{"id":7346,"date":"2024-03-29T12:01:02","date_gmt":"2024-03-29T04:01:02","guid":{"rendered":""},"modified":"2024-03-29T12:01:02","modified_gmt":"2024-03-29T04:01:02","slug":"EfficientNet\u7f51\u7edc\u8be6\u89e3","status":"publish","type":"post","link":"https:\/\/mushiming.com\/7346.html","title":{"rendered":"EfficientNet\u7f51\u7edc\u8be6\u89e3"},"content":{"rendered":"

\n <\/path> \n<\/svg> <\/p>\n

\u539f\u8bba\u6587\u540d\u79f0\uff1aEfficientNet: Rethinking Model Scaling for Convolutional Neural Networks<\/strong>
\u8bba\u6587\u4e0b\u8f7d\u5730\u5740\uff1ahttps:\/\/arxiv.org\/abs\/1905.11946
\u539f\u8bba\u6587\u63d0\u4f9b\u4ee3\u7801\uff1ahttps:\/\/github.com\/tensorflow\/tpu\/tree\/master\/models\/official\/efficientnet
\u81ea\u5df1\u4f7f\u7528Pytorch\u5b9e\u73b0\u7684\u4ee3\u7801\uff1a pytorch_classification\/Test9_efficientNet
\u81ea\u5df1\u4f7f\u7528Tensorflow\u5b9e\u73b0\u7684\u4ee3\u7801\uff1a tensorflow_classification\/Test9_efficientNet
\u4e0d\u60f3\u770b\u6587\u7ae0\u7684\u53ef\u4ee5\u770b\u4e0b\u6211\u5728bilibili\u4e0a\u5f55\u5236\u7684\u89c6\u9891\uff1ahttps:\/\/www.bilibili.com\/video\/BV1XK4y1U7PX<\/p>\n

\n
\n

\u76ee\u5f55<\/h4>\n
    \n
  • \n
      \n
    • 0 \u524d\u8a00<\/li>\n
    • 1 \u8bba\u6587\u601d\u60f3<\/li>\n
    • 2 \u7f51\u7edc\u8be6\u7ec6\u7ed3\u6784<\/li>\n
    • \n
        \n
      • 2.1 MBConv\u7ed3\u6784<\/li>\n
      • 2.2 EfficientNet(B0-B7)\u53c2\u6570<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<\/div>\n
        \n

        0 \u524d\u8a00<\/h3>\n

        \u5728\u4e4b\u524d\u7684\u4e00\u4e9b\u624b\u5de5\u8bbe\u8ba1\u7f51\u7edc\u4e2d(AlexNet\uff0cVGG\uff0cResNet\u7b49\u7b49)\u7ecf\u5e38\u6709\u4eba\u95ee\uff0c\u4e3a\u4ec0\u4e48\u8f93\u5165\u56fe\u50cf\u5206\u8fa8\u7387\u8981\u56fa\u5b9a\u4e3a224\uff0c\u4e3a\u4ec0\u4e48\u5377\u79ef\u7684\u4e2a\u6570\u8981\u8bbe\u7f6e\u4e3a\u8fd9\u4e2a\u503c\uff0c\u4e3a\u4ec0\u4e48\u7f51\u7edc\u7684\u6df1\u5ea6\u8bbe\u4e3a\u8fd9\u4e48\u6df1\uff1f\u8fd9\u4e9b\u95ee\u9898\u4f60\u8981\u95ee\u8bbe\u8ba1\u4f5c\u8005\u7684\u8bdd\uff0c\u4f30\u8ba1\u56de\u590d\u5c31\u56db\u4e2a\u5b57\u2014\u2014\u5de5\u7a0b\u7ecf\u9a8c\u3002\u800c\u8fd9\u7bc7\u8bba\u6587\u4e3b\u8981\u662f\u7528NAS\uff08Neural Architecture Search\uff09\u6280\u672f\u6765\u641c\u7d22\u7f51\u7edc\u7684\u56fe\u50cf\u8f93\u5165\u5206\u8fa8\u7387 r r <\/span><\/span>r<\/span><\/span><\/span><\/span><\/span>\uff0c\u7f51\u7edc\u7684\u6df1\u5ea6 d e p t h depth <\/span><\/span>d<\/span>e<\/span>p<\/span>t<\/span>h<\/span><\/span><\/span><\/span><\/span>\u4ee5\u53cachannel<\/code>\u7684\u5bbd\u5ea6 w i d t h width <\/span><\/span>w<\/span>i<\/span>d<\/span>t<\/span>h<\/span><\/span><\/span><\/span><\/span>\u4e09\u4e2a\u53c2\u6570\u7684\u5408\u7406\u5316\u914d\u7f6e\u3002\u5728\u4e4b\u524d\u7684\u4e00\u4e9b\u8bba\u6587\u4e2d\uff0c\u57fa\u672c\u90fd\u662f\u901a\u8fc7\u6539\u53d8\u4e0a\u8ff03\u4e2a\u53c2\u6570\u4e2d\u7684\u4e00\u4e2a\u6765\u63d0\u5347\u7f51\u7edc\u7684\u6027\u80fd\uff0c\u800c\u8fd9\u7bc7\u8bba\u6587\u5c31\u662f\u540c\u65f6\u6765\u63a2\u7d22\u8fd9\u4e09\u4e2a\u53c2\u6570\u7684\u5f71\u54cd\u3002\u5728\u8bba\u6587\u4e2d\u63d0\u5230\uff0c\u672c\u6587\u63d0\u51fa\u7684EfficientNet-B7<\/strong>\u5728Imagenet top-1<\/strong>\u4e0a\u8fbe\u5230\u4e86\u5f53\u5e74\u6700\u9ad8\u51c6\u786e\u738784.3%<\/strong>\uff0c\u4e0e\u4e4b\u524d\u51c6\u786e\u7387\u6700\u9ad8\u7684GPipe<\/strong>\u76f8\u6bd4\uff0c\u53c2\u6570\u6570\u91cf\uff08Params\uff09\u4ec5\u4e3a\u51761\/8.4<\/strong>\uff0c\u63a8\u7406\u901f\u5ea6\u63d0\u5347\u4e866.1<\/strong>\u500d\uff08\u770b\u4e0a\u53bb\u53c8\u5feb\u53c8\u8f7b\u91cf\uff0c\u4f46\u4e2a\u4eba\u5b9e\u9645\u4f7f\u7528\u8d77\u6765\u53d1\u73b0\u5f88\u5403\u663e\u5b58\uff09\u3002\u4e0b\u56fe\u662fEfficientNet\u4e0e\u5176\u4ed6\u7f51\u7edc\u7684\u5bf9\u6bd4\uff08\u6ce8\u610f\uff0c\u53c2\u6570\u6570\u91cf\u5c11\u5e76\u4e0d\u610f\u5473\u63a8\u7406\u901f\u5ea6\u5c31\u5feb<\/strong>\uff09\u3002<\/p>\n

        \"EfficientNet\u7f51\u7edc\u8be6\u89e3<\/p>\n

        \n

        1 \u8bba\u6587\u601d\u60f3<\/h3>\n

        \u5728\u4e4b\u524d\u7684\u4e00\u4e9b\u8bba\u6587\u4e2d\uff0c\u6709\u7684\u4f1a\u901a\u8fc7\u589e\u52a0\u7f51\u7edc\u7684width<\/code>\u5373\u589e\u52a0\u5377\u79ef\u6838\u7684\u4e2a\u6570\uff08\u589e\u52a0\u7279\u5f81\u77e9\u9635\u7684channels<\/code>\uff09\u6765\u63d0\u5347\u7f51\u7edc\u7684\u6027\u80fd\u5982\u56fe(b)\u6240\u793a\uff0c\u6709\u7684\u4f1a\u901a\u8fc7\u589e\u52a0\u7f51\u7edc\u7684\u6df1\u5ea6\u5373\u4f7f\u7528\u66f4\u591a\u7684\u5c42\u7ed3\u6784\u6765\u63d0\u5347\u7f51\u7edc\u7684\u6027\u80fd\u5982\u56fe(c)\u6240\u793a\uff0c\u6709\u7684\u4f1a\u901a\u8fc7\u589e\u52a0\u8f93\u5165\u7f51\u7edc\u7684\u5206\u8fa8\u7387\u6765\u63d0\u5347\u7f51\u7edc\u7684\u6027\u80fd\u5982\u56fe(d)\u6240\u793a\u3002\u800c\u5728\u672c\u7bc7\u8bba\u6587\u4e2d\u4f1a\u540c\u65f6\u589e\u52a0\u7f51\u7edc\u7684width<\/code>\u3001\u7f51\u7edc\u7684\u6df1\u5ea6\u4ee5\u53ca\u8f93\u5165\u7f51\u7edc\u7684\u5206\u8fa8\u7387\u6765\u63d0\u5347\u7f51\u7edc\u7684\u6027\u80fd\u5982\u56fe(e)\u6240\u793a\uff1a<\/p>\n

