{"id":88202,"date":"2026-07-31T17:56:44","date_gmt":"2026-07-31T09:56:44","guid":{"rendered":"https:\/\/www.wsisp.com\/helps\/88202.html"},"modified":"2026-07-31T17:56:44","modified_gmt":"2026-07-31T09:56:44","slug":"%e6%b7%b1%e5%ba%a6%e5%ad%a6%e4%b9%a0%e5%85%ab%e8%82%a1%ef%bc%8c%ef%bc%88%e5%9b%be%e5%83%8f%ef%bc%89","status":"publish","type":"post","link":"https:\/\/www.wsisp.com\/helps\/88202.html","title":{"rendered":"\u6df1\u5ea6\u5b66\u4e60\u516b\u80a1\uff0c\uff08\u56fe\u50cf\uff09"},"content":{"rendered":"<h3>\u4e00\u3001\u6df1\u5ea6\u5b66\u4e60\u57fa\u7840\u6838\u5fc3\u6982\u5ff5<\/h3>\n<h4>1. \u6df1\u5ea6\u5b66\u4e60 vs \u673a\u5668\u5b66\u4e60 vs \u4eba\u5de5\u667a\u80fd\u7684\u5173\u7cfb<\/h4>\n<ul>\n<li>\u4eba\u5de5\u667a\u80fd&#xff08;AI&#xff09;&#xff1a;\u6700\u5927\u8303\u7574&#xff0c;\u76ee\u6807\u662f\u8ba9\u673a\u5668\u5177\u5907\u4eba\u7c7b\u667a\u80fd\u80fd\u529b&#xff0c;\u6db5\u76d6\u673a\u5668\u5b66\u4e60\u3001\u4e13\u5bb6\u7cfb\u7edf\u3001\u77e5\u8bc6\u56fe\u8c31\u7b49\u3002<\/li>\n<li>\u673a\u5668\u5b66\u4e60&#xff08;ML&#xff09;&#xff1a;AI \u7684\u5b50\u96c6&#xff0c;\u901a\u8fc7\u6570\u636e\u548c\u7b97\u6cd5\u8ba9\u673a\u5668\u81ea\u52a8\u5b66\u4e60\u89c4\u5f8b&#xff0c;\u65e0\u9700\u4eba\u5de5\u786c\u7f16\u7801\u89c4\u5219&#xff0c;\u5305\u542b\u4f20\u7edf\u673a\u5668\u5b66\u4e60&#xff08;SVM\u3001\u51b3\u7b56\u6811\u3001LR \u7b49&#xff09;\u548c\u6df1\u5ea6\u5b66\u4e60\u3002<\/li>\n<li>\u6df1\u5ea6\u5b66\u4e60&#xff08;DL&#xff09;&#xff1a;\u673a\u5668\u5b66\u4e60\u7684\u5b50\u96c6&#xff0c;\u6838\u5fc3\u662f\u591a\u5c42\u795e\u7ecf\u7f51\u7edc&#xff0c;\u901a\u8fc7\u591a\u5c42\u975e\u7ebf\u6027\u53d8\u6362\u81ea\u52a8\u63d0\u53d6\u7279\u5f81&#xff0c;\u7aef\u5230\u7aef\u5b8c\u6210\u4efb\u52a1&#xff0c;\u65e0\u9700\u4eba\u5de5\u7279\u5f81\u5de5\u7a0b\u3002<\/li>\n<\/ul>\n<h4>2. \u6df1\u5ea6\u5b66\u4e60\u7684\u6838\u5fc3\u4f18\u52bf<\/h4>\n<li>\u81ea\u52a8\u7279\u5f81\u63d0\u53d6&#xff1a;\u4ece\u539f\u59cb\u6570\u636e&#xff08;\u50cf\u7d20\u3001\u6587\u672c&#xff09;\u4e2d\u81ea\u52a8\u5b66\u4e60\u5206\u5c42\u7279\u5f81&#xff0c;\u66ff\u4ee3\u4eba\u5de5\u7279\u5f81\u5de5\u7a0b\u3002<\/li>\n<li>\u5927\u6570\u636e\u4e0b\u6027\u80fd\u4e0a\u9650\u9ad8&#xff1a;\u6570\u636e\u91cf\u8d8a\u5927&#xff0c;\u6df1\u5ea6\u5b66\u4e60\u6548\u679c\u6301\u7eed\u63d0\u5347&#xff0c;\u4f20\u7edf\u673a\u5668\u5b66\u4e60\u6613\u8fdb\u5165\u74f6\u9888\u3002<\/li>\n<li>\u7aef\u5230\u7aef\u5b66\u4e60&#xff1a;\u8f93\u5165\u539f\u59cb\u6570\u636e&#xff0c;\u76f4\u63a5\u8f93\u51fa\u6700\u7ec8\u7ed3\u679c&#xff0c;\u7b80\u5316\u4efb\u52a1\u6d41\u7a0b\u3002<\/li>\n<li>\u6cdb\u5316\u80fd\u529b\u5f3a&#xff1a;\u9884\u8bad\u7ec3 &#043; \u5fae\u8c03\u8303\u5f0f\u53ef\u5feb\u901f\u8fc1\u79fb\u5230\u4e0d\u540c\u4efb\u52a1\u3002<\/li>\n<h4>3. \u611f\u77e5\u673a&#xff08;Perceptron&#xff09;<\/h4>\n<ul>\n<li>\u5b9a\u4e49&#xff1a;\u6700\u7b80\u5355\u7684\u4eba\u5de5\u795e\u7ecf\u5143&#xff0c;\u63a5\u6536\u591a\u4e2a\u8f93\u5165&#xff0c;\u52a0\u6743\u6c42\u548c\u540e\u7ecf\u8fc7\u9636\u8dc3\u51fd\u6570\u8f93\u51fa 0\/1\u3002<\/li>\n<li>\u5c40\u9650&#xff1a;\u53ea\u80fd\u89e3\u51b3\u7ebf\u6027\u53ef\u5206\u95ee\u9898&#xff0c;\u65e0\u6cd5\u5904\u7406\u5f02\u6216\u95ee\u9898&#xff1b;\u591a\u5c42\u611f\u77e5\u673a&#xff08;MLP&#xff09;\u901a\u8fc7\u5f15\u5165\u9690\u85cf\u5c42\u548c\u975e\u7ebf\u6027\u6fc0\u6d3b\u51fd\u6570\u89e3\u51b3\u8be5\u95ee\u9898\u3002<\/li>\n<\/ul>\n<h4>4. \u591a\u5c42\u611f\u77e5\u673a&#xff08;MLP&#xff09;<\/h4>\n<ul>\n<li>\u7ed3\u6784&#xff1a;\u8f93\u5165\u5c42 &#043; \u82e5\u5e72\u9690\u85cf\u5c42 &#043; \u8f93\u51fa\u5c42&#xff0c;\u5c42\u4e0e\u5c42\u4e4b\u95f4\u5168\u8fde\u63a5\u3002<\/li>\n<li>\u672c\u8d28&#xff1a;\u901a\u8fc7\u591a\u5c42\u975e\u7ebf\u6027\u53d8\u6362&#xff0c;\u5c06\u8f93\u5165\u6570\u636e\u6620\u5c04\u5230\u9ad8\u7ef4\u7a7a\u95f4&#xff0c;\u5b9e\u73b0\u590d\u6742\u51fd\u6570\u62df\u5408\u3002<\/li>\n<li>\u7f3a\u70b9&#xff1a;\u5168\u8fde\u63a5\u53c2\u6570\u8fc7\u591a&#xff1b;\u5bf9\u7a7a\u95f4 \/ \u65f6\u5e8f\u4fe1\u606f\u5229\u7528\u5dee&#xff1b;\u6613\u8fc7\u62df\u5408\u3002<\/li>\n<\/ul>\n<hr \/>\n<h3>\u4e8c\u3001\u795e\u7ecf\u7f51\u7edc\u57fa\u7840\u4e0e\u53cd\u5411\u4f20\u64ad<\/h3>\n<h4>1. \u524d\u5411\u4f20\u64ad&#xff08;Forward Propagation&#xff09;<\/h4>\n<ul>\n<li>\u6d41\u7a0b&#xff1a;\u8f93\u5165\u6570\u636e\u4ece\u8f93\u5165\u5c42\u8fdb\u5165&#xff0c;\u4f9d\u6b21\u7ecf\u8fc7\u6bcf\u4e00\u5c42\u7684\u7ebf\u6027\u53d8\u6362&#xff08;\\\\(Z&#061;WX&#043;b\\\\)&#xff09;\u548c\u975e\u7ebf\u6027\u6fc0\u6d3b&#xff08;\\\\(A&#061;\\\\sigma(Z)\\\\)&#xff09;&#xff0c;\u6700\u7ec8\u5728\u8f93\u51fa\u5c42\u5f97\u5230\u9884\u6d4b\u7ed3\u679c\u3002<\/li>\n<li>\u6838\u5fc3&#xff1a;\u8ba1\u7b97\u7f51\u7edc\u7684\u9884\u6d4b\u503c&#xff0c;\u4e3a\u540e\u7eed\u635f\u5931\u8ba1\u7b97\u63d0\u4f9b\u4f9d\u636e\u3002<\/li>\n<\/ul>\n<h4>2. \u635f\u5931\u51fd\u6570&#xff08;Loss Function&#xff09;<\/h4>\n<ul>\n<li>\u4f5c\u7528&#xff1a;\u8861\u91cf\u9884\u6d4b\u503c\u4e0e\u771f\u5b9e\u503c\u7684\u5dee\u5f02&#xff0c;\u662f\u53cd\u5411\u4f20\u64ad\u7684 \u201c\u6307\u6325\u68d2\u201d\u3002<\/li>\n<li>\u5206\u7c7b&#xff1a;\n<ul>\n<li>\u5206\u7c7b\u4efb\u52a1&#xff1a;\u4ea4\u53c9\u71b5\u635f\u5931\u3001Focal Loss \u7b49<\/li>\n<li>\u56de\u5f52\u4efb\u52a1&#xff1a;MSE\u3001MAE\u3001Smooth L1 \u7b49<\/li>\n<li>\u8be6\u89c1\u540e\u6587\u300c\u635f\u5931\u51fd\u6570\u5168\u89e3\u6790\u300d\u7ae0\u8282<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<h4>3. \u53cd\u5411\u4f20\u64ad&#xff08;Back