{"id":550,"date":"2021-08-31T16:36:53","date_gmt":"2021-08-31T08:36:53","guid":{"rendered":"https:\/\/aif.amtbbs.org\/?p=550"},"modified":"2021-08-31T16:36:53","modified_gmt":"2021-08-31T08:36:53","slug":"resnet%e5%88%b0%e5%ba%95%e5%9c%a8%e8%a7%a3%e5%86%b3%e4%b8%80%e4%b8%aa%e4%bb%80%e4%b9%88%e9%97%ae%e9%a2%98%e5%91%a2%ef%bc%9f","status":"publish","type":"post","link":"https:\/\/aif.amtbbs.org\/index.php\/2021\/08\/31\/550\/","title":{"rendered":"ResNet\u5230\u5e95\u5728\u89e3\u51b3\u4e00\u4e2a\u4ec0\u4e48\u95ee\u9898\u5462\uff1f"},"content":{"rendered":"

ResNet YYDS\uff01<\/strong><\/p>\n

ResNet\u53d1\u5e03\u4e8e2015\u5e74\uff0c\u76ee\u524d\u4ecd\u6709\u5927\u91cfCV\u4efb\u52a1\u7528\u5176\u4f5c\u4e3abackbone\uff08\u5c24\u5176\u662f\u9876\u4f1a\u5b9e\u9a8c\u6bd4\u8f83\uff09\uff0c\u800c\u4e14\u5f53\u524d\u5f88\u591a\u7f51\u7edc\u90fd\u5728\u4f7f\u7528\u6b8b\u5dee\u6a21\u5757\u3002<\/p>\n

Deep Residual Learning for Image Recognition<\/strong><\/h1>\n

\u8bba\u6587\uff1ahttps:\/\/arxiv.org\/abs\/1512.03385<\/p>\n

\u622a\u6b622021\u5e748\u670819\u65e5\uff0cResNet\u7684\u5f15\u7528\u91cf\u5df2\u8fbe\u523086531<\/strong>\uff08\u6570\u636e\u6765\u81ea\u8c37\u6b4c\u5b66\u672f\uff09\u3002\u76f2\u731c2021\u5e74\u5e95\uff0c\u5f15\u7528\u91cf\u80fd\u783410\u4e07\uff01\u672c\u6587\u6839\u636e\u77e5\u4e4e\u4e0a\u7684\u8bdd\u9898\uff0c\u6765\u6df1\u5165\u63a2\u8ba8\u4e00\u4e0b\u5f3a\u608d\u7684ResNet\u3002\u800c\u4e14\u6b63\u503c\u62db\u8058\/\u6c42\u804c\u5b63\uff0cResNet\u4e5f\u662f\u4e00\u4e2a\u9ad8\u9891\u9762\u8bd5\u91cd\u70b9\uff0c\u65b9\u4fbf\u5927\u5bb6\u67e5\u6f0f\u8865\u7f3a\u3002<\/p>\n

\"\u56fe\u7247\"<\/p>\n

Resnet\u5230\u5e95\u5728\u89e3\u51b3\u4e00\u4e2a\u4ec0\u4e48\u95ee\u9898\u5462\uff1f<\/strong><\/h2>\n

\u65e2\u7136\u53ef\u4ee5\u901a\u8fc7\u521d\u8bd5\u5316\u548c\u5f52\u4e00\u5316\uff08BN\u5c42\uff09\u89e3\u51b3\u68af\u5ea6\u5f25\u6563\u6216\u7206\u70b8\u7684\u95ee\u9898\uff0c\u90a3Resnet\u63d0\u51fa\u7684\u90a3\u6761\u901a\u8def\u662f\u5728\u89e3\u51b3\u4ec0\u4e48\u95ee\u9898\u5462\uff1f<\/p>\n

\u5728He\u7684\u539f\u6587\u4e2d\u6709\u63d0\u5230\u662f\u89e3\u51b3\u6df1\u5c42\u7f51\u7edc\u7684\u4e00\u79cd\u9000\u5316\u95ee\u9898\uff0c\u4f46\u5e76\u660e\u786e\u8bf4\u660e\u662f\u4ec0\u4e48\u95ee\u9898\uff01<\/p>\n

\u4f5c\u8005\uff1a\u738b\u5cf0
\nhttps:\/\/www.zhihu.com\/question\/64494691\/answer\/220989469<\/strong><\/p>\n

17\u5e742\u6708\u4efd\u6709\u7bc7\u6587\u7ae0\uff0c\u6b63\u597d\u8ddf\u8fd9\u4e2a\u95ee\u9898\u4e00\u6837\u3002<\/strong><\/p>\n

The Shattered Gradients Problem: If resnets are the answer, then what is the question?<\/strong><\/p>\n

\u5927\u610f\u662f\u795e\u7ecf\u7f51\u7edc\u8d8a\u6765\u8d8a\u6df1\u7684\u65f6\u5019\uff0c\u53cd\u4f20\u56de\u6765\u7684\u68af\u5ea6\u4e4b\u95f4\u7684\u76f8\u5173\u6027\u4f1a\u8d8a\u6765\u8d8a\u5dee\uff0c\u6700\u540e\u63a5\u8fd1\u767d\u566a\u58f0\u3002\u56e0\u4e3a\u6211\u4eec\u77e5\u9053\u56fe\u50cf\u662f\u5177\u5907\u5c40\u90e8\u76f8\u5173\u6027\u7684\uff0c\u90a3\u5176\u5b9e\u53ef\u4ee5\u8ba4\u4e3a\u68af\u5ea6\u4e5f\u5e94\u8be5\u5177\u5907\u7c7b\u4f3c\u7684\u76f8\u5173\u6027\uff0c\u8fd9\u6837\u66f4\u65b0\u7684\u68af\u5ea6\u624d\u6709\u610f\u4e49\uff0c\u5982\u679c\u68af\u5ea6\u63a5\u8fd1\u767d\u566a\u58f0\uff0c\u90a3\u68af\u5ea6\u66f4\u65b0\u53ef\u80fd\u6839\u672c\u5c31\u662f\u5728\u505a\u968f\u673a\u6270\u52a8\u3002<\/p>\n

\u6709\u4e86\u68af\u5ea6\u76f8\u5173\u6027\u8fd9\u4e2a\u6307\u6807\u4e4b\u540e\uff0c\u4f5c\u8005\u5206\u6790\u4e86\u4e00\u7cfb\u5217\u7684\u7ed3\u6784\u548c\u6fc0\u6d3b\u51fd\u6570\uff0c\u53d1\u73b0resnet\u5728\u4fdd\u6301\u68af\u5ea6\u76f8\u5173\u6027\u65b9\u9762\u5f88\u4f18\u79c0\uff08\u76f8\u5173\u6027\u8870\u51cf\u4ece\u00a0\"\u56fe\u7247\"\u00a0\u5230\u4e86\u00a0\"\u56fe\u7247\"\u00a0\uff09\u3002\u8fd9\u4e00\u70b9\u5176\u5b9e\u4e5f\u5f88\u597d\u7406\u89e3\uff0c\u4ece\u68af\u5ea6\u6d41\u6765\u770b\uff0c\u6709\u4e00\u8def\u68af\u5ea6\u662f\u4fdd\u6301\u539f\u6837\u4e0d\u52a8\u5730\u5f80\u56de\u4f20\uff0c\u8fd9\u90e8\u5206\u7684\u76f8\u5173\u6027\u662f\u975e\u5e38\u5f3a\u7684\u3002<\/p>\n

\u4f5c\u8005\uff1a\u7070\u7070
\nhttps:\/\/www.zhihu.com\/question\/64494691\/answer\/271335912<\/strong><\/p>\n

\u4e00\u65b9\u9762: ResNet\u89e3\u51b3\u7684\u4e0d\u662f\u68af\u5ea6\u5f25\u6563\u6216\u7206\u70b8\u95ee\u9898\uff0ckaiming\u7684\u8bba\u6587\u4e2d\u4e5f\u8bf4\u4e86:\u81ed\u540d\u662d\u8457\u7684\u68af\u5ea6\u5f25\u6563\/\u7206\u70b8\u95ee\u9898\u5df2\u7ecf\u5f88\u5927\u7a0b\u5ea6\u4e0a\u88abnormalized initialization \u00a0and intermediate normalization layers\u89e3\u51b3\u4e86;<\/p>\n

\u53e6\u4e00\u65b9\u9762: \u7531\u4e8e\u76f4\u63a5\u589e\u52a0\u7f51\u7edc\u6df1\u5ea6\u7684(plain)\u7f51\u7edc\u5728\u8bad\u7ec3\u96c6\u4e0a\u4f1a\u6709\u66f4\u9ad8\u7684\u9519\u8bef\u7387\uff0c\u6240\u4ee5\u66f4\u6df1\u7684\u7f51\u7edc\u5e76\u6ca1\u6709\u8fc7\u62df\u5408\uff0c\u4e5f\u5c31\u662f\u8bf4\u66f4\u6df1\u7684\u7f51\u7edc\u6548\u679c\u4e0d\u597d\uff0c\u662f\u56e0\u4e3a\u7f51\u7edc\u6ca1\u6709\u88ab\u8bad\u7ec3\u597d\uff0c\u81f3\u4e8e\u4e3a\u5565\u6ca1\u6709\u88ab\u8bad\u7ec3\u597d\uff0c\u4e2a\u4eba\u5f88\u8d5e\u540c\u524d\u9762\u738b\u5cf0\u7684\u7b54\u6848\u4e2d\u7684\u89e3\u91ca\u3002<\/p>\n

\u5728ResNet\u4e2d,building block:<\/p>\n

\"\u56fe\u7247\"<\/figure>\n

H(x)\u662f\u671f\u671b\u62df\u5408\u7684\u7279\u5f81\u56fe,\u8fd9\u91cc\u53eb\u505adesired underlying mapping<\/p>\n

\u4e00\u4e2abuilding block\u8981\u62df\u5408\u7684\u5c31\u662f\u8fd9\u4e2a\u6f5c\u5728\u7684\u7279\u5f81\u56fe<\/p>\n

\u5f53\u6ca1\u6709\u4f7f\u7528\u6b8b\u5dee\u7f51\u7edc\u7ed3\u6784\u65f6,building block\u7684\u6620\u5c04F(x)\u9700\u8981\u505a\u7684\u5c31\u662f\u62df\u5408H(x)<\/p>\n