        \"EfficientNet\u7f51\u7edc\u8be6\u89e3<\/p>\n

          \n
        • \u6839\u636e\u4ee5\u5f80\u7684\u7ecf\u9a8c\uff0c\u589e\u52a0\u7f51\u7edc\u7684\u6df1\u5ea6depth<\/code>\u80fd\u591f\u5f97\u5230\u66f4\u52a0\u4e30\u5bcc\u3001\u590d\u6742\u7684\u7279\u5f81\u5e76\u4e14\u80fd\u591f\u5f88\u597d\u7684\u5e94\u7528\u5230\u5176\u5b83\u4efb\u52a1\u4e2d\u3002\u4f46\u7f51\u7edc\u7684\u6df1\u5ea6\u8fc7\u6df1\u4f1a\u9762\u4e34\u68af\u5ea6\u6d88\u5931\uff0c\u8bad\u7ec3\u56f0\u96be\u7684\u95ee\u9898\u3002<\/li>\n<\/ul>\n
          \n

          The intuition is that deeper ConvNet can capture richer and more complex features, and generalize well on new tasks. However, deeper networks are also more difficult to train due to the vanishing gradient problem<\/p>\n<\/blockquote>\n

            \n
          • \u589e\u52a0\u7f51\u7edc\u7684width<\/code>\u80fd\u591f\u83b7\u5f97\u66f4\u9ad8\u7ec6\u7c92\u5ea6\u7684\u7279\u5f81\u5e76\u4e14\u4e5f\u66f4\u5bb9\u6613\u8bad\u7ec3\uff0c\u4f46\u5bf9\u4e8ewidth<\/code>\u5f88\u5927\u800c\u6df1\u5ea6\u8f83\u6d45\u7684\u7f51\u7edc\u5f80\u5f80\u5f88\u96be\u5b66\u4e60\u5230\u66f4\u6df1\u5c42\u6b21\u7684\u7279\u5f81\u3002<\/li>\n<\/ul>\n
            \n

            wider networks tend to be able to capture more fine-grained features and are easier to train. However, extremely wide but shallow networks tend to have difficulties in capturing higher level features.<\/p>\n<\/blockquote>\n

              \n
            • \u589e\u52a0\u8f93\u5165\u7f51\u7edc\u7684\u56fe\u50cf\u5206\u8fa8\u7387\u80fd\u591f\u6f5c\u5728\u5f97\u83b7\u5f97\u66f4\u9ad8\u7ec6\u7c92\u5ea6\u7684\u7279\u5f81\u6a21\u677f\uff0c\u4f46\u5bf9\u4e8e\u975e\u5e38\u9ad8\u7684\u8f93\u5165\u5206\u8fa8\u7387\uff0c\u51c6\u786e\u7387\u7684\u589e\u76ca\u4e5f\u4f1a\u51cf\u5c0f\u3002\u5e76\u4e14\u5927\u5206\u8fa8\u7387\u56fe\u50cf\u4f1a\u589e\u52a0\u8ba1\u7b97\u91cf\u3002<\/li>\n<\/ul>\n
              \n

              With higher resolution input images, ConvNets can potentially capture more fine-grained patterns. but the accuracy gain diminishes for very high resolutions.<\/p>\n<\/blockquote>\n

              \u4e0b\u56fe\u5c55\u793a\u4e86\u5728\u57fa\u51c6EfficientNetB-0<\/strong>\u4e0a\u5206\u522b\u589e\u52a0width<\/code>\u3001depth<\/code>\u4ee5\u53caresolution<\/code>\u540e\u5f97\u5230\u7684\u7edf\u8ba1\u7ed3\u679c\u3002\u901a\u8fc7\u4e0b\u56fe\u53ef\u4ee5\u770b\u51fa\u5927\u6982\u5728Accuracy\u8fbe\u523080%\u65f6\u5c31\u8d8b\u4e8e\u9971\u548c\u4e86\u3002
              \"EfficientNet\u7f51\u7edc\u8be6\u89e3
              \u63a5\u7740\u4f5c\u8005\u53c8\u505a\u4e86\u4e00\u4e2a\u5b9e\u9a8c\uff0c\u91c7\u7528\u4e0d\u540c\u7684 d , r d, r <\/span><\/span>d<\/span>,<\/span><\/span>r<\/span><\/span><\/span><\/span><\/span>\u7ec4\u5408\uff0c\u7136\u540e\u4e0d\u65ad\u6539\u53d8\u7f51\u7edc\u7684width<\/code>\u5c31\u5f97\u5230\u4e86\u5982\u4e0b\u56fe\u6240\u793a\u76844\u6761\u66f2\u7ebf\uff0c\u901a\u8fc7\u5206\u6790\u53ef\u4ee5\u53d1\u73b0\u5728\u76f8\u540c\u7684FLOPs\u4e0b\uff0c\u540c\u65f6\u589e\u52a0 d d <\/span><\/span>d<\/span><\/span><\/span><\/span><\/span>\u548c r r <\/span><\/span>r<\/span><\/span><\/span><\/span><\/span>\u7684\u6548\u679c\u6700\u597d\u3002
              \"EfficientNet\u7f51\u7edc\u8be6\u89e3
              \u4e3a\u4e86\u65b9\u4fbf\u540e\u7eed\u7406\u89e3\uff0c\u6211\u4eec\u5148\u770b\u4e0b\u8bba\u6587\u4e2d\u901a\u8fc7 NAS\uff08Neural Architecture Search\uff09<\/strong> \u6280\u672f\u641c\u7d22\u5f97\u5230\u7684EfficientNetB0\u7684\u7ed3\u6784\uff0c\u5982\u4e0b\u56fe\u6240\u793a\uff0c\u6574\u4e2a\u7f51\u7edc\u6846\u67b6\u7531\u4e00\u7cfb\u5217Stage<\/code>\u7ec4\u6210\uff0c F ^ i \\widehat{F}_i <\/span><\/span><\/span>F<\/span><\/span><\/span><\/span>
              \n \n <\/path> \n <\/svg><\/span><\/span><\/span><\/span><\/span><\/span><\/span>i<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u8868\u793a\u5bf9\u5e94Stage<\/code>\u7684\u8fd0\u7b97\u64cd\u4f5c\uff0c L ^ i \\widehat{L}_i <\/span><\/span><\/span>L<\/span><\/span><\/span><\/span>
              \n \n <\/path> \n <\/svg><\/span><\/span><\/span><\/span><\/span><\/span><\/span>i<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u8868\u793a\u5728\u8be5Stage<\/code>\u4e2d\u91cd\u590d F ^ i \\widehat{F}_i <\/span><\/span><\/span>F<\/span><\/span><\/span><\/span>
              \n \n <\/path> \n <\/svg><\/span><\/span><\/span><\/span><\/span><\/span><\/span>i<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u7684\u6b21\u6570\uff1a
              \"EfficientNet\u7f51\u7edc\u8be6\u89e3
              \u4f5c\u8005\u5728\u8bba\u6587\u4e2d\u5bf9\u6574\u4e2a\u7f51\u7edc\u7684\u8fd0\u7b97\u8fdb\u884c\u62bd\u8c61\uff1a
              N ( d , w , r ) = \u2299 i = 1... s F i L i ( X \u27e8 H i , W i , C i \u27e9 ) N(d,w,r)=\\underset{i=1...s}{\\odot} {F}_i^{L_i}(X_{\\left\\langle{
              \n {H}_i, {W}_i, {C}_i } \\right\\rangle}) <\/span><\/span>N<\/span>(<\/span>d<\/span>,<\/span><\/span>w<\/span>,<\/span><\/span>r<\/span>)<\/span><\/span>=<\/span><\/span><\/span><\/span><\/span>i<\/span>=<\/span>1<\/span>.<\/span>.<\/span>.<\/span>s<\/span><\/span><\/span><\/span><\/span>\u2299<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>F<\/span><\/span><\/span>i<\/span><\/span><\/span><\/span>L<\/span><\/span>i<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>(<\/span>X<\/span><\/span>\u27e8<\/span><\/span>H<\/span><\/span><\/span>i<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>,<\/span>W<\/span><\/span><\/span>i<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>,<\/span>C<\/span><\/span><\/span>i<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u27e9<\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>)<\/span><\/span><\/span><\/span><\/span><\/span>
              \u5176\u4e2d\uff1a<\/p>\n