Propagation, BP&#xff09;<\/h4>\n<h5>\u6838\u5fc3\u539f\u7406<\/h5>\n<p>\u57fa\u4e8e\u94fe\u5f0f\u6c42\u5bfc\u6cd5\u5219&#xff0c;\u4ece\u8f93\u51fa\u5c42\u5411\u8f93\u5165\u5c42\u9010\u5c42\u8ba1\u7b97\u635f\u5931\u51fd\u6570\u5bf9\u6bcf\u4e2a\u53c2\u6570\u7684\u68af\u5ea6&#xff0c;\u518d\u901a\u8fc7\u4f18\u5316\u5668\u66f4\u65b0\u53c2\u6570&#xff0c;\u6700\u5c0f\u5316\u635f\u5931\u3002<\/p>\n<h5>\u8ba1\u7b97\u6b65\u9aa4<\/h5>\n<li>\u524d\u5411\u4f20\u64ad\u8ba1\u7b97\u5404\u5c42\u8f93\u51fa\u4e0e\u6700\u7ec8\u635f\u5931<\/li>\n<li>\u8ba1\u7b97\u8f93\u51fa\u5c42\u635f\u5931\u5bf9\u8f93\u51fa\u7684\u68af\u5ea6<\/li>\n<li>\u4ece\u540e\u5f80\u524d\u9010\u5c42\u8ba1\u7b97&#xff1a;\u635f\u5931\u5bf9\u5f53\u524d\u5c42\u6743\u91cdW\u3001\u504f\u7f6eb\u7684\u68af\u5ea6&#xff0c;\u4ee5\u53ca\u635f\u5931\u5bf9\u4e0a\u4e00\u5c42\u6fc0\u6d3b\u503c\u7684\u68af\u5ea6<\/li>\n<li>\u4f18\u5316\u5668\u6839\u636e\u68af\u5ea6\u66f4\u65b0\u6240\u6709\u53c2\u6570<\/li>\n<h5>\u8ba1\u7b97\u56fe&#xff08;Computational Graph&#xff09;<\/h5>\n<p>\u5c06\u795e\u7ecf\u7f51\u7edc\u62c6\u89e3\u4e3a\u4e00\u4e2a\u4e2a\u57fa\u7840\u8fd0\u7b97\u8282\u70b9&#xff0c;\u5f62\u6210\u6709\u5411\u65e0\u73af\u56fe&#xff1b;\u524d\u5411\u4f20\u64ad\u8ba1\u7b97\u8282\u70b9\u503c&#xff0c;\u53cd\u5411\u4f20\u64ad\u6cbf\u53cd\u65b9\u5411\u8ba1\u7b97\u68af\u5ea6\u3002<\/p>\n<h4>4. \u68af\u5ea6\u6d88\u5931\u4e0e\u68af\u5ea6\u7206\u70b8<\/h4>\n<h5>\u4ea7\u751f\u539f\u56e0<\/h5>\n<p>\u6df1\u5c42\u7f51\u7edc\u4e2d&#xff0c;\u53cd\u5411\u4f20\u64ad\u65f6\u68af\u5ea6\u901a\u8fc7\u94fe\u5f0f\u6cd5\u5219\u5c42\u5c42\u76f8\u4e58&#xff1a;<\/p>\n<ul>\n<li>\u82e5\u6bcf\u5c42\u68af\u5ea6\u90fd\u5c0f\u4e8e 1&#xff0c;\u591a\u6b21\u76f8\u4e58\u540e\u68af\u5ea6\u6307\u6570\u7ea7\u8870\u51cf\u2192\u68af\u5ea6\u6d88\u5931&#xff0c;\u6d45\u5c42\u7f51\u7edc\u53c2\u6570\u51e0\u4e4e\u4e0d\u66f4\u65b0\u3002<\/li>\n<li>\u82e5\u6bcf\u5c42\u68af\u5ea6\u90fd\u5927\u4e8e 1&#xff0c;\u591a\u6b21\u76f8\u4e58\u540e\u68af\u5ea6\u6307\u6570\u7ea7\u589e\u5927\u2192\u68af\u5ea6\u7206\u70b8&#xff0c;\u53c2\u6570\u66f4\u65b0\u5267\u70c8&#xff0c;\u6a21\u578b\u9707\u8361\u4e0d\u6536\u655b\u3002<\/li>\n<\/ul>\n<h5>\u6839\u672c\u8bf1\u56e0<\/h5>\n<ul>\n<li>\u6fc0\u6d3b\u51fd\u6570\u9009\u62e9\u4e0d\u5f53&#xff08;\u5982 sigmoid\/tanh \u6613\u5bfc\u81f4\u68af\u5ea6\u6d88\u5931&#xff09;<\/li>\n<li>\u6743\u91cd\u521d\u59cb\u5316\u4e0d\u5408\u7406&#xff08;\u521d\u59cb\u6743\u91cd\u8fc7\u5927 \/ \u8fc7\u5c0f&#xff09;<\/li>\n<li>\u7f51\u7edc\u5c42\u6570\u8fc7\u6df1<\/li>\n<\/ul>\n<h5>\u89e3\u51b3\u65b9\u6848<\/h5>\n<p>\u8868\u683c<\/p>\n<table>\n<tr>\u95ee\u9898\u89e3\u51b3\u65b9\u6848<\/tr>\n<tbody>\n<tr>\n<td>\u68af\u5ea6\u6d88\u5931<\/td>\n<td>1. \u4f7f\u7528 ReLU \u7cfb\u5217\u6fc0\u6d3b\u51fd\u6570&#xff1b;2. \u6b8b\u5dee\u8fde\u63a5&#xff08;ResNet&#xff09;&#xff1b;3. BatchNorm\/LayerNorm&#xff1b;4. \u5408\u7406\u7684\u6743\u91cd\u521d\u59cb\u5316&#xff08;He \u521d\u59cb\u5316&#xff09;&#xff1b;5. \u95e8\u63a7\u673a\u5236&#xff08;LSTM\/GRU&#xff09;<\/td>\n<\/tr>\n<tr>\n<td>\u68af\u5ea6\u7206\u70b8<\/td>\n<td>1. \u68af\u5ea6\u88c1\u526a&#xff08;Gradient Clipping&#xff09;&#xff1b;2. \u6743\u91cd\u6b63\u5219\u5316&#xff1b;3. BatchNorm&#xff1b;4. \u964d\u4f4e\u5b66\u4e60\u7387<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<hr \/>\n<h3>\u4e09\u3001\u6fc0\u6d3b\u51fd\u6570\u4e13\u9898<\/h3>\n<h4>\u6838\u5fc3\u8981\u6c42<\/h4>\n<p>\u6fc0\u6d3b\u51fd\u6570\u5fc5\u987b\u975e\u7ebf\u6027&#xff08;\u5426\u5219\u591a\u5c42\u7f51\u7edc\u9000\u5316\u4e3a\u5355\u5c42\u7ebf\u6027\u53d8\u6362&#xff09;\u3001\u53ef\u5bfc&#xff08;\u652f\u6301\u53cd\u5411\u4f20\u64ad&#xff09;\u3001\u8ba1\u7b97\u9ad8\u6548\u3002<\/p>\n<h4>1. Sigmoid<\/h4>\n<ul>\n<li>\u516c\u5f0f&#xff1a;\\\\(\\\\sigma(x) &#061; \\\\frac{1}{1&#043;e^{-x}}\\\\)<\/li>\n<li>\u8f93\u51fa\u8303\u56f4&#xff1a;\\\\((0,1)\\\\)&#xff0c;\u53ef\u8868\u793a\u6982\u7387\u6216\u505a\u5f52\u4e00\u5316<\/li>\n<li>\u7f3a\u70b9&#xff1a;\n<li>\u6613\u5bfc\u81f4\u68af\u5ea6\u6d88\u5931&#xff1a;\u8f93\u5165\u7edd\u5bf9\u503c\u8f83\u5927\u65f6&#xff0c;\u5bfc\u6570\u8d8b\u8fd1\u4e8e 0<\/li>\n<li>\u8f93\u51fa\u4e0d\u662f 0 \u5747\u503c&#xff08;zero-centered&#xff09;&#xff0c;\u4f1a\u5bfc\u81f4\u6743\u91cd\u66f4\u65b0\u65b9\u5411\u504f\u79fb<\/li>\n<li>\u5305\u542b\u6307\u6570\u8fd0\u7b97&#xff0c;\u8ba1\u7b97\u8f83\u6162<\/li>\n<\/li>\n<\/ul>\n<h4>2. Tanh<\/h4>\n<ul>\n<li>\u516c\u5f0f&#xff1a;\\\\(\\\\tanh(x) &#061; \\\\frac{e^x &#8211; e^{-x}}{e^x &#043; e^{-x}}\\\\)<\/li>\n<li>\u8f93\u51fa\u8303\u56f4&#xff1a;\\\\((-1,1)\\\\)&#xff0c;\u96f6\u5747\u503c<\/li>\n<li>\u7f3a\u70b9&#xff1a;\u4ecd\u5b58\u5728\u68af\u5ea6\u6d88\u5931\u95ee\u9898&#xff1b;\u4ecd\u6709\u6307\u6570\u8fd0\u7b97<\/li>\n<\/ul>\n<h4>3. ReLU&#xff08;Rectified Linear Unit&#xff09;<\/h4>\n<ul>\n<li>\u516c\u5f0f&#xff1a;\\\\(ReLU(x) &#061; max(0, x)\\\\)<\/li>\n<li>\u4f18\u70b9&#xff1a;\n<li>\u6b63\u533a\u95f4\u65e0\u68af\u5ea6\u6d88\u5931\u95ee\u9898<\/li>\n<li>\u8ba1\u7b97\u6781\u5feb&#xff0c;\u65e0\u6307\u6570\u8fd0\u7b97<\/li>\n<li>\u5e26\u6765\u7a00\u758f\u6027&#xff0c;\u90e8\u5206\u795e\u7ecf\u5143\u5931\u6d3b&#xff0c;\u964d\u4f4e\u8fc7\u62df\u5408\u98ce\u9669<\/li>\n<\/li>\n<li>\u7f3a\u70b9&#xff1a;\n<li>Dead ReLU \u95ee\u9898&#xff1a;\u8d1f\u533a\u95f4\u68af\u5ea6\u6052\u4e3a 0&#xff0c;\u795e\u7ecf\u5143\u4e00\u65e6\u8fdb\u5165\u8d1f\u533a\u95f4\u53ef\u80fd\u6c38\u4e45\u5931\u6d3b<\/li>\n<li>\u8f93\u51fa\u975e\u96f6\u5747\u503c<\/li>\n<li>\u65e0\u4e0a\u9650&#xff0c;\u53ef\u80fd\u5bfc\u81f4\u6570\u503c\u4e0d\u7a33\u5b9a<\/li>\n<\/li>\n<\/ul>\n<h4>4. Leaky ReLU<\/h4>\n<ul>\n<li>\u516c\u5f0f&#xff1a;\\\\(LeakyReLU(x) &#061; max(\\\\alpha x, x)\\\\)&#xff0c;\\\\(\\\\alpha\\\\)\u901a\u5e38\u53d6 0.01<\/li>\n<li>\u6539\u8fdb&#xff1a;\u8d1f\u533a\u95f4\u4fdd\u7559\u5c0f\u68af\u5ea6&#xff0c;\u89e3\u51b3 Dead ReLU \u95ee\u9898<\/li>\n<li>\u7f3a\u70b9&#xff1a;\\\\(\\\\alpha\\\\)\u4e3a\u56fa\u5b9a\u8d85\u53c2\u6570&#xff0c;\u6548\u679c\u4e0d\u4e00\u5b9a\u7a33\u5b9a<\/li>\n<\/ul>\n<h4>5. PReLU&#xff08;Parametric