\u5f53\u4f7f\u7528\u4e86\u6b8b\u5dee\u7f51\u7edc\u65f6,\u5c31\u662f\u52a0\u5165\u4e86skip connection \u7ed3\u6784,\u8fd9\u65f6\u5019\u7531\u4e00\u4e2abuilding block \u7684\u4efb\u52a1\u7531: F(x) := H(x)\uff0c\u53d8\u6210\u4e86F(x) := H(x)-x<\/p>\n

\u5bf9\u6bd4\u8fd9\u4e24\u4e2a\u5f85\u62df\u5408\u7684\u51fd\u6570,\u6587\u4e2d\u8bf4\u5047\u8bbe\u62df\u5408\u6b8b\u5dee\u56fe\u66f4\u5bb9\u6613\u4f18\u5316,\u4e5f\u5c31\u662f\u8bf4:F(x) := H(x)-x\u6bd4F(x) := H(x)\u66f4\u5bb9\u6613\u4f18\u5316,\u63a5\u4e0b\u6765\u4e3e\u4e86\u4e00\u4e2a\u4f8b\u5b50,\u6781\u7aef\u60c5\u51b5\u4e0b:desired underlying mapping\u8981\u62df\u5408\u7684\u662fidentity mapping\uff0c\u8fd9\u65f6\u5019\u6b8b\u5dee\u7f51\u7edc\u7684\u4efb\u52a1\u5c31\u662f\u62df\u5408F(x): 0,\u800c\u539f\u672c\u7684plain\u7ed3\u6784\u7684\u8bdd\u5c31\u662fF(x) : x,\u800cF(x): 0\u4efb\u52a1\u4f1a\u66f4\u5bb9\u6613,\u539f\u56e0\u662f:resnet\uff08\u6b8b\u5dee\u7f51\u7edc\uff09\u7684F\uff08x\uff09\u7a76\u7adf\u957f\u4ec0\u4e48\u6837\u5b50\uff1f\u4e2dtheone\u7684\u7b54\u6848:<\/p>\n

F\u662f\u6c42\u548c\u524d\u7f51\u7edc\u6620\u5c04\uff0cH\u662f\u4ece\u8f93\u5165\u5230\u6c42\u548c\u540e\u7684\u7f51\u7edc\u6620\u5c04\u3002\u6bd4\u5982\u628a5\u6620\u5c04\u52305.1\uff0c\u90a3\u4e48\u5f15\u5165\u6b8b\u5dee\u524d\u662fF'(5)=5.1\uff0c\u5f15\u5165\u6b8b\u5dee\u540e\u662fH(5)=5.1, H(5)=F(5)+5, F(5)=0.1\u3002\u8fd9\u91cc\u7684F’\u548cF\u90fd\u8868\u793a\u7f51\u7edc\u53c2\u6570\u6620\u5c04\uff0c\u5f15\u5165\u6b8b\u5dee\u540e\u7684\u6620\u5c04\u5bf9\u8f93\u51fa\u7684\u53d8\u5316\u66f4\u654f\u611f\u3002\u6bd4\u5982s\u8f93\u51fa\u4ece5.1\u53d8\u52305.2\uff0c\u6620\u5c04F’\u7684\u8f93\u51fa\u589e\u52a0\u4e861\/51=2%\uff0c\u800c\u5bf9\u4e8e\u6b8b\u5dee\u7ed3\u6784\u8f93\u51fa\u4ece5.1\u52305.2\uff0c\u6620\u5c04F\u662f\u4ece0.1\u52300.2\uff0c\u589e\u52a0\u4e86100%\u3002\u660e\u663e\u540e\u8005\u8f93\u51fa\u53d8\u5316\u5bf9\u6743\u91cd\u7684\u8c03\u6574\u4f5c\u7528\u66f4\u5927\uff0c\u6240\u4ee5\u6548\u679c\u66f4\u597d\u3002\u6b8b\u5dee\u7684\u601d\u60f3\u90fd\u662f\u53bb\u6389\u76f8\u540c\u7684\u4e3b\u4f53\u90e8\u5206\uff0c\u4ece\u800c\u7a81\u51fa\u5fae\u5c0f\u7684\u53d8\u5316\uff0c\u770b\u5230\u6b8b\u5dee\u7f51\u7edc\u6211\u7b2c\u4e00\u53cd\u5e94\u5c31\u662f\u5dee\u5206\u653e\u5927\u5668<\/p>\n

\u540e\u7eed\u7684\u5b9e\u9a8c\u4e5f\u662f\u8bc1\u660e\u4e86\u5047\u8bbe\u7684, \u6b8b\u5dee\u7f51\u7edc\u6bd4plain\u7f51\u7edc\u66f4\u597d\u8bad\u7ec3\u3002\u56e0\u6b64\uff0cResNet\u89e3\u51b3\u7684\u662f\u66f4\u597d\u5730\u8bad\u7ec3\u7f51\u7edc\u7684\u95ee\u9898\uff0c\u738b\u5cf0\u7684\u7b54\u6848\u7b97\u662f\u5bf9ResNet\u4e4b\u6240\u4ee5\u597d\u7684\u4e00\u4e2a\u7406\u8bba\u8bba\u8bc1\u5427.<\/p>\n

\u4f5c\u8005\uff1a\u5468\u7ffc\u5357
\nhttps:\/\/www.zhihu.com\/question\/64494691\/answer\/325492405<\/strong><\/p>\n

\u6700\u8fd1\u4e5f\u5728\u7814\u7a76residual\u7ed3\u6784\uff0c\u8bf4\u70b9\u81ea\u5df1\u7684\u7406\u89e3\u3002<\/p>\n

residual\u7ed3\u6784\u5f88\u5927\u7a0b\u5ea6\u4e0a\u89e3\u51b3\u4e86relu\u4e4b\u540e\u7684\u7a7a\u7f3a\u95ee\u9898\u3002<\/p>\n

\u9996\u5148\uff0c\u4e00\u4e2a3×3\u7684\u5377\u79ef+relu\uff0c\u4e0d\u53ef\u80fd\u5bf9\u4e8e\u6240\u6709\u7684\u56fe\u7247\u90fd\u6709\u54cd\u5e94\u7684\u3002\u56fe\u7247A\u7ecf\u8fc7\u8fd9\u4e2a\u5377\u79ef\u6709\u6b63\u54cd\u5e94\uff0c\u56fe\u7247B\u53ef\u80fd\u5c31\u6ca1\u6709\u3002<\/p>\n

\u90a3\u80fd\u4e0d\u80fd\u586b\u4e0a\u8fd9\u4e2a\u7a7a\u7f3a\uff0c\u53ef\u4ee5\u7684\u3002<\/p>\n

resnet\u7ed3\u6784\u7684\u673a\u5236\u5927\u6982\u662f\u8fd9\u6837\uff0c3×3\u611f\u53d7\u91ce\u6ca1\u6709\u63d0\u51fa\u6765\u7684\u7279\u5f81\uff0c\u6211\u75285×5\u7684\u611f\u53d7\u91ce\u53bb\u586b\uff0c\u8fd8\u4e0d\u884c\u5c31\u75287×7\u53bb\u586b\u3002<\/p>\n

\u8fd9\u4e2a\u586b\u7a7a\u7f3a\u7684\u601d\u60f3\u5728resnext\u4e0a\u8868\u73b0\u7684\u66f4\u4e3a\u660e\u663e\uff0c\u4ed6\u7528\u4e86\u4e00\u58065×5\u53bb\u586b\u3002<\/p>\n

\u6211\u662f\u505a\u4eba\u8138\u7684\uff0c\u6211\u8bbe\u8ba1\u8fc7rrsnet\u7ed3\u6784\u548c\u975eresnet\u7ed3\u6784\u5e76\u4e14\u5bf9\u6bcf\u4e00\u5c42layer\u505a\u8fc7\u53ef\u89c6\u5316\uff0c\u53d1\u73b0resnet\u51fa\u6765\u7684channel\u8981\u6bd4\u975eresnet\u7ed3\u6784\u51fa\u6765\u7684channel\u9971\u6ee1\u3002\u800c\u4e14bypass\u4e0a\u6b63\u597d\u5c31\u505a\u4e86\u586b\u5145\u5de5\u4f5c\u3002<\/p>\n

\u4f5c\u8005\uff1a\u5218\u6615\u5bb8
\nhttps:\/\/www.zhihu.com\/question\/64494691\/answer\/1545759510<\/strong><\/p>\n

\u6700\u8fd1\u53c8\u91cd\u8bfb\u4e86ResNet\u8fd9\u7bc7\u8bba\u6587\uff0c\u5e76\u8be6\u7ec6\u5199\u4e86\u4e00\u4e9b\u81ea\u5df1\u7684\u7406\u89e3\uff0c\u94fe\u63a5\u89c1\u672c\u56de\u7b54\u672b\u5c3e\u3002<\/p>\n

\n

ResNet\u89e3\u51b3\u7684\u5176\u5b9e\u662f\u4e00\u4e2a\u4e0d\u7b26\u5408\u9884\u671f\u7684\u73b0\u8c61<\/strong>\u3002<\/p>\n

\u505a\u5b9e\u9a8c\u5982\u679c\u51fa\u73b0\u4e86\u4e0d\u7b26\u5408\u9884\u671f\u7684\u73b0\u8c61\uff0c\u5e76\u4e14\u5728\u6211\u4eec\u7684\u9884\u671f\u6ca1\u4ec0\u4e48\u95ee\u9898\u7684\u60c5\u51b5\u4e0b\uff0c\u90a3\u4e48\u4e00\u5b9a\u662f\u54ea\u91cc\u51fa\u4e86\u95ee\u9898\uff0c\u5e76\u4e14\u8fd9\u4e2a\u95ee\u9898\u975e\u5e38\u503c\u5f97\u89e3\u51b3<\/strong>\uff01<\/p><\/blockquote>\n

\u524d\u63d0\uff1a<\/strong><\/p>\n

\u7f51\u7edc\u7684\u4e0d\u540c\u5c42\u53ef\u4ee5\u63d0\u53d6\u4e0d\u540c\u62bd\u8c61\u5c42\u6b21\u7684\u7279\u5f81\uff0c\u8d8a\u6df1\u7684\u5c42\u63d0\u53d6\u7684\u7279\u5f81\u8d8a\u62bd\u8c61\u3002\u56e0\u6b64\u6df1\u5ea6\u7f51\u7edc\u53ef\u4ee5\u6574\u5408low-medium-high\u5404\u79cd\u5c42\u6b21\u7684\u7279\u5f81\uff0c\u589e\u5f3a\u7f51\u7edc\u8868\u5f81\u80fd\u529b\u3002<\/p>\n