                \n
              • \u2299 i = 1... s \\underset{i=1...s}{\\odot} <\/span><\/span><\/span>i<\/span>=<\/span>1<\/span>.<\/span>.<\/span>.<\/span>s<\/span><\/span><\/span><\/span><\/span>\u2299<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u8868\u793a\u8fde\u4e58\u8fd0\u7b97\u3002<\/li>\n
              • F i {F}_i <\/span><\/span>F<\/span><\/span><\/span>i<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u8868\u793a\u4e00\u4e2a\u8fd0\u7b97\u64cd\u4f5c\uff08\u5982\u4e0a\u56fe\u4e2d\u7684Operator<\/code>\uff09\uff0c\u90a3\u4e48 F i L i {F}_i^{L_i} <\/span><\/span>F<\/span><\/span><\/span>i<\/span><\/span><\/span><\/span>L<\/span><\/span>i<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u8868\u793a\u5728 S t a g e i {\\rm Stage}i <\/span><\/span>S<\/span>t<\/span>a<\/span>g<\/span>e<\/span><\/span><\/span>i<\/span><\/span><\/span><\/span><\/span>\u4e2d F i {F}_i <\/span><\/span>F<\/span><\/span><\/span>i<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u8fd0\u7b97\u88ab\u91cd\u590d\u6267\u884c L i L_i <\/span><\/span>L<\/span><\/span>i<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u6b21\u3002<\/li>\n
              • X X <\/span><\/span>X<\/span><\/span><\/span><\/span><\/span>\u8868\u793a\u8f93\u5165 S t a g e i {\\rm Stage}i <\/span><\/span>S<\/span>t<\/span>a<\/span>g<\/span>e<\/span><\/span><\/span>i<\/span><\/span><\/span><\/span><\/span>\u7684\u7279\u5f81\u77e9\u9635(input tensor<\/code>)\u3002<\/li>\n
              • \u27e8 H i , W i , C i \u27e9 {\\left\\langle{
                \n {H}_i, {W}_i, {C}_i } \\right\\rangle} <\/span><\/span>\u27e8<\/span>H<\/span><\/span><\/span>i<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>,<\/span><\/span>W<\/span><\/span><\/span>i<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>,<\/span><\/span>C<\/span><\/span><\/span>i<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u27e9<\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u8868\u793a X X <\/span><\/span>X<\/span><\/span><\/span><\/span><\/span>\u7684\u9ad8\u5ea6\uff0c\u5bbd\u5ea6\uff0c\u4ee5\u53caChannels\uff08shape<\/code>\uff09\u3002<\/li>\n<\/ul>\n

                \u4e3a\u4e86\u63a2\u7a76 d , r , w d, r, w <\/span><\/span>d<\/span>,<\/span><\/span>r<\/span>,<\/span><\/span>w<\/span><\/span><\/span><\/span><\/span>\u8fd9\u4e09\u4e2a\u56e0\u5b50\u5bf9\u6700\u7ec8\u51c6\u786e\u7387\u7684\u5f71\u54cd\uff0c\u5219\u5c06 d , r , w d, r, w <\/span><\/span>d<\/span>,<\/span><\/span>r<\/span>,<\/span><\/span>w<\/span><\/span><\/span><\/span><\/span>\u52a0\u5165\u5230\u516c\u5f0f\u4e2d\uff0c\u6211\u4eec\u53ef\u4ee5\u5f97\u5230\u62bd\u8c61\u5316\u540e\u7684\u4f18\u5316\u95ee\u9898\uff08\u5728\u6307\u5b9a\u8d44\u6e90\u9650\u5236\u4e0b\uff09\uff0c\u5176\u4e2d s . t . s.t. <\/span><\/span>s<\/span>.<\/span>t<\/span>.<\/span><\/span><\/span><\/span><\/span>\u4ee3\u8868\u9650\u5236\u6761\u4ef6\uff1a<\/p>\n

                \n

                Our target is to maximize the model accuracy for any given resource constraints, which can be formulated as an optimization problem:<\/p>\n<\/blockquote>\n

                m a x d , w , r       A c c u r a c y ( N ( d , w , r ) ) s . t .      N ( d , w , r ) = \u2299 i = 1... s F ^ i d \u22c5 L ^ i ( X \u27e8 r \u22c5 H ^ i ,   r \u22c5 W ^ i ,   w \u22c5 C ^ i \u27e9 ) M e m o r y ( N ) \u2264 t a r g e t _ m e m o r y            F L O P s ( N ) \u2264 t a r g e t _ f l o p s          ( 2 ) \\underset{d, w, r}{max} \\ \\ \\ \\ \\ Accuracy(N(d, w, r)) \\\\ s.t. \\ \\ \\ \\ N(d,w,r)=\\underset{i=1...s}{\\odot} \\widehat{F}_i^{d \\cdot \\widehat{L}_i}(X_{\\left\\langle{r \\cdot \\widehat{H}_i, \\ r \\cdot \\widehat{W}_i, \\ w \\cdot \\widehat{C}_i } \\right\\rangle}) \\\\ Memory(N) \\leq {\\rm target\\_memory} \\\\ \\ \\ \\ \\ \\ \\ \\ \\ \\ \\ FLOPs(N) \\leq {\\rm target\\_flops} \\ \\ \\ \\ \\ \\ \\ \\ (2) <\/span><\/span><\/span>d<\/span>,<\/span>w<\/span>,<\/span>r<\/span><\/span><\/span><\/span><\/span>ma<\/span>x<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span> <\/span> <\/span> <\/span> <\/span> <\/span>A<\/span>c<\/span>c<\/span>u<\/span>r<\/span>a<\/span>c<\/span>y<\/span>(<\/span>N<\/span>(<\/span>d<\/span>,<\/span><\/span>w<\/span>,<\/span><\/span>r<\/span>)<\/span>)<\/span><\/span><\/span><\/span>s<\/span>.<\/span>t<\/span>.<\/span> <\/span> <\/span> <\/span> <\/span>N<\/span>(<\/span>d<\/span>,<\/span><\/span>w<\/span>,<\/span><\/span>r<\/span>)<\/span><\/span>=<\/span><\/span><\/span><\/span><\/span>i<\/span>=<\/span>1<\/span>.<\/span>.<\/span>.<\/span>s<\/span><\/span><\/span><\/span><\/span>\u2299<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>F<\/span><\/span><\/span><\/span>
                \n \n <\/path> \n <\/svg><\/span><\/span><\/span><\/span><\/span><\/span><\/span>i<\/span><\/span><\/span><\/span>d<\/span>\u22c5<\/span><\/span>L<\/span><\/span><\/span><\/span>
                \n \n <\/path> \n <\/svg><\/span><\/span><\/span><\/span><\/span><\/span><\/span>i<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>(<\/span>X<\/span><\/span>\u27e8<\/span><\/span>r<\/span>\u22c5<\/span><\/span>H<\/span><\/span><\/span><\/span>
                \n \n <\/path> \n <\/svg><\/span><\/span><\/span><\/span><\/span><\/span><\/span>i<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>,<\/span> <\/span><\/span>r<\/span>\u22c5<\/span><\/span>W<\/span><\/span><\/span><\/span>
                \n \n <\/path> \n <\/svg><\/span><\/span><\/span><\/span><\/span><\/span><\/span>i<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>,<\/span> <\/span><\/span>w<\/span>\u22c5<\/span><\/span>C<\/span><\/span><\/span><\/span>
                \n \n <\/path> \n <\/svg><\/span><\/span><\/span><\/span><\/span><\/span><\/span>i<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u27e9<\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>)<\/span><\/span><\/span><\/span>M<\/span>e<\/span>m<\/span>o<\/span>r<\/span>y<\/span>(<\/span>N<\/span>)<\/span><\/span>\u2264<\/span><\/span><\/span><\/span>t<\/span>a<\/span>r<\/span>g<\/span>e<\/span>t<\/span>_<\/span>m<\/span>e<\/span>m<\/span>o<\/span>r<\/span>y<\/span><\/span><\/span><\/span><\/span><\/span> <\/span> <\/span> <\/span> <\/span> <\/span> <\/span> <\/span> <\/span> <\/span> <\/span>F<\/span>L<\/span>O<\/span>P<\/span>s<\/span>(<\/span>N<\/span>)<\/span><\/span>\u2264<\/span><\/span><\/span><\/span>t<\/span>a<\/span>r<\/span>g<\/span>e<\/span>t<\/span>_<\/span>f<\/span>l<\/span>o<\/span>p<\/span>s<\/span><\/span><\/span> <\/span> <\/span> <\/span> <\/span> <\/span> <\/span> <\/span> <\/span>(<\/span>2<\/span>)<\/span><\/span><\/span><\/span><\/span><\/span>
                \u5176\u4e2d\uff1a<\/p>\n

