u91cf\uff0c\u5f97\u5230Flatten patch\uff1b<\/p>\n\n\n\n

\u5982\u679c\u56fe\u7247\u662fH\u00d7W\u00d7C<\/strong>\u7ef4\u7684\uff0c\u5c31\u7528P\u00d7P<\/strong>\u5927\u5c0f\u7684patch\u53bb\u5206\u5272\u56fe\u7247\u53ef\u4ee5\u5f97\u5230N\u4e2apatch<\/strong>\uff0c\u90a3\u4e48\u6bcf\u4e2apatch\u7684\u5927\u5c0f\u5c31\u662fP\u00d7P\u00d7C<\/strong>\uff0c\u5c06N\u4e2apatch<\/strong> \u91cd\u7ec4\u540e\u7684\u5411\u91cfconcat\u5728\u4e00\u8d77\u5c31\u5f97\u5230\u4e86\u4e00\u4e2aN\u00d7P\u00d7P\u00d7C<\/strong>\u7684\u4e8c\u7ef4\u77e9\u9635\uff0c\u76f8\u5f53\u4e8eNLP\u4e2d\u8f93\u5165Transformer\u7684\u8bcd\u5411\u91cf\uff1b<\/p>\n\n\n\n

Linear Projection of Flatten Patches\uff0c\u4f5c\u7528\u662f\u5c06\u56fe\u50cf\u8865\u4e01\u8f6c\u6362\u4e3a Transformer \u53ef\u5904\u7406\u7684\u5d4c\u5165\u5411\u91cf\uff1b<\/p>\n\n\n\n

\u7b2c\u4e8c\u90e8\u5206-Position Embedding<\/h3>\n\n\n\n
\"\"<\/div><\/figure>\n\n\n\n

\u4e4b\u524d\u5b66\u4e60Transformer\u53ef\u4ee5\u77e5\u9053\uff0cTransformer\u6a21\u578b\u672c\u8eab\u662f\u4e0d\u77e5\u9053\u6bcf\u4e2a\u201c\u8bcd\u201d\u7684\u4f4d\u7f6e\u4fe1\u606f\u7684\uff0c\u56e0\u6b64\u5c31\u9700\u8981Position Embedding\u5c06\u4f4d\u7f6e\u4fe1\u606f\u52a0\u5230\u6a21\u578b\u4e2d\uff1b<\/p>\n\n\n\n

\u5982\u4e0a\u56fe\u6240\u793a\uff0c\u7f16\u53f7\u67090-9\u7684\u7d2b\u8272\u6846\u8868\u793a\u5404\u4e2a\u4f4d\u7f6e\u7684position embedding\uff0c\u800c\u7d2b\u8272\u6846\u65c1\u8fb9\u7684\u7c89\u8272\u6846\u5219\u662f\u7ecf\u8fc7linear projection\u4e4b\u540e\u7684flattened patch\u5411\u91cf\uff08\u5c31\u76f8\u5f53\u4e8eNLP\u4e2d\u8bcd\u5411\u91cf\u52a0\u4e0a\u4f4d\u7f6e\u7f16\u7801\uff09\uff1b<\/p>\n\n\n\n

\u6587\u4e2d\u91c7\u7528\u5c06position embedding\uff08\u5373\u56fe\u4e2d\u7d2b\u8272\u6846\uff09<\/strong>\u548cpatch embedding\uff08\u5373\u56fe\u4e2d\u7c89\u8272\u6846\uff09<\/strong>\u76f8\u52a0\u7684\u65b9\u5f0f\u7ed3\u5408position\u4fe1\u606f\u3002<\/p>\n\n\n\n

\u7b2c\u4e09\u90e8\u5206-learnable embedding<\/h3>\n\n\n\n
\"\"<\/div><\/figure>\n\n\n\n

\u5c06 patch \u8f93\u5165\u4e00\u4e2a Linear Projection of Flattened Patches \u8fd9\u4e2a Embedding \u5c42\uff0c\u5c31\u4f1a\u5f97\u5230\u4e00\u4e2a\u4e2a\u5411\u91cf\uff0c\u901a\u5e38\u5c31\u79f0\u4f5c tokens<\/strong>\u3002tokens\u5305\u542bposition\u4fe1\u606f\u4ee5\u53ca\u56fe\u50cf\u4fe1\u606f\u3002<\/p>\n\n\n\n

\u7d27\u63a5\u7740\u5728\u4e00\u7cfb\u5217 token \u7684\u524d\u9762\u52a0\u4e0a\u52a0\u4e0a\u4e00\u4e2a\u65b0\u7684 token\uff0c\u53eb\u505aclass token\uff0c\u4e5f\u5c31\u662f\u4e0a\u56fe\u5e26\u661f\u53f7\u7684\u7c89\u8272\u6846\uff08\u53730\u53f7\u7d2b\u8272\u6846\u53f3\u8fb9\u7684\u90a3\u4e2a\uff09\uff0c\u6ce8\u610f\u8fd9\u4e2a\u4e0d\u662f\u901a\u8fc7\u67d0\u4e2apatch\u4ea7\u751f\u7684\u3002\u5176\u4f5c\u7528\u7c7b\u4f3c\u4e8eBERT\u4e2d\u7684[class] token\u3002\u5728BERT\u4e2d\uff0c[class] token\u7ecf\u8fc7encoder\u540e\u5bf9\u5e94\u7684\u7ed3\u679c\u4f5c\u4e3a\u6574\u4e2a\u53e5\u5b50\u7684\u8868\u793a\uff1bclass token\u4e5f\u662f\u5176\u4ed6\u6240\u6709token\u505a\u5168\u5c40\u5e73\u5747\u6c60\u5316\uff0c\u6548\u679c\u4e00\u6837\uff1b<\/p>\n\n\n\n

\u5b83\u662f\u6a21\u578b\u8bad\u7ec3\u8fc7\u7a0b\u4e2d \u901a\u8fc7\u53cd\u5411\u4f20\u64ad\u4e0d\u65ad\u66f4\u65b0\u53c2\u6570<\/strong> \u7684\u5d4c\u5165\uff0c\u4f1a\u6839\u636e\u8bad\u7ec3\u6570\u636e\uff08\u56fe\u50cf\u53ca\u5bf9\u5e94\u6807\u7b7e\uff09\u5b66\u4e60\u5230\u5bf9\u5206\u7c7b\u4efb\u52a1\u66f4\u6709\u6548\u7684\u8868\u793a\uff0c\u6240\u4ee5\u5c5e\u4e8e \u201clearnable embedding\u201d \uff1b<\/p>\n\n\n\n

\u7b2c\u56db\u90e8\u5206-Transformer Encoder<\/h3>\n\n\n\n
\"\"<\/div><\/figure>\n\n\n\n
\"\"<\/div><\/figure>\n\n\n\n

\u5c06\u524d\u9762\u5904\u7406\u540e\u7684\u5185\u5bb9\u8f93\u5165\u5230Transformer Encoder\u4e2d\uff0c\u5c06 block <\/strong>\uff08Transformer Encoder\u5185\u90e8\u7684\u6838\u5fc3\u91cd\u590d\u5355\u5143<\/strong>\uff09\u91cd\u590d\u5806\u53e0 L \u6b21\uff0c\u6574\u4e2a\u6a21\u578b\u4e5f\u5c31\u5305\u62ec L \u4e2a Transformer<\/strong>\u3002Transformer Encoder\u7ed3\u6784\u548cNLP\u4e2dTransformer\u7ed3\u6784\u57fa\u672c\u4e0a\u76f8\u540c\uff0cclass embedding \u5bf9\u5e94\u7684\u8f93\u51fa\u7ecf\u8fc7 MLP Head \u8fdb\u884c\u7c7b\u522b\u5224\u65ad\u3002<\/p>\n\n\n\n

3.2 \u5fae\u8c03\u4e0e\u66f4\u9ad8\u5206\u8fa8\u7387<\/h2>\n\n\n\n
\u7ffb\u8bd1<\/summary>\n

\u901a\u5e38\uff0c\u6211\u4eec\u4f1a\u5728\u5927\u578b\u6570\u636e\u96c6\u4e0a\u5bf9ViT\u8fdb\u884c\u9884\u8bad\u7ec3\uff0c\u5e76\u5bf9\uff08\u8f83\u5c0f\u7684\uff09\u4e0b\u6e38\u4efb\u52a1\u8fdb\u884c\u5fae\u8c03\u3002\u4e3a\u6b64\uff0c\u6211\u4eec\u53bb\u6389\u4e86\u9884\u5148\u8bad\u7ec3\u597d\u7684\u9884\u6d4b\u5934\uff0c\u5e76\u9644\u52a0\u4e86\u4e00\u4e2a\u96f6\u521d\u59cb\u5316\u7684D\u00d7K\u524d\u9988\u5c42\uff0c\u5176\u4e2dK\u662f\u4e0b\u6e38\u7c7b\u7684\u6570\u91cf\u3002\u4e0e\u8bad\u7ec3\u524d\u76f8\u6bd4\uff0c\u4ee5\u66f4\u9ad8\u7684\u5206\u8fa8\u7387\u8fdb\u884c\u5fae\u8c03\u5f80\u5f80\u6709\u76ca\u7684(Touvron\u7b49\u4eba\uff0c2019\uff1b\u79d1\u5217\u65af\u5c3c\u79d1\u592b\u7b49\u4eba\uff0c2020\u5e74)\u3002\u5f53\u8f93\u5165\u66f4\u9ad8\u5206\u8fa8\u7387\u7684\u56fe\u50cf\u65f6\uff0c\u6211\u4eec\u4fdd\u6301patch\u5927\u5c0f\u4e0d\u53d8\uff0c\u8fd9\u5bfc\u81f4\u66f4\u5927\u7684\u6709\u6548\u5e8f\u5217\u957f\u5ea6\u3002\u89c6\u89c9\u8f6c\u6362\u5668\u53ef\u4ee5\u5904\u7406\u4efb\u610f\u7684\u5e8f\u5217\u957f\u5ea6\uff08\u76f4\u5230\u5185\u5b58\u9650\u5236\uff09\uff0c\u7136\u800c\uff0c\u9884\u5148\u8bad\u7ec3\u7684Position embeding\u53ef\u80fd\u4e0d\u518d\u6709\u610f\u4e49\u3002\u56e0\u6b64\uff0c\u6211\u4eec\u6839\u636e\u9884\u5148\u8bad\u7ec3\u597d\u7684\u4f4d\u7f6eembeding\u5728\u539f\u59cb\u56fe\u50cf\u4e2d\u7684\u4f4d\u7f6e\uff0c\u8fdb\u884c\u4e8c\u7ef4\u63d2\u503c\u3002\u8bf7\u6ce8\u610f\uff0c\u8fd9\u79cd\u5206\u8fa8\u7387\u8c03\u6574\u548cpatch\u63d0\u53d6\u662f\u5c06\u56fe\u50cf\u4e8c\u7ef4\u7ed3\u6784\u7684\u611f\u5e94\u504f\u5dee\u624b\u52a8\u6ce8\u5165\u89c6\u89c9transformer\u7684\u552f\u4e00\u70b9\u3002<\/p>\n<\/details>\n\n\n\n