\u9884\u671f\uff1a<\/strong><\/p>\n

\u8d8a\u6df1\u7684\u7f51\u7edc\uff0c\u8868\u5f81\u80fd\u529b\u5e94\u8be5\u8d8a\u5f3a\uff0c\u505a\u5206\u7c7b\u7b49\u4efb\u52a1\u6548\u679c\u5e94\u8be5\u8d8a\u597d\u3002<\/p>\n

\u73b0\u5b9e\uff1a<\/strong><\/p>\n

\"\u56fe\u7247\"<\/figure>\n

\u66f4\u6df1\u768456-layer\u7f51\u7edc\u5728\u8bad\u7ec3\u548c\u6d4b\u8bd5\u7ed3\u679c\u4e0a\u5747\u4e0d\u598220-layer\u7f51\u7edc\uff01<\/p>\n

\u4e0d\u7b26\u5408\u9884\u671f\u7684\u73b0\u8c61\u51fa\u73b0\u4e86\uff01\uff01\uff01<\/strong><\/p>\n

\u8fd9\u4e2a\u5c31\u662f\u8bba\u6587\u4e2d\u63d0\u5230\u7684\u9000\u5316\u95ee\u9898\uff0c\u800cResNet\u5219\u6b63\u662f\u4e3a\u4e86\u89e3\u51b3\u8fd9\u4e2a\u95ee\u9898\u3002<\/strong><\/p>\n

\u4e0b\u9762\u8fdb\u5165\u788e\u788e\u5ff5\u5206\u6790\uff1a<\/strong><\/p>\n

deeper network\u51c6\u786e\u7387\u9971\u548c\u540e\uff0c\u5f88\u5feb\u5c31\u9000\u5316\u4e86<\/p>\n

\u4e3a\u4ec0\u4e48\u4f1a\u8fd9\u6837\u5462\uff1f<\/p>\n

\u9996\u5148\uff0c\u6df1\u5ea6\u7f51\u7edc\u4f18\u5316\u662f\u6bd4\u8f83\u56f0\u96be\u7684\uff0c\u6bd4\u5982\u4f1a\u51fa\u73b0\u68af\u5ea6\u7206\u70b8\/\u68af\u5ea6\u6d88\u5931<\/strong>\u7b49\u95ee\u9898\u3002\u4e0d\u8fc7\uff0c\u8fd9\u4e2a\u95ee\u9898\u5df2\u7ecf\u88abnormalized initialization\u548cbatch normalization\u7b49\u63aa\u65bd\u89e3\u51b3\u5f97\u5dee\u4e0d\u591a\u4e86\u3002<\/p>\n

\u5176\u6b21\uff0c\u7f51\u7edc\u66f4\u6df1\u4e86\uff0c\u53c2\u6570\u66f4\u591a\u4e86\uff0c\u5e94\u8be5\u62df\u5408\u80fd\u529b\u66f4\u5f3a\u4e86\u624d\u5bf9\u554a\uff01\u5662\uff0c\u4e00\u5b9a\u662f\u8fc7\u62df\u5408<\/strong>\u4e86\u3002<\/p>\n

\u4f46\u4f3c\u4e4e\u4e5f\u4e0d\u662f\u8fc7\u62df\u5408\u7684\u95ee\u9898\uff1a<\/p>\n

\u56e0\u4e3a56-layer\u7f51\u7edc\uff08\u4e0a\u56fe\u7ea2\u7ebf\uff09\u7684training error\uff08\u4e0a\u5de6\u56fe\uff09\u4e5f\u6bd420-layer\u7f51\u7edc\uff08\u4e0a\u56fe\u9ec4\u7ebf\uff09\u8981\u9ad8\uff0c\u8fd9\u5c31\u5e94\u8be5\u4e0d\u662f\u8fc7\u62df\u5408\u4e86\u554a\uff01<\/p>\n

\u90a3\u4e48\u7a76\u7adf\u662f\u4ec0\u4e48\u539f\u56e0\u5bfc\u81f4\u4e86deeper network degradation\u95ee\u9898\u5462\uff1f<\/p>\n

\u73b0\u5728\uff0c\u6211\u4eec\u6362\u4e00\u79cd\u601d\u8def\u6765\u6784\u5efadeeper network\uff1a<\/p>\n

\u5047\u8bbe\u6d45\u5c42\u7f51\u7edc\u5df2\u7ecf\u53ef\u4ee5\u5f97\u5230\u4e00\u4e2a\u4e0d\u9519\u7684\u7ed3\u679c\u4e86\uff0c\u90a3\u6211\u63a5\u4e0b\u6765\u65b0\u589e\u52a0\u7684\u5c42\u5565\u4e5f\u4e0d\u5e72\uff0c\u53ea\u662f\u62df\u5408\u4e00\u4e2aidentity mapping\uff0c\u8f93\u51fa\u5c31\u62df\u5408\u8f93\u5165\uff0c\u8fd9\u6837\u603b\u53ef\u4ee5\u5427\u3002<\/p>\n

\u8fd9\u6837\u7684\u8bdd\uff0c\u6211\u4eec\u89c9\u5f97\uff1a\u8fd9\u6837\u6784\u5efa\u7684\u6df1\u5c42\u7f51\u7edc\u81f3\u5c11\u4e0d\u5e94\u8be5\u6bd4\u5b83\u5bf9\u5e94\u7684\u6d45\u5c42training error\u8981\u9ad8\u3002\u5bf9\u5427\u3002<\/p>\n

\u4f46\u662f\u5b9e\u9a8c\u53c8\u65e0\u60c5\u5730\u6253\u8138\u3002<\/p>\n

\u770b\u6765\uff0c\u4e00\u5b9a\u662f\u5728\u4f18\u5316\u6df1\u5c42\u7f51\u7edc\u65f6\u51fa\u4e86\u95ee\u9898\uff01\u8fd9\u79cd\u95ee\u9898\u5bfc\u81f4\u4e86\u6df1\u5c42\u7f51\u7edc\u7684\u6027\u80fd\u9000\u5316\uff0cResNet\u5c31\u662f\u8981\u89e3\u51b3\u8fd9\u4e2a\u95ee\u9898\uff01<\/strong><\/p>\n

\u600e\u4e48\u89e3\u51b3\uff1f<\/strong><\/p>\n

ResNet\u7684\u6838\u5fc3\u5c31\u662fresidual learning<\/strong>\u548cidentity learning by shortcut<\/strong><\/p>\n

\u6548\u679c\u600e\u4e48\u6837\u5462\uff1f<\/strong><\/p>\n

ResNet\u89e3\u51b3\u4e86\u9000\u5316\u95ee\u9898\uff0c\u83b7\u5f97\u4e86\u66f4\u9ad8\u7684\u51c6\u786e\u7387\uff0c\u5b9e\u73b0\u4e86\u66f4\u5feb\u7684\u6536\u655b\u901f\u5ea6\u3002<\/p>\n

\u4f5c\u8005\uff1a[\u5df2\u91cd\u7f6e]
\nhttps:\/\/www.zhihu.com\/question\/64494691\/answer\/511981382<\/strong><\/p>\n

\u8fd9\u4e48\u7b80\u5355\u7684\u95ee\u9898\u9700\u8981\u641e\u90a3\u4e48\u590d\u6742\u4e48\uff1f<\/p>\n

dy(x)\/dx=F(x)=> y(x)=x+F(x)*dt\uff0c\u8fd9\u4e0d\u5c31\u662f\u4e2a\u5fae\u5206\u65b9\u7a0b\u7684\u5dee\u5206\u5316\u4e48\uff1f\u6240\u4ee5ResNet\u5c31\u662f\u4e2a\u5fae\u5206\u65b9\u7a0b\u554a\u3002<\/p>\n

\u81f3\u4e8e\u8bf4\u5176\u6536\u655b\u6027\u597d\uff0c\u5c31\u662f\u901a\u8fc7\u8fd9\u79cd\u64cd\u4f5c\u4f7f\u5f97\u7f51\u7edc\u7684\u7ed3\u6784\u66f4\u5747\u5300\uff0c\u6574\u4e2a\u6620\u5c04\u5c31\u662f\u7531\u51e0\u4e4e\u5b8c\u5168\u4e00\u6837\u7684\u6a21\u5757\u6765\u6784\u9020\uff0c\u800c\u4e14\u5f53F(x)=0\u7684\u65f6\u5019\uff0c\u76f8\u5e94\u7684\u6a21\u5757\u5c31\u76f8\u5f53\u4e8e\u4e0d\u5b58\u5728\uff0c\u6216\u8005\u662f\u5f53F(x)\u6309\u7167\u4e00\u5b9a\u6bd4\u4f8bscaling\u7684\u65f6\u5019\uff0c\u7b49\u4ef7\u4e8e\u8c03\u6574\u6b65\u957fdx, \u8fd9\u79cd\u7ed3\u6784\u4f7f\u5f97\u6620\u5c04\u7a7a\u95f4\u7684\u66f2\u7387\u66f4\u5e73\u6ed1\uff0c\u57fa\u672c\u662f\u5747\u5300\u7684\uff0c\u5e76\u4e14\u53ef\u4ee5\u6216\u8005\u901a\u8fc7\u8df3\u8fc7\u67d0\u4e9bF(x)\u6a21\u5757\u6216\u8005\u901a\u8fc7\u5bf9F(x)\u7684scaling\u6765\u8c03\u6574\u6620\u5c04\u590d\u6742\u5ea6\uff0c\u8fbe\u5230\u6a21\u578b\u590d\u6742\u5ea6\u548c\u95ee\u9898\u590d\u6742\u5ea6\u7684\u66f4\u597d\u7684\u5339\u914d\uff0c\u8fd9\u4e00\u70b9\u5728\u666e\u901a\u7684CNN\u4e2d\u662f\u5f88\u96be\u5b8c\u6210\u7684\u3002<\/p>\n

\u7b80\u5355\u8bf4\uff0cResNet\u6784\u9020\u4e86\u4e00\u4e2a\u5e73\u6ed1\u7684\u89e3\u7a7a\u95f4\u6d41\u5f62\uff0c\u5728\u8fd9\u4e2a\u5e73\u6ed1\u7684\u6d41\u5f62\u4e0a\u66f4\u5bb9\u6613\u627e\u5230\u89e3\u3002<\/p>\n

\u4f5c\u8005\uff1a\u85b0\u98ce\u521d\u5165\u5f26
\nhttps:\/\/www.zhihu.com\/question\/64494691\/answer\/786270699<\/strong><\/p>\n