                  \n
                • d d <\/span><\/span>d<\/span><\/span><\/span><\/span><\/span>\u7528\u6765\u7f29\u653e\u6df1\u5ea6 L ^ i \\widehat{L}_i <\/span><\/span><\/span>L<\/span><\/span><\/span><\/span>
                  \n \n <\/path> \n <\/svg><\/span><\/span><\/span><\/span><\/span><\/span><\/span>i<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/li>\n
                • r r <\/span><\/span>r<\/span><\/span><\/span><\/span><\/span>\u7528\u6765\u7f29\u653e\u5206\u8fa8\u7387\u5373\u5f71\u54cd H ^ i \\widehat{H}_i <\/span><\/span><\/span>H<\/span><\/span><\/span><\/span>
                  \n \n <\/path> \n <\/svg><\/span><\/span><\/span><\/span><\/span><\/span><\/span>i<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u548c W ^ i \\widehat{W}_i <\/span><\/span><\/span>W<\/span><\/span><\/span><\/span>
                  \n \n <\/path> \n <\/svg><\/span><\/span><\/span><\/span><\/span><\/span><\/span>i<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/li>\n
                • w w <\/span><\/span>w<\/span><\/span><\/span><\/span><\/span>\u5c31\u662f\u7528\u6765\u7f29\u653e\u7279\u5f81\u77e9\u9635\u7684channel<\/code>\u5373 C ^ i \\widehat{C}_i <\/span><\/span><\/span>C<\/span><\/span><\/span><\/span>
                  \n \n <\/path> \n <\/svg><\/span><\/span><\/span><\/span><\/span><\/span><\/span>i<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/li>\n
                • target_memory<\/code>\u4e3amemory<\/code>\u9650\u5236<\/li>\n
                • target_flops<\/code>\u4e3aFLOPs\u9650\u5236<\/li>\n<\/ul>\n

                  \u63a5\u7740\u4f5c\u8005\u53c8\u63d0\u51fa\u4e86\u4e00\u4e2a\u6df7\u5408\u7f29\u653e\u65b9\u6cd5 ( compound scaling method)<\/strong> \u5728\u8fd9\u4e2a\u65b9\u6cd5\u4e2d\u4f7f\u7528\u4e86\u4e00\u4e2a\u6df7\u5408\u56e0\u5b50 \u03d5 \\phi <\/span><\/span>\u03d5<\/span><\/span><\/span><\/span><\/span>\u53bb\u7edf\u4e00\u7684\u7f29\u653ewidth\uff0cdepth\uff0cresolution<\/code>\u53c2\u6570\uff0c\u5177\u4f53\u7684\u8ba1\u7b97\u516c\u5f0f\u5982\u4e0b\uff0c\u5176\u4e2d s . t . s.t. <\/span><\/span>s<\/span>.<\/span>t<\/span>.<\/span><\/span><\/span><\/span><\/span>\u4ee3\u8868\u9650\u5236\u6761\u4ef6\uff1a
                  d e p t h : d = \u03b1 \u03d5 w i d t h : w = \u03b2 \u03d5        r e s o l u t i o n : r = \u03b3 \u03d5            ( 3 ) s . t .         \u03b1 \u22c5 \u03b2 2 \u22c5 \u03b3 2 \u2248 2 \u03b1 \u2265 1 , \u03b2 \u2265 1 , \u03b3 \u2265 1        depth: d={\\alpha}^{\\phi} \\\\ width: w={\\beta}^{\\phi} \\\\ \\ \\ \\ \\ \\ \\ resolution: r={\\gamma}^{\\phi} \\ \\ \\ \\ \\ \\ \\ \\ \\ \\ (3) \\\\ s.t. \\ \\ \\ \\ \\ \\ \\ {\\alpha} \\cdot {\\beta}^{2} \\cdot {\\gamma}^{2} \\approx 2 \\\\ \\alpha \\geq 1, \\beta \\geq 1, \\gamma \\geq 1 \\ \\ \\ \\ \\ \\ <\/span><\/span>d<\/span>e<\/span>p<\/span>t<\/span>h<\/span><\/span>:<\/span><\/span><\/span><\/span>d<\/span><\/span>=<\/span><\/span><\/span><\/span>\u03b1<\/span><\/span><\/span>\u03d5<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>w<\/span>i<\/span>d<\/span>t<\/span>h<\/span><\/span>:<\/span><\/span><\/span><\/span>w<\/span><\/span>=<\/span><\/span><\/span><\/span>\u03b2<\/span><\/span><\/span>\u03d5<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span> <\/span> <\/span> <\/span> <\/span> <\/span> <\/span>r<\/span>e<\/span>s<\/span>o<\/span>l<\/span>u<\/span>t<\/span>i<\/span>o<\/span>n<\/span><\/span>:<\/span><\/span><\/span><\/span>r<\/span><\/span>=<\/span><\/span><\/span><\/span>\u03b3<\/span><\/span><\/span>\u03d5<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span> <\/span> <\/span> <\/span> <\/span> <\/span> <\/span> <\/span> <\/span> <\/span> <\/span>(<\/span>3<\/span>)<\/span><\/span><\/span><\/span>s<\/span>.<\/span>t<\/span>.<\/span> <\/span> <\/span> <\/span> <\/span> <\/span> <\/span> <\/span>\u03b1<\/span><\/span><\/span>\u22c5<\/span><\/span><\/span><\/span>\u03b2<\/span><\/span><\/span>2<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u22c5<\/span><\/span><\/span><\/span>\u03b3<\/span><\/span><\/span>2<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u2248<\/span><\/span><\/span><\/span>2<\/span><\/span><\/span><\/span>\u03b1<\/span><\/span>\u2265<\/span><\/span><\/span><\/span>1<\/span>,<\/span><\/span>\u03b2<\/span><\/span>\u2265<\/span><\/span><\/span><\/span>1<\/span>,<\/span><\/span>\u03b3<\/span><\/span>\u2265<\/span><\/span><\/span><\/span>1<\/span> <\/span> <\/span> <\/span> <\/span> <\/span> <\/span><\/span><\/span><\/span><\/span><\/span><\/p>\n