\u5b58\u5728\u7684\u95ee\u9898\uff1a<\/strong>\u5728\u4e0b\u6e38\u4efb\u52a1\u7684\u56fe\u50cf\u7684\u5206\u8fa8\u7387\u4f1a\u53d8\u5927\uff08\u5bfc\u81f4\u56fe\u50cf\u7684\u957f\u548c\u5bbd\u4f1a\u589e\u52a0\uff09\uff0c\u672c\u6587\u4fdd\u6301Patch\u5927\u5c0f\u4e0d\u53d8\uff0c\u5bfc\u81f4\u6700\u540e\u5f97\u5230\u7684Patch\u4e2a\u6570\u589e\u52a0\uff0c\u8fd9\u6837\u7684\u8bdd\u8fd9\u91cc\u7684Patch\u4e0ePosition embedding\u5904\u7406\u5f97\u5230\u7684Patch\u4e2a\u6570\u4e0d\u540c\uff0c\u4e0d\u80fd\u4fdd\u8bc1\u5176\u4e2d\u7684\u4f4d\u7f6e\u8bed\u4e49\u4fe1\u606f\u8fd8\u6709\u5de6\u53f3\uff0c\u56e0\u6b64\u53ef\u80fd\u5bfc\u81f4Position embedding\u5931\u53bb\u610f\u4e49\uff1b<\/p>\n\n\n\n

\u89e3\u51b3\u65b9\u6cd5\uff1a<\/strong>\u6839\u636e\u9884\u5148\u8bad\u7ec3\u597d\u7684\u4f4d\u7f6eembeding\u5728\u539f\u59cb\u56fe\u50cf\u4e2d\u7684\u4f4d\u7f6e\uff0c\u8fdb\u884c\u4e8c\u7ef4\u63d2\u503c\uff0c\u57fa\u4e8e\u539f\u56fe\u4e2d\u7684\u4f4d\u7f6e\u4fe1\u606f\uff0c\u5c06\u9884\u8bad\u7ec3\u4e2d\u7684 position embedding\u7684\u6570\u91cf\u7ecf\u8fc7\u63d2\u503c\u53d8\u6210\u589e\u52a0\u4e4b\u540e\u7684\u3002\u8fd9\u6837\u5728\u5f97\u5230\u5bf9\u5e94\u6570\u76ee\u7684position embedding\u7684\u540c\u65f6\u4e5f\u4fdd\u8bc1\u4e86position embedding\u7684\u8bed\u4e49\u4fe1\u606f\uff1b<\/p>\n\n\n\n

4 \u5b9e\u9a8c<\/h1>\n\n\n\n
\u7ffb\u8bd1<\/summary>\n

\u6211\u4eec\u8bc4\u4f30\u4e86 ResNet\u3001\u89c6\u89c9 Transformer\uff08ViT\uff09\u4ee5\u53ca\u6df7\u5408\u6a21\u578b\u7684\u8868\u5f81\u5b66\u4e60\u80fd\u529b\u3002\u4e3a\u4e86\u63a2\u7a76\u6bcf\u4e2a\u6a21\u578b\u5bf9\u6570\u636e\u7684\u9700\u6c42\uff0c\u6211\u4eec\u5728\u4e0d\u540c\u89c4\u6a21\u7684\u6570\u636e\u96c6\u4e0a\u8fdb\u884c\u9884\u8bad\u7ec3\uff0c\u5e76\u5728\u591a\u4e2a\u57fa\u51c6\u4efb\u52a1\u4e0a\u5c55\u5f00\u8bc4\u4f30\u3002\u4ece\u9884\u8bad\u7ec3\u7684\u8ba1\u7b97\u6210\u672c\u6765\u770b\uff0cViT \u8868\u73b0\u5341\u5206\u4f18\u5f02 \u2014\u2014 \u5b83\u80fd\u4ee5\u66f4\u4f4e\u7684\u9884\u8bad\u7ec3\u6210\u672c\uff0c\u5728\u5927\u591a\u6570\u8bc6\u522b\u57fa\u51c6\u4efb\u52a1\u4e0a\u8fbe\u5230\u6700\u5148\u8fdb\u6c34\u5e73\u3002\u6700\u540e\uff0c\u6211\u4eec\u5f00\u5c55\u4e86\u4e00\u9879\u5173\u4e8e\u81ea\u76d1\u7763\u5b66\u4e60\u7684\u5c0f\u578b\u5b9e\u9a8c\uff0c\u7ed3\u679c\u8868\u660e\u81ea\u76d1\u7763 ViT \u5728\u672a\u6765\u5177\u6709\u826f\u597d\u7684\u5e94\u7528\u524d\u666f\u3002<\/p>\n<\/details>\n\n\n\n

4.1 \u5b9e\u9a8c\u8bbe\u7f6e<\/h2>\n\n\n\n
\u7ffb\u8bd1<\/summary>\n

Datasets. \u63a2\u7d22\u6a21\u578b\u7684\u53ef\u4f38\u7f29\u6027\uff0c\u6211\u4eec\u4f7f\u7528ILSVRC-2012\u56fe\u50cf\u6570\u636e\u96c61k\u7c7b\u548c1.3M\u56fe\u50cf(\u6211\u4eec\u79f0\u4e4b\u4e3a\u201cImageNet\u5982\u4e0b)\uff0c\u5176\u8d85\u96c6ImageNet-21k 21k\u7c7b\u548c14M\u56fe\u50cf\uff08\u9093\u7b49\uff0c2009\uff09\uff0c\u548cJFT\uff08sun\u7b49\uff0c2017\uff0918k\u7c7b\u548c303\u9ad8\u5206\u8fa8\u7387\u56fe\u50cf\u3002\u6211\u4eec\u5c06\u8bad\u7ec3\u524d\u7684\u6570\u636e\u96c6w.r.t.\u5728Kolesnikov\u7b49\u4eba\uff082020\u5e74\uff09\u4e4b\u540e\u7684\u4e0b\u6e38\u4efb\u52a1\u7684\u6d4b\u8bd5\u96c6\u3002\u6211\u4eec\u5c06\u5728\u8fd9\u4e9b\u6570\u636e\u96c6\u4e0a\u8bad\u7ec3\u7684\u6a21\u578b\u8f6c\u79fb\u5230\u51e0\u4e2a\u57fa\u51c6\u4efb\u52a1\u4e2d\uff1a\u539f\u59cb\u9a8c\u8bc1\u6807\u7b7e\u548c\u6e05\u7406\u540e\u7684ReaL\u6807\u7b7e(Beyer\u7b49\u4eba\uff0c2020\u5e74)\uff0cCIFAR-10\/100\uff08\u514b\u91cc\u70ed\u592b\u65af\u57fa\uff0c2009\u5e74\uff09\uff0cOxford-IIIT Pets (Parkhi et al., 2012), and Oxford Flowers-102 (Nilsback & Zisserman, 2008)\u3002\u5bf9\u4e8e\u8fd9\u4e9b\u6570\u636e\u96c6\uff0c\u9884\u5904\u7406\u9075\u5faaKolesnikov\u7b49\u4eba\uff082020\u5e74\uff09\u3002<\/p>\n\n\n\n

\u6211\u4eec\u8fd8\u8bc4\u4f30\u4e86\u5305\u542b19\u4e2a\u4efb\u52a1\u7684VTAB\u5206\u7c7b\u5957\u4ef6(Zhai\u7b49\u4eba\uff0c2019b)\u3002VTAB\u4e3a\u6bcf\u4e2a\u4efb\u52a1\u4f7f\u75281 000 \u4e2a\u8bad\u7ec3\u793a\u4f8b\u3002\u8fd9\u4e9b\u4efb\u52a1\u88ab\u5206\u4e3a\u4e09\u7ec4\uff1aNatural \u2013 tasks \uff0c\u5982\u4e0a\u8ff0\u4efb\u52a1\uff0c\u5ba0\u7269\uff0cCIFAR\u7b49\u3002\u4e13\u4e1a\u7684\u533b\u7597\u548c\u536b\u661f\u56fe\u50cf\uff0c\u4ee5\u53ca\u7ed3\u6784\u5316\u7684\u4efb\u52a1\uff0c\u9700\u8981\u51e0\u4f55\u7406\u89e3\uff0c\u5982\u5c40\u90e8\u5316\u3002

Model Variants. \u6211\u4eec\u57fa\u4e8eBERT\u4e2d\u4f7f\u7528\u7684ViT\u914d\u7f6e(Devlinetal.\uff0c2019)\uff0c\u5982\u88681\u6240\u793a\u3002\u201cbase\u201d\u548c\u201cLarge\u201d\u6a21\u578b\u76f4\u63a5\u91c7\u7528\u4e86BERT\uff0c\u6211\u4eec\u6dfb\u52a0\u4e86\u66f4\u5927\u7684\u201chuge\u201d\u6a21\u578b\u3002\u5728\u63a5\u4e0b\u6765\u7684\u5185\u5bb9\u4e2d\uff0c\u6211\u4eec\u4f7f\u7528\u7b80\u77ed\u7684\u7b26\u53f7\u6765\u8868\u793a\u6a21\u578b\u5927\u5c0f\u548c\u8f93\u5165patch\u5927\u5c0f\uff1a\u4f8b\u5982\uff0cViT-L\/16\u8868\u793a\u5177\u670916\u00d716\u8f93\u5165patch\u5927\u5c0f\u7684\u201cLarge\u201d\u53d8\u4f53\u3002\u8bf7\u6ce8\u610f\uff0ctransformer\u7684\u5e8f\u5217\u957f\u5ea6\u4e0epatch\u5927\u5c0f\u7684\u5e73\u65b9\u6210\u53cd\u6bd4\uff0c\u56e0\u6b64\u5177\u6709\u8f83\u5c0fpatch\u5927\u5c0f\u7684\u6a21\u578b\u7684\u8ba1\u7b97\u6210\u672c\u66f4\u9ad8\u3002