\u770b\u4e86\u8fd9\u4e2a\u95ee\u9898\u4e4b\u540e\u6211\u601d\u8003\u4e86\u5f88\u4e45\uff0c\u4e8e\u662f\u5199\u51fa\u4e86\u8fd9\u7bc7\u4e13\u680f\uff0c\u73b0\u5728\u8d34\u8fc7\u6765\u5f53\u7b54\u6848\u3002<\/p>\n

\u9996\u5148\u662f\u8ddf\u7740\u8bba\u6587\u7684\u601d\u8def\u8d70\uff0c\u4e86\u89e3\u4f5c\u8005\u63d0\u51faresnet\u7684\u201c\u5fc3\u8def\u5386\u7a0b\u201d\uff0c\u6700\u540e\u4e5f\u6709\u4e9b\u4e2a\u4eba\u6574\u7406\u7684\u7406\u89e3\u3002<\/p>\n

ps\uff1a\u6b22\u8fce\u5173\u6ce8\u6211\u7684\u4e13\u680f\uff0c\u8fd9\u6bb5\u65f6\u95f4\u6211\u4f1a\u6301\u7eed\u66f4\u65b0\uff0c\u5e76\u4e14\u5728\u66f4\u5b8c\u7ea6\u83ab\u5341\u51e0\u7bc7\u8bba\u6587\u9605\u8bfb\u540e\uff0c\u4f1a\u518d\u5199\u4e00\u4e9b\u6a21\u578b\u5b9e\u73b0\/\u4ee3\u7801\u65b9\u9762\u7684\u7406\u89e3\u3002<\/p>\n

https:\/\/zhuanlan.zhihu.com\/IsonomiaCS<\/p>\n

\u4e00\u3001\u5f15\u8a00\uff1a\u4e3a\u4ec0\u4e48\u4f1a\u6709ResNet\uff1fWhy ResNet\uff1f<\/strong><\/h2>\n

\u795e\u7ecf\u7f51\u7edc\u53e0\u7684\u8d8a\u6df1\uff0c\u5219\u5b66\u4e60\u51fa\u7684\u6548\u679c\u5c31\u4e00\u5b9a\u4f1a\u8d8a\u597d\u5417\uff1f<\/p><\/blockquote>\n

\u7b54\u6848\u65e0\u7591\u662f\u5426\u5b9a\u7684\uff0c\u4eba\u4eec\u53d1\u73b0\u5f53\u6a21\u578b\u5c42\u6570\u589e\u52a0\u5230\u67d0\u79cd\u7a0b\u5ea6\uff0c\u6a21\u578b\u7684\u6548\u679c\u5c06\u4f1a\u4e0d\u5347\u53cd\u964d\u3002\u4e5f\u5c31\u662f\u8bf4\uff0c\u6df1\u5ea6\u6a21\u578b\u53d1\u751f\u4e86\u9000\u5316\uff08degradation\uff09\u60c5\u51b5\u3002<\/p>\n

\u90a3\u4e48\uff0c\u4e3a\u4ec0\u4e48\u4f1a\u51fa\u73b0\u8fd9\u79cd\u60c5\u51b5\uff1f<\/p>\n

1. \u8fc7\u62df\u5408\uff1fOverfitting\uff1f<\/strong><\/h3>\n

\u9996\u5148\u5370\u5165\u8111\u6d77\u7684\u5c31\u662fAndrew Ng\u673a\u5668\u5b66\u4e60\u516c\u5f00\u8bfe[1]\u7684\u8fc7\u62df\u5408\u95ee\u9898<\/p>\n

\"\u56fe\u7247\"
Andrew Ng\u7684\u8bfe\u4ef6\u622a\u56fe<\/figcaption><\/figure>\n

\u5728\u8fd9\u4e2a\u591a\u9879\u5f0f\u56de\u5f52\u95ee\u9898\u4e2d\uff0c\u5de6\u8fb9\u7684\u6a21\u578b\u662f\u6b20\u62df\u5408\uff08under fit\uff09\u7684\u6b64\u65f6\u6709\u5f88\u9ad8\u7684\u504f\u5dee\uff08high bias\uff09\uff0c\u4e2d\u95f4\u7684\u62df\u5408\u6bd4\u8f83\u6210\u529f\uff0c\u800c\u53f3\u8fb9\u5219\u662f\u5178\u578b\u7684\u8fc7\u62df\u5408\uff08overfit\uff09\uff0c\u6b64\u65f6\u7531\u4e8e\u6a21\u578b\u8fc7\u4e8e\u590d\u6742<\/strong>\uff0c\u5bfc\u81f4\u4e86\u9ad8\u65b9\u5dee\uff08high variance\uff09\u3002<\/p>\n

\u7136\u800c\uff0c\u5f88\u660e\u663e\u5f53\u524dCNN\u9762\u4e34\u7684\u6548\u679c\u9000\u5316\u4e0d\u662f\u56e0\u4e3a\u8fc7\u62df\u5408\uff0c\u56e0\u4e3a\u8fc7\u62df\u5408\u7684\u73b0\u8c61\u662f”\u9ad8\u65b9\u5dee\uff0c\u4f4e\u504f\u5dee”\uff0c\u5373\u6d4b\u8bd5\u8bef\u5dee\u5927\u800c\u8bad\u7ec3\u8bef\u5dee\u5c0f\u3002\u4f46\u5b9e\u9645\u4e0a\uff0c\u6df1\u5c42CNN\u7684\u8bad\u7ec3\u8bef\u5dee\u548c\u6d4b\u8bd5\u8bef\u5dee\u90fd\u5f88\u5927\u3002<\/p>\n

\"\u56fe\u7247\"
\uff08a\uff09\u6b20\u62df\u5408\u4e0e\u8fc7\u62df\u5408 \uff08b\uff09\u6a21\u578b\u9000\u5316<\/figcaption><\/figure>\n

2.\u00a0\u68af\u5ea6\u7206\u70b8\/\u6d88\u5931\uff1fGradient Exploding\/Vanishing\uff1f<\/strong><\/h3>\n

\u9664\u6b64\u4e4b\u5916\uff0c\u6700\u53d7\u4eba\u8ba4\u53ef\u7684\u539f\u56e0\u5c31\u662f\u201c\u68af\u5ea6\u7206\u70b8\/\u6d88\u5931\uff08\u5f25\u6563\uff09\u201d\u4e86\u3002\u4e3a\u4e86\u7406\u89e3\u4ec0\u4e48\u662f\u68af\u5ea6\u5f25\u6563\uff0c\u9996\u5148\u56de\u987e\u4e00\u4e0b\u53cd\u5411\u4f20\u64ad\u7684\u77e5\u8bc6\u3002<\/p>\n

\u5047\u8bbe\u6211\u4eec\u73b0\u5728\u9700\u8981\u8ba1\u7b97\u4e00\u4e2a\u51fd\u6570\u00a0\"\u56fe\u7247\"\u00a0\uff0c\u00a0\"\u56fe\u7247\"\u00a0,\"\u56fe\u7247\"\u00a0,\"\u56fe\u7247\"\u5728\u65f6\u7684\u68af\u5ea6\uff0c\u90a3\u4e48\u9996\u5148\u53ef\u4ee5\u505a\u51fa\u5982\u4e0b\u6240\u793a\u7684\u8ba1\u7b97\u56fe\u3002<\/p>\n

\u5c06\"\u56fe\u7247\"\u00a0,\"\u56fe\u7247\"\u00a0,\"\u56fe\u7247\"\u5e26\u5165\uff0c\u5176\u4e2d\uff0c\u4ee4\u00a0\"\u56fe\u7247\"\u00a0,\u4e00\u6b65\u6b65\u8ba1\u7b97\uff0c\u5f88\u5bb9\u6613\u5c31\u80fd\u5f97\u51fa\u00a0\"\u56fe\u7247\"\u00a0\u3002<\/p>\n

\u8fd9\u5c31\u662f\u524d\u5411\u4f20\u64ad\uff08\u8ba1\u7b97\u56fe\u4e0a\u90e8\u5206\u7eff\u8272\u6253\u5370\u5b57\u4f53\u4e0e\u84dd\u8272\u624b\u5199\u5b57\u4f53\uff09\uff0c\u5373\uff1a<\/p>\n

\"\u56fe\u7247\"<\/figure>\n

\u524d\u5411\u4f20\u64ad\u662f\u4ece\u8f93\u5165\u4e00\u6b65\u6b65\u5411\u524d\u8ba1\u7b97\u8f93\u51fa\uff0c\u800c\u53cd\u5411\u4f20\u64ad\u5219\u662f\u4ece\u8f93\u51fa\u53cd\u5411\u4e00\u70b9\u70b9\u63a8\u51fa\u8f93\u5165\u7684\u68af\u5ea6(\u8ba1\u7b97\u56fe\u4e0b\u7ea2\u8272\u7684\u90e8\u5206)\u3002<\/p>\n

\"\u56fe\u7247\"<\/figure>\n
\"\u56fe\u7247\"
\u539f\u8c05\u6211\u5b57\u4e11\u2026\u2026<\/figcaption><\/figure>\n

\u6ce8\uff1a\u8fd9\u91cc\u7684\u53cd\u5411\u4f20\u64ad\u5047\u8bbe\u8f93\u51fa\u7aef\u63a5\u53d7\u4e4b\u524d\u56de\u4f20\u7684\u68af\u5ea6\u4e3a1\uff08\u4e5f\u53ef\u4ee5\u662f\u8f93\u51fa\u5bf9\u8f93\u51fa\u6c42\u5bfc=1\uff09<\/p>\n

\u89c2\u5bdf\u4e0a\u8ff0\u53cd\u5411\u4f20\u64ad\uff0c\u4e0d\u96be\u53d1\u73b0\uff0c\u5728\u8f93\u51fa\u7aef\u68af\u5ea6\u7684\u6a21\u503c\uff0c\u7ecf\u8fc7\u56de\u4f20\u6269\u5927\u4e863~4\u500d\u3002<\/p>\n

\u8fd9\u662f\u7531\u4e8e\u53cd\u5411\u4f20\u64ad\u7ed3\u679c\u7684\u6570\u503c\u5927\u5c0f<\/strong>\u4e0d\u6b62\u53d6\u51b3\u4e8e\u6c42\u5bfc\u7684\u5f0f\u5b50\uff0c\u5f88\u5927\u7a0b\u5ea6\u4e0a\u4e5f\u53d6\u51b3\u4e8e\u8f93\u5165\u7684\u6a21\u503c\u3002<\/strong>\u5f53\u8ba1\u7b97\u56fe\u6bcf\u6b21\u8f93\u5165\u7684\u6a21\u503c\u90fd\u5927\u4e8e1\uff0c\u90a3\u4e48\u7ecf\u8fc7\u5f88\u591a\u5c42\u56de\u4f20\uff0c\u68af\u5ea6\u5c06\u4e0d\u53ef\u907f\u514d\u5730\u5448\u51e0\u4f55\u500d\u6570\u589e\u957f\uff08\u6bcf\u6b21\u90fd\u53d8\u62103~4\u500d\uff0c\u91cd\u590d\u4e0a\u4e07\u6b21\uff0c\u60f3\u8c61\u4e00\u4e0b310000\u6709\u591a\u5927\u2026\u2026\uff09\uff0c\u76f4\u5230Nan\u3002\u8fd9\u5c31\u662f\u68af\u5ea6\u7206\u70b8\u73b0\u8c61\u3002<\/p>\n