                  \u6ce8\u610f\uff1a<\/p>\n

                    \n
                  • FLOPs\uff08\u7406\u8bba\u8ba1\u7b97\u91cf\uff09\u4e0edepth<\/code>\u7684\u5173\u7cfb\u662f\uff1a\u5f53depth<\/code>\u7ffb\u500d\uff0cFLOPs\u4e5f\u7ffb\u500d\u3002<\/li>\n
                  • FLOPs\u4e0ewidth<\/code>\u7684\u5173\u7cfb\u662f\uff1a\u5f53width<\/code>\u7ffb\u500d\uff08\u5373channal<\/code>\u7ffb\u500d\uff09\uff0cFLOPs\u4f1a\u7ffb4\u500d\uff0c\u56e0\u4e3a\u5377\u79ef\u5c42\u7684FLOPs\u7ea6\u7b49\u4e8e f e a t u r e w \u00d7 f e a t u r e h \u00d7 f e a t u r e c \u00d7 k e r n e l w \u00d7 k e r n e l h \u00d7 k e r n e l n u m b e r feature_w \\times feature_h \\times feature_c \\times kernel_w \\times kernel_h \\times kernel_{number} <\/span><\/span>f<\/span>e<\/span>a<\/span>t<\/span>u<\/span>r<\/span>e<\/span><\/span>w<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u00d7<\/span><\/span><\/span><\/span>f<\/span>e<\/span>a<\/span>t<\/span>u<\/span>r<\/span>e<\/span><\/span>h<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u00d7<\/span><\/span><\/span><\/span>f<\/span>e<\/span>a<\/span>t<\/span>u<\/span>r<\/span>e<\/span><\/span>c<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u00d7<\/span><\/span><\/span><\/span>k<\/span>e<\/span>r<\/span>n<\/span>e<\/span>l<\/span><\/span>w<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u00d7<\/span><\/span><\/span><\/span>k<\/span>e<\/span>r<\/span>n<\/span>e<\/span>l<\/span><\/span>h<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u00d7<\/span><\/span><\/span><\/span>k<\/span>e<\/span>r<\/span>n<\/span>e<\/span>l<\/span><\/span>n<\/span>u<\/span>m<\/span>b<\/span>e<\/span>r<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\uff08\u5047\u8bbe\u8f93\u5165\u8f93\u51fa\u7279\u5f81\u77e9\u9635\u7684\u9ad8\u5bbd\u4e0d\u53d8\uff09\uff0c\u5f53width<\/code>\u7ffb\u500d\uff0c\u8f93\u5165\u7279\u5f81\u77e9\u9635\u7684channels\uff08 f e a t u r e c feature_c <\/span><\/span>f<\/span>e<\/span>a<\/span>t<\/span>u<\/span>r<\/span>e<\/span><\/span>c<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\uff09\u548c\u8f93\u51fa\u7279\u5f81\u77e9\u9635\u7684channels\u6216\u5377\u79ef\u6838\u7684\u4e2a\u6570\uff08 k e r n e l n u m b e r kernel_{number} <\/span><\/span>k<\/span>e<\/span>r<\/span>n<\/span>e<\/span>l<\/span><\/span>n<\/span>u<\/span>m<\/span>b<\/span>e<\/span>r<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\uff09\u90fd\u4f1a\u7ffb\u500d\uff0c\u6240\u4ee5FLOPs\u4f1a\u7ffb4\u500d<\/li>\n
                  • FLOPs\u4e0eresolution<\/code>\u7684\u5173\u7cfb\u662f\uff1a\u5f53resolution<\/code>\u7ffb\u500d\uff0cFLOPs\u4e5f\u4f1a\u7ffb4\u500d\uff0c\u548c\u4e0a\u9762\u7c7b\u4f3c\u56e0\u4e3a\u7279\u5f81\u77e9\u9635\u7684\u5bbd\u5ea6 f e a t u r e w feature_w <\/span><\/span>f<\/span>e<\/span>a<\/span>t<\/span>u<\/span>r<\/span>e<\/span><\/span>w<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u548c\u7279\u5f81\u77e9\u9635\u7684\u9ad8\u5ea6 f e a t u r e h feature_h <\/span><\/span>f<\/span>e<\/span>a<\/span>t<\/span>u<\/span>r<\/span>e<\/span><\/span>h<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u90fd\u4f1a\u7ffb\u500d\u3002<\/li>\n<\/ul>\n

                    \u6240\u4ee5\u603b\u7684FLOPs\u500d\u7387\u53ef\u4ee5\u7528\u8fd1\u4f3c\u7528 ( \u03b1 \u22c5 \u03b2 2 \u22c5 \u03b3 2 ) \u03d5 (\\alpha \\cdot \\beta^{2} \\cdot \\gamma^{2})^{\\phi} <\/span><\/span>(<\/span>\u03b1<\/span><\/span>\u22c5<\/span><\/span><\/span><\/span>\u03b2<\/span><\/span>2<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u22c5<\/span><\/span><\/span><\/span>\u03b3<\/span><\/span>2<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>)<\/span><\/span>\u03d5<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u6765\u8868\u793a\uff0c\u5f53\u9650\u5236 \u03b1 \u22c5 \u03b2 2 \u22c5 \u03b3 2 \u2248 2 \\alpha \\cdot \\beta^{2} \\cdot \\gamma^{2} \\approx 2 <\/span><\/span>\u03b1<\/span><\/span>\u22c5<\/span><\/span><\/span><\/span>\u03b2<\/span><\/span>2<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u22c5<\/span><\/span><\/span><\/span>\u03b3<\/span><\/span>2<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u2248<\/span><\/span><\/span><\/span>2<\/span><\/span><\/span><\/span><\/span>\u65f6\uff0c\u5bf9\u4e8e\u4efb\u610f\u4e00\u4e2a \u03d5 \\phi <\/span><\/span>\u03d5<\/span><\/span><\/span><\/span><\/span>\u800c\u8a00FLOPs\u76f8\u5f53\u589e\u52a0\u4e86 2 \u03d5 2^{\\phi} <\/span><\/span>2<\/span><\/span>\u03d5<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u500d\u3002<\/p>\n

                    \u63a5\u4e0b\u6765\u4f5c\u8005\u5728\u57fa\u51c6\u7f51\u7edcEfficientNetB-0\uff08\u5728\u540e\u9762\u7684\u7f51\u7edc\u8be6\u7ec6\u7ed3\u6784<\/strong>\u7ae0\u8282\u4f1a\u8be6\u7ec6\u8bb2\uff09\u4e0a\u4f7f\u7528NAS<\/strong>\u6765\u641c\u7d22 \u03b1 , \u03b2 , \u03b3 \\alpha, \\beta, \\gamma <\/span><\/span>\u03b1<\/span>,<\/span><\/span>\u03b2<\/span>,<\/span><\/span>\u03b3<\/span><\/span><\/span><\/span><\/span>\u8fd9\u4e09\u4e2a\u53c2\u6570\u3002<\/p>\n