\u5bf9\u4e8e\u57fa\u7ebfcnn\uff0c\u6211\u4eec\u4f7f\u7528ResNet(He\u7b49\u4eba\uff0c2016\u5e74)\uff0c\u4f46\u5c06\u6279\u5904\u7406\u5f52\u4e00\u5316\u5c42(Wu&He)\u66ff\u6362\u4e3a\u7ec4\u5f52\u4e00\u5316(Ioffe&\u03df\uff0c2015\u5e74)\uff0c\u5e76\u4f7f\u7528\u6807\u51c6\u5316\u5377\u79ef(Qiao\u7b49\u4eba\uff0c2019\u5e74)\u3002\u8fd9\u4e9b\u4fee\u6539\u6539\u5584\u4e86\u8f6c\u79fb(\u79d1\u5217\u65afnikov\u7b49\u4eba\uff0c2020\u5e74)\uff0c\u6211\u4eec\u5c06\u4fee\u6539\u540e\u7684\u6a21\u578b\u8868\u793a\u4e3a\u201cResNet(BiT)\u201d\u3002\u5bf9\u4e8e\u6df7\u5408\u4f53\uff0c\u6211\u4eec\u5c06\u4e2d\u95f4\u7279\u5f81\u6620\u5c04\u8f93\u5165\u5177\u6709\u4e00\u4e2a\u201c\u50cf\u7d20\u201d\u7684patch\u5927\u5c0f\u7684ViT\u3002\u4e3a\u4e86\u5b9e\u9a8c\u4e0d\u540c\u7684\u5e8f\u5217\u957f\u5ea6\uff0c\u6211\u4eec\u8981\u4e48(i)\u53d6\u5e38\u89c4ResNet50\u7684\u7b2c4\u9636\u6bb5\u7684\u8f93\u51fa\uff0c\u6216\u8005(ii)\u5220\u9664\u7b2c4\u9636\u6bb5\uff0c\u5728\u7b2c3\u9636\u6bb5\u4e2d\u653e\u7f6e\u76f8\u540c\u6570\u91cf\u7684\u5c42\uff08\u4fdd\u6301\u5c42\u7684\u603b\u6570\uff09\uff0c\u5e76\u53d6\u8fd9\u4e2a\u6269\u5c55\u9636\u6bb53\u7684\u8f93\u51fa\u3002\u9009\u9879(ii)\u5bfc\u81f4\u4e00\u4e2a4\u500d\u957f\u7684\u5e8f\u5217\u957f\u5ea6\uff0c\u548c\u4e00\u4e2a\u66f4\u6602\u8d35\u7684ViT\u6a21\u578b\u3002

Training & Fine-tuning. \u6211\u4eec\u4f7f\u7528Adam(Kingma&Ba\uff0c2015)\u548c\u03b21=0.9\uff0c\u03b22=0.999\uff0c\u6279\u5904\u7406\u5927\u5c0f\u4e3a4096\uff0c\u5e76\u5e94\u75280.1\u7684\u9ad8weight decay\uff0c\u6211\u4eec\u53d1\u73b0\u8fd9\u5bf9\u6240\u6709\u6a21\u578b\u7684\u8f6c\u79fb\u90fd\u5f88\u6709\u7528(\u9644\u5f55D.1\u8868\u660e\uff0c\u4e0e\u5e38\u89c1\u7684\u505a\u6cd5\u76f8\u6bd4\uff0cAdam\u5bf9resnet\u7684\u8bbe\u7f6e\u6548\u679c\u6bd4SGD\u7a0d\u597d)\u3002\u6211\u4eec\u4f7f\u7528\u7ebf\u6027\u5b66\u4e60\u7387\u9884\u70ed\u548c\u8870\u51cf\uff0c\u89c1\u9644\u5f55B.1\u7684\u7ec6\u8282\u3002\u5bf9\u4e8e\u5fae\u8c03\uff0c\u6211\u4eec\u4f7f\u7528\u5177\u6709\u52a8\u91cf\u7684SGD\uff0c\u6279\u5904\u7406\u5927\u5c0f\u4e3a512\uff0c\u5bf9\u4e8e\u6240\u6709\u6a21\u578b\uff0c\u8bf7\u53c2\u89c1\u9644\u5f55B.1.1\u3002\u5bf9\u4e8e\u88682\u4e2d\u7684ImageNet\u7ed3\u679c\uff0c\u6211\u4eec\u4ee5\u66f4\u9ad8\u7684\u5206\u8fa8\u7387\u8fdb\u884c\u4e86\u5fae\u8c03\uff1aViT-L\/16\u4e3a512\uff0cViT-H\/14\u4e3a518\uff0c\u540c\u65f6\u8fd8\u4f7f\u7528\u4e86Polyak&\u6731\u8fea\u8328\u57fa\uff081992\uff09\uff0c\u5e73\u5747\u7cfb\u6570\u4e3a0.9999\uff08\u62c9\u66fc\u94b1\u5fb7\u5170\u7b49\u4eba\uff0c2019\uff1bWang\u7b49\u4eba\uff0c2020b)\u3002

Metrics. \u6211\u4eec\u901a\u8fc7few-shot\u6216\u5fae\u8c03\u7cbe\u5ea6\u6765\u62a5\u544a\u4e0b\u6e38\u6570\u636e\u96c6\u7684\u7ed3\u679c\u3002\u5fae\u8c03\u7cbe\u5ea6\u53ef\u4ee5\u5728\u5bf9\u5404\u81ea\u7684\u6570\u636e\u96c6\u8fdb\u884c\u5fae\u8c03\u540e\u6355\u83b7\u6bcf\u4e2a\u6a21\u578b\u7684\u6027\u80fd\u3002few-shot\u7cbe\u5ea6\u662f\u901a\u8fc7\u6c42\u89e3\u4e00\u4e2a\u6b63\u5219\u5316\u6700\u5c0f\u4e8c\u4e58\u56de\u5f52\u95ee\u9898\u6765\u83b7\u5f97\u7684\uff0c\u8be5\u95ee\u9898\u5c06\u8bad\u7ec3\u56fe\u50cf\u5b50\u96c6\u7684\uff08\u51bb\u7ed3\uff09\u8868\u793a\u6620\u5c04\u5230{\u22121,1}^K\u4e2a\u76ee\u6807\u77e2\u91cf\u3002\u8fd9\u4e2a\u516c\u5f0f\u4f7f\u6211\u4eec\u80fd\u591f\u6062\u590d\u5c01\u95ed\u5f62\u5f0f\u7684\u7cbe\u786e\u89e3\u3002\u867d\u7136\u6211\u4eec\u4e3b\u8981\u5173\u6ce8\u5fae\u8c03\u6027\u80fd\uff0c\u4f46\u6211\u4eec\u6709\u65f6\u4f1a\u4f7f\u7528\u7ebf\u6027\u7684few-shot\u7cbe\u5ea6\u6765\u8fdb\u884c\u5feb\u901f\u7684\u52a8\u6001\u8bc4\u4f30\uff0c\u56e0\u4e3a\u5fae\u8c03\u6210\u672c\u592a\u9ad8\u3002<\/p>\n<\/details>\n\n\n\n

\u9884\u8bad\u7ec3\u6570\u636e\u96c6<\/strong>\uff1a<\/p>\n\n\n\n

ILSVRC-2012 ImageNet\uff08\u542b 1k \u7c7b\u522b\u3001130 \u4e07\u5f20\u56fe\u50cf\uff0c\u4e0b\u6587\u7b80\u79f0 ImageNet\uff09\uff1b<\/p>\n\n\n\n

\u5176\u8d85\u96c6 ImageNet-21k\uff08\u542b 21k \u7c7b\u522b\u30011400 \u4e07\u5f20\u56fe\u50cf\uff0cDeng \u7b49\u4eba\uff0c2009\uff09\uff1b<\/p>\n\n\n\n

JFT \u6570\u636e\u96c6\uff08\u542b 18k \u7c7b\u522b\u30013.03 \u4ebf\u5f20\u9ad8\u5206\u8fa8\u7387\u56fe\u50cf\uff0cSun \u7b49\u4eba\uff0c2017\uff09\u3002<\/p>\n\n\n\n

\u6a21\u578b\u8bc4\u4f30\u6570\u636e\u96c6\uff1a<\/strong><\/p>\n\n\n\n

ImageNet\uff08\u539f\u59cb\u9a8c\u8bc1\u6807\u7b7e\u53ca\u6e05\u7406\u540e\u7684 ReaL \u6807\u7b7e\uff0cBeyer \u7b49\u4eba\uff0c2020\uff09\uff1b<\/p>\n\n\n\n

CIFAR-10\/100\uff08Krizhevsky\uff0c2009\uff09\uff1b<\/p>\n\n\n\n

Oxford-IIIT Pets\uff08Parkhi \u7b49\u4eba\uff0c2012\uff09\uff1b<\/p>\n\n\n\n

Oxford Flowers-102\uff08Nilsback & Zisserman\uff0c2008\uff09\uff1b<\/p>\n\n\n\n

\u6a21\u578b\uff1a<\/strong><\/p>\n\n\n\n

\u8bba\u6587\u5171\u8bbe\u8ba1\u4e86Base\u3001Large\u548cHuge\u4e09\u6b3e\u4e0d\u540c\u5927\u5c0f\u7684ViT\u6a21\u578b\uff0c\u5206\u522b\u8868\u793a\u57fa\u7840\u6a21\u578b\u3001\u5927\u6a21\u578b\u548c\u8d85\u5927\u6a21\u578b\uff0c\u4e09\u6b3e\u6a21\u578b\u7684\u5404\u53c2\u6570\u5982\u4e0b\u8868\u6240\u793a\uff1a<\/p>\n\n\n\n

\"\"<\/div><\/figure>\n\n\n\n

\u57fa\u4e8eBERT\u4e2d\u4f7f\u7528\u7684ViT\u914d\u7f6e(Devlinetal.\uff0c2019)\uff0c\u201cbase\u201d\u548c\u201cLarge\u201d\u6a21\u578b\u76f4\u63a5\u91c7\u7528\u4e86BERT\uff0c\u53c8\u8bbe\u8ba1\u4e00\u4e2ahuge\u8d85\u5927\uff1a<\/p>\n\n\n\n

\u8bad\u7ec3\u3001\u5fae\u8c03\uff1a<\/strong><\/p>\n\n\n\n

Adam\u8bad\u7ec3\uff0c\u8bad\u7ec3\u6279\u6b214096\uff1b<\/p>\n\n\n\n

SGD\u5fae\u8c03\uff0c\u6279\u6b21512<\/p>\n\n\n\n

\u8861\u91cf\u6807\u51c6\uff1a<\/strong><\/p>\n\n\n\n

\u5c11\u6570\u955c\u5934\u6216\u5fae\u8c03\u7cbe\u5ea6\uff1b<\/p>\n\n\n\n

4.2 \u4e0e\u6700\u5148\u8fdb\u6a21\u578b\u7684\u5bf9\u6bd4<\/h2>\n\n\n\n
\u7ffb\u8bd1<\/summary>\n