\u5f53\u7136\u53cd\u8fc7\u6765\uff0c\u5982\u679c\u6211\u4eec\u6bcf\u4e2a\u9636\u6bb5\u8f93\u5165\u7684\u6a21\u6052\u5c0f\u4e8e1\uff0c\u90a3\u4e48\u68af\u5ea6\u4e5f\u5c06\u4e0d\u53ef\u907f\u514d\u5730\u5448\u51e0\u4f55\u500d\u6570\u4e0b\u964d\uff08\u6bd4\u5982\u6bcf\u6b21\u90fd\u53d8\u6210\u539f\u6765\u7684\u4e09\u5206\u4e4b\u4e00\uff0c\u91cd\u590d\u4e00\u4e07\u6b21\u5c31\u662f3-10000\uff09\uff0c\u76f4\u52300\u3002\u8fd9\u5c31\u662f\u68af\u5ea6\u6d88\u5931\u73b0\u8c61\u3002\u503c\u5f97\u4e00\u63d0\u7684\u662f\uff0c\u7531\u4e8e\u4eba\u4e3a\u7684\u53c2\u6570\u8bbe\u7f6e\uff0c\u68af\u5ea6\u66f4\u503e\u5411\u4e8e\u6d88\u5931\u800c\u4e0d\u662f\u7206\u70b8\u3002<\/p>\n

\u7531\u4e8e\u81f3\u4eca\u795e\u7ecf\u7f51\u7edc\u90fd\u4ee5\u53cd\u5411\u4f20\u64ad\u4e3a\u53c2\u6570\u66f4\u65b0\u7684\u57fa\u7840\uff0c\u6240\u4ee5\u68af\u5ea6\u6d88\u5931\u95ee\u9898\u542c\u8d77\u6765\u5f88\u6709\u9053\u7406\u3002\u7136\u800c\uff0c\u4e8b\u5b9e\u4e5f\u5e76\u975e\u5982\u6b64\uff0c\u81f3\u5c11\u4e0d\u6b62\u5982\u6b64\u3002<\/p>\n

\u6211\u4eec\u73b0\u5728\u65e0\u8bba\u7528Pytorch\u8fd8\u662fTensorflow\uff0c\u90fd\u4f1a\u81ea\u7136\u800c\u7136\u5730\u52a0\u4e0aBacth Normalization(\u7b80\u79f0BN)\uff0c\u800cBN\u7684\u4f5c\u7528\u672c\u8d28\u4e0a\u4e5f\u662f\u63a7\u5236\u6bcf\u5c42\u8f93\u5165\u7684\u6a21\u503c\uff0c<\/strong>\u56e0\u6b64\u68af\u5ea6\u7684\u7206\u70b8\/\u6d88\u5931\u73b0\u8c61\u7406\u5e94\u5728\u5f88\u65e9\u5c31\u88ab\u89e3\u51b3\u4e86\uff08\u81f3\u5c11\u89e3\u51b3\u4e86\u5927\u534a\uff09\u3002<\/p>\n

\u4e0d\u662f\u8fc7\u62df\u5408\uff0c\u4e5f\u4e0d\u662f\u68af\u5ea6\u6d88\u5931\uff0c\u8fd9\u5c31\u5f88\u5c34\u5c2c\u4e86\u2026\u2026CNN\u6ca1\u6709\u9047\u5230\u6211\u4eec\u719f\u77e5\u7684\u4e24\u4e2a\u8001\u5927\u96be\u95ee\u9898\uff0c\u5374\u8fd8\u662f\u968f\u7740\u6a21\u578b\u7684\u52a0\u6df1\u800c\u5bfc\u81f4\u6548\u679c\u9000\u5316\u3002\u65e0\u9700\u4efb\u4f55\u6570\u5b66\u8bba\u8bc1\uff0c\u6211\u4eec\u90fd\u4f1a\u89c9\u5f97\u8fd9\u4e0d\u7b26\u5408\u5e38\u7406\u3002\u7b49\u7b49\uff0c\u4e0d\u7b26\u5408\u5e38\u7406\u2026\u2026<\/p>\n

3.\u00a0\u4e3a\u4ec0\u4e48\u6a21\u578b\u9000\u5316\u4e0d\u7b26\u5408\u5e38\u7406\uff1f<\/strong><\/h3>\n

\u6309\u7406\u8bf4\uff0c\u5f53\u6211\u4eec\u5806\u53e0\u4e00\u4e2a\u6a21\u578b\u65f6\uff0c\u7406\u6240\u5f53\u7136\u7684\u4f1a\u8ba4\u4e3a\u6548\u679c\u4f1a\u8d8a\u5806\u8d8a\u597d\u3002\u56e0\u4e3a\uff0c\u5047\u8bbe\u4e00\u4e2a\u6bd4\u8f83\u6d45\u7684\u7f51\u7edc\u5df2\u7ecf\u53ef\u4ee5\u8fbe\u5230\u4e0d\u9519\u7684\u6548\u679c\uff0c\u90a3\u4e48\u5373\u4f7f\u4e4b\u540e\u5806\u4e0a\u53bb\u7684\u7f51\u7edc\u4ec0\u4e48\u4e5f\u4e0d\u505a\uff0c\u6a21\u578b\u7684\u6548\u679c\u4e5f\u4e0d\u4f1a\u53d8\u5dee<\/strong>\u3002<\/p>\n

\u7136\u800c\u4e8b\u5b9e\u4e0a\uff0c\u8fd9\u5374\u662f\u95ee\u9898\u6240\u5728\u3002\u201c\u4ec0\u4e48\u90fd\u4e0d\u505a\u201d\u6070\u597d\u662f\u5f53\u524d\u795e\u7ecf\u7f51\u7edc\u6700\u96be\u505a\u5230\u7684\u4e1c\u897f\u4e4b\u4e00\u3002<\/em><\/p>\n

MobileNet V2\u7684\u8bba\u6587[2]\u4e5f\u63d0\u5230\u8fc7\u7c7b\u4f3c\u7684\u73b0\u8c61\uff0c\u7531\u4e8e\u975e\u7ebf\u6027\u6fc0\u6d3b\u51fd\u6570Relu\u7684\u5b58\u5728\uff0c\u6bcf\u6b21\u8f93\u5165\u5230\u8f93\u51fa\u7684\u8fc7\u7a0b\u90fd\u51e0\u4e4e\u662f\u4e0d\u53ef\u9006\u7684\uff08\u4fe1\u606f\u635f\u5931\uff09\u3002\u6211\u4eec\u5f88\u96be\u4ece\u8f93\u51fa\u53cd\u63a8\u56de\u5b8c\u6574\u7684\u8f93\u5165\u3002<\/p>\n

\"\u56fe\u7247\"
Mobilenet v2\u662f\u8003\u8651\u7684\u7ed3\u679c\u662f\u53bb\u6389\u4f4e\u7ef4\u7684Relu\u4ee5\u4fdd\u7559\u4fe1\u606f<\/figcaption><\/figure>\n

\u4e5f\u8bb8\u8d4b\u4e88\u795e\u7ecf\u7f51\u7edc\u65e0\u9650\u53ef\u80fd\u6027\u7684\u201c\u975e\u7ebf\u6027\u201d\u8ba9\u795e\u7ecf\u7f51\u7edc\u6a21\u578b\u8d70\u5f97\u592a\u8fdc\uff0c\u5374\u4e5f\u8ba9\u5b83\u5fd8\u8bb0\u4e86\u4e3a\u4ec0\u4e48\u51fa\u53d1\uff08\u60f3\u60f3\u8fd8\u633a\u54f2\u5b66\uff09\u3002\u8fd9\u4e5f\u4f7f\u5f97\u7279\u5f81\u968f\u7740\u5c42\u5c42\u524d\u5411\u4f20\u64ad\u5f97\u5230\u5b8c\u6574\u4fdd\u7559\uff08\u4ec0\u4e48\u4e5f\u4e0d\u505a\uff09\u7684\u53ef\u80fd\u6027\u90fd\u5fae\u4e4e\u5176\u5fae\u3002<\/p>\n

\u7528\u5b66\u672f\u70b9\u7684\u8bdd\u8bf4\uff0c\u8fd9\u79cd\u795e\u7ecf\u7f51\u7edc\u4e22\u5931\u7684\u201c\u4e0d\u5fd8\u521d\u5fc3\u201d\/\u201c\u4ec0\u4e48\u90fd\u4e0d\u505a\u201d\u7684\u54c1\u8d28\u53eb\u505a\u6052\u7b49\u6620\u5c04\uff08identity mapping\uff09<\/strong>\u3002<\/p>\n

\u56e0\u6b64\uff0c\u53ef\u4ee5\u8ba4\u4e3aResidual Learning\u7684\u521d\u8877\uff0c\u5176\u5b9e\u662f\u8ba9\u6a21\u578b\u7684\u5185\u90e8\u7ed3\u6784\u81f3\u5c11\u6709\u6052\u7b49\u6620\u5c04\u7684\u80fd\u529b\u3002\u4ee5\u4fdd\u8bc1\u5728\u5806\u53e0\u7f51\u7edc\u7684\u8fc7\u7a0b\u4e2d\uff0c\u7f51\u7edc\u81f3\u5c11\u4e0d\u4f1a\u56e0\u4e3a\u7ee7\u7eed\u5806\u53e0\u800c\u4ea7\u751f\u9000\u5316\uff01<\/p>\n

\u4e8c\u3001\u6df1\u5ea6\u6b8b\u5dee\u5b66\u4e60 Deep Residual Learning<\/strong><\/h2>\n