                      \n
                    • \uff08step1\uff09\u9996\u5148\u56fa\u5b9a \u03d5 = 1 \\phi=1 <\/span><\/span>\u03d5<\/span><\/span>=<\/span><\/span><\/span><\/span>1<\/span><\/span><\/span><\/span><\/span>\uff0c\u5e76\u57fa\u4e8e\u4e0a\u9762\u7ed9\u51fa\u7684\u516c\u5f0f(2)\u548c(3)\u8fdb\u884c\u641c\u7d22\uff0c\u4f5c\u8005\u53d1\u73b0\u5bf9\u4e8eEfficientNetB-0\u6700\u4f73\u53c2\u6570\u4e3a \u03b1 = 1.2 , \u03b2 = 1.1 , \u03b3 = 1.15 \\alpha=1.2, \\beta=1.1, \\gamma=1.15 <\/span><\/span>\u03b1<\/span><\/span>=<\/span><\/span><\/span><\/span>1<\/span>.<\/span>2<\/span>,<\/span><\/span>\u03b2<\/span><\/span>=<\/span><\/span><\/span><\/span>1<\/span>.<\/span>1<\/span>,<\/span><\/span>\u03b3<\/span><\/span>=<\/span><\/span><\/span><\/span>1<\/span>.<\/span>1<\/span>5<\/span><\/span><\/span><\/span><\/span><\/li>\n
                    • \uff08step2\uff09\u63a5\u7740\u56fa\u5b9a \u03b1 = 1.2 , \u03b2 = 1.1 , \u03b3 = 1.15 \\alpha=1.2, \\beta=1.1, \\gamma=1.15 <\/span><\/span>\u03b1<\/span><\/span>=<\/span><\/span><\/span><\/span>1<\/span>.<\/span>2<\/span>,<\/span><\/span>\u03b2<\/span><\/span>=<\/span><\/span><\/span><\/span>1<\/span>.<\/span>1<\/span>,<\/span><\/span>\u03b3<\/span><\/span>=<\/span><\/span><\/span><\/span>1<\/span>.<\/span>1<\/span>5<\/span><\/span><\/span><\/span><\/span>\uff0c\u5728EfficientNetB-0\u7684\u57fa\u7840\u4e0a\u4f7f\u7528\u4e0d\u540c\u7684 \u03d5 \\phi <\/span><\/span>\u03d5<\/span><\/span><\/span><\/span><\/span>\u5206\u522b\u5f97\u5230EfficientNetB-1\u81f3EfficientNetB-7\uff08\u5728\u540e\u9762\u7684EfficientNet(B0-B7)\u53c2\u6570<\/strong>\u7ae0\u8282\u6709\u7ed9\u51fa\u8be6\u7ec6\u53c2\u6570\uff09<\/li>\n<\/ul>\n

                      \u9700\u8981\u6ce8\u610f\u7684\u662f\uff0c\u5bf9\u4e8e\u4e0d\u540c\u7684\u57fa\u51c6\u7f51\u7edc\u641c\u7d22\u51fa\u7684 \u03b1 , \u03b2 , \u03b3 \\alpha, \\beta, \\gamma <\/span><\/span>\u03b1<\/span>,<\/span><\/span>\u03b2<\/span>,<\/span><\/span>\u03b3<\/span><\/span><\/span><\/span><\/span>\u4e5f\u4e0d\u5b9a\u76f8\u540c\u3002\u8fd8\u9700\u8981\u6ce8\u610f\u7684\u662f\uff0c\u5728\u539f\u8bba\u6587\u4e2d\uff0c\u4f5c\u8005\u4e5f\u8bf4\u4e86\uff0c\u5982\u679c\u76f4\u63a5\u5728\u5927\u6a21\u578b\u4e0a\u53bb\u641c\u7d22 \u03b1 , \u03b2 , \u03b3 \\alpha, \\beta, \\gamma <\/span><\/span>\u03b1<\/span>,<\/span><\/span>\u03b2<\/span>,<\/span><\/span>\u03b3<\/span><\/span><\/span><\/span><\/span>\u53ef\u80fd\u83b7\u5f97\u66f4\u597d\u7684\u7ed3\u679c\uff0c\u4f46\u662f\u5728\u8f83\u5927\u7684\u6a21\u578b\u4e2d\u641c\u7d22\u6210\u672c\u592a\u5927\uff08Google\u5927\u5382\u5c45\u7136\u8bf4\u8fd9\u79cd\u8bdd\uff09\uff0c\u6240\u4ee5\u8fd9\u7bc7\u6587\u7ae0\u5c31\u5728\u6bd4\u8f83\u5c0f\u7684EfficientNetB-0\u6a21\u578b\u4e0a\u8fdb\u884c\u641c\u7d22\u7684\u3002<\/p>\n

                      \n

                      Notably, it is possible to achieve even better performance by searching for \u03b1, \u03b2, \u03b3 directly around a large model, but the search cost becomes prohibitively more expensive on larger models. Our method solves this issue by only doing search once on the small baseline network (step 1), and then use the same scaling coefficients for all other models (step 2).<\/p>\n<\/blockquote>\n

                      \n

                      2 \u7f51\u7edc\u8be6\u7ec6\u7ed3\u6784<\/h3>\n

                      \u4e0b\u8868\u4e3aEfficientNet-B0\u7684\u7f51\u7edc\u6846\u67b6\uff08B1-B7\u5c31\u662f\u5728B0\u7684\u57fa\u7840\u4e0a\u4fee\u6539Resolution<\/code>\uff0cChannels<\/code>\u4ee5\u53caLayers<\/code>\uff09\uff0c\u53ef\u4ee5\u770b\u51fa\u7f51\u7edc\u603b\u5171\u5206\u6210\u4e869\u4e2aStage<\/code>\uff0c\u7b2c\u4e00\u4e2aStage<\/code>\u5c31\u662f\u4e00\u4e2a\u5377\u79ef\u6838\u5927\u5c0f\u4e3a3x3<\/code>\u6b65\u8ddd\u4e3a2\u7684\u666e\u901a\u5377\u79ef\u5c42\uff08\u5305\u542bBN\u548c\u6fc0\u6d3b\u51fd\u6570Swish\uff09\uff0cStage2\uff5eStage8<\/code>\u90fd\u662f\u5728\u91cd\u590d\u5806\u53e0MBConv<\/code>\u7ed3\u6784\uff08\u6700\u540e\u4e00\u5217\u7684Layers<\/code>\u8868\u793a\u8be5Stage<\/code>\u91cd\u590dMBConv<\/code>\u7ed3\u6784\u591a\u5c11\u6b21\uff09\uff0c\u800cStage9<\/code>\u7531\u4e00\u4e2a\u666e\u901a\u76841x1<\/code>\u7684\u5377\u79ef\u5c42\uff08\u5305\u542bBN\u548c\u6fc0\u6d3b\u51fd\u6570Swish\uff09\u4e00\u4e2a\u5e73\u5747\u6c60\u5316\u5c42\u548c\u4e00\u4e2a\u5168\u8fde\u63a5\u5c42\u7ec4\u6210\u3002\u8868\u683c\u4e2d\u6bcf\u4e2aMBConv<\/code>\u540e\u4f1a\u8ddf\u4e00\u4e2a\u6570\u5b571\u62166\uff0c\u8fd9\u91cc\u76841\u62166\u5c31\u662f\u500d\u7387\u56e0\u5b50n<\/code>\u5373MBConv<\/code>\u4e2d\u7b2c\u4e00\u4e2a1x1<\/code>\u7684\u5377\u79ef\u5c42\u4f1a\u5c06\u8f93\u5165\u7279\u5f81\u77e9\u9635\u7684channels<\/code>\u6269\u5145\u4e3an<\/code>\u500d\uff0c\u5176\u4e2dk3x3<\/code>\u6216k5x5<\/code>\u8868\u793aMBConv<\/code>\u4e2dDepthwise Conv<\/code>\u6240\u91c7\u7528\u7684\u5377\u79ef\u6838\u5927\u5c0f\u3002Channels<\/code>\u8868\u793a\u901a\u8fc7\u8be5Stage<\/code>\u540e\u8f93\u51fa\u7279\u5f81\u77e9\u9635\u7684Channels<\/code>\u3002
                      \"EfficientNet\u7f51\u7edc\u8be6\u89e3<\/p>\n