\u6211\u4eec\u9996\u5148\u5c06\u6700\u5927\u89c4\u6a21\u7684\u6a21\u578b \u2014\u2014ViT-H\/14 \u548c ViT-L\/16\uff0c\u4e0e\u6587\u732e\u4e2d\u6700\u5148\u8fdb\u7684\u5377\u79ef\u795e\u7ecf\u7f51\u7edc\uff08CNNs\uff09\u8fdb\u884c\u5bf9\u6bd4\u3002\u5bf9\u6bd4\u5bf9\u8c61\u5305\u62ec\u4e24\u4e2a\uff1a\u4e00\u662f \u201cBig Transfer\uff08BiT\uff09\u201d\uff08Kolesnikov \u7b49\u4eba\uff0c2020\uff09\uff0c\u5b83\u57fa\u4e8e\u5927\u578b ResNet \u8fdb\u884c\u6709\u76d1\u7763\u8fc1\u79fb\u5b66\u4e60\uff1b\u4e8c\u662f \u201cNoisy Student\u201d\uff08Xie \u7b49\u4eba\uff0c2020\uff09\uff0c\u8fd9\u662f\u4e00\u4e2a\u5927\u578b EfficientNet\uff0c\u901a\u8fc7\u534a\u76d1\u7763\u5b66\u4e60\u5728 ImageNet \u548c\u53bb\u9664\u6807\u7b7e\u7684 JFT-300M \u6570\u636e\u96c6\u4e0a\u8bad\u7ec3\u3002\u76ee\u524d\uff0cNoisy Student \u662f ImageNet \u4e0a\u7684\u6700\u5148\u8fdb\u6a21\u578b\uff0c\u800c BiT-L \u5219\u5728\u672c\u6587\u62a5\u544a\u7684\u5176\u4ed6\u6570\u636e\u96c6\u4e0a\u8868\u73b0\u6700\u4f18\u3002\u6240\u6709\u6a21\u578b\u5747\u5728 TPUv3 \u786c\u4ef6\u4e0a\u8bad\u7ec3\uff0c\u6211\u4eec\u540c\u65f6\u62a5\u544a\u4e86\u6bcf\u4e2a\u6a21\u578b\u9884\u8bad\u7ec3\u6240\u9700\u7684 TPUv3 \u6838\u5fc3\u5929\u6570\uff08\u5373\u8bad\u7ec3\u6240\u7528\u7684 TPUv3 \u6838\u5fc3\u6570\uff08\u6bcf\u82af\u7247 2 \u4e2a\u6838\u5fc3\uff09\u4e58\u4ee5\u8bad\u7ec3\u5929\u6570\uff09\u3002<\/p>\n\n\n\n

\u8868 2 \u5c55\u793a\u4e86\u5bf9\u6bd4\u7ed3\u679c\u3002\u5728 JFT-300M \u6570\u636e\u96c6\u4e0a\u9884\u8bad\u7ec3\u7684\u8f83\u5c0f\u6a21\u578b ViT-L\/16\uff0c\u5728\u6240\u6709\u4efb\u52a1\u4e0a\u7684\u8868\u73b0\u5747\u4f18\u4e8e\u540c\u6837\u5728\u8be5\u6570\u636e\u96c6\u4e0a\u9884\u8bad\u7ec3\u7684 BiT-L\uff0c\u4e14\u8bad\u7ec3\u6240\u9700\u7684\u8ba1\u7b97\u8d44\u6e90\u663e\u8457\u66f4\u5c11\u3002\u66f4\u5927\u7684\u6a21\u578b ViT-H\/14 \u8fdb\u4e00\u6b65\u63d0\u5347\u4e86\u6027\u80fd\uff0c\u5c24\u5176\u5728\u66f4\u5177\u6311\u6218\u6027\u7684\u6570\u636e\u96c6\uff08ImageNet\u3001CIFAR-100 \u548c VTAB \u5957\u4ef6\uff09\u4e0a\u8868\u73b0\u7a81\u51fa\u3002\u6709\u8da3\u7684\u662f\uff0c\u8be5\u6a21\u578b\u7684\u9884\u8bad\u7ec3\u8ba1\u7b97\u91cf\u4ecd\u663e\u8457\u4f4e\u4e8e\u6b64\u524d\u7684\u6700\u5148\u8fdb\u6a21\u578b\u3002\u4e0d\u8fc7\u9700\u8981\u6ce8\u610f\u7684\u662f\uff0c\u9884\u8bad\u7ec3\u6548\u7387\u4e0d\u4ec5\u53d7\u67b6\u6784\u9009\u62e9\u5f71\u54cd\uff0c\u8fd8\u4e0e\u5176\u4ed6\u53c2\u6570\uff08\u5982\u8bad\u7ec3\u8c03\u5ea6\u3001\u4f18\u5316\u5668\u3001\u6743\u91cd\u8870\u51cf\u7b49\uff09\u76f8\u5173\u3002\u6211\u4eec\u5c06\u5728 4.4 \u8282\u4e2d\u5bf9\u4e0d\u540c\u67b6\u6784\u7684 \u201c\u6027\u80fd – \u8ba1\u7b97\u91cf\u201d \u5173\u7cfb\u8fdb\u884c\u5bf9\u7167\u7814\u7a76\u3002\u6700\u540e\uff0c\u5728\u516c\u5f00\u7684 ImageNet-21k \u6570\u636e\u96c6\u4e0a\u9884\u8bad\u7ec3\u7684 ViT-L\/16 \u6a21\u578b\u5728\u5927\u591a\u6570\u6570\u636e\u96c6\u4e0a\u4e5f\u8868\u73b0\u826f\u597d\uff0c\u4e14\u9884\u8bad\u7ec3\u8d44\u6e90\u66f4\u5c11\uff1a\u4f7f\u7528\u6807\u51c6\u7684 8 \u6838\u4e91 TPUv3\uff0c\u5927\u7ea6 30 \u5929\u5373\u53ef\u5b8c\u6210\u8bad\u7ec3\u3002<\/p>\n<\/details>\n\n\n\n

\"\"<\/div><\/figure>\n\n\n\n

\u88682\u662f\u5c06\u6700\u5927\u7684\u6a21\u578bViT-H\/14\u548cViT-L\/16\uff0c\u4e0eBiT\u548cNoisy Student\u8fdb\u884c\u6bd4\u8f83\uff1b<\/p>\n\n\n\n

\u7ed3\u8bba\uff1a<\/strong>\u6211\u4eec\u7684\u6a21\u578b\u8ba1\u7b97\u91cf\u8981\u5c0f\u5f97\u591a\uff0c\u5e76\u4e14\u7cbe\u5ea6\u66f4\u9ad8<\/p>\n\n\n\n

\"\"<\/div><\/figure>\n\n\n\n

\u56fe2\u5c06 VTAB \u4efb\u52a1\u5206\u89e3\u4e3a\u5404\u81ea\u7684\u7ec4\uff0c\u5e76\u5728\u6b64\u57fa\u51c6\u6d4b\u8bd5\u4e2d\u4e0e\u4e4b\u524d\u7684 SOTA \u65b9\u6cd5\u8fdb\u884c\u6bd4\u8f83<\/p>\n\n\n\n

\u7ed3\u8bba\uff1a<\/strong>ViT\u548cBiT\u7c7b\u4f3c\uff0c\u5e76\u9ad8\u4e8e\u522b\u7684\u65b9\u6cd5<\/p>\n\n\n\n

VTAB\u4efb\u52a1<\/strong>\u5373\u89c6\u89c9\u4efb\u52a1\u9002\u5e94\u6027\u57fa\u51c6\uff08Visual Task Adaptation Benchmark\uff09\u4efb\u52a1\uff0c\u5b83\u662fGoogle AI Research\u63d0\u51fa\u7684\u4e00\u4e2a\u7528\u4e8e\u8bc4\u4f30\u89c6\u89c9\u6a21\u578b\u8868\u793a\u80fd\u529b\u7684\u57fa\u51c6\u6d4b\u8bd5\uff0c\u65e8\u5728\u8861\u91cf\u6a21\u578b\u5728\u672a\u77e5\u4efb\u52a1\u548c\u6709\u9650\u6570\u636e\u4e0a\u7684\u6027\u80fd\uff0c\u63a8\u52a8\u901a\u7528\u89c6\u89c9\u7b97\u6cd5\u7684\u53d1\u5c55\uff1b<\/p>\n\n\n\n

VTAB\u4efb\u52a1\u76f8\u5173\u4ecb\u7ecd<\/summary>\n

\u4efb\u52a1\u6784\u6210<\/strong>\uff1aVTAB \u5305\u542b 19 \u4e2a\u8bc4\u4f30\u4efb\u52a1\uff0c\u8de8\u8d8a\u591a\u4e2a\u9886\u57df\uff0c\u5206\u4e3a\u81ea\u7136\uff08Natural\uff09\u3001\u4e13\u4e1a\uff08Specialized\uff09\u548c\u7ed3\u6784\u5316\uff08Structured\uff09\u4e09\u7ec4\u3002\u81ea\u7136\u56fe\u50cf\u4efb\u52a1\u5305\u62ec\u901a\u8fc7\u6807\u51c6\u76f8\u673a\u6355\u83b7\u7684\u81ea\u7136\u754c\u56fe\u50cf\uff0c\u4ee3\u8868\u901a\u7528\u5bf9\u8c61\u3001\u7ec6\u7c92\u5ea6\u7c7b\u6216\u62bd\u8c61\u6982\u5ff5\u7b49\uff1b\u4e13\u95e8\u4efb\u52a1\u5229\u7528\u4e13\u4e1a\u8bbe\u5907\u6355\u83b7\u7684\u56fe\u50cf\uff0c\u4f8b\u5982\u533b\u5b66\u56fe\u50cf\u6216\u9065\u611f\u56fe\u50cf\uff1b\u7ed3\u6784\u5316\u4efb\u52a1\u901a\u5e38\u6e90\u81ea\u65e8\u5728\u7406\u89e3\u56fe\u50cf\u4e4b\u95f4\u7279\u5b9a\u53d8\u5316\u7684\u4eba\u5de5\u73af\u5883\uff0c\u5982\u9884\u6d4b\u5230 3D \u573a\u666f\u4e2d\u67d0\u4e2a\u5bf9\u8c61\u7684\u8ddd\u79bb\u3001\u8ba1\u6570\u5bf9\u8c61\u6216\u68c0\u6d4b\u65b9\u5411\u7b49\u3002<\/p>\n\n\n\n