1. \u6b8b\u5dee\u5b66\u4e60 Residual Learning<\/strong><\/p>\n

\u524d\u9762\u5206\u6790\u5f97\u51fa\uff0c\u5982\u679c\u6df1\u5c42\u7f51\u7edc\u540e\u9762\u7684\u5c42\u90fd\u662f\u662f\u6052\u7b49\u6620\u5c04<\/strong>\uff0c\u90a3\u4e48\u6a21\u578b\u5c31\u53ef\u4ee5\u8f6c\u5316\u4e3a\u4e00\u4e2a\u6d45\u5c42\u7f51\u7edc\u3002\u90a3\u73b0\u5728\u7684\u95ee\u9898\u5c31\u662f\u5982\u4f55\u5f97\u5230\u6052\u7b49\u6620\u5c04<\/strong>\u4e86\u3002<\/p>\n

\u4e8b\u5b9e\u4e0a\uff0c\u5df2\u6709\u7684\u795e\u7ecf\u7f51\u7edc\u5f88\u96be\u62df\u5408\u6f5c\u5728\u7684\u6052\u7b49\u6620\u5c04\u51fd\u6570H(x) = x\u3002<\/p>\n

\u4f46\u5982\u679c\u628a\u7f51\u7edc\u8bbe\u8ba1\u4e3aH(x) = F(x) + x\uff0c\u5373\u76f4\u63a5\u628a\u6052\u7b49\u6620\u5c04\u4f5c\u4e3a\u7f51\u7edc\u7684\u4e00\u90e8\u5206<\/strong>\u3002\u5c31\u53ef\u4ee5\u628a\u95ee\u9898\u8f6c\u5316\u4e3a\u5b66\u4e60\u4e00\u4e2a\u6b8b\u5dee\u51fd\u6570F(x) = H(x) – x.<\/strong><\/p>\n

\u53ea\u8981F(x)=0\uff0c\u5c31\u6784\u6210\u4e86\u4e00\u4e2a\u6052\u7b49\u6620\u5c04H(x) = x\u3002\u800c\u4e14\uff0c\u62df\u5408\u6b8b\u5dee\u81f3\u5c11\u6bd4\u62df\u5408\u6052\u7b49\u6620\u5c04\u5bb9\u6613\u5f97\u591a\u3002<\/p>\n

\u4e8e\u662f\uff0c\u5c31\u6709\u4e86\u8bba\u6587[3]\u4e2d\u7684Residual block\u7ed3\u6784<\/p>\n

\"\u56fe\u7247\"
Residual Block\u7684\u7ed3\u6784<\/figcaption><\/figure>\n

\u56fe\u4e2d\u53f3\u4fa7\u7684\u66f2\u7ebf\u53eb\u505a\u8df3\u63a5\uff08shortcut connection\uff09\uff0c\u901a\u8fc7\u8df3\u63a5\u5728\u6fc0\u6d3b\u51fd\u6570\u524d\uff0c<\/strong>\u5c06\u4e0a\u4e00\u5c42\uff08\u6216\u51e0\u5c42\uff09\u4e4b\u524d\u7684\u8f93\u51fa\u4e0e\u672c\u5c42<\/strong>\u8ba1\u7b97\u7684\u8f93\u51fa\u76f8\u52a0<\/strong>\uff0c\u5c06\u6c42\u548c\u7684\u7ed3\u679c\u8f93\u5165\u5230\u6fc0\u6d3b\u51fd\u6570\u4e2d\u505a\u4e3a\u672c\u5c42\u7684\u8f93\u51fa\u3002<\/p>\n

\u7528\u6570\u5b66\u8bed\u8a00\u63cf\u8ff0\uff0c\u5047\u8bbeResidual Block\u7684\u8f93\u5165\u4e3a\u00a0\"\u56fe\u7247\"\u00a0\uff0c\u5219\u8f93\u51fa\u00a0\"\u56fe\u7247\"\u00a0\u7b49\u4e8e\uff1a<\/p>\n

\"\u56fe\u7247\"<\/figure>\n

\u5176\u4e2d\u00a0\"\u56fe\u7247\"\u00a0\u662f\u6211\u4eec\u5b66\u4e60\u7684\u76ee\u6807\uff0c\u5373\u8f93\u51fa\u8f93\u5165\u7684\u6b8b\u5dee\u00a0\"\u56fe\u7247\"\u00a0\u3002\u4ee5\u4e0a\u56fe\u4e3a\u4f8b\uff0c\u6b8b\u5dee\u90e8\u5206\u662f\u4e2d\u95f4\u6709\u4e00\u4e2aRelu\u6fc0\u6d3b\u7684\u53cc\u5c42\u6743\u91cd\uff0c\u5373\uff1a<\/p>\n

\"\u56fe\u7247\"<\/figure>\n

\u5176\u4e2d\u00a0\"\u56fe\u7247\"\u00a0\u6307\u4ee3Relu\uff0c\u800c\u00a0\"\u56fe\u7247\"\u00a0\u6307\u4ee3\u4e24\u5c42\u6743\u91cd\u3002<\/p>\n

\u987a\u5e26\u4e00\u63d0\uff0c\u8fd9\u91cc\u4e00\u4e2aBlock\u4e2d\u5fc5\u987b\u81f3\u5c11\u542b\u6709\u4e24\u4e2a\u5c42<\/strong>\uff0c\u5426\u5219\u5c31\u4f1a\u51fa\u73b0\u5f88\u6ed1\u7a3d\u7684\u60c5\u51b5\uff1a<\/em><\/p>\n

\"\u56fe\u7247\"<\/figure>\n

\u663e\u7136\u8fd9\u6837\u52a0\u4e86\u548c\u6ca1\u52a0\u5dee\u4e0d\u591a\u2026\u2026<\/em><\/p>\n

2.\u7f51\u7edc\u7ed3\u6784\u4e0e\u7ef4\u5ea6\u95ee\u9898<\/strong><\/p>\n

\"\u56fe\u7247\"
ResNet\u7ed3\u6784\u793a\u610f\u56fe\uff08\u5de6\u5230\u53f3\u5206\u522b\u662fVGG\uff0c\u6ca1\u6709\u6b8b\u5dee\u7684PlainNet\uff0c\u6709\u6b8b\u5dee\u7684ResNet\uff09<\/figcaption><\/figure>\n

\u8bba\u6587\u4e2d\u539f\u59cb\u7684ResNet34\u4e0eVGG\u7684\u7ed3\u6784\u5982\u4e0a\u56fe\u6240\u793a\uff0c\u53ef\u4ee5\u770b\u5230\u5373\u4f7f\u662f\u5f53\u5e74\u53f7\u79f0\u201cVery Deep\u201d\u7684VGG\uff0c\u548c\u6700\u57fa\u7840\u7684Resnet\u5728\u6df1\u5ea6\u4e0a\u76f8\u6bd4\u90fd\u662f\u4e2a\u5f1f\u5f1f\u3002<\/p>\n

\u53ef\u80fd\u6709\u597d\u5947\u5fc3\u5b9d\u5b9d\u53d1\u73b0\u4e86\uff0c\u8df3\u63a5\u7684\u66f2\u7ebf\u4e2d\u5927\u90e8\u5206\u662f\u5b9e\u73b0\uff0c\u4f46\u4e5f\u6709\u5c11\u90e8\u5206\u865a\u7ebf\u3002\u8fd9\u4e9b\u865a\u7ebf\u7684\u4ee3\u8868\u8fd9\u4e9bBlock\u524d\u540e\u7684\u7ef4\u5ea6\u4e0d\u4e00\u81f4\uff0c\u56e0\u4e3a\u53bb\u6389\u6b8b\u5dee\u7ed3\u6784\u7684Plain\u7f51\u7edc\u8fd8\u662f\u53c2\u7167\u4e86VGG\u7ecf\u5178\u7684\u8bbe\u8ba1\u601d\u8def\uff1a\u6bcf\u9694x\u5c42\uff0c\u7a7a\u95f4\u4e0a\/2\uff08\u4e0b\u91c7\u6837\uff09\u4f46\u6df1\u5ea6\u7ffb\u500d\u3002<\/strong><\/p>\n

\u4e5f\u5c31\u662f\u8bf4\uff0c\u7ef4\u5ea6\u4e0d\u4e00\u81f4\u4f53\u73b0\u5728\u4e24\u4e2a\u5c42\u9762<\/strong>\uff1a<\/p>\n

    \n
  • \u7a7a\u95f4\u4e0a<\/strong>\u4e0d\u4e00\u81f4<\/li>\n
  • \u6df1\u5ea6\u4e0a<\/strong>\u4e0d\u4e00\u81f4<\/li>\n<\/ul>\n

    \u7a7a\u95f4\u4e0a<\/strong>\u4e0d\u4e00\u81f4\u5f88\u7b80\u5355\uff0c\u53ea\u9700\u8981\u5728\u8df3\u63a5\u7684\u90e8\u5206\u7ed9\u8f93\u5165x\u52a0\u4e0a\u4e00\u4e2a\u7ebf\u6027\u6620\u5c04\u00a0\"\u56fe\u7247\"\u00a0,\u5373\uff1a<\/p>\n

    \"\u56fe\u7247\"<\/figure>\n

    \u800c\u5bf9\u4e8e\u6df1\u5ea6\u4e0a<\/strong>\u7684\u4e0d\u4e00\u81f4\uff0c\u5219\u6709\u4e24\u79cd\u89e3\u51b3\u529e\u6cd5\uff0c\u4e00\u79cd\u662f\u5728\u8df3\u63a5\u8fc7\u7a0b\u4e2d\u52a0\u4e00\u4e2a1*1\u7684\u5377\u79ef\u5c42\u8fdb\u884c\u5347\u7ef4\uff0c\u53e6\u4e00\u79cd\u5219\u662f\u76f4\u63a5\u7b80\u5355\u7c97\u66b4\u5730\u8865\u96f6\u3002\u4e8b\u5b9e\u8bc1\u660e\u4e24\u79cd\u65b9\u6cd5\u90fd\u884c\u5f97\u901a\u3002<\/p>\n

    \u6ce8\uff1a\u6df1\u5ea6\u4e0a\u548c\u7a7a\u95f4\u4e0a\u7ef4\u5ea6\u7684\u4e0d\u4e00\u81f4\u662f\u5206\u5f00\u5904\u7406\u7684\uff0c\u4f46\u5f88\u591a\u4eba\u5c06\u4e24\u8005\u6df7\u4e3a\u4e00\u8c08\uff08\u5305\u62ec\u76ee\u524d\u67d0\u4e4e\u4e00\u4e9b\u9ad8\u8d5e\u6587\u7ae0\uff09\uff0c\u8fd9\u5bfc\u81f4\u4e86\u4e00\u4e9b\u4eba\u5728\u6a21\u578b\u7684\u5b9e\u73b0\u4e0a\u611f\u5230\u56f0\u60d1\uff08\u6bd4\u5982\u5f53\u5e74\u7684\u6211\uff09\u3002<\/em><\/p>\n