ReLU&#xff09;<\/h4>\n<ul>\n<li>\u516c\u5f0f&#xff1a;\\\\(PReLU(x) &#061; max(\\\\alpha x, x)\\\\)&#xff0c;\\\\(\\\\alpha\\\\)\u4e3a\u53ef\u5b66\u4e60\u53c2\u6570<\/li>\n<li>\u6539\u8fdb&#xff1a;\u7f51\u7edc\u81ea\u52a8\u5b66\u4e60\u8d1f\u533a\u95f4\u659c\u7387&#xff0c;\u9002\u914d\u4e0d\u540c\u6570\u636e\u5206\u5e03<\/li>\n<\/ul>\n<h4>6. ELU&#xff08;Exponential Linear Unit&#xff09;<\/h4>\n<ul>\n<li>\u516c\u5f0f&#xff1a;\\\\(ELU(x) &#061; \\\\begin{cases} x, &amp; x&gt;0 \\\\\\\\ \\\\alpha(e^x-1), &amp; x\\\\le0 \\\\end{cases}\\\\)<\/li>\n<li>\u7279\u70b9&#xff1a;\u8d1f\u533a\u95f4\u5e73\u6ed1\u8fc7\u6e21&#xff0c;\u8f93\u51fa\u5747\u503c\u8d8b\u8fd1\u4e8e 0&#xff1b;\u4ecd\u6709\u6307\u6570\u8fd0\u7b97\u5f00\u9500<\/li>\n<\/ul>\n<h4>7. GELU&#xff08;Gaussian Error Linear Unit&#xff09;<\/h4>\n<ul>\n<li>\u516c\u5f0f&#xff1a;\\\\(GELU(x) &#061; x \\\\cdot \\\\Phi(x)\\\\)&#xff0c;\\\\(\\\\Phi(x)\\\\)\u4e3a\u6807\u51c6\u6b63\u6001\u5206\u5e03\u7d2f\u79ef\u5206\u5e03\u51fd\u6570<\/li>\n<li>\u7279\u70b9&#xff1a;\n<li>\u5e73\u6ed1\u7684\u975e\u7ebf\u6027&#xff0c;\u517c\u5177\u6b63\u5219\u6548\u679c<\/li>\n<li>\u662f Transformer\u3001BERT\u3001GPT \u7cfb\u5217\u7684\u9ed8\u8ba4\u6fc0\u6d3b\u51fd\u6570<\/li>\n<li>\u8fd1\u4f3c\u516c\u5f0f&#xff1a;\\\\(0.5x(1&#043;\\\\tanh(\\\\sqrt{2\/\\\\pi}(x&#043;0.044715x^3)))\\\\)&#xff0c;\u52a0\u901f\u8ba1\u7b97<\/li>\n<\/li>\n<\/ul>\n<h4>8. Swish \/ SiLU<\/h4>\n<ul>\n<li>\u516c\u5f0f&#xff1a;\\\\(Swish(x) &#061; x \\\\cdot \\\\sigma(\\\\beta x)\\\\)&#xff0c;\\\\(\\\\beta\\\\)\u4e3a\u53ef\u5b66\u4e60\u53c2\u6570\u6216\u56fa\u5b9a\u4e3a 1<\/li>\n<li>\u7279\u70b9&#xff1a;\u5e73\u6ed1\u3001\u65e0\u4e0a\u754c\u3001\u6709\u4e0b\u754c&#xff1b;\u5728\u6df1\u5c42\u7f51\u7edc\u4e2d\u6548\u679c\u5e38\u4f18\u4e8e ReLU<\/li>\n<\/ul>\n<h4>9. \u5e38\u89c1\u5bf9\u6bd4<\/h4>\n<p>\u8868\u683c<\/p>\n<table>\n<tr>\u6fc0\u6d3b\u51fd\u6570\u68af\u5ea6\u6d88\u5931\u96f6\u5747\u503c\u8ba1\u7b97\u901f\u5ea6\u9002\u7528\u573a\u666f<\/tr>\n<tbody>\n<tr>\n<td>Sigmoid<\/td>\n<td>\u4e25\u91cd<\/td>\n<td>\u5426<\/td>\n<td>\u6162<\/td>\n<td>\u8f93\u51fa\u5c42\u4e8c\u5206\u7c7b\u3001\u95e8\u63a7\u5355\u5143<\/td>\n<\/tr>\n<tr>\n<td>Tanh<\/td>\n<td>\u8f83\u4e25\u91cd<\/td>\n<td>\u662f<\/td>\n<td>\u6162<\/td>\n<td>\u5faa\u73af\u795e\u7ecf\u7f51\u7edc\u9690\u85cf\u5c42<\/td>\n<\/tr>\n<tr>\n<td>ReLU<\/td>\n<td>\u6b63\u533a\u95f4\u65e0<\/td>\n<td>\u5426<\/td>\n<td>\u6781\u5feb<\/td>\n<td>CNN\u3001MLP \u9690\u85cf\u5c42&#xff08;\u6700\u5e38\u7528&#xff09;<\/td>\n<\/tr>\n<tr>\n<td>GELU<\/td>\n<td>\u65e0<\/td>\n<td>\u8fd1\u4f3c<\/td>\n<td>\u8f83\u5feb<\/td>\n<td>Transformer\u3001\u5927\u6a21\u578b<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<hr \/>\n<h3>\u56db\u3001\u5377\u79ef\u795e\u7ecf\u7f51\u7edc&#xff08;CNN&#xff09;\u6838\u5fc3\u516b\u80a1<\/h3>\n<h4>1. \u5377\u79ef\u8fd0\u7b97\u7684\u672c\u8d28<\/h4>\n<ul>\n<li>\u5377\u79ef\u6838&#xff08;\u6ee4\u6ce2\u5668&#xff09;\u5728\u8f93\u5165\u7279\u5f81\u56fe\u4e0a\u6ed1\u52a8&#xff0c;\u5bf9\u5e94\u4f4d\u7f6e\u5143\u7d20\u76f8\u4e58\u6c42\u548c&#xff0c;\u63d0\u53d6\u5c40\u90e8\u7279\u5f81\u3002<\/li>\n<li>\u6838\u5fc3\u601d\u60f3&#xff1a;\u5c40\u90e8\u611f\u53d7\u91ce &#043; \u53c2\u6570\u5171\u4eab &#043; \u7a7a\u95f4\u5c42\u6b21\u5316&#xff0c;\u5927\u5e45\u51cf\u5c11\u53c2\u6570\u91cf&#xff0c;\u4fdd\u7559\u7a7a\u95f4\u7ed3\u6784\u4fe1\u606f\u3002<\/li>\n<\/ul>\n<h4>2. \u5377\u79ef\u6838\u5fc3\u53c2\u6570<\/h4>\n<ul>\n<li>\u5377\u79ef\u6838\u5927\u5c0f&#xff08;Kernel Size&#xff09;&#xff1a;\u5e38\u7528 3&#215;3\u30011&#215;1\u30015&#215;5&#xff1b;3&#215;3 \u662f\u4e3b\u6d41&#xff0c;\u5806\u53e0\u4e24\u4e2a 3&#215;3 \u7b49\u4ef7\u4e8e\u4e00\u4e2a 5&#215;5 \u7684\u611f\u53d7\u91ce&#xff0c;\u53c2\u6570\u91cf\u66f4\u5c11\u3001\u975e\u7ebf\u6027\u66f4\u5f3a\u3002<\/li>\n<li>\u6b65\u957f&#xff08;Stride&#xff09;&#xff1a;\u5377\u79ef\u6838\u6bcf\u6b21\u6ed1\u52a8\u7684\u50cf\u7d20\u6570&#xff1b;\u6b65\u957f &gt; 1 \u53ef\u5b9e\u73b0\u4e0b\u91c7\u6837&#xff0c;\u7f29\u5c0f\u7279\u5f81\u56fe\u5c3a\u5bf8\u3002<\/li>\n<li>\u586b\u5145&#xff08;Padding&#xff09;&#xff1a;\u5728\u7279\u5f81\u56fe\u8fb9\u7f18\u8865 0&#xff1b;\u4f5c\u7528\u662f&#xff1a;\n<li>\u4fdd\u6301\u8f93\u5165\u8f93\u51fa\u5c3a\u5bf8\u4e00\u81f4&#xff08;Same Padding&#xff09;<\/li>\n<li>\u907f\u514d\u8fb9\u7f18\u4fe1\u606f\u4e22\u5931<\/li>\n<\/li>\n<li>\u8f93\u51fa\u5c3a\u5bf8\u8ba1\u7b97\u516c\u5f0f&#xff1a; \\\\(Output\\\\_size &#061; \\\\lfloor \\\\frac{Input\\\\_size &#043; 2\\\\times Padding &#8211; Kernel\\\\_size}{Stride} \\\\rfloor &#043; 1\\\\)<\/li>\n<\/ul>\n<h4>3. \u6c60\u5316&#xff08;Pooling&#xff09;<\/h4>\n<ul>\n<li>\u4f5c\u7528&#xff1a;\u4e0b\u91c7\u6837&#xff0c;\u7f29\u5c0f\u7279\u5f81\u56fe\u5c3a\u5bf8\u3001\u51cf\u5c11\u53c2\u6570\u91cf\u3001\u6269\u5927\u611f\u53d7\u91ce&#xff1b;\u5e26\u6765\u4e00\u5b9a\u7684\u5e73\u79fb\u4e0d\u53d8\u6027\u3002<\/li>\n<li>\u5e38\u89c1\u7c7b\u578b&#xff1a;\n<ul>\n<li>\u6700\u5927\u6c60\u5316&#xff08;Max Pooling&#xff09;&#xff1a;\u53d6\u7a97\u53e3\u5185\u6700\u5927\u503c&#xff0c;\u4fdd\u7559\u7eb9\u7406\u8fb9\u7f18\u7279\u5f81&#xff0c;\u6700\u5e38\u7528<\/li>\n<li>\u5e73\u5747\u6c60\u5316&#xff08;Average Pooling&#xff09;&#xff1a;\u53d6\u7a97\u53e3\u5185\u5e73\u5747\u503c&#xff0c;\u4fdd\u7559\u6574\u4f53\u80cc\u666f\u4fe1\u606f<\/li>\n<li>\u5168\u5c40\u5e73\u5747\u6c60\u5316&#xff08;GAP&#xff09;&#xff1a;\u5bf9\u6574\u4e2a\u7279\u5f81\u56fe\u53d6\u5747\u503c&#xff0c;\u66ff\u4ee3\u5168\u8fde\u63a5\u5c42&#xff0c;\u5927\u5e45\u51cf\u5c11\u53c2\u6570<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<h4>4. \u611f\u53d7\u91ce&#xff08;Receptive Field&#xff09;<\/h4>\n<ul>\n<li>\u5b9a\u4e49&#xff1a;\u7279\u5f81\u56fe\u4e0a\u4e00\u4e2a\u50cf\u7d20\u70b9&#xff0c;\u5bf9\u5e94\u539f\u59cb\u8f93\u5165\u56fe\u50cf\u4e0a\u7684\u533a\u57df\u5927\u5c0f\u3002<\/li>\n<li>\u8ba1\u7b97\u65b9\u5f0f&#xff08;\u4ece\u540e\u5f80\u524d&#xff09;&#xff1a; \\\\(RF_i &#061; (RF_{i&#043;1} &#8211; 1) \\\\times Stride_i &#043; Kernel\\\\_size_i\\\\) \u9876\u5c42\u8f93\u51fa\u7684\u521d\u59cb\u611f\u53d7\u91ce\u4e3a 1\u3002<\/li>\n<li>\u610f\u4e49&#xff1a;\u611f\u53d7\u91ce\u8d8a\u5927&#xff0c;\u80fd\u6355\u6349\u7684\u4e0a\u4e0b\u6587\u4fe1\u606f\u8d8a\u4e30\u5bcc&#xff1b;\u6df1\u5c42\u7f51\u7edc\u611f\u53d7\u91ce\u66f4\u5927\u3002<\/li>\n<\/ul>\n<h4>5. 