\u8bc4\u4f30\u65b9\u5f0f<\/strong>\uff1a\u8be5\u57fa\u51c6\u6d4b\u8bd5\u671f\u671b\u5c06\u7ecf\u8fc7\u9884\u8bad\u7ec3\u7684\u6a21\u578b\u4f5c\u4e3a\u8f93\u5165\uff0c\u7ed9\u5b9a\u6a21\u578b\u9488\u5bf9\u4e0d\u540c\u7684\u89c6\u89c9\u4efb\u52a1\u7ecf\u8fc7\u72ec\u7acb\u5fae\u8c03\uff0c\u6765\u89e3\u51b3\u4e0a\u8ff0\u4efb\u52a1\u3002\u4e3a\u4e86\u8bc4\u4f30\u5bf9\u6570\u636e\u6709\u9650\u7684\u65b0\u4efb\u52a1\u7684\u7b97\u6cd5\u6982\u62ec\u80fd\u529b\uff0c\u6bcf\u4e2a\u4efb\u52a1\u4ec5\u4f7f\u7528 1000 \u4e2a\u793a\u4f8b\u6765\u8bc4\u4f30\u6027\u80fd\uff0c\u6240\u6709\u4efb\u52a1\u7684\u5e73\u5747\u51c6\u786e\u5ea6\u7528\u4e8e\u8861\u91cf\u6a21\u578b\u7684\u6027\u80fd\u3002<\/p>\n\n\n\n

\u4efb\u52a1\u76ee\u7684<\/strong>\uff1aVTAB \u65e8\u5728\u4e3a\u89c6\u89c9\u8868\u793a\u5b66\u4e60\u63d0\u4f9b\u4e00\u4e2a\u7edf\u4e00\u7684\u8bc4\u4f30\u6807\u51c6\u3002\u901a\u8fc7\u8fd9\u4e2a\u57fa\u51c6\uff0c\u53ef\u4ee5\u5e2e\u52a9\u7814\u7a76\u4eba\u5458\u533a\u5206\u54ea\u4e9b\u6a21\u578b\u9002\u7528\u4e8e\u901a\u7528\u89c6\u89c9\u4efb\u52a1\uff0c\u54ea\u4e9b\u4e0d\u9002\u7528\uff0c\u4ece\u800c\u4e3a\u5c06\u6765\u7684\u89c6\u89c9\u8bc4\u4f30\u65b9\u5411\u63d0\u4f9b\u6307\u5f15\uff0c\u4fc3\u8fdb\u901a\u7528\u548c\u5b9e\u7528\u7684\u89c6\u89c9\u7b97\u6cd5\u7684\u53d1\u5c55\u3002<\/p>\n<\/details>\n\n\n\n

4.3 \u9884\u8bad\u7ec3\u6570\u636e\u9700\u6c42<\/h2>\n\n\n\n
\u7ffb\u8bd1<\/summary>\n

Vision Transformer \u5728\u5927\u89c4\u6a21 JFT-300M \u6570\u636e\u96c6\u4e0a\u9884\u8bad\u7ec3\u65f6\u8868\u73b0\u4f18\u5f02\u3002\u7531\u4e8e\u5176\u89c6\u89c9\u76f8\u5173\u7684\u5f52\u7eb3\u504f\u7f6e\u6bd4 ResNet \u66f4\u5c11\uff0c\u90a3\u4e48\u6570\u636e\u96c6\u5927\u5c0f\u5bf9\u5176\u5f71\u54cd\u6709\u591a\u5173\u952e\uff1f\u6211\u4eec\u901a\u8fc7\u4e24\u7ec4\u5b9e\u9a8c\u8fdb\u884c\u63a2\u7a76\u3002<\/p>\n\n\n\n

\u7b2c\u4e00\u7ec4\u5b9e\u9a8c\u4e2d\uff0c\u6211\u4eec\u5728\u89c4\u6a21\u9012\u589e\u7684\u6570\u636e\u96c6\uff08ImageNet\u3001ImageNet-21k\u3001JFT-300M\uff09\u4e0a\u9884\u8bad\u7ec3 ViT \u6a21\u578b\u3002\u4e3a\u63d0\u5347\u5728\u8f83\u5c0f\u6570\u636e\u96c6\u4e0a\u7684\u6027\u80fd\uff0c\u6211\u4eec\u4f18\u5316\u4e86\u4e09\u4e2a\u57fa\u7840\u6b63\u5219\u5316\u53c2\u6570 \u2014\u2014 \u6743\u91cd\u8870\u51cf\u3001dropout \u548c\u6807\u7b7e\u5e73\u6ed1\u3002\u56fe 3 \u5c55\u793a\u4e86\u5728 ImageNet \u4e0a\u5fae\u8c03\u540e\u7684\u7ed3\u679c\uff08\u5176\u4ed6\u6570\u636e\u96c6\u7684\u7ed3\u679c\u89c1\u8868 5\uff09\u3002\u5728\u6700\u5c0f\u7684\u6570\u636e\u96c6 ImageNet \u4e0a\u9884\u8bad\u7ec3\u65f6\uff0c\u5c3d\u7ba1\u4f7f\u7528\u4e86\uff08\u9002\u5ea6\u7684\uff09\u6b63\u5219\u5316\uff0cViT-Large \u6a21\u578b\u7684\u6027\u80fd\u4ecd\u4f4e\u4e8e ViT-Base \u6a21\u578b\uff1b\u5728 ImageNet-21k \u4e0a\u9884\u8bad\u7ec3\u65f6\uff0c\u4e24\u8005\u6027\u80fd\u63a5\u8fd1\uff1b\u53ea\u6709\u5728 JFT-300M \u4e0a\u9884\u8bad\u7ec3\u65f6\uff0c\u66f4\u5927\u6a21\u578b\u7684\u4f18\u52bf\u624d\u5b8c\u5168\u663e\u73b0\u3002\u56fe 3 \u8fd8\u5c55\u793a\u4e86\u4e0d\u540c\u89c4\u6a21 BiT \u6a21\u578b\u7684\u6027\u80fd\u8303\u56f4\uff1aBiT \u5377\u79ef\u795e\u7ecf\u7f51\u7edc\u5728 ImageNet \u4e0a\u7684\u8868\u73b0\u4f18\u4e8e ViT\uff0c\u4f46\u968f\u7740\u6570\u636e\u96c6\u89c4\u6a21\u589e\u5927\uff0cViT \u5b9e\u73b0\u4e86\u53cd\u8d85\u3002<\/p>\n\n\n\n

\u7b2c\u4e8c\u7ec4\u5b9e\u9a8c\u4e2d\uff0c\u6211\u4eec\u5728 JFT-300M \u7684\u968f\u673a\u5b50\u96c6\uff08900 \u4e07\u30013000 \u4e07\u30019000 \u4e07\u5f20\u56fe\u50cf\uff09\u4ee5\u53ca\u5b8c\u6574\u6570\u636e\u96c6\u4e0a\u8bad\u7ec3\u6a21\u578b\u3002\u6240\u6709\u8bbe\u7f6e\u5747\u4f7f\u7528\u76f8\u540c\u8d85\u53c2\u6570\uff0c\u672a\u5bf9\u5c0f\u5b50\u96c6\u989d\u5916\u65bd\u52a0\u6b63\u5219\u5316\uff0c\u4ee5\u6b64\u8bc4\u4f30\u6a21\u578b\u7684\u5185\u5728\u5c5e\u6027\u800c\u975e\u6b63\u5219\u5316\u7684\u5f71\u54cd\u3002\u4e0d\u8fc7\uff0c\u6211\u4eec\u91c7\u7528\u4e86\u65e9\u505c\u7b56\u7565\uff0c\u5e76\u62a5\u544a\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u8fbe\u5230\u7684\u6700\u4f73\u9a8c\u8bc1\u51c6\u786e\u7387\u3002\u4e3a\u8282\u7701\u8ba1\u7b97\u8d44\u6e90\uff0c\u6211\u4eec\u4f7f\u7528\u5c11\u6837\u672c\u7ebf\u6027\u51c6\u786e\u7387\u800c\u975e\u5b8c\u6574\u5fae\u8c03\u51c6\u786e\u7387\u4f5c\u4e3a\u8bc4\u4f30\u6307\u6807\uff0c\u7ed3\u679c\u5982\u56fe 4 \u6240\u793a\u3002\u5728\u8f83\u5c0f\u6570\u636e\u96c6\u4e0a\uff0c\u4e0e\u8ba1\u7b97\u6210\u672c\u76f8\u5f53\u7684 ResNet \u76f8\u6bd4\uff0cVision Transformer \u66f4\u5bb9\u6613\u8fc7\u62df\u5408\u3002\u4f8b\u5982\uff0cViT-B\/32 \u7684\u8bad\u7ec3\u901f\u5ea6\u7565\u5feb\u4e8e ResNet50\uff0c\u5728 900 \u4e07\u56fe\u50cf\u5b50\u96c6\u4e0a\u8868\u73b0\u5dee\u5f88\u591a\uff0c\u4f46\u5728 9000 \u4e07\u4ee5\u4e0a\u5b50\u96c6\u4e0a\u8868\u73b0\u66f4\u4f18\uff1bResNet152x2 \u4e0e ViT-L\/16 \u7684\u5bf9\u6bd4\u4e5f\u5448\u73b0\u76f8\u540c\u89c4\u5f8b\u3002\u8fd9\u4e00\u7ed3\u679c\u5370\u8bc1\u4e86\u4e00\u4e2a\u76f4\u89c9\uff1a\u5377\u79ef\u5f52\u7eb3\u504f\u7f6e\u5728\u8f83\u5c0f\u6570\u636e\u96c6\u4e0a\u66f4\u6709\u7528\uff0c\u4f46\u5bf9\u4e8e\u66f4\u5927\u7684\u6570\u636e\u96c6\uff0c\u76f4\u63a5\u4ece\u6570\u636e\u4e2d\u5b66\u4e60\u76f8\u5173\u6a21\u5f0f\u5c31\u8db3\u591f\u4e86\uff0c\u751a\u81f3\u66f4\u6709\u5229\u3002<\/p>\n\n\n\n

\u603b\u4f53\u800c\u8a00\uff0cImageNet \u4e0a\u7684\u5c11\u6837\u672c\u7ed3\u679c\uff08\u56fe 4\uff09\u4ee5\u53ca VTAB \u4e0a\u7684\u4f4e\u6570\u636e\u91cf\u7ed3\u679c\uff08\u8868 2\uff09\uff0c\u8868\u660e ViT \u5728\u6781\u4f4e\u6570\u636e\u91cf\u8fc1\u79fb\u573a\u666f\u4e2d\u5177\u6709\u6f5c\u529b\u3002\u5bf9 ViT \u5c11\u6837\u672c\u7279\u6027\u7684\u8fdb\u4e00\u6b65\u5206\u6790\uff0c\u5c06\u662f\u672a\u6765\u4ee4\u4eba\u671f\u5f85\u7684\u7814\u7a76\u65b9\u5411\u3002<\/p>\n<\/details>\n\n\n\n

\"\"<\/div><\/figure>\n\n\n\n

\u56fe3\u5c55\u793a\u4e86\u6a21\u578b\u5728 ImageNet \u6570\u636e\u96c6\u4e0a\u7684\u6027\u80fd\uff1b<\/p>\n\n\n\n