    3. torchvision\u4e2d\u7684\u5b98\u65b9\u5b9e\u73b0<\/strong><\/h3>\n

    \u4e8b\u5b9e\u4e0a\u8bba\u6587\u4e2d\u7684ResNet\u5e76\u4e0d\u662f\u6700\u5e38\u7528\u7684\uff0c\u6211\u4eec\u53ef\u4ee5\u5728Torchvision\u7684\u6a21\u578b\u5e93\u4e2d\u627e\u5230\u4e00\u4e9b\u5f88\u4e0d\u9519\u7684\u4f8b\u5b50\uff0c\u8fd9\u91cc\u62ffResnet18\u4e3a\u4f8b\uff1a<\/p>\n

    \"\u56fe\u7247\"<\/figure>\n

    \u8fd0\u884c\u4ee3\u7801\uff1a<\/p>\n

    \n
    \n
    \n
      \n
    • <\/li>\n
    • <\/li>\n
    • <\/li>\n<\/ul>\n
      import torchvision<\/span><\/code>model = torchvision.models.resnet18(pretrained=False) #\u6211\u4eec\u4e0d\u4e0b\u8f7d\u9884\u8bad\u7ec3\u6743\u91cd<\/span><\/code>print(model)<\/span><\/code><\/pre>\n<\/section>\n<\/section>\n<\/section>\n
      <\/pre>\n
      \u5f97\u5230\u8f93\u51fa\uff1a<\/p>\n
      \n
      \n
      \n
        \n
      • <\/li>\n
      • <\/li>\n
      • <\/li>\n
      • <\/li>\n
      • <\/li>\n
      • <\/li>\n
      • <\/li>\n
      • <\/li>\n
      • <\/li>\n
      • <\/li>\n
      • <\/li>\n
      • <\/li>\n
      • <\/li>\n
      • <\/li>\n
      • <\/li>\n
      • <\/li>\n
      • <\/li>\n
      • <\/li>\n
      • <\/li>\n
      • <\/li>\n
      • <\/li>\n
      • <\/li>\n
      • <\/li>\n
      • <\/li>\n
      • <\/li>\n
      • <\/li>\n
      • <\/li>\n
      • <\/li>\n
      • <\/li>\n
      • <\/li>\n
      • <\/li>\n
      • <\/li>\n
      • <\/li>\n
      • <\/li>\n
      • <\/li>\n
      • <\/li>\n
      • <\/li>\n
      • <\/li>\n
      • <\/li>\n
      • <\/li>\n
      • <\/li>\n
      • <\/li>\n
      • <\/li>\n
      • <\/li>\n
      • <\/li>\n
      • <\/li>\n
      • <\/li>\n
      • <\/li>\n
      • <\/li>\n
      • <\/li>\n
      • <\/li>\n
      • <\/li>\n
      • <\/li>\n
      • <\/li>\n
      • <\/li>\n
      • <\/li>\n
      • <\/li>\n
      • <\/li>\n
      • <\/li>\n
      • <\/li>\n
      • <\/li>\n
      • <\/li>\n
      • <\/li>\n
      • <\/li>\n
      • <\/li>\n
      • <\/li>\n
      • <\/li>\n
      • <\/li>\n
      • <\/li>\n
      • <\/li>\n
      • <\/li>\n
      • <\/li>\n
      • <\/li>\n
      • <\/li>\n
      • <\/li>\n
      • <\/li>\n
      • <\/li>\n
      • <\/li>\n
      • <\/li>\n
      • <\/li>\n
      • <\/li>\n
      • <\/li>\n
      • <\/li>\n
      • <\/li>\n<\/ul>\n
        ResNet(<\/span><\/code>  (conv1): Conv2d(3, 64, kernel_size=(7, 7), stride=(2, 2), padding=(3, 3), bias=False)<\/span><\/code>  (bn1): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)<\/span><\/code>  (relu): ReLU(inplace)<\/span><\/code>  (maxpool): MaxPool2d(kernel_size=3, stride=2, padding=1, dilation=1, ceil_mode=False)<\/span><\/code>  (layer1): Sequential(<\/span><\/code>    (0): BasicBlock(<\/span><\/code>      (conv1): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)<\/span><\/code>      (bn1): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)<\/span><\/code>      (relu): ReLU(inplace)<\/span><\/code>      (conv2): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)<\/span><\/code>      (bn2): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)<\/span><\/code>    )<\/span><\/code>    (1): BasicBlock(<\/span><\/code>      (conv1): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)<\/span><\/code>      (bn1): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)<\/span><\/code>      (relu): ReLU(inplace)<\/span><\/code>      (conv2): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)<\/span><\/code>      (bn2): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)<\/span><\/code>    )<\/span><\/code>  )<\/span><\/code>  (layer2): Sequential(<\/span><\/code>    (0): BasicBlock(<\/span><\/code>      (conv1): Conv2d(64, 128, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)<\/span><\/code>      (bn1): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)<\/span><\/code>      (relu): ReLU(inplace)<\/span><\/code>      (conv2): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)<\/span><\/code>      (bn2): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)<\/span><\/code>      (downsample): Sequential(<\/span><\/code>        (0): Conv2d(64, 128, kernel_size=(1, 1), stride=(2, 2), bias=False)<\/span><\/code>        (1): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)<\/span><\/code>      )<\/span><\/code>    )<\/span><\/code>    (1): BasicBlock(<\/span><\/code>      (conv1): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)<\/span><\/code>      (bn1): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)<\/span><\/code>      (relu): ReLU(inplace)<\/span><\/code>      (conv2): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)<\/span><\/code>      (bn2): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)<\/span><\/code>    )<\/span><\/code>  )<\/span><\/code>  (layer3): Sequential(<\/span><\/code>    (0): BasicBlock(<\/span><\/code>      (conv1): Conv2d(128, 256, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)<\/span><\/code>      (bn1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)<\/span><\/code>      (relu): ReLU(inplace)<\/span><\/code>      (conv2): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)<\/span><\/code>      (bn2): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)<\/span><\/code>      (downsample): Sequential(<\/span><\/code>        (0): Conv2d(128, 256, kernel_size=(1, 1), stride=(2, 2), bias=False)<\/span><\/code>        (1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)<\/span><\/code>      )<\/span><\/code>    )<\/span><\/code>    (1): BasicBlock(<\/span><\/code>      (conv1): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)<\/span><\/code>      (bn1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)<\/span><\/code>      (relu): ReLU(inplace)<\/span><\/code>      (conv2): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)<\/span><\/code>      (bn2): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)<\/span><\/code>    )<\/span><\/code>  )<\/span><\/code>  (layer4): Sequential(<\/span><\/code>    (0): BasicBlock(<\/span><\/code>      (conv1): Conv2d(256, 512, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)<\/span><\/code>      (bn1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)<\/span><\/code>      (relu): ReLU(inplace)<\/span><\/code>      (conv2): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)<\/span><\/code>      (bn2): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)<\/span><\/code>      (downsample): Sequential(<\/span><\/code>        (0): Conv2d(256, 512, kernel_size=(1, 1), stride=(2, 2), bias=False)<\/span><\/code>        (1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)<\/span><\/code>      )<\/span><\/code>    )<\/span><\/code>    (1): BasicBlock(<\/span><\/code>      (conv1): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)<\/span><\/code>      (bn1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)<\/span><\/code>      (relu): ReLU(inplace)<\/span><\/code>      (conv2): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)<\/span><\/code>      (bn2): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)<\/span><\/code>    )<\/span><\/code>  )<\/span><\/code>  (avgpool): AvgPool2d(kernel_size=7, stride=1, padding=0)<\/span><\/code>  (fc): Linear(in_features=512, out_features=1000, bias=True)<\/span><\/code>)<\/span><\/code><\/pre>\n<\/section>\n<\/section>\n<\/section>\n
        <\/pre>\n
        \u00a0<\/code><\/pre>\n
        \u85b0\u98ce\u8bf4 Thinkings<\/strong><\/h2>\n