1\u00d71 \u5377\u79ef\u7684\u4f5c\u7528<\/h4>\n<li>\u5347\u7ef4 \/ \u964d\u7ef4&#xff1a;\u901a\u8fc7\u8c03\u6574\u8f93\u51fa\u901a\u9053\u6570&#xff0c;\u6539\u53d8\u7279\u5f81\u7ef4\u5ea6&#xff0c;\u51cf\u5c11\u8ba1\u7b97\u91cf<\/li>\n<li>\u589e\u52a0\u975e\u7ebf\u6027&#xff1a;\u5f15\u5165\u6fc0\u6d3b\u51fd\u6570&#xff0c;\u63d0\u5347\u7f51\u7edc\u8868\u8fbe\u80fd\u529b<\/li>\n<li>\u8de8\u901a\u9053\u4fe1\u606f\u4ea4\u4e92&#xff1a;\u878d\u5408\u4e0d\u540c\u901a\u9053\u7684\u7279\u5f81<\/li>\n<li>\u662f Inception\u3001ResNet \u7b49\u7f51\u7edc\u7684\u6838\u5fc3\u7ec4\u4ef6<\/li>\n<h4>6. \u6df1\u5ea6\u53ef\u5206\u79bb\u5377\u79ef&#xff08;Depthwise Separable Convolution&#xff09;<\/h4>\n<ul>\n<li>\u62c6\u5206&#xff1a;\u6df1\u5ea6\u5377\u79ef&#xff08;Depthwise Conv&#xff09;&#043; \u9010\u70b9\u5377\u79ef&#xff08;Pointwise Conv&#xff0c;\u5373 1&#215;1 \u5377\u79ef&#xff09;\n<li>\u6df1\u5ea6\u5377\u79ef&#xff1a;\u6bcf\u4e2a\u5377\u79ef\u6838\u53ea\u8d1f\u8d23\u4e00\u4e2a\u8f93\u5165\u901a\u9053&#xff0c;\u9010\u901a\u9053\u5377\u79ef<\/li>\n<li>\u9010\u70b9\u5377\u79ef&#xff1a;\u7528 1&#215;1 \u5377\u79ef\u878d\u5408\u6240\u6709\u901a\u9053\u4fe1\u606f<\/li>\n<\/li>\n<li>\u4f18\u52bf&#xff1a;\u53c2\u6570\u91cf\u548c\u8ba1\u7b97\u91cf\u8fdc\u5c0f\u4e8e\u6807\u51c6\u5377\u79ef&#xff0c;\u662f\u8f7b\u91cf\u5316\u7f51\u7edc&#xff08;MobileNet&#xff09;\u7684\u6838\u5fc3\u3002<\/li>\n<\/ul>\n<h4>7. \u7a7a\u6d1e\u5377\u79ef&#xff08;Dilated Convolution&#xff09;<\/h4>\n<ul>\n<li>\u7279\u70b9&#xff1a;\u5377\u79ef\u6838\u5185\u90e8\u63d2\u5165\u7a7a\u6d1e&#xff0c;\u5728\u4e0d\u589e\u52a0\u53c2\u6570\u91cf\u7684\u524d\u63d0\u4e0b\u6269\u5927\u611f\u53d7\u91ce\u3002<\/li>\n<li>\u9002\u7528\u573a\u666f&#xff1a;\u56fe\u50cf\u5206\u5272\u3001\u76ee\u6807\u68c0\u6d4b\u7b49\u9700\u8981\u5927\u611f\u53d7\u91ce\u53c8\u4e0d\u60f3\u4e0b\u91c7\u6837\u7684\u4efb\u52a1\u3002<\/li>\n<li>\u95ee\u9898&#xff1a;\u6805\u683c\u6548\u5e94&#xff08;\u4fe1\u606f\u4e0d\u8fde\u7eed&#xff09;&#xff0c;\u53ef\u901a\u8fc7\u591a\u5c42\u4e0d\u540c\u81a8\u80c0\u7387\u7684\u7a7a\u6d1e\u5377\u79ef\u53e0\u52a0\u7f13\u89e3\u3002<\/li>\n<\/ul>\n<h4>8. \u8f6c\u7f6e\u5377\u79ef&#xff08;Transposed Convolution \/ \u53cd\u5377\u79ef&#xff09;<\/h4>\n<ul>\n<li>\u4f5c\u7528&#xff1a;\u4e0a\u91c7\u6837&#xff0c;\u5c06\u5c0f\u5c3a\u5bf8\u7279\u5f81\u56fe\u6062\u590d\u4e3a\u5927\u5c3a\u5bf8&#xff0c;\u5e38\u7528\u4e8e\u8bed\u4e49\u5206\u5272\u3001\u751f\u6210\u6a21\u578b\u3002<\/li>\n<li>\u6ce8\u610f&#xff1a;\u4e0d\u662f\u5377\u79ef\u7684\u9006\u8fd0\u7b97&#xff0c;\u53ea\u662f\u5b9e\u73b0\u5c3a\u5bf8\u4e0a\u7684\u653e\u5927&#xff1b;\u53ef\u5b66\u4e60\u53c2\u6570\u3002<\/li>\n<\/ul>\n<h4>9. \u5206\u7ec4\u5377\u79ef&#xff08;Group Convolution&#xff09;<\/h4>\n<ul>\n<li>\u505a\u6cd5&#xff1a;\u5c06\u8f93\u5165\u901a\u9053\u5206\u6210\u82e5\u5e72\u7ec4&#xff0c;\u6bcf\u7ec4\u5185\u72ec\u7acb\u505a\u5377\u79ef&#xff0c;\u6700\u540e\u62fc\u63a5\u7ed3\u679c\u3002<\/li>\n<li>\u4f18\u52bf&#xff1a;\u51cf\u5c11\u53c2\u6570\u91cf\u548c\u8ba1\u7b97\u91cf&#xff1b;\u901a\u9053\u5206\u7ec4\u5e26\u6765\u4e00\u5b9a\u7684\u6b63\u5219\u6548\u679c\u3002<\/li>\n<li>\u6781\u7aef\u60c5\u51b5&#xff1a;\u5206\u7ec4\u6570 &#061; \u901a\u9053\u6570 \u2192 \u6df1\u5ea6\u5377\u79ef\u3002<\/li>\n<\/ul>\n<h4>10. CNN \u4e3a\u4ec0\u4e48\u9002\u5408\u56fe\u50cf\u4efb\u52a1&#xff1f;<\/h4>\n<li>\u53c2\u6570\u5171\u4eab&#xff1a;\u4e00\u4e2a\u5377\u79ef\u6838\u5728\u5168\u56fe\u6ed1\u52a8&#xff0c;\u53c2\u6570\u91cf\u8fdc\u5c0f\u4e8e\u5168\u8fde\u63a5&#xff0c;\u964d\u4f4e\u8fc7\u62df\u5408<\/li>\n<li>\u5c40\u90e8\u8fde\u63a5&#xff1a;\u7b26\u5408\u89c6\u89c9\u4fe1\u606f\u5c40\u90e8\u76f8\u5173\u6027&#xff08;\u50cf\u7d20\u53ea\u548c\u90bb\u8fd1\u50cf\u7d20\u5f3a\u76f8\u5173&#xff09;<\/li>\n<li>\u5c42\u6b21\u5316\u7279\u5f81&#xff1a;\u6d45\u5c42\u63d0\u53d6\u8fb9\u7f18\u3001\u7eb9\u7406&#xff0c;\u4e2d\u5c42\u63d0\u53d6\u5f62\u72b6&#xff0c;\u9ad8\u5c42\u63d0\u53d6\u8bed\u4e49\u4fe1\u606f<\/li>\n<li>\u5e73\u79fb\u7b49\u53d8\u6027&#xff1a;\u76ee\u6807\u5e73\u79fb\u540e&#xff0c;\u7279\u5f81\u8f93\u51fa\u4e5f\u5bf9\u5e94\u5e73\u79fb&#xff0c;\u7279\u5f81\u68c0\u6d4b\u66f4\u7a33\u5b9a<\/li>\n<hr \/>\n<h3>\u4e94\u3001\u5faa\u73af\u795e\u7ecf\u7f51\u7edc&#xff08;RNN\/GRU\/LSTM&#xff09;<\/h3>\n<h4>1. RNN \u57fa\u672c\u539f\u7406<\/h4>\n<ul>\n<li>\u6838\u5fc3&#xff1a;\u9690\u85cf\u5c42\u72b6\u6001\u4e0d\u4ec5\u53d6\u51b3\u4e8e\u5f53\u524d\u8f93\u5165&#xff0c;\u8fd8\u53d6\u51b3\u4e8e\u4e0a\u4e00\u65f6\u523b\u7684\u9690\u85cf\u72b6\u6001&#xff0c;\u4ece\u800c\u6355\u6349\u65f6\u5e8f\u4f9d\u8d56\u3002<\/li>\n<li>\u516c\u5f0f&#xff1a;\\\\(h_t &#061; \\\\tanh(W_{xh}x_t &#043; W_{hh}h_{t-1} &#043; b_h)\\\\)<\/li>\n<li>\u7f3a\u70b9&#xff1a;\n<li>\u957f\u5e8f\u5217\u4e0b\u4e25\u91cd\u7684\u68af\u5ea6\u6d88\u5931&#xff0c;\u65e0\u6cd5\u6355\u6349\u957f\u8ddd\u79bb\u4f9d\u8d56<\/li>\n<li>\u4e32\u884c\u8ba1\u7b97&#xff0c;\u65e0\u6cd5\u5e76\u884c&#xff0c;\u901f\u5ea6\u6162<\/li>\n<\/li>\n<\/ul>\n<h4>2. LSTM&#xff08;\u957f\u77ed\u671f\u8bb0\u5fc6\u7f51\u7edc&#xff09;<\/h4>\n<p>\u901a\u8fc7\u4e09\u95e8\u4e00\u7ec6\u80de\u72b6\u6001\u89e3\u51b3\u957f\u5e8f\u5217\u68af\u5ea6\u6d88\u5931\u95ee\u9898&#xff1a;<\/p>\n<ul>\n<li>\u9057\u5fd8\u95e8&#xff1a;\u51b3\u5b9a\u4e0a\u4e00\u65f6\u523b\u7ec6\u80de\u72b6\u6001\u4fdd\u7559\u591a\u5c11\u4fe1\u606f<\/li>\n<li>\u8f93\u5165\u95e8&#xff1a;\u51b3\u5b9a\u5f53\u524d\u65f6\u523b\u65b0\u4fe1\u606f\u5b58\u5165\u7ec6\u80de\u72b6\u6001\u7684\u591a\u5c11<\/li>\n<li>\u8f93\u51fa\u95e8&#xff1a;\u51b3\u5b9a\u7ec6\u80de\u72b6\u6001\u8f93\u51fa\u591a\u5c11\u5230\u9690\u85cf\u72b6\u6001<\/li>\n<li>\u7ec6\u80de\u72b6\u6001&#xff08;Cell