                      \n

                      2.1 MBConv\u7ed3\u6784<\/h4>\n

                      MBConv<\/code>\u5176\u5b9e\u5c31\u662fMobileNetV3\u7f51\u7edc\u4e2d\u7684InvertedResidualBlock<\/code>\uff0c\u4f46\u4e5f\u6709\u4e9b\u8bb8\u533a\u522b\u3002\u4e00\u4e2a\u662f\u91c7\u7528\u7684\u6fc0\u6d3b\u51fd\u6570\u4e0d\u4e00\u6837\uff08EfficientNet\u7684MBConv<\/code>\u4e2d\u4f7f\u7528\u7684\u90fd\u662fSwish<\/strong>\u6fc0\u6d3b\u51fd\u6570\uff09\uff0c\u53e6\u4e00\u4e2a\u662f\u5728\u6bcf\u4e2aMBConv<\/code>\u4e2d\u90fd\u52a0\u5165\u4e86SE\uff08Squeeze-and-Excitation<\/strong>\uff09\u6a21\u5757\u3002\u4e0b\u56fe\u662f\u6211\u81ea\u5df1\u7ed8\u5236\u7684MBConv<\/code>\u7ed3\u6784\u3002<\/p>\n

                      \"EfficientNet\u7f51\u7edc\u8be6\u89e3<\/p>\n

                      \u5982\u56fe\u6240\u793a\uff0cMBConv<\/code>\u7ed3\u6784\u4e3b\u8981\u7531\u4e00\u4e2a1x1<\/code>\u7684\u666e\u901a\u5377\u79ef\uff08\u5347\u7ef4\u4f5c\u7528\uff0c\u5305\u542bBN\u548cSwish\uff09\uff0c\u4e00\u4e2akxk<\/code>\u7684Depthwise Conv<\/code>\u5377\u79ef\uff08\u5305\u542bBN\u548cSwish\uff09k<\/code>\u7684\u5177\u4f53\u503c\u53ef\u770bEfficientNet-B0\u7684\u7f51\u7edc\u6846\u67b6\u4e3b\u8981\u67093x3<\/code>\u548c5x5<\/code>\u4e24\u79cd\u60c5\u51b5\uff0c\u4e00\u4e2aSE<\/code>\u6a21\u5757\uff0c\u4e00\u4e2a1x1<\/code>\u7684\u666e\u901a\u5377\u79ef\uff08\u964d\u7ef4\u4f5c\u7528\uff0c\u5305\u542bBN\uff09\uff0c\u4e00\u4e2aDroupout<\/code>\u5c42\u6784\u6210\u3002\u642d\u5efa\u8fc7\u7a0b\u4e2d\u8fd8\u9700\u8981\u6ce8\u610f\u51e0\u70b9\uff1a<\/p>\n

                        \n
                      • \u7b2c\u4e00\u4e2a\u5347\u7ef4\u76841x1<\/code>\u5377\u79ef\u5c42\uff0c\u5b83\u7684\u5377\u79ef\u6838\u4e2a\u6570\u662f\u8f93\u5165\u7279\u5f81\u77e9\u9635channel<\/code>\u7684 n n <\/span><\/span>n<\/span><\/span><\/span><\/span><\/span>\u500d\uff0c n \u2208 { 1 , 6 } n \\in \\left\\{1, 6\\right\\} <\/span><\/span>n<\/span><\/span>\u2208<\/span><\/span><\/span><\/span>{
                        \n <\/span>1<\/span>,<\/span><\/span>6<\/span>}<\/span><\/span><\/span><\/span><\/span><\/span>\u3002<\/li>\n
                      • \u5f53 n = 1 n=1 <\/span><\/span>n<\/span><\/span>=<\/span><\/span><\/span><\/span>1<\/span><\/span><\/span><\/span><\/span>\u65f6\uff0c\u4e0d\u8981\u7b2c\u4e00\u4e2a\u5347\u7ef4\u76841x1<\/code>\u5377\u79ef\u5c42\uff0c\u5373Stage2<\/code>\u4e2d\u7684MBConv<\/code>\u7ed3\u6784\u90fd\u6ca1\u6709\u7b2c\u4e00\u4e2a\u5347\u7ef4\u76841x1<\/code>\u5377\u79ef\u5c42\uff08\u8fd9\u548cMobileNetV3\u7f51\u7edc\u7c7b\u4f3c\uff09\u3002<\/li>\n
                      • \u5173\u4e8eshortcut<\/code>\u8fde\u63a5\uff0c\u4ec5\u5f53\u8f93\u5165MBConv<\/code>\u7ed3\u6784\u7684\u7279\u5f81\u77e9\u9635\u4e0e\u8f93\u51fa\u7684\u7279\u5f81\u77e9\u9635shape<\/code>\u76f8\u540c\u65f6\u624d\u5b58\u5728\uff08\u4ee3\u7801\u4e2d\u53ef\u901a\u8fc7stride==1 and inputc_channels==output_channels<\/code>\u6761\u4ef6\u6765\u5224\u65ad\uff09\u3002<\/li>\n
                      • SE\u6a21\u5757\u5982\u4e0b\u6240\u793a\uff0c\u7531\u4e00\u4e2a\u5168\u5c40\u5e73\u5747\u6c60\u5316\uff0c\u4e24\u4e2a\u5168\u8fde\u63a5\u5c42\u7ec4\u6210\u3002\u7b2c\u4e00\u4e2a\u5168\u8fde\u63a5\u5c42\u7684\u8282\u70b9\u4e2a\u6570\u662f\u8f93\u5165\u8be5MBConv<\/code>\u7279\u5f81\u77e9\u9635channels<\/code>\u7684 1 4 \\frac{1}{4} <\/span><\/span><\/span><\/span>4<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>1<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\uff0c\u4e14\u4f7f\u7528Swish\u6fc0\u6d3b\u51fd\u6570\u3002\u7b2c\u4e8c\u4e2a\u5168\u8fde\u63a5\u5c42\u7684\u8282\u70b9\u4e2a\u6570\u7b49\u4e8eDepthwise Conv<\/code>\u5c42\u8f93\u51fa\u7684\u7279\u5f81\u77e9\u9635channels<\/code>\uff0c\u4e14\u4f7f\u7528Sigmoid\u6fc0\u6d3b\u51fd\u6570\u3002<\/li>\n
                      • Dropout\u5c42\u7684dropout_rate<\/code>\u5728tensorflow\u7684keras\u6e90\u7801\u4e2d\u5bf9\u5e94\u7684\u662fdrop_connect_rate<\/code>\u540e\u9762\u4f1a\u7ec6\u8bb2\uff08\u6ce8\u610f\uff0c\u5728\u6e90\u7801\u5b9e\u73b0\u4e2d\u53ea\u6709\u4f7f\u7528shortcut\u7684\u65f6\u5019\u624d\u6709Dropout\u5c42<\/strong>\uff09\u3002<\/li>\n<\/ul>\n