        \u4e0a\u8ff0\u7684\u5185\u5bb9\u662f\u6211\u4ee5\u81ea\u5df1\u7684\u89d2\u5ea6\u601d\u8003\u4f5c\u8005\u63d0\u51faResNet\u7684\u5fc3\u8def\u5386\u7a0b\uff0c\u6211\u6bd4\u4f5c\u8005\u8521\u5f88\u591a\uff0c\u6240\u4ee5\u96be\u514d\u51fa\u73b0\u601d\u8003\u4e0d\u5168\u7684\u5730\u65b9\u3002<\/p>\n
        ResNet\u662f\u5982\u6b64\u7b80\u6d01\u9ad8\u6548\uff0c\u4ee5\u81f3\u4e8e\u6a21\u578b\u63d0\u51fa\u540e\u8fd8\u6709\u65e0\u6570\u8bba\u6587\u8ba8\u8bba\u201cResNet\u5230\u5e95\u89e3\u51b3\u4e86\u4ec0\u4e48\u95ee\u9898(The Shattered Gradients Problem: If resnets are the answer, then what is the question?)\u201d[4]<\/p>\n
        \u8bba\u6587[4]\u8ba4\u4e3a\uff0c\u5373\u4f7fBN\u8fc7\u540e\u68af\u5ea6\u7684\u6a21\u7a33\u5b9a\u5728\u4e86\u6b63\u5e38\u8303\u56f4\u5185\uff0c\u4f46\u68af\u5ea6\u7684\u76f8\u5173\u6027\u5b9e\u9645\u4e0a\u662f\u968f\u7740\u5c42\u6570\u589e\u52a0\u6301\u7eed\u8870\u51cf\u7684<\/strong>\u3002\u800c\u7ecf\u8fc7\u8bc1\u660e\uff0cResNet\u53ef\u4ee5\u6709\u6548\u51cf\u5c11\u8fd9\u79cd\u76f8\u5173\u6027\u7684\u8870\u51cf\u3002<\/p>\n
        \u5bf9\u4e8e\u00a0\"\u56fe\u7247\"\u00a0\u5c42\u7684\u7f51\u7edc\u6765\u8bf4\uff0c\u6ca1\u6709\u6b8b\u5dee\u8868\u793a\u7684Plain Net\u68af\u5ea6\u76f8\u5173\u6027\u7684\u8870\u51cf\u5728\u00a0\"\u56fe\u7247\"\u00a0\uff0c\u800cResNet\u7684\u8870\u51cf\u5374\u53ea\u6709\u00a0\"\u56fe\u7247\"\u00a0\u3002\u8fd9\u4e5f\u9a8c\u8bc1\u4e86ResNet\u8bba\u6587\u672c\u8eab\u7684\u89c2\u70b9\uff0c\u7f51\u7edc\u8bad\u7ec3\u96be\u5ea6\u968f\u7740\u5c42\u6570\u589e\u957f\u7684\u901f\u5ea6\u4e0d\u662f\u7ebf\u6027\uff0c\u800c\u81f3\u5c11\u662f\u591a\u9879\u5f0f\u7b49\u7ea7\u7684\u589e\u957f\uff08\u5982\u679c\u8be5\u8bba\u6587\u5c5e\u5b9e\uff0c\u5219\u53ef\u80fd\u662f\u6307\u6570\u7ea7\u589e\u957f\u7684\uff09<\/p>\n
        \u800c\u5bf9\u4e8e\u201c\u68af\u5ea6\u5f25\u6563\u201d\u89c2\u70b9\u6765\u8bf4\uff0c\u5728\u8f93\u51fa\u5f15\u5165\u4e00\u4e2a\u8f93\u5165x\u7684\u6052\u7b49\u6620\u5c04\uff0c\u5219\u68af\u5ea6\u4e5f\u4f1a\u5bf9\u5e94\u5730\u5f15\u5165\u4e00\u4e2a\u5e38\u65701\uff0c\u8fd9\u6837\u7684\u7f51\u7edc\u7684\u786e\u4e0d\u5bb9\u6613\u51fa\u73b0\u68af\u5ea6\u503c\u5f02\u5e38\uff0c\u5728\u67d0\u79cd\u610f\u4e49\u4e0a\uff0c\u8d77\u5230\u4e86\u7a33\u5b9a\u68af\u5ea6\u7684\u4f5c\u7528\u3002<\/p>\n
        \u9664\u6b64\u4e4b\u5916\uff0cshortcut\u7c7b\u4f3c\u7684\u65b9\u6cd5\u4e5f\u5e76\u4e0d\u662f\u7b2c\u4e00\u6b21\u63d0\u51fa\uff0c\u4e4b\u524d\u5c31\u6709\u201cHighway Networks\u201d\u3002\u53ef\u4ee5\u53ea\u7ba1\u7406\u89e3\u4e3a\uff0c\u4ee5\u5f80\u53c2\u6570\u8981\u5f97\u5230\u68af\u5ea6\uff0c\u9700\u8981\u5feb\u9012\u5458\u5c06\u68af\u5ea6\u4e00\u5c42\u4e00\u5c42\u4e2d\u8f6c\u5230\u53c2\u6570\u624b\u4e2d\uff08\u5c31\u50cf\u6211\u53d6\u4e2a\u5feb\u9012\uff0c\u90fd\u663e\u793a\u8981\u4ece\u201c\u4e0a\u6d77\u5e02\u201d\u53d1\u5f80\u201c\u95f5\u884c\u5206\u62e3\u4e2d\u5fc3\u201d\uff0c\u95f5\u5927\u8352\u65e5\u5e38\u88ab\u8e22\u51fa\u4e0a\u6d77\u7c4d\uff09\u3002\u800c\u8df3\u63a5\u5b9e\u9645\u4e0a\u7ed9\u68af\u5ea6\u5f00\u4e86\u4e00\u6761\u201c\u9ad8\u901f\u516c\u8def\u201d\uff08\u53d6\u5feb\u9012\u53ef\u4ee5\u76f4\u63a5\u7528\u65e0\u4eba\u673a\u7a7a\u6295\u5230\u6211\u624b\u91cc\u4e86\uff09\uff0c\u6548\u7387\u81ea\u7136\u5927\u5e45\u63d0\u9ad8\uff0c\u4e0d\u8fc7\u8fd9\u53ea\u662f\u4e2a\u6bd4\u8f83\u60f3\u5f53\u7136\u7684\u7406\u7531\u3002<\/p>\n
        \"\u56fe\u7247\"<\/figure>\n
        \u4e0a\u9762\u7684\u7406\u89e3\u5f88\u591a\u8bba\u6587\u90fd\u8bb2\u8fc7\uff0c\u4f46\u6211\u4e2a\u4eba\u6700\u559c\u6b22\u4e0b\u9762\u4e24\u4e2a\u7406\u89e3\u3002<\/p>\n
        \u7b2c\u4e00\u4e2a\u5df2\u7ecf\u7531Feature Pyramid Network[5]\u63d0\u51fa\u4e86\uff0c\u90a3\u5c31\u662f\u8df3\u8fde\u63a5\u76f8\u52a0\u53ef\u4ee5\u5b9e\u73b0\u4e0d\u540c\u5206\u8fa8\u7387\u7279\u5f81\u7684\u7ec4\u5408\uff0c\u56e0\u4e3a\u6d45\u5c42\u5bb9\u6613\u6709\u9ad8\u5206\u8fa8\u7387\u4f46\u662f\u4f4e\u7ea7\u8bed\u4e49\u7684\u7279\u5f81\uff0c\u800c\u6df1\u5c42\u7684\u7279\u5f81\u6709\u9ad8\u7ea7\u8bed\u4e49\uff0c\u4f46\u5206\u8fa8\u7387\u5c31\u5f88\u4f4e\u4e86\u3002<\/p>\n
        \u7b2c\u4e8c\u4e2a\u7406\u89e3\u5219\u662f\u8bf4\uff0c\u5f15\u5165\u8df3\u63a5\u5b9e\u9645\u4e0a\u8ba9\u6a21\u578b\u81ea\u8eab\u6709\u4e86\u66f4\u52a0\u201c\u7075\u6d3b\u201d\u7684\u7ed3\u6784<\/strong>\uff0c\u5373\u5728\u8bad\u7ec3\u8fc7\u7a0b\u672c\u8eab\uff0c\u6a21\u578b\u53ef\u4ee5\u9009\u62e9\u5728\u6bcf\u4e00\u4e2a\u90e8\u5206\u662f\u201c\u66f4\u591a\u8fdb\u884c\u5377\u79ef\u4e0e\u975e\u7ebf\u6027\u53d8\u6362\u201d\u8fd8\u662f\u201c\u66f4\u591a\u503e\u5411\u4e8e\u4ec0\u4e48\u90fd\u4e0d\u505a\u201d\uff0c\u6291\u6216\u662f\u5c06\u4e24\u8005\u7ed3\u5408\u3002\u6a21\u578b\u5728\u8bad\u7ec3\u4fbf\u53ef\u4ee5\u81ea\u9002\u5e94\u672c\u8eab\u7684\u7ed3\u6784\uff0c\u8fd9\u542c\u8d77\u6765\u662f\u591a\u4e48\u9177\u7684\u4e00\u4ef6\u4e8b\u554a\uff01<\/p>\n
        \u6709\u7684\u4eba\u4e5f\u8bb8\u4f1a\u7eb3\u95f7\uff0c\u6211\u4eec\u5df2\u7ecf\u77e5\u9053\u4e00\u4e2a\u6a21\u578b\u7684\u6765\u9f99\u53bb\u8109\u4e86\uff0c\u90a3\u4e48\u5728\u4e00\u4e2a\u5ba2\u89c2\u4e0a\u5df2\u7ecf\u5341\u5206\u4f18\u79c0\u7684\u6a21\u578b\uff0c\u5f3a\u52a0\u90a3\u4e48\u591a\u4e3b\u89c2\u7684\u4e2a\u4eba\u5224\u65ad\u6709\u610f\u601d\u5417\uff1f<\/p>\n
        \u7136\u800c\u7b14\u8005\u8fd8\u662f\u76f8\u4fe1\uff0c\u66f4\u591a\u89d2\u5ea6\u7684\u601d\u8003\u6709\u52a9\u4e8e\u6211\u4eec\u53d1\u73b0\u73b0\u6709\u6a21\u578b\u7684\u4e0d\u8db3\uff0c\u4ee5\u53ca\u503c\u5f97\u6539\u8fdb\u7684\u70b9\u3002\u6bd4\u5982\u6211\u6700\u559c\u6b22\u7684\u4e24\u4e2a\u7406\u89e3\u5c31\u53ef\u4ee5\u5f15\u7533\u51fa\u8fd9\u6837\u7684\u95ee\u9898\u201c\u867d\u7136\u8df3\u63a5\u53ef\u4ee5\u7ed3\u5408\u4e0d\u540c\u5206\u8fa8\u7387\uff0c\u4f46ResNet\u663e\u7136\u6ca1\u6709\u5145\u5206\u5229\u7528\u8fd9\u4e2a\u4f18\u70b9\uff0c\u56e0\u4e3a\u6bcf\u4e2ashortcut\u9876\u591a\u8de8\u8d8a\u4e00\u79cd\u5206\u8fa8\u7387\uff08\u5927\u90e8\u5206\u8fd8\u4e0d\u4f1a\u53d1\u751f\u8de8\u8d8a\uff09\u201d\u3002<\/p>\n
        \u90a3\u4e48\u201c\u5982\u679c\u7528\u8df3\u63a5\u7ec4\u5408\u66f4\u591a\u5206\u8fa8\u7387\u7684\u7279\u5f81\uff0c\u6a21\u578b\u7684\u6548\u679c\u4f1a\u4e0d\u4f1a\u66f4\u597d\uff1f\u201d\u8fd9\u5c31\u662fDenseNet\u56de\u7b54\u6211\u4eec\u7684\u95ee\u9898\u4e86\u3002<\/p>\n
        <\/div>\n
        <\/g><\/svg><\/i> \"Loading\"<\/p>\n
        <\/div>\n","protected":false},"excerpt":{"rendered":"
        ResNet YYDS\uff01 ResNet\u53d1\u5e03\u4e8e2015\u5e74\uff0c\u76ee\u524d\u4ecd\u6709\u5927\u91cfCV\u4efb\u52a1\u7528\u5176\u4f5c\u4e3abackbone\uff08\u5c24\u5176\u662f\u9876 […]<\/p>\n
        <\/div>\n
        <\/g><\/svg><\/i> \"Loading\"<\/p>\n
        <\/div>\n","protected":false},"author":1,"featured_media":551,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[23],"tags":[],"class_list":["post-550","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-23"],"_links":{"self":[{"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/posts\/550","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/comments?post=550"}],"version-history":[{"count":1,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/posts\/550\/revisions"}],"predecessor-version":[{"id":563,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/posts\/550\/revisions\/563"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/media\/551"}],"wp:attachment":[{"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/media?parent=550"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/categories?post=550"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/tags?post=550"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}