State&#xff09;&#xff1a;\u4fe1\u606f\u4e3b\u5e72\u9053&#xff0c;\u68af\u5ea6\u53ef\u987a\u7545\u4f20\u9012&#xff0c;\u7f13\u89e3\u68af\u5ea6\u6d88\u5931<\/li>\n<\/ul>\n<h4>3. GRU&#xff08;\u95e8\u63a7\u5faa\u73af\u5355\u5143&#xff09;<\/h4>\n<ul>\n<li>LSTM \u7684\u7b80\u5316\u7248&#xff0c;\u5c06\u9057\u5fd8\u95e8\u548c\u8f93\u5165\u95e8\u5408\u5e76\u4e3a\u66f4\u65b0\u95e8&#xff0c;\u65b0\u589e\u91cd\u7f6e\u95e8&#xff1b;\u53bb\u6389\u7ec6\u80de\u72b6\u6001&#xff0c;\u4ec5\u4fdd\u7559\u9690\u85cf\u72b6\u6001\u3002<\/li>\n<li>\u53c2\u6570\u91cf\u66f4\u5c11&#xff0c;\u8bad\u7ec3\u66f4\u5feb&#xff1b;\u6548\u679c\u4e0e LSTM \u76f8\u8fd1&#xff0c;\u6570\u636e\u91cf\u5c0f\u65f6\u66f4\u4f18\u3002<\/li>\n<\/ul>\n<h4>4. LSTM vs GRU<\/h4>\n<p>\u8868\u683c<\/p>\n<table>\n<tr>\u7ef4\u5ea6LSTMGRU<\/tr>\n<tbody>\n<tr>\n<td>\u95e8\u6570\u91cf<\/td>\n<td>3 \u4e2a&#xff08;\u9057\u5fd8\u3001\u8f93\u5165\u3001\u8f93\u51fa&#xff09;<\/td>\n<td>2 \u4e2a&#xff08;\u66f4\u65b0\u3001\u91cd\u7f6e&#xff09;<\/td>\n<\/tr>\n<tr>\n<td>\u72b6\u6001<\/td>\n<td>\u7ec6\u80de\u72b6\u6001 &#043; \u9690\u85cf\u72b6\u6001<\/td>\n<td>\u4ec5\u9690\u85cf\u72b6\u6001<\/td>\n<\/tr>\n<tr>\n<td>\u53c2\u6570\u91cf<\/td>\n<td>\u591a<\/td>\n<td>\u5c11<\/td>\n<\/tr>\n<tr>\n<td>\u8bad\u7ec3\u901f\u5ea6<\/td>\n<td>\u6162<\/td>\n<td>\u5feb<\/td>\n<\/tr>\n<tr>\n<td>\u957f\u5e8f\u5217\u6548\u679c<\/td>\n<td>\u901a\u5e38\u66f4\u7a33\u5b9a<\/td>\n<td>\u76f8\u8fd1&#xff0c;\u5c0f\u6570\u636e\u66f4\u4f18<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<h4>5. RNN \u7cfb\u5217\u4e3a\u4ec0\u4e48\u4f1a\u68af\u5ea6\u6d88\u5931&#xff1f;\u548c CNN \u7684\u68af\u5ea6\u6d88\u5931\u6709\u4f55\u4e0d\u540c&#xff1f;<\/h4>\n<ul>\n<li>RNN&#xff1a;\u540c\u4e00\u6743\u91cd\u77e9\u9635\u5728\u4e0d\u540c\u65f6\u523b\u53cd\u590d\u76f8\u4e58&#xff0c;\u957f\u5e8f\u5217\u4e0b\u6307\u6570\u7ea7\u8870\u51cf&#xff0c;\u662f\u65f6\u95f4\u7ef4\u5ea6\u7684\u68af\u5ea6\u6d88\u5931\u3002<\/li>\n<li>CNN&#xff1a;\u4e0d\u540c\u5c42\u6743\u91cd\u77e9\u9635\u76f8\u4e58&#xff0c;\u662f\u6df1\u5ea6\u7ef4\u5ea6\u7684\u68af\u5ea6\u6d88\u5931\u3002<\/li>\n<li>\u89e3\u51b3\u601d\u8def&#xff1a;RNN \u9760\u95e8\u63a7\u673a\u5236&#xff0c;CNN \u9760\u6b8b\u5dee\u8fde\u63a5 &#043; ReLU\u3002<\/li>\n<\/ul>\n<hr \/>\n<h3>\u516d\u3001Transformer \u4e0e\u6ce8\u610f\u529b\u673a\u5236&#xff08;\u9762\u8bd5\u9ad8\u9891\u91cd\u70b9&#xff09;<\/h3>\n<h4>1. \u81ea\u6ce8\u610f\u529b\u673a\u5236&#xff08;Self-Attention&#xff09;<\/h4>\n<h5>\u6838\u5fc3\u601d\u60f3<\/h5>\n<p>\u8ba1\u7b97\u5e8f\u5217\u4e2d\u6bcf\u4e2a\u4f4d\u7f6e\u4e0e\u6240\u6709\u4f4d\u7f6e\u7684\u76f8\u5173\u6027\u6743\u91cd&#xff0c;\u52a0\u6743\u6c42\u548c\u5f97\u5230\u5f53\u524d\u4f4d\u7f6e\u7684\u7279\u5f81&#xff0c;\u4e00\u6b65\u6355\u6349\u957f\u8ddd\u79bb\u4f9d\u8d56\u3002<\/p>\n<h5>Q\u3001K\u3001V \u542b\u4e49<\/h5>\n<ul>\n<li>Query&#xff08;\u67e5\u8be2&#xff09;&#xff1a;\u5f53\u524d\u4f4d\u7f6e\u7684\u7279\u5f81\u5411\u91cf&#xff0c;\u7528\u4e8e \u201c\u67e5\u8be2\u201d \u5176\u4ed6\u4f4d\u7f6e<\/li>\n<li>Key&#xff08;\u952e&#xff09;&#xff1a;\u5176\u4ed6\u4f4d\u7f6e\u7684\u7279\u5f81\u5411\u91cf&#xff0c;\u7528\u4e8e\u88ab\u67e5\u8be2\u5339\u914d<\/li>\n<li>Value&#xff08;\u503c&#xff09;&#xff1a;\u5176\u4ed6\u4f4d\u7f6e\u7684\u7279\u5f81\u5411\u91cf&#xff0c;\u7528\u4e8e\u52a0\u6743\u6c42\u548c<\/li>\n<li>\u4e09\u8005\u5747\u7531\u8f93\u5165\u7279\u5f81\u4e58\u4e0d\u540c\u7684\u53ef\u5b66\u4e60\u77e9\u9635\u5f97\u5230\u3002<\/li>\n<\/ul>\n<h5>\u8ba1\u7b97\u6b65\u9aa4<\/h5>\n<li>\u8f93\u5165\u5411\u91cf\u5206\u522b\u4e0e\\\\(W_Q, W_K, W_V\\\\)\u76f8\u4e58&#xff0c;\u5f97\u5230\\\\(Q, K, V\\\\)<\/li>\n<li>\u8ba1\u7b97Q\u4e0eK\u7684\u70b9\u79ef&#xff0c;\u9664\u4ee5\\\\(\\\\sqrt{d_k}\\\\)&#xff08;\u7f29\u653e\u56e0\u5b50&#xff0c;\u9632\u6b62\u70b9\u79ef\u8fc7\u5927\u5bfc\u81f4 softmax \u68af\u5ea6\u6d88\u5931&#xff09;<\/li>\n<li>\u7ecf\u8fc7 softmax \u5f97\u5230\u6ce8\u610f\u529b\u6743\u91cd<\/li>\n<li>\u6743\u91cd\u4e0eV\u52a0\u6743\u6c42\u548c&#xff0c;\u5f97\u5230\u6700\u7ec8\u8f93\u51fa<\/li>\n<h5>\u516c\u5f0f<\/h5>\n<p>\\\\(Attention(Q,K,V) &#061; softmax(\\\\frac{QK^T}{\\\\sqrt{d_k}})V\\\\)<\/p>\n<h4>2. \u591a\u5934\u6ce8\u610f\u529b&#xff08;Multi-Head Attention&#xff09;<\/h4>\n<ul>\n<li>\u505a\u6cd5&#xff1a;\u5c06\\\\(Q,K,V\\\\)\u62c6\u5206\u6210\u591a\u4e2a\u5934&#xff0c;\u6bcf\u4e2a\u5934\u72ec\u7acb\u505a\u81ea\u6ce8\u610f\u529b&#xff0c;\u6700\u540e\u62fc\u63a5\u7ed3\u679c\u3002<\/li>\n<li>\u4f18\u52bf&#xff1a;\n<li>\u6355\u6349\u4e0d\u540c\u5b50\u7a7a\u95f4\u7684\u7279\u5f81\u4fe1\u606f&#xff0c;\u8868\u8fbe\u80fd\u529b\u66f4\u5f3a<\/li>\n<li>\u591a\u4e2a\u5934\u5e76\u884c\u8ba1\u7b97&#xff0c;\u6548\u7387\u9ad8<\/li>\n<\/li>\n<\/ul>\n<h4>3. Transformer \u6574\u4f53\u7ed3\u6784<\/h4>\n<ul>\n<li>Encoder&#xff08;\u7f16\u7801\u5668&#xff09;&#xff1a;\u5806\u53e0 N \u5c42&#xff0c;\u6bcf\u5c42\u5305\u542b&#xff1a;\n<li>\u591a\u5934\u81ea\u6ce8\u610f\u529b &#043; \u6b8b\u5dee\u8fde\u63a5 &#043; LayerNorm<\/li>\n<li>\u524d\u9988\u7f51\u7edc&#xff08;FFN&#xff0c;\u4e24\u5c42\u5168\u8fde\u63a5 &#043; \u6fc0\u6d3b&#xff09; &#043; \u6b8b\u5dee\u8fde\u63a5 &#043; LayerNorm<\/li>\n<\/li>\n<li>Decoder&#xff08;\u89e3\u7801\u5668&#xff09;&#xff1a;\u5806\u53e0 N \u5c42&#xff0c;\u6bcf\u5c42\u5305\u542b&#xff1a;\n<li>\u63a9\u7801\u591a\u5934\u81ea\u6ce8\u610f\u529b&#xff08;Masked