\u56fe4\u5c55\u793a\u4e86\u5728 JFT300M \u6570\u636e\u96c6\u7684\u968f\u673a\u5b50\u96c6\u4ee5\u53ca\u5b8c\u6574\u6570\u636e\u96c6\u4e0a\u8fdb\u884c\u4e86\u6a21\u578b\u8bad\u7ec3\u7684\u7ed3\u679c\uff1b<\/p>\n\n\n\n

\u7ed3\u8bba\uff1a<\/strong>\u5377\u79ef\u5f52\u7eb3\u504f\u7f6e\u5bf9\u4e8e\u89c4\u6a21\u8f83\u5c0f\u7684\u6570\u636e\u96c6\u8f83\u4e3a\u6709\u7528\uff0c\u4f46\u5bf9\u4e8e\u8f83\u5927\u7684\u6570\u636e\u96c6\u800c\u8a00\uff0c\u5b66\u4e60\u76f8\u5173\u6a21\u5f0f\u5c31\u8db3\u591f\u4e86\uff0c\u751a\u81f3\u66f4\u52a0\u6709\u6548\u3002<\/p>\n\n\n\n

4.4 \u6a21\u578b\u6269\u5c55\u7814\u7a76<\/h2>\n\n\n\n
\u7ffb\u8bd1<\/summary>\n

\u6211\u4eec\u901a\u8fc7\u5bf9\u4e0d\u540c\u6a21\u578b\u8fdb\u884c\u4e86\u53d7\u63a7\u7f29\u653e\u7814\u7a76\u8bc4\u4f30 JFT-300M \u7684\u8fc1\u79fb\u6027\u80fd\u3002 \u5728\u8fd9\u79cd\u60c5\u51b5\u4e0b\uff0c\u6570\u636e\u5927\u5c0f\u4e0d\u4f1a\u9650\u5236\u6a21\u578b\u7684\u6027\u80fd\uff0c\u6211\u4eec\u8bc4\u4f30\u6bcf\u4e2a\u6a21\u578b\u7684\u6027\u80fd\u4e0e\u9884\u8bad\u7ec3\u6210\u672c\u3002 \u6a21\u578b\u96c6\u5305\u62ec\uff1a7\u4e2aResNets\uff0cR50x1\u3001R50x2\u3001R101x1\u3001R152x1\u3001R152x2\uff0c\u9884\u8bad\u7ec3\u4e867\u4e2aepochs\uff0c\u52a0\u4e0a\u9884\u8bad\u7ec3\u4e8614\u4e2aepochs\u7684R152x2\u548cR200x3\uff1b 6\u4e2aVision Transformers\uff0cViT-B\/32\u3001B\/16\u3001L\/32\u3001L\/16\uff0c\u9884\u8bad\u7ec37\u4e2aepochs\uff0c\u52a0\u4e0aL\/16\u548cH\/14\u9884\u8bad\u7ec314\u4e2aepochs\uff1b \u548c 5 \u4e2a\u6df7\u5408\uff0cR50+ViT B\/32\u3001B\/16\u3001L\/32\u3001L\/16 \u9884\u8bad\u7ec3\u4e86 7 \u4e2a epoch\uff0c\u52a0\u4e0a R50+ViT-L\/16 \u9884\u8bad\u7ec3\u4e86 14 \u4e2a epoch\uff08\u5bf9\u4e8e\u6df7\u5408\uff0c\u6700\u540e\u7684\u6570\u5b57 \u6a21\u578b\u540d\u79f0\u4e0d\u4ee3\u8868patch\u5927\u5c0f\uff0c\u800c\u662f\u4ee3\u8868 ResNet \u4e3b\u5e72\u4e2d\u7684\u603b\u4e0b\u91c7\u6837\u7387\uff09\u3002<\/p>\n\n\n\n

\u56fe 5 \u5305\u542b\u8fc1\u79fb\u6027\u80fd\u4e0e\u603b\u9884\u8bad\u7ec3\u8ba1\u7b97\uff08\u6709\u5173\u8ba1\u7b97\u6210\u672c\u7684\u8be6\u7ec6\u4fe1\u606f\uff0c\u8bf7\u53c2\u89c1\u9644\u5f55 D.5\uff09\u3002 \u6bcf\u4e2a\u6a21\u578b\u7684\u8be6\u7ec6\u7ed3\u679c\u5728\u9644\u5f55\u7684\u8868 6 \u4e2d\u63d0\u4f9b\u3002 \u53ef\u4ee5\u89c2\u5bdf\u5230\u4e00\u4e9b\u6a21\u5f0f\u3002 \u9996\u5148\uff0cvision transformer\u5728\u6027\u80fd\/\u8ba1\u7b97\u6743\u8861\u65b9\u9762\u4e3b\u5bfc\u4e86 ResNet\u3002 ViT \u4f7f\u7528\u5927\u7ea6 \u5c11\u4e862 – 4 \u500d\u7684\u8ba1\u7b97\u6765\u83b7\u5f97\u76f8\u540c\u7684\u6027\u80fd\uff085 \u4e2a\u6570\u636e\u96c6\u7684\u5e73\u5747\u503c\uff09\u3002 \u5176\u6b21\uff0c\u6df7\u5408\u6a21\u578b\u5728\u8f83\u5c0f\u7684\u8ba1\u7b97\u9884\u7b97\u4e0b\u7565\u80dc\u4e8e ViT\uff0c\u4f46\u5bf9\u4e8e\u8f83\u5927\u7684\u6a21\u578b\uff0c\u8fd9\u79cd\u5dee\u5f02\u6d88\u5931\u4e86\u3002 \u8fd9\u4e2a\u7ed3\u679c\u6709\u4e9b\u51fa\u4eba\u610f\u6599\uff0c\u56e0\u4e3a\u4eba\u4eec\u53ef\u80fd\u671f\u671b\u5377\u79ef\u5c40\u90e8\u7279\u5f81\u5904\u7406\u80fd\u591f\u5728\u4efb\u4f55\u89c4\u6a21\u4e0b\u8f85\u52a9 ViT\u3002 \u7b2c\u4e09\uff0cVision Transformers \u4f3c\u4e4e\u4e0d\u4f1a\u5728\u5c1d\u8bd5\u7684\u8303\u56f4\u5185\u9971\u548c\uff0c\u4ece\u800c\u63a8\u52a8\u672a\u6765\u7684\u6269\u5c55\u5de5\u4f5c\u3002<\/p>\n\n\n\n

\u5176\u6b21\uff0c\u6211\u4eec\u57289M\u300130M\u548c90M\u7684\u968f\u673a\u5b50\u96c6\u4ee5\u53ca\u5b8c\u6574\u7684JFT-300M\u6570\u636e\u96c6\u4e0a\u8bad\u7ec3\u6211\u4eec\u7684\u6a21\u578b\u3002\u6211\u4eec\u4e0d\u5bf9\u8f83\u5c0f\u7684\u5b50\u96c6\u6267\u884c\u989d\u5916\u7684\u6b63\u5219\u5316\uff0c\u800c\u662f\u5bf9\u6240\u6709\u8bbe\u7f6e\u4f7f\u7528\u76f8\u540c\u7684\u8d85\u53c2\u6570\u3002\u8fd9\u6837\uff0c\u6211\u4eec\u5c31\u53ef\u4ee5\u8bc4\u4f30\u6a21\u578b\u7684\u5185\u5728\u6027\u8d28\uff0c\u800c\u4e0d\u662f\u6b63\u5219\u5316\u7684\u5f71\u54cd\u3002\u7136\u800c\uff0c\u6211\u4eec\u786e\u5b9e\u4f7f\u7528\u4e86early stoping\uff0c\u5e76\u62a5\u544a\u4e86\u5728\u8bad\u7ec3\u671f\u95f4\u83b7\u5f97\u7684\u6700\u4f73\u9a8c\u8bc1\u51c6\u786e\u6027\u3002\u4e3a\u4e86\u8282\u7701\u8ba1\u7b97\u91cf\uff0c\u6211\u4eec\u62a5\u544a\u4e86Few-shot\u7684\u7ebf\u6027\u7cbe\u5ea6\uff0c\u800c\u4e0d\u662f\u5168\u5fae\u8c03\u7cbe\u5ea6\u3002\u56fe4\u5305\u542b\u4e86\u8fd9\u4e9b\u7ed3\u679c\u3002\u5728\u8f83\u5c0f\u7684\u6570\u636e\u96c6\u4e0a\uff0cVision Transformer\u6bd4ResNets\u8fc7\u62df\u5408\u66f4\u591a\u3002\u4f8b\u5982\uff0cViT-B\/32\u6bd4ResNet50\u7565\u5feb\uff1b\u5b83\u57289M\u5b50\u96c6\u4e0a\u8868\u73b0\u5f97\u66f4\u5dee\uff0c\u4f46\u572890M+\u5b50\u96c6\u4e0a\u8868\u73b0\u5f97\u66f4\u597d\u3002ResNet152x2\u548cViT-L\/16\u4e5f\u662f\u5982\u6b64\u3002\u8fd9\u4e00\u7ed3\u679c\u5f3a\u5316\u4e86\u4e00\u79cd\u76f4\u89c9\uff0c\u5373\u5377\u79ef\u5f52\u7eb3\u504f\u5dee\u5bf9\u4e8e\u8f83\u5c0f\u7684\u6570\u636e\u96c6\u662f\u6709\u7528\u7684\uff0c\u4f46\u5bf9\u4e8e\u8f83\u5927\u7684\u6570\u636e\u96c6\uff0c\u76f4\u63a5\u4ece\u6570\u636e\u4e2d\u5b66\u4e60\u76f8\u5173\u6a21\u5f0f\u5c31\u8db3\u591f\u4e86\uff0c\u751a\u81f3\u662f\u6709\u76ca\u7684\u3002<\/p>\n\n\n\n

\u603b\u7684\u6765\u8bf4\uff0cImageNet\u4e0a\u7684Few-shot\u7ed3\u679c\uff08\u56fe4\uff09\uff0c\u4ee5\u53caVTAB\u4e0a\u7684\u4f4e\u6570\u636e\u7ed3\u679c\uff08\u88682\uff09\u4f3c\u4e4e\u5f88\u9002\u5408\u975e\u5e38\u4f4e\u7684\u6570\u636e\u4f20\u8f93\u3002\u8fdb\u4e00\u6b65\u5206\u6790ViT\u7684Few-shot\u7279\u6027\u662f\u672a\u6765\u5de5\u4f5c\u7684\u4e00\u4e2a\u4ee4\u4eba\u5174\u594b\u7684\u65b9\u5411\u3002 <\/p>\n<\/details>\n\n\n\n