                        \"EfficientNet\u7f51\u7edc\u8be6\u89e3<\/p>\n

                        \n

                        2.2 EfficientNet(B0-B7)\u53c2\u6570<\/h4>\n

                        \u8fd8\u662f\u5148\u7ed9\u51faEfficientNetB0\u7684\u7f51\u7edc\u7ed3\u6784\uff0c\u65b9\u4fbf\u540e\u9762\u7406\u89e3\u3002
                        \"EfficientNet\u7f51\u7edc\u8be6\u89e3
                        \u901a\u8fc7\u4e0a\u9762\u7684\u5185\u5bb9\uff0c\u6211\u4eec\u662f\u53ef\u4ee5\u642d\u5efa\u51faEfficientNetB0\u7f51\u7edc\u7684\uff0c\u5176\u4ed6\u7248\u672c\u7684\u8be6\u7ec6\u53c2\u6570\u53ef\u89c1\u4e0b\u8868\uff1a<\/p>\n\n\n\n\n\n\n\n\n\n\n\n\n
                        Model<\/th>\ninput_size<\/th>\nwidth_coefficient<\/th>\ndepth_coefficient<\/th>\ndrop_connect_rate<\/th>\ndropout_rate<\/th>\n<\/tr>\n<\/thead>\n
                        EfficientNetB0<\/td>\n224x224<\/td>\n1.0<\/td>\n1.0<\/td>\n0.2<\/td>\n0.2<\/td>\n<\/tr>\n
                        EfficientNetB1<\/td>\n240x240<\/td>\n1.0<\/td>\n1.1<\/td>\n0.2<\/td>\n0.2<\/td>\n<\/tr>\n
                        EfficientNetB2<\/td>\n260x260<\/td>\n1.1<\/td>\n1.2<\/td>\n0.2<\/td>\n0.3<\/td>\n<\/tr>\n
                        EfficientNetB3<\/td>\n300x300<\/td>\n1.2<\/td>\n1.4<\/td>\n0.2<\/td>\n0.3<\/td>\n<\/tr>\n
                        EfficientNetB4<\/td>\n380x380<\/td>\n1.4<\/td>\n1.8<\/td>\n0.2<\/td>\n0.4<\/td>\n<\/tr>\n
                        EfficientNetB5<\/td>\n456x456<\/td>\n1.6<\/td>\n2.2<\/td>\n0.2<\/td>\n0.4<\/td>\n<\/tr>\n
                        EfficientNetB6<\/td>\n528x528<\/td>\n1.8<\/td>\n2.6<\/td>\n0.2<\/td>\n0.5<\/td>\n<\/tr>\n
                        EfficientNetB7<\/td>\n600x600<\/td>\n2.0<\/td>\n3.1<\/td>\n0.2<\/td>\n0.5<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n
                          \n
                        • input_size<\/code>\u4ee3\u8868\u8bad\u7ec3\u7f51\u7edc\u65f6\u8f93\u5165\u7f51\u7edc\u7684\u56fe\u50cf\u5927\u5c0f<\/li>\n
                        • width_coefficient<\/code>\u4ee3\u8868channel<\/code>\u7ef4\u5ea6\u4e0a\u7684\u500d\u7387\u56e0\u5b50\uff0c\u6bd4\u5982\u5728 EfficientNetB0\u4e2dStage1<\/code>\u76843x3<\/code>\u5377\u79ef\u5c42\u6240\u4f7f\u7528\u7684\u5377\u79ef\u6838\u4e2a\u6570\u662f32\uff0c\u90a3\u4e48\u5728B6\u4e2d\u5c31\u662f 32 \u00d7 1.8 = 57.6 32 \\times 1.8=57.6 <\/span><\/span>3<\/span>2<\/span><\/span>\u00d7<\/span><\/span><\/span><\/span>1<\/span>.<\/span>8<\/span><\/span>=<\/span><\/span><\/span><\/span>5<\/span>7<\/span>.<\/span>6<\/span><\/span><\/span><\/span><\/span>\u63a5\u7740\u53d6\u6574\u5230\u79bb\u5b83\u6700\u8fd1\u76848\u7684\u6574\u6570\u500d\u537356\uff0c\u5176\u5b83Stage<\/code>\u540c\u7406\u3002<\/li>\n
                        • depth_coefficient<\/code>\u4ee3\u8868depth<\/code>\u7ef4\u5ea6\u4e0a\u7684\u500d\u7387\u56e0\u5b50\uff08\u4ec5\u9488\u5bf9Stage2<\/code>\u5230Stage8<\/code>\uff09\uff0c\u6bd4\u5982\u5728EfficientNetB0\u4e2dStage7<\/code>\u7684 L ^ i = 4 {\\widehat L}_i=4 <\/span><\/span><\/span>L<\/span><\/span><\/span>
                          \n \n <\/path> \n <\/svg><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>i<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>=<\/span><\/span><\/span><\/span>4<\/span><\/span><\/span><\/span><\/span>\uff0c\u90a3\u4e48\u5728B6\u4e2d\u5c31\u662f 4 \u00d7 2.6 = 10.4 4 \\times 2.6=10.4 <\/span><\/span>4<\/span><\/span>\u00d7<\/span><\/span><\/span><\/span>2<\/span>.<\/span>6<\/span><\/span>=<\/span><\/span><\/span><\/span>1<\/span>0<\/span>.<\/span>4<\/span><\/span><\/span><\/span><\/span>\u63a5\u7740\u5411\u4e0a\u53d6\u6574\u537311.<\/li>\n
                        • drop_connect_rate<\/code>\u662f\u5728MBConv<\/code>\u7ed3\u6784\u4e2ddropout\u5c42\u4f7f\u7528\u7684drop_rate<\/code>\uff0c\u5728\u5b98\u65b9keras\u6a21\u5757\u7684\u5b9e\u73b0\u4e2dMBConv<\/code>\u7ed3\u6784\u7684drop_rate<\/code>\u662f\u4ece0\u9012\u589e\u5230drop_connect_rate<\/code>\u7684\uff08\u5177\u4f53\u5b9e\u73b0\u53ef\u4ee5\u770b\u4e0b\u5b98\u65b9\u6e90\u7801\uff0c\u6ce8\u610f\uff0c\u5728\u6e90\u7801\u5b9e\u73b0\u4e2d\u53ea\u6709\u4f7f\u7528shortcut\u7684\u65f6\u5019\u624d\u6709Dropout\u5c42<\/strong>\uff09\u3002\u8fd8\u9700\u8981\u6ce8\u610f\u7684\u662f\uff0c\u8fd9\u91cc\u7684Dropout\u5c42\u662fStochastic Depth<\/code>\uff0c\u5373\u4f1a\u968f\u673a\u4e22\u6389\u6574\u4e2ablock\u7684\u4e3b\u5206\u652f\uff08\u53ea\u5269\u6377\u5f84\u5206\u652f\uff0c\u76f8\u5f53\u4e8e\u76f4\u63a5\u8df3\u8fc7\u4e86\u8fd9\u4e2ablock\uff09\u4e5f\u53ef\u4ee5\u7406\u89e3\u4e3a\u51cf\u5c11\u4e86\u7f51\u7edc\u7684\u6df1\u5ea6\u3002\u5177\u4f53\u53ef\u53c2\u8003Deep Networks with Stochastic Depth<\/code>\u8fd9\u7bc7\u6587\u7ae0\u3002<\/li>\n
                        • dropout_rate<\/code>\u662f\u6700\u540e\u4e00\u4e2a\u5168\u8fde\u63a5\u5c42\u524d\u7684dropout<\/code>\u5c42\uff08\u5728stage9<\/code>\u7684Pooling\u4e0eFC\u4e4b\u95f4\uff09\u7684dropout_rate<\/code>\u3002<\/li>\n<\/ul>\n

                          \u6700\u540e\u7ed9\u51fa\u539f\u8bba\u6587\u4e2d\u5173\u4e8eEfficientNet\u4e0e\u5f53\u65f6\u4e3b\u6d41\u7f51\u7edc\u7684\u6027\u80fd\u53c2\u6570\u5bf9\u6bd4\uff1a<\/p>\n

                          \"EfficientNet\u7f51\u7edc\u8be6\u89e3<\/p>\n","protected":false},"excerpt":{"rendered":"EfficientNet\u7f51\u7edc\u8be6\u89e3\u76ee\u5f55\u524d\u8a00\u8bba\u6587\u601d\u60f3\u7f51\u7edc\u8be6\u7ec6\u7ed3\u6784MBConv\u7ed3\u6784EfficientNet(B0-B7)\u53c2\u6570\u524d\u8a00\u539f\u8bba\u6587\u540d\u79f0\uff1aEfficientNet:Rethink...","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[],"tags":[],"_links":{"self":[{"href":"https:\/\/mushiming.com\/wp-json\/wp\/v2\/posts\/7346"}],"collection":[{"href":"https:\/\/mushiming.com\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/mushiming.com\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/mushiming.com\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/mushiming.com\/wp-json\/wp\/v2\/comments?post=7346"}],"version-history":[{"count":0,"href":"https:\/\/mushiming.com\/wp-json\/wp\/v2\/posts\/7346\/revisions"}],"wp:attachment":[{"href":"https:\/\/mushiming.com\/wp-json\/wp\/v2\/media?parent=7346"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/mushiming.com\/wp-json\/wp\/v2\/categories?post=7346"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/mushiming.com\/wp-json\/wp\/v2\/tags?post=7346"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}