Self-Attention&#xff0c;\u9632\u6b62\u770b\u5230\u672a\u6765\u4fe1\u606f&#xff09;<\/li>\n<li>\u4ea4\u53c9\u6ce8\u610f\u529b&#xff08;Encoder-Decoder Attention&#xff0c;Q \u6765\u81ea\u89e3\u7801\u5668&#xff0c;K\/V \u6765\u81ea\u7f16\u7801\u5668&#xff09;<\/li>\n<li>\u524d\u9988\u7f51\u7edc &#043; \u6b8b\u5dee &#043; LayerNorm<\/li>\n<\/li>\n<\/ul>\n<h4>4. \u4f4d\u7f6e\u7f16\u7801&#xff08;Positional Encoding&#xff09;<\/h4>\n<ul>\n<li>\u95ee\u9898&#xff1a;\u81ea\u6ce8\u610f\u529b\u672c\u8eab\u65e0\u5e8f&#xff0c;\u65e0\u6cd5\u6355\u6349\u5e8f\u5217\u987a\u5e8f\u4fe1\u606f\u3002<\/li>\n<li>\u89e3\u51b3\u65b9\u6848&#xff1a;\u7ed9\u8f93\u5165\u5d4c\u5165\u52a0\u4e0a\u4f4d\u7f6e\u7f16\u7801&#xff0c;\u6ce8\u5165\u4f4d\u7f6e\u4fe1\u606f\u3002<\/li>\n<li>\u5e38\u89c1\u5b9e\u73b0&#xff1a;\n<li>\u6b63\u5f26\u4f59\u5f26\u4f4d\u7f6e\u7f16\u7801&#xff1a;\u4e0d\u540c\u9891\u7387\u7684\u6b63\u4f59\u5f26\u51fd\u6570&#xff0c;\u53ef\u5916\u63a8\u5230\u66f4\u957f\u5e8f\u5217<\/li>\n<li>\u53ef\u5b66\u4e60\u4f4d\u7f6e\u7f16\u7801&#xff1a;\u76f4\u63a5\u8bad\u7ec3\u4f4d\u7f6e\u5d4c\u5165\u5411\u91cf&#xff0c;\u7b80\u5355\u6709\u6548&#xff0c;ViT\u3001BERT \u5e38\u7528<\/li>\n<\/li>\n<\/ul>\n<h4>5. Layer Normalization vs Batch Normalization<\/h4>\n<p>\u8868\u683c<\/p>\n<table>\n<tr>\u7ef4\u5ea6LayerNormBatchNorm<\/tr>\n<tbody>\n<tr>\n<td>\u5f52\u4e00\u5316\u7ef4\u5ea6<\/td>\n<td>\u5bf9\u6bcf\u4e2a\u6837\u672c\u7684\u6240\u6709\u7279\u5f81\u505a\u5f52\u4e00\u5316<\/td>\n<td>\u5bf9\u4e00\u4e2a\u6279\u6b21\u5185\u540c\u4e00\u7ef4\u5ea6\u7684\u7279\u5f81\u505a\u5f52\u4e00\u5316<\/td>\n<\/tr>\n<tr>\n<td>\u4f9d\u8d56\u6279\u6b21<\/td>\n<td>\u4e0d\u4f9d\u8d56&#xff0c;\u5355\u6837\u672c\u5373\u53ef\u8ba1\u7b97<\/td>\n<td>\u4f9d\u8d56\u6279\u6b21\u5927\u5c0f&#xff0c;\u5c0f\u6279\u6b21\u6548\u679c\u5dee<\/td>\n<\/tr>\n<tr>\n<td>\u9002\u7528\u573a\u666f<\/td>\n<td>NLP\u3001Transformer\u3001\u53d8\u957f\u5e8f\u5217<\/td>\n<td>CNN\u3001\u56fa\u5b9a\u7ef4\u5ea6\u8f93\u5165<\/td>\n<\/tr>\n<tr>\n<td>\u8bad\u7ec3 \/ \u6d4b\u8bd5\u5dee\u5f02<\/td>\n<td>\u65e0\u5dee\u5f02&#xff0c;\u63a8\u7406\u76f4\u63a5\u7528<\/td>\n<td>\u8bad\u7ec3\u7528\u6279\u6b21\u5747\u503c\u65b9\u5dee&#xff0c;\u6d4b\u8bd5\u7528\u79fb\u52a8\u5e73\u5747<\/td>\n<\/tr>\n<tr>\n<td>\u7ef4\u5ea6\u4f4d\u7f6e<\/td>\n<td>\u901a\u5e38\u5728\u6ce8\u610f\u529b \/ FFN \u4e4b\u524d&#xff08;Pre-Norm&#xff09;<\/td>\n<td>\u901a\u5e38\u5728\u5377\u79ef \/ \u6fc0\u6d3b\u4e4b\u540e<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<h4>6. Transformer \u4e3a\u4ec0\u4e48\u7528 LayerNorm \u800c\u4e0d\u662f BatchNorm&#xff1f;<\/h4>\n<li>\u5e8f\u5217\u957f\u5ea6\u4e0d\u56fa\u5b9a&#xff0c;\u4e0d\u540c\u6837\u672c\u957f\u5ea6\u4e0d\u540c&#xff0c;BN \u5728\u5e8f\u5217\u7ef4\u5ea6\u5f52\u4e00\u5316\u65e0\u610f\u4e49<\/li>\n<li>NLP \u6279\u6b21\u901a\u5e38\u8f83\u5c0f&#xff0c;BN \u7edf\u8ba1\u91cf\u4e0d\u7a33\u5b9a<\/li>\n<li>LN \u5bf9\u6bcf\u4e2a\u6837\u672c\u72ec\u7acb\u5f52\u4e00\u5316&#xff0c;\u4e0d\u53d7\u6279\u6b21\u548c\u5e8f\u5217\u957f\u5ea6\u5f71\u54cd&#xff0c;\u66f4\u7a33\u5b9a<\/li>\n<h4>7. \u6b8b\u5dee\u8fde\u63a5&#xff08;Residual Connection&#xff09;<\/h4>\n<ul>\n<li>\u516c\u5f0f&#xff1a;\\\\(x &#043; F(x)\\\\)<\/li>\n<li>\u4f5c\u7528&#xff1a;\n<li>\u7f13\u89e3\u68af\u5ea6\u6d88\u5931&#xff0c;\u68af\u5ea6\u53ef\u901a\u8fc7\u6052\u7b49\u8def\u5f84\u76f4\u63a5\u56de\u4f20<\/li>\n<li>\u8ba9\u6df1\u5c42\u7f51\u7edc\u66f4\u5bb9\u6613\u8bad\u7ec3<\/li>\n<li>\u4fdd\u7559\u539f\u59cb\u4fe1\u606f&#xff0c;\u4fc3\u8fdb\u7279\u5f81\u590d\u7528<\/li>\n<\/li>\n<\/ul>\n<h4>8. Transformer \u76f8\u6bd4 RNN \u7684\u4f18\u52bf<\/h4>\n<li>\u5e76\u884c\u8ba1\u7b97&#xff1a;\u81ea\u6ce8\u610f\u529b\u53ef\u4e00\u6b65\u8ba1\u7b97\u6240\u6709\u4f4d\u7f6e&#xff0c;RNN \u5fc5\u987b\u4e32\u884c<\/li>\n<li>\u957f\u8ddd\u79bb\u4f9d\u8d56&#xff1a;\u81ea\u6ce8\u610f\u529b\u4e00\u6b65\u6355\u6349\u5168\u5c40\u4f9d\u8d56&#xff0c;RNN \u968f\u8ddd\u79bb\u589e\u957f\u8870\u51cf<\/li>\n<li>\u6a21\u578b\u80fd\u529b\u66f4\u5f3a&#xff0c;\u652f\u6301\u5927\u89c4\u6a21\u9884\u8bad\u7ec3<\/li>\n<ul>\n<li>\u7f3a\u70b9&#xff1a;\u8ba1\u7b97\u590d\u6742\u5ea6\u4e3a\\\\(O(n^2)\\\\)&#xff0c;\u957f\u5e8f\u5217\u4e0b\u8ba1\u7b97\u91cf\u7206\u70b8<\/li>\n<\/ul>\n<h4>9. Self-Attention \u7684\u590d\u6742\u5ea6<\/h4>\n<ul>\n<li>\u65f6\u95f4 \/ \u7a7a\u95f4\u590d\u6742\u5ea6&#xff1a;\\\\(O(n^2 \\\\cdot d)\\\\)&#xff0c;n\u4e3a\u5e8f\u5217\u957f\u5ea6&#xff0c;d\u4e3a\u7279\u5f81\u7ef4\u5ea6<\/li>\n<li>\u74f6\u9888&#xff1a;\u5e8f\u5217\u957f\u5ea6 n \u5e73\u65b9\u589e\u957f&#xff0c;\u957f\u5e8f\u5217\u573a\u666f&#xff08;\u5982\u957f\u6587\u672c\u3001\u9ad8\u5206\u8fa8\u7387\u56fe\u50cf&#xff09;\u5f00\u9500\u6781\u5927<\/li>\n<\/ul>\n<h4>10. \u63a9\u7801\u6ce8\u610f\u529b&#xff08;Masked Attention&#xff09;<\/h4>\n<ul>\n<li>\u4f5c\u7528&#xff1a;\u5728\u89e3\u7801\u5668\u4e2d&#xff0c;\u9632\u6b62\u5f53\u524d\u4f4d\u7f6e\u770b\u5230\u672a\u6765\u65f6\u523b\u7684\u4fe1\u606f&#xff0c;\u4fdd\u8bc1\u81ea\u56de\u5f52\u751f\u6210\u7684\u5408\u7406\u6027\u3002<\/li>\n<li>\u5b9e\u73b0&#xff1a;\u8ba1\u7b97\u6ce8\u610f\u529b\u5206\u6570\u540e&#xff0c;\u5c06\u672a\u6765\u4f4d\u7f6e\u7684\u5206\u6570\u8bbe\u4e3a\u8d1f\u65e0\u7a77&#xff0c;softmax \u540e\u6743\u91cd\u8d8b\u8fd1\u4e8e 0\u3002<\/li>\n<\/ul>\n<h4>11. Flash Attention \u6838\u5fc3\u601d\u60f3<\/h4>\n<ul>\n<li>\u76ee\u6807&#xff1a;\u5728\u4fdd\u8bc1\u7cbe\u5ea6\u7684\u524d\u63d0\u4e0b&#xff0c;\u52a0\u901f\u6ce8\u610f\u529b\u8ba1\u7b97\u3001\u964d\u4f4e\u663e\u5b58\u5360\u7528\u3002<\/li>\n<li>\u6838\u5fc3\u4f18\u5316&#xff1a;\n<li>\u5206\u5757\u8ba1\u7b97&#xff1a;\u5c06 Q\/K\/V \u5206\u5757&#xff0c;\u9010\u5757\u8ba1\u7b97\u6ce8\u610f\u529b&#xff0c;\u907f\u514d\u4e00\u6b21\u6027\u52a0\u8f7d\u5168\u90e8\u6570\u636e\u5230\u663e\u5b58<\/li>\n<li>\u7b97\u5b50\u878d\u5408&#xff1a;\u5c06 