\"\"<\/div><\/figure>\n\n\n\n
\"\"<\/div><\/figure>\n\n\n\n

\u56fe5\u5c55\u793a\u4e86\u6a21\u578b\u5728\u4e0d\u540c\u9884\u8bad\u7ec3\u8ba1\u7b97\u6210\u672c\u60c5\u51b5\u4e0b\u7684\u8fc1\u79fb\u6027\u80fd<\/p>\n\n\n\n

\u88686\u5c55\u793a\u4e86\u6a21\u578b\u7f29\u653e\u5b9e\u9a8c\u7684\u8be6\u7ec6\u7ed3\u679c\uff1b<\/p>\n\n\n\n

\u7ed3\u8bba\uff1a<\/strong>ViT\u5728\u6027\u80fd \/ \u7b97\u529b\u6743\u8861\u4e2d\u663e\u8457\u4f18\u4e8e ResNet\u3002<\/p>\n\n\n\n

\u7f29\u653e\u5b9e\u9a8c<\/summary>\n

\u7406\u89e3\u6a21\u578b\u5728\u4e0d\u540c\u89c4\u6a21\u4e0b\u7684\u884c\u4e3a\u548c\u6027\u80fd\u53d8\u5316\uff0c\u4ece\u800c\u6307\u5bfc\u6a21\u578b\u8bbe\u8ba1\u548c\u9009\u62e9\u5408\u9002\u7684\u6a21\u578b\u914d\u7f6e\u3001\u5bfb\u627e\u6a21\u578b\u6027\u80fd\u3001\u8ba1\u7b97\u6210\u672c\u548c\u6570\u636e\u9700\u6c42\u4e4b\u95f4\u7684\u6700\u4f73\u5e73\u8861\u70b9\uff0c\u4ee5\u63d0\u9ad8\u6a21\u578b\u7684\u6548\u7387\u548c\u6cdb\u5316\u80fd\u529b\uff1b<\/p>\n<\/details>\n\n\n\n

4.5 Vision Transformer \u5185\u90e8\u673a\u5236\u5206\u6790<\/h2>\n\n\n\n
\u7ffb\u8bd1<\/summary>\n

\u4e3a\u4e86\u5f00\u59cb\u7406\u89e3\u89c6\u89c9\u8f6c\u6362\u5668\u5982\u4f55\u5904\u7406\u56fe\u50cf\u6570\u636e\uff0c\u6211\u4eec\u5206\u6790\u4e86\u5b83\u7684\u5185\u90e8\u8868\u793a\u3002Vision Transformer\u7684\u7b2c\u4e00\u5c42\u7ebf\u6027\u5730\u6295\u5c04\u4e86\u6241\u5e73\u7684patch\u5230\u4e00\u4e2a\u4f4e\u7ef4\u7a7a\u95f4(Eq. 1).\u56fe7\uff08\u5de6\uff09\u663e\u793a\u4e86\u5b66\u4e60\u5230\u7684embeding\u8fc7\u6ee4\u5668\u7684\u9876\u90e8\u4e3b\u6210\u5206\u3002\u8fd9\u4e9b\u7ec4\u4ef6\u7c7b\u4f3c\u4e8e\u6bcf\u4e2apatch\u5185\u7cbe\u7ec6\u7ed3\u6784\u7684\u4f4e\u7ef4\u8868\u793a\u7684\u53ef\u4fe1\u57fa\u51fd\u6570\u3002<\/p>\n\n\n\n

\u6295\u5f71\u7ed3\u675f\u540e\uff0c\u5c06\u5b66\u4e60\u5230\u7684\u4f4d\u7f6eembeding\u6dfb\u52a0\u5230patch\u8868\u793a\u4e2d\u3002\u4ece\u56fe7\uff08\u4e2d\uff09\u53ef\u4ee5\u770b\u51fa\uff0c\u6a21\u578b\u901a\u8fc7\u4f4d\u7f6eembeding\u7684\u76f8\u4f3c\u6027\u6765\u5b66\u4e60\u5bf9\u56fe\u50cf\u5185\u7684\u8ddd\u79bb\u8fdb\u884c\u7f16\u7801\uff0c\u5373\u8d8a\u8fd1\u7684\u6591\u5757\u5f80\u5f80\u5177\u6709\u66f4\u591a\u76f8\u4f3c\u7684\u4f4d\u7f6eembeding\u3002\u6b64\u5916\uff0c\u8fd8\u4f1a\u51fa\u73b0\u884c-\u5217\u7ed3\u6784\uff1b\u540c\u4e00\u884c\/\u5217\u4e2d\u7684patch\u5177\u6709\u7c7b\u4f3c\u7684embeding\u3002\u6700\u540e\uff0c\u5bf9\u4e8e\u8f83\u5927\u7684\u7f51\u683c\uff0c\u6b63\u5f26\u7ed3\u6784\u6709\u65f6\u662f\u660e\u663e\u7684(\u9644\u5f55D)\u3002\u4f4d\u7f6eembeding\u5b66\u4e60\u4e86\u8868\u793a\u4e8c\u7ef4\u56fe\u50cf\u62d3\u6251\uff0c\u8fd9\u89e3\u91ca\u4e86\u4e3a\u4ec0\u4e48\u624b\u5de5\u5236\u4f5c\u7684\u4e8c\u7ef4\u611f\u77e5\u5d4c\u5165\u53d8\u4f53\u6ca1\u6709\u4ea7\u751f\u6539\u8fdb(\u9644\u5f55D.4)\u3002

\u81ea\u6211\u6ce8\u610f\u5141\u8bb8ViT\u6574\u5408\u6574\u4e2a\u56fe\u50cf\u7684\u4fe1\u606f\uff0c\u5373\u4f7f\u662f\u5728\u6700\u4f4e\u7684\u5c42\u3002\u6211\u4eec\u5c06\u7814\u7a76\u8be5\u7f51\u7edc\u5728\u591a\u5927\u7a0b\u5ea6\u4e0a\u5229\u7528\u4e86\u8fd9\u79cd\u80fd\u529b\u3002\u5177\u4f53\u6765\u8bf4\uff0c\u6211\u4eec\u6839\u636e\u6ce8\u610f\u529b\u6743\u91cd\u8ba1\u7b97\u4fe1\u606f\u5728\u88ab\u6574\u5408\u8fc7\u7684\u56fe\u50cf\u7a7a\u95f4\u4e2d\u7684\u5e73\u5747\u8ddd\u79bb\uff08\u56fe7\uff0c\u53f3\uff09\u3002\u8fd9\u79cd\u201c\u6ce8\u610f\u8ddd\u79bb\u201d\u7c7b\u4f3c\u4e8ecnn\u4e2d\u7684\u63a5\u53d7\u91ce\u5927\u5c0f\u3002\u6211\u4eec\u53d1\u73b0\uff0c\u4e00\u4e9b\u5934\u90e8\u5173\u6ce8\u4e86\u5df2\u7ecf\u5728\u6700\u4f4e\u5c42\u7684\u5927\u90e8\u5206\u56fe\u50cf\uff0c\u8fd9\u8868\u660e\u6a21\u578b\u786e\u5b9e\u4f7f\u7528\u4e86\u5168\u5c40\u96c6\u6210\u4fe1\u606f\u7684\u80fd\u529b\u3002\u5176\u4ed6\u7684\u6ce8\u610f\u529b\u5934\u8111\u5728\u4f4e\u5c42\u7684\u6ce8\u610f\u8ddd\u79bb\u4e00\u76f4\u5f88\u5c0f\u3002\u8fd9\u79cd\u9ad8\u5ea6\u5c40\u90e8\u5316\u7684\u6ce8\u610f\u5728transformer\u4e4b\u524d\u5e94\u7528ResNet\u7684\u6df7\u5408\u6a21\u578b\u4e2d\u4e0d\u90a3\u4e48\u660e\u663e\uff08\u56fe7\uff0c\u53f3\uff09\uff0c\u8fd9\u8868\u660e\u5b83\u53ef\u80fd\u5728cnn\u4e2d\u5177\u6709\u4e0e\u65e9\u671f\u5377\u79ef\u5c42\u7c7b\u4f3c\u7684\u529f\u80fd\u3002\u6b64\u5916\uff0c\u6ce8\u610f\u8ddd\u79bb\u968f\u7740\u7f51\u7edc\u6df1\u5ea6\u7684\u589e\u52a0\u800c\u589e\u52a0\u3002\u5728\u5168\u7403\u8303\u56f4\u5185\uff0c\u6211\u4eec\u53d1\u73b0\u8be5\u6a21\u578b\u5173\u6ce8\u7684\u662f\u4e0e\u5206\u7c7b\u6709\u8bed\u4e49\u76f8\u5173\u7684\u56fe\u50cf\u533a\u57df\uff08\u56fe6\uff09\u3002

<\/p>\n<\/details>\n\n\n\n

\u56fe7\uff1a\u5de6\uff1aViT-L\/32 RGB \u503c\u7684\u521d\u59cb\u7ebf\u6027embedding\u6ee4\u6ce2\u5668\u3002 \u4e2d\uff1aViT-L\/32 \u7684position embedding\u7684\u76f8\u4f3c\u6027\u3002 \u53f3\uff1a\u6309heads\u548c\u7f51\u7edc\u6df1\u5ea6\u5212\u5206\u7684\u53c2\u4e0e\u533a\u57df\u5927\u5c0f\u3002<\/p>\n\n\n\n

\"\"<\/div><\/figure>\n\n\n\n

\u56fe6\uff1a\u4ece\u8f93\u51fatoken\u5230\u8f93\u5165\u7a7a\u95f4\u7684\u6ce8\u610f\u529b\u7684\u4ee3\u8868\u6027\u793a\u4f8b<\/p>\n\n\n\n

\"\"<\/div><\/figure>\n\n\n\n

\u7ed3\u8bba\uff1a<\/strong><\/p>\n\n\n\n

1\u3001\u6a21\u578b\u4f7f\u7528\u4e86\u5168\u5c40\u96c6\u6210\u4fe1\u606f\u7684\u80fd\u529b\u3002\u5176\u4ed6\u6ce8\u610f\u529bhead\u5728\u4f4e\u5c42\u4e2d\u59cb\u7ec8\u5177\u6709\u8f83\u5c0f\u7684\u6ce8\u610f\u529b\u8ddd\u79bb\uff1b<\/p>\n\n\n\n

2\u3001\u8be5\u6a21\u578b\u5173\u6ce8\u4e0e\u5206\u7c7b\u8bed\u4e49\u76f8\u5173\u7684\u56fe\u50cf\u533a\u57df\u3002<\/p>\n\n\n\n