softmax\u3001dropout\u3001\u52a0\u6743\u6c42\u548c\u7b49\u64cd\u4f5c\u878d\u5408&#xff0c;\u51cf\u5c11\u663e\u5b58\u8bfb\u5199<\/li>\n<li>\u5229\u7528 SRAM \u9ad8\u901f\u7f13\u5b58&#xff0c;\u63d0\u5347\u8ba1\u7b97\u6548\u7387<\/li>\n<\/li>\n<li>\u73b0\u72b6&#xff1a;\u5927\u6a21\u578b\u8bad\u7ec3\u63a8\u7406\u7684\u6807\u914d\u6280\u672f\u3002<\/li>\n<\/ul>\n<hr \/>\n<h3>\u4e03\u3001\u4f18\u5316\u7b97\u6cd5\u4e0e\u4f18\u5316\u5668<\/h3>\n<h4>1. \u68af\u5ea6\u4e0b\u964d\u6cd5\u4e09\u5927\u7c7b<\/h4>\n<h5>\u6279\u91cf\u68af\u5ea6\u4e0b\u964d&#xff08;BGD&#xff09;<\/h5>\n<ul>\n<li>\u6bcf\u6b21\u7528\u5168\u90e8\u6570\u636e\u8ba1\u7b97\u68af\u5ea6&#xff0c;\u66f4\u65b0\u4e00\u6b21\u53c2\u6570<\/li>\n<li>\u4f18\u70b9&#xff1a;\u6536\u655b\u7a33\u5b9a&#xff0c;\u5168\u5c40\u6700\u4f18\u65b9\u5411\u51c6\u786e<\/li>\n<li>\u7f3a\u70b9&#xff1a;\u6570\u636e\u91cf\u5927\u65f6\u6781\u6162&#xff0c;\u65e0\u6cd5\u5728\u7ebf\u66f4\u65b0<\/li>\n<\/ul>\n<h5>\u968f\u673a\u68af\u5ea6\u4e0b\u964d&#xff08;SGD&#xff09;<\/h5>\n<ul>\n<li>\u6bcf\u6b21\u7528 1 \u4e2a\u6837\u672c\u8ba1\u7b97\u68af\u5ea6&#xff0c;\u66f4\u65b0\u53c2\u6570<\/li>\n<li>\u4f18\u70b9&#xff1a;\u901f\u5ea6\u5feb&#xff0c;\u652f\u6301\u5728\u7ebf\u5b66\u4e60<\/li>\n<li>\u7f3a\u70b9&#xff1a;\u66f4\u65b0\u65b9\u5411\u9707\u8361&#xff0c;\u6536\u655b\u4e0d\u7a33\u5b9a<\/li>\n<\/ul>\n<h5>\u5c0f\u6279\u91cf\u68af\u5ea6\u4e0b\u964d&#xff08;Mini-batch SGD&#xff09;<\/h5>\n<ul>\n<li>\u6bcf\u6b21\u7528\u4e00\u5c0f\u6279\u6837\u672c\u8ba1\u7b97\u68af\u5ea6&#xff0c;\u517c\u987e\u901f\u5ea6\u4e0e\u7a33\u5b9a\u6027<\/li>\n<li>\u6df1\u5ea6\u5b66\u4e60\u4e3b\u6d41\u8bad\u7ec3\u65b9\u5f0f<\/li>\n<\/ul>\n<h4>2. \u5e26\u52a8\u91cf\u7684 SGD&#xff08;SGD with Momentum&#xff09;<\/h4>\n<ul>\n<li>\u601d\u60f3&#xff1a;\u5f15\u5165\u52a8\u91cf\u9879&#xff0c;\u7d2f\u79ef\u5386\u53f2\u68af\u5ea6&#xff0c;\u5e73\u6ed1\u66f4\u65b0\u65b9\u5411<\/li>\n<li>\u516c\u5f0f&#xff1a;\\\\(v_t &#061; \\\\beta v_{t-1} &#043; (1-\\\\beta)g_t\\\\)&#xff0c;\\\\(\\\\theta_t &#061; \\\\theta_{t-1} &#8211; \\\\alpha v_t\\\\)<\/li>\n<li>\u4f18\u70b9&#xff1a;\n<li>\u52a0\u901f\u6536\u655b&#xff0c;\u51cf\u5c11\u9707\u8361<\/li>\n<li>\u4e00\u5b9a\u7a0b\u5ea6\u4e0a\u51b2\u51fa\u5c40\u90e8\u6700\u4f18<\/li>\n<\/li>\n<\/ul>\n<h4>3. NAG&#xff08;Nesterov Accelerated Gradient&#xff09;<\/h4>\n<ul>\n<li>\u6539\u8fdb&#xff1a;\u5148\u6309\u52a8\u91cf\u65b9\u5411\u8d70\u4e00\u6b65&#xff0c;\u518d\u8ba1\u7b97\u8be5\u4f4d\u7f6e\u7684\u68af\u5ea6&#xff0c;\u76f8\u5f53\u4e8e \u201c\u63d0\u524d\u9884\u5224\u201d<\/li>\n<li>\u6536\u655b\u901f\u5ea6\u901a\u5e38\u4f18\u4e8e\u666e\u901a\u52a8\u91cf SGD<\/li>\n<\/ul>\n<h4>4. AdaGrad<\/h4>\n<ul>\n<li>\u601d\u60f3&#xff1a;\u81ea\u9002\u5e94\u5b66\u4e60\u7387&#xff0c;\u9891\u7e41\u66f4\u65b0\u7684\u53c2\u6570\u5b66\u4e60\u7387\u66f4\u5c0f&#xff0c;\u7a00\u758f\u53c2\u6570\u5b66\u4e60\u7387\u66f4\u5927<\/li>\n<li>\u7f3a\u70b9&#xff1a;\u5b66\u4e60\u7387\u5355\u8c03\u9012\u51cf&#xff0c;\u8bad\u7ec3\u540e\u671f\u5b66\u4e60\u7387\u6781\u5c0f&#xff0c;\u6a21\u578b\u505c\u6b62\u5b66\u4e60<\/li>\n<\/ul>\n<h4>5. RMSProp<\/h4>\n<ul>\n<li>\u6539\u8fdb&#xff1a;\u5bf9\u5386\u53f2\u68af\u5ea6\u5e73\u65b9\u505a\u6307\u6570\u79fb\u52a8\u5e73\u5747&#xff0c;\u66ff\u4ee3 AdaGrad \u7684\u7d2f\u52a0&#xff0c;\u89e3\u51b3\u5b66\u4e60\u7387\u6301\u7eed\u4e0b\u964d\u95ee\u9898<\/li>\n<li>\u516c\u5f0f&#xff1a;\\\\(s_t &#061; \\\\beta s_{t-1} &#043; (1-\\\\beta)g_t^2\\\\)&#xff0c;\\\\(\\\\theta_t &#061; \\\\theta_{t-1} &#8211; \\\\alpha \\\\frac{g_t}{\\\\sqrt{s_t}&#043;\\\\epsilon}\\\\)<\/li>\n<\/ul>\n<h4>6. Adam&#xff08;Adaptive Moment Estimation&#xff09;<\/h4>\n<ul>\n<li>\u672c\u8d28&#xff1a;\u52a8\u91cf &#043; RMSProp \u7684\u7ed3\u5408<\/li>\n<li>\u7ef4\u62a4\u4e24\u4e2a\u4e00\u9636\u77e9&#xff08;\u52a8\u91cf&#xff09;\u548c\u4e8c\u9636\u77e9&#xff08;\u68af\u5ea6\u5e73\u65b9&#xff09;&#xff0c;\u5e76\u505a\u504f\u5dee\u4fee\u6b63<\/li>\n<li>\u4f18\u70b9&#xff1a;\u81ea\u9002\u5e94\u5b66\u4e60\u7387&#xff0c;\u6536\u655b\u5feb&#xff0c;\u5bf9\u8d85\u53c2\u6570\u4e0d\u654f\u611f&#xff0c;\u901a\u7528\u6027\u5f3a<\/li>\n<li>\u7f3a\u70b9&#xff1a;\n<li>\u540e\u671f\u6613\u9707\u8361&#xff0c;\u6cdb\u5316\u80fd\u529b\u6709\u65f6\u4e0d\u5982 SGD&#043;Momentum<\/li>\n<li>\u53ef\u80fd\u5b58\u5728\u81ea\u9002\u5e94\u5b66\u4e60\u7387\u8fc7\u5927\u5bfc\u81f4\u7684\u6cdb\u5316\u95ee\u9898<\/li>\n<\/li>\n<\/ul>\n<h4>7. AdamW<\/h4>\n<ul>\n<li>\u6539\u8fdb&#xff1a;\u5c06\u6743\u91cd\u8870\u51cf\u4ece\u68af\u5ea6\u4e2d\u89e3\u8026&#xff0c;\u76f4\u63a5\u5bf9\u6743\u91cd\u505a\u8870\u51cf&#xff0c;\u800c\u975e\u52a0\u5230\u68af\u5ea6\u91cc<\/li>\n<li>\u4f18\u52bf&#xff1a;\u4fee\u6b63\u4e86 Adam \u4e2d L2 \u6b63\u5219\u5b9e\u73b0\u4e0d\u5408\u7406\u7684\u95ee\u9898&#xff0c;\u6cdb\u5316\u80fd\u529b\u66f4\u5f3a<\/li>\n<li>\u73b0\u72b6&#xff1a;Transformer\u3001\u5927\u6a21\u578b\u7684\u9ed8\u8ba4\u4f18\u5316\u5668<\/li>\n<\/ul>\n<h4>8. SGD vs Adam \u600e\u4e48\u9009&#xff1f;<\/h4>\n<p>\u8868\u683c<\/p>\n<table>\n<tr>\u573a\u666f\u63a8\u8350\u4f18\u5316\u5668<\/tr>\n<tbody>\n<tr>\n<td>\u8ffd\u6c42\u6781\u81f4\u6cdb\u5316\u3001\u7b97\u529b\u5145\u8db3<\/td>\n<td>SGD &#043; Momentum<\/td>\n<\/tr>\n<tr>\n<td>\u5feb\u901f\u8fed\u4ee3\u3001\u6570\u636e\u91cf\u5927\u3001\u8c03\u53c2\u5c11<\/td>\n<td>Adam \/ AdamW<\/td>\n<\/tr>\n<tr>\n<td>Transformer\u3001\u5927\u6a21\u578b<\/td>\n<td>AdamW<\/td>\n<\/tr>\n<tr>\n<td>\u7a00\u758f\u6570\u636e\u3001NLP \u4efb\u52a1<\/td>\n<td>Adam \/ AdamW<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<h4>9. \u5b66\u4e60\u7387\u8c03\u5ea6\u5668&#xff08;Learning Rate