4.6 \u81ea\u76d1\u7763\u5b66\u4e60<\/h2>\n\n\n\n
\u7ffb\u8bd1<\/summary>\n

transformer\u5728NLP\u4efb\u52a1\u4e0a\u7684\u8868\u73b0\u51fa\u4e86\u4ee4\u4eba\u5370\u8c61\u6df1\u523b\u7684\u8868\u73b0\u3002\u7136\u800c\uff0c\u4ed6\u4eec\u7684\u6210\u529f\u4e0d\u4ec5\u6765\u81ea\u4e8e\u4ed6\u4eec\u4f18\u79c0\u7684\u53ef\u4f38\u7f29\u6027\uff0c\u800c\u4e14\u8fd8\u6765\u81ea\u4e8e\u5927\u89c4\u6a21\u7684\u81ea\u6211\u76d1\u7763\u7684\u9884\u8bad\u7ec3.\u6211\u4eec\u8fd8\u6a21\u62df\u4e86BERT\u4e2d\u4f7f\u7528\u7684\u63a9\u7801\u8bed\u8a00\u5efa\u6a21\u4efb\u52a1\uff0c\u5bf9\u81ea\u6211\u76d1\u7763\u7684\u63a9\u7801patch\u9884\u6d4b\u8fdb\u884c\u4e86\u521d\u6b65\u7684\u63a2\u7d22\u3002\u901a\u8fc7\u81ea\u6211\u76d1\u7763\u9884\u8bad\u7ec3\uff0c\u6211\u4eec\u8f83\u5c0f\u7684ViT-B\/16\u6a21\u578b\u5728ImageNet\u4e0a\u8fbe\u5230\u4e8679.9%\u7684\u51c6\u786e\u7387\uff0c\u6bd4\u4ece\u5934\u5f00\u59cb\u8bad\u7ec3\u663e\u8457\u63d0\u9ad8\u4e862%\uff0c\u4f46\u4ecd\u6bd4\u76d1\u7763\u9884\u8bad\u7ec3\u4f4e4%\u3002\u9644\u5f55B.1.2\u5305\u542b\u4e86\u66f4\u591a\u7684\u7ec6\u8282\u3002\u6211\u4eec\u5c06\u5bf9\u6bd4\u9884\u8bad\u7ec3\u7684\u63a2\u7d22\uff08Chen\u7b49\u4eba\uff0c2020b\uff1b\u4ed6\u7b49\u4eba\uff0c2020\u5e74\uff1bBachman\u7b49\u4eba\uff0c2019\u5e74\uff1bH\u2018enaff\u7b49\u4eba\uff0c2020\u5e74\uff09\u7559\u7ed9\u672a\u6765\u7684\u5de5\u4f5c\u3002<\/p>\n<\/details>\n\n\n\n

\u4f5c\u8005\u6a21\u4eff\u4e86 BERT \u4e2d\u4f7f\u7528\u7684masked\u8bed\u8a00\u5efa\u6a21\u4efb\u52a1\uff0c\u6bd4\u4ece\u5934\u8bad\u7ec3\u7cbe\u786e\u5ea6\u9ad8\uff0c\u4f46\u6548\u679c\u8fd8\u4e0d\u5982\u76d1\u7763\u8bad\u7ec3\uff0c\u8fd9\u5c31\u7559\u7ed9\u672a\u6765\u7684\u63a2\u7d22\u4e86\u3002<\/p>\n\n\n\n

5 \u7ed3\u8bba<\/h1>\n\n\n\n
\u7ffb\u8bd1<\/summary>\n

\u6211\u4eec\u63a2\u7d22\u4e86transformer\u5728\u56fe\u50cf\u8bc6\u522b\u4e2d\u7684\u76f4\u63a5\u5e94\u7528\u3002\u4e0e\u4e4b\u524d\u5728\u8ba1\u7b97\u673a\u89c6\u89c9\u4e2d\u4f7f\u7528\u81ea\u6ce8\u610f\u7684\u5de5\u4f5c\u4e0d\u540c\uff0c\u9664\u4e86\u6700\u521d\u7684patch\u63d0\u53d6\u6b65\u9aa4\u5916\uff0c\u6211\u4eec\u6ca1\u6709\u5728\u67b6\u6784\u4e2d\u5f15\u5165\u7279\u5b9a\u4e8e\u56fe\u50cf\u7684\u5f52\u7eb3\u504f\u5dee\u3002\u76f8\u53cd\uff0c\u6211\u4eec\u5c06\u56fe\u50cf\u89e3\u91ca\u4e3a\u4e00\u7cfb\u5217patch\uff0c\u5e76\u4f7f\u7528NLP\u4e2d\u4f7f\u7528\u7684\u6807\u51c6\u53d8\u538b\u5668\u7f16\u7801\u5668\u5bf9\u5176\u8fdb\u884c\u5904\u7406\u3002\u8fd9\u79cd\u7b80\u5355\u4f46\u53ef\u6269\u5c55\u7684\u7b56\u7565\u5728\u7ed3\u5408\u5927\u578b\u6570\u636e\u96c6\u4e0a\u7684\u9884\u8bad\u7ec3\u65f6\uff0c\u5de5\u4f5c\u5f97\u60ca\u4eba\u5730\u597d\u3002\u56e0\u6b64\uff0cVision Transformer\u5728\u8bb8\u591a\u56fe\u50cf\u5206\u7c7b\u6570\u636e\u96c6\u4e0a\u5339\u914d\u6216\u8d85\u8fc7\u4e86\u6700\u5148\u8fdb\u7684\u72b6\u6001\uff0c\u800c\u662f\u76f8\u5bf9\u4fbf\u5b9c\u7684\u9884\u8bad\u7ec3\u3002<\/p>\n\n\n\n

\u867d\u7136\u8fd9\u4e9b\u521d\u6b65\u7ed3\u679c\u4ee4\u4eba\u9f13\u821e\uff0c\u4f46\u4ecd\u5b58\u5728\u8bb8\u591a\u6311\u6218\u3002\u4e00\u662f\u5c06ViT\u5e94\u7528\u4e8e\u5176\u4ed6\u8ba1\u7b97\u673a\u89c6\u89c9\u4efb\u52a1\uff0c\u5982\u68c0\u6d4b\u548c\u5206\u5272\u3002\u6211\u4eec\u7684\u7814\u7a76\u7ed3\u679c\uff0c\u52a0\u4e0aCarion\u7b49\u4eba\uff082020\u5e74\uff09\u7684\u7814\u7a76\u7ed3\u679c\uff0c\u8868\u660e\u4e86\u8fd9\u79cd\u65b9\u6cd5\u7684\u524d\u666f\u3002\u53e6\u4e00\u4e2a\u6311\u6218\u662f\u7ee7\u7eed\u63a2\u7d22\u81ea\u6211\u76d1\u7763\u7684\u8bad\u7ec3\u524d\u65b9\u6cd5\u3002\u6211\u4eec\u6700\u521d\u7684\u5b9e\u9a8c\u8868\u660e\uff0c\u81ea\u6211\u76d1\u7763\u9884\u8bad\u7ec3\u6709\u6240\u6539\u8fdb\uff0c\u4f46\u81ea\u6211\u76d1\u7763\u9884\u8bad\u7ec3\u548c\u5927\u89c4\u6a21\u76d1\u7763\u9884\u8bad\u7ec3\u4e4b\u95f4\u4ecd\u5b58\u5728\u5f88\u5927\u7684\u5dee\u8ddd\u3002\u6700\u540e\uff0cViT\u7684\u8fdb\u4e00\u6b65\u6269\u5c55\u53ef\u80fd\u4f1a\u5bfc\u81f4\u6027\u80fd\u7684\u63d0\u9ad8\u3002<\/p>\n<\/details>\n\n\n\n

\u672c\u6587\u5de5\u4f5c\uff1a<\/strong><\/p>\n\n\n\n

\u672c\u6587\u4e3b\u8981\u5c06\u56fe\u7247\u5904\u7406\u6210 patch \u5e8f\u5217\uff0c\u7136\u540e\u4f7f\u7528 Transformer \u53bb\u5904\u7406<\/strong>\uff0c\u53d6\u5f97\u4e86\u63a5\u8fd1\u6216\u8d85\u8fc7\u5377\u79ef\u795e\u7ecf\u7f51\u7edc\u7684\u7ed3\u679c\uff0c\u540c\u65f6\u8bad\u7ec3\u8d77\u6765\u4e5f\u66f4\u5feb\u3002<\/p>\n\n\n\n

\u672a\u6765\u5c55\u671b\uff1a<\/strong><\/p>\n\n\n\n

1\u3001\u5c06ViT\u5e94\u7528\u4e8e\u5176\u4ed6\u8ba1\u7b97\u673a\u89c6\u89c9\u4efb\u52a1\uff0c\u5982\u68c0\u6d4b\u548c\u5206\u5272\u3002<\/p>\n\n\n\n

2\u3001\u7ee7\u7eed\u63a2\u7d22\u81ea\u6211\u76d1\u7763\u7684\u9884\u8bad\u7ec3\u65b9\u6cd5\u3002<\/p>\n\n\n\n

3\u3001\u53ef\u4ee5\u8fdb\u4e00\u6b65\u6269\u5927ViT\u7684\u89c4\u6a21\uff0c\u968f\u7740\u6a21\u578b\u5c3a\u5bf8\u7684\u589e\u52a0\uff0c\u6027\u80fd\u4f3c\u4e4e\u8fd8\u6ca1\u6709\u9971\u548c\u3002<\/p>\n\n\n\n

<\/p>\n","protected":false},"excerpt":{"rendered":"

arXiv\u8bba\u6587\u5730\u5740 \u300a\u5c06 Transformer \u67b6\u6784\u5e94\u7528\u4e8e\u56fe\u50cf\u8bc6\u522b\u4efb\u52a1\u300b Submitted on 22 O […]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[],"class_list":["post-330","post","type-post","status-publish","format-standard","hentry","category-message"],"_links":{"self":[{"href":"http:\/\/anyone.anuoblog.cn\/index.php\/wp-json\/wp\/v2\/posts\/330","targetHints":{"allow":["GET"]}}],"collection":[{"href":"http:\/\/anyone.anuoblog.cn\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"http:\/\/anyone.anuoblog.cn\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"http:\/\/anyone.anuoblog.cn\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"http:\/\/anyone.anuoblog.cn\/index.php\/wp-json\/wp\/v2\/comments?post=330"}],"version-history":[{"count":3,"href":"http:\/\/anyone.anuoblog.cn\/index.php\/wp-json\/wp\/v2\/posts\/330\/revisions"}],"predecessor-version":[{"id":462,"href":"http:\/\/anyone.anuoblog.cn\/index.php\/wp-json\/wp\/v2\/posts\/330\/revisions\/462"}],"wp:attachment":[{"href":"http:\/\/anyone.anuoblog.cn\/index.php\/wp-json\/wp\/v2\/media?parent=330"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/anyone.anuoblog.cn\/index.php\/wp-json\/wp\/v2\/categories?post=330"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/anyone.anuoblog.cn\/index.php\/wp-json\/wp\/v2\/tags?post=330"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}