Scheduler&#xff09;<\/h4>\n<ul>\n<li>\u4f5c\u7528&#xff1a;\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u52a8\u6001\u8c03\u6574\u5b66\u4e60\u7387&#xff0c;\u524d\u671f\u5927\u6b65\u5feb\u8d70&#xff0c;\u540e\u671f\u5c0f\u6b65\u6536\u655b<\/li>\n<li>\u5e38\u89c1\u7c7b\u578b&#xff1a;\n<li>StepLR&#xff1a;\u56fa\u5b9a\u6b65\u957f\u6309\u6bd4\u4f8b\u8870\u51cf<\/li>\n<li>CosineAnnealingLR&#xff1a;\u4f59\u5f26\u9000\u706b&#xff0c;\u5468\u671f\u6027\u5347\u964d\u5b66\u4e60\u7387&#xff0c;\u8df3\u51fa\u5c40\u90e8\u6700\u4f18<\/li>\n<li>ReduceLROnPlateau&#xff1a;\u76d1\u63a7\u6307\u6807\u4e0d\u518d\u4e0b\u964d\u65f6\u964d\u4f4e\u5b66\u4e60\u7387<\/li>\n<li>Warmup&#xff1a;\u8bad\u7ec3\u521d\u671f\u5b66\u4e60\u7387\u4ece 0 \u7ebf\u6027\u589e\u957f\u5230\u521d\u59cb\u503c&#xff0c;\u907f\u514d\u521d\u671f\u6a21\u578b\u4e0d\u7a33\u5b9a<\/li>\n<\/li>\n<\/ul>\n<hr \/>\n<h3>\u516b\u3001\u6b63\u5219\u5316\u4e0e\u6cdb\u5316\u80fd\u529b<\/h3>\n<h4>1. \u8fc7\u62df\u5408 vs \u6b20\u62df\u5408<\/h4>\n<p>\u8868\u683c<\/p>\n<table>\n<tr>\u73b0\u8c61\u8868\u73b0\u539f\u56e0<\/tr>\n<tbody>\n<tr>\n<td>\u6b20\u62df\u5408<\/td>\n<td>\u8bad\u7ec3\u96c6\u3001\u9a8c\u8bc1\u96c6\u6548\u679c\u90fd\u5dee<\/td>\n<td>\u6a21\u578b\u590d\u6742\u5ea6\u4e0d\u8db3\u3001\u7279\u5f81\u592a\u5c11\u3001\u8bad\u7ec3\u4e0d\u8db3<\/td>\n<\/tr>\n<tr>\n<td>\u8fc7\u62df\u5408<\/td>\n<td>\u8bad\u7ec3\u96c6\u6548\u679c\u6781\u597d&#xff0c;\u9a8c\u8bc1\u96c6 \/ \u6d4b\u8bd5\u96c6\u6548\u679c\u5dee<\/td>\n<td>\u6a21\u578b\u592a\u590d\u6742\u3001\u6570\u636e\u592a\u5c11\u3001\u8bad\u7ec3\u8fc7\u5ea6<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<h4>2. \u89e3\u51b3\u6b20\u62df\u5408<\/h4>\n<li>\u589e\u52a0\u6a21\u578b\u590d\u6742\u5ea6&#xff08;\u52a0\u6df1\u52a0\u5bbd\u7f51\u7edc&#xff09;<\/li>\n<li>\u589e\u52a0\u7279\u5f81\u7ef4\u5ea6<\/li>\n<li>\u51cf\u5c11\u6b63\u5219\u5316\u5f3a\u5ea6<\/li>\n<li>\u5ef6\u957f\u8bad\u7ec3\u65f6\u95f4<\/li>\n<h4>3. \u89e3\u51b3\u8fc7\u62df\u5408&#xff08;\u6838\u5fc3\u6b63\u5219\u5316\u624b\u6bb5&#xff09;<\/h4>\n<h5>&#xff08;1&#xff09;\u6570\u636e\u5c42\u9762<\/h5>\n<ul>\n<li>\u589e\u52a0\u6570\u636e\u91cf<\/li>\n<li>\u6570\u636e\u589e\u5f3a&#xff08;\u56fe\u50cf&#xff1a;\u7ffb\u8f6c\u3001\u88c1\u526a\u3001\u65cb\u8f6c\u3001\u8272\u5f69\u6296\u52a8&#xff1b;\u6587\u672c&#xff1a;\u540c\u4e49\u8bcd\u66ff\u6362\u3001\u56de\u8bd1\u3001\u63a9\u7801&#xff09;<\/li>\n<\/ul>\n<h5>&#xff08;2&#xff09;\u6a21\u578b\u5c42\u9762<\/h5>\n<ul>\n<li>L1\/L2 \u6b63\u5219\u5316&#xff08;\u6743\u91cd\u8870\u51cf&#xff09;<\/li>\n<li>Dropout \/ DropConnect<\/li>\n<li>\u65e9\u505c&#xff08;Early Stopping&#xff09;<\/li>\n<li>\u6279\u91cf\u5f52\u4e00\u5316 \/ \u5c42\u5f52\u4e00\u5316<\/li>\n<li>\u6807\u7b7e\u5e73\u6ed1&#xff08;Label Smoothing&#xff09;<\/li>\n<li>\u964d\u4f4e\u6a21\u578b\u590d\u6742\u5ea6<\/li>\n<\/ul>\n<h5>&#xff08;3&#xff09;\u8bad\u7ec3\u5c42\u9762<\/h5>\n<ul>\n<li>\u4ea4\u53c9\u9a8c\u8bc1<\/li>\n<li>\u591a\u6a21\u578b\u96c6\u6210<\/li>\n<\/ul>\n<h4>4. L1 vs L2 \u6b63\u5219\u5316<\/h4>\n<p>\u8868\u683c<\/p>\n<table>\n<tr>\u7ef4\u5ea6L1 \u6b63\u5219&#xff08;Lasso&#xff09;L2 \u6b63\u5219&#xff08;Ridge&#xff09;<\/tr>\n<tbody>\n<tr>\n<td>\u5f62\u5f0f<\/td>\n<td>\u6743\u91cd\u7edd\u5bf9\u503c\u4e4b\u548c<\/td>\n<td>\u6743\u91cd\u5e73\u65b9\u548c<\/td>\n<\/tr>\n<tr>\n<td>\u6548\u679c<\/td>\n<td>\u4ea7\u751f\u7a00\u758f\u89e3&#xff0c;\u53ef\u7528\u4e8e\u7279\u5f81\u9009\u62e9<\/td>\n<td>\u6743\u91cd\u6574\u4f53\u53d8\u5c0f&#xff0c;\u66f4\u7a33\u5b9a<\/td>\n<\/tr>\n<tr>\n<td>\u5bfc\u6570<\/td>\n<td>\u4e0d\u53ef\u5bfc&#xff08;\u6b21\u68af\u5ea6&#xff09;<\/td>\n<td>\u5904\u5904\u53ef\u5bfc<\/td>\n<\/tr>\n<tr>\n<td>\u6297\u5f02\u5e38\u503c<\/td>\n<td>\u8f83\u5f3a<\/td>\n<td>\u8f83\u5f31<\/td>\n<\/tr>\n<tr>\n<td>\u573a\u666f<\/td>\n<td>\u7279\u5f81\u9009\u62e9\u3001\u7a00\u758f\u5316<\/td>\n<td>\u901a\u7528\u6743\u91cd\u8870\u51cf<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<h4>5. Dropout<\/h4>\n<ul>\n<li>\u539f\u7406&#xff1a;\u8bad\u7ec3\u65f6\u968f\u673a\u8ba9\u4e00\u90e8\u5206\u795e\u7ecf\u5143\u5931\u6d3b&#xff08;\u8f93\u51fa\u7f6e 0&#xff09;&#xff0c;\u6d4b\u8bd5\u65f6\u5168\u90e8\u795e\u7ecf\u5143\u542f\u7528&#xff0c;\u6743\u91cd\u7f29\u653e<\/li>\n<li>\u4e3a\u4ec0\u4e48\u80fd\u9632\u6b62\u8fc7\u62df\u5408&#xff1f;\n<li>\u51cf\u5c11\u795e\u7ecf\u5143\u4e4b\u95f4\u7684\u5171\u9002\u5e94&#xff0c;\u5f3a\u8feb\u7f51\u7edc\u5b66\u4e60\u66f4\u9c81\u68d2\u7684\u7279\u5f81<\/li>\n<li>\u76f8\u5f53\u4e8e\u96c6\u6210\u4e86\u5927\u91cf\u4e0d\u540c\u7ed3\u6784\u7684\u5b50\u7f51\u7edc<\/li>\n<\/li>\n<li>\u6ce8\u610f&#xff1a;\u6d4b\u8bd5\u65f6\u4e0d\u542f\u7528&#xff0c;\u9700\u5bf9\u6743\u91cd\u4e58\u4ee5\u4fdd\u7559\u6982\u7387&#xff0c;\u4fdd\u8bc1\u8f93\u51fa\u5c3a\u5ea6\u4e00\u81f4<\/li>\n<\/ul>\n<h4>6. DropConnect<\/h4>\n<ul>\n<li>\u6539\u8fdb&#xff1a;\u968f\u673a\u8ba9\u6743\u91cd\u8fde\u63a5\u5931\u6d3b&#xff0c;\u800c\u975e\u795e\u7ecf\u5143\u5931\u6d3b<\/li>\n<li>\u6b63\u5219\u6548\u679c\u66f4\u5f3a&#xff0c;\u4f46\u8ba1\u7b97\u66f4\u590d\u6742&#xff0c;\u4f7f\u7528\u8f83\u5c11<\/li>\n<\/ul>\n<h4>7. \u65e9\u505c&#xff08;Early Stopping&#xff09;<\/h4>\n<ul>\n<li>\u539f\u7406&#xff1a;\u8bad\u7ec3\u4e2d\u76d1\u63a7\u9a8c\u8bc1\u96c6\u6307\u6807&#xff0c;\u6307\u6807\u6301\u7eed\u4e0d\u63d0\u5347\u65f6\u505c\u6b62\u8bad\u7ec3<\/li>\n<li>\u6700\u7b80\u5355\u6709\u6548\u7684\u6b63\u5219\u5316\u624b\u6bb5&#xff0c;\u51e0\u4e4e\u6240\u6709\u8bad\u7ec3\u4efb\u52a1\u90fd\u4f1a\u4f7f\u7528<\/li>\n<\/ul>\n<h4>8. \u6807\u7b7e\u5e73\u6ed1&#xff08;Label Smoothing&#xff09;<\/h4>\n<ul>\n<li>\u539f\u7406&#xff1a;\u5c06\u786c\u6807\u7b7e&#xff08;0\/1&#xff09;\u8f6f\u5316&#xff0c;\u5982\u5c06 1 \u6539\u4e3a 0.9&#xff0c;0 \u6539\u4e3a 0.1<\/li>\n<li>\u4f5c\u7528&#xff1a;\u964d\u4f4e\u6a21\u578b\u5bf9\u6807\u7b7e\u7684\u7f6e\u4fe1\u5ea6&#xff0c;\u9632\u6b62\u8fc7\u5ea6\u81ea\u4fe1&#xff0c;\u63d0\u5347\u6cdb\u5316\u80fd\u529b<\/li>\n<li>\u5e38\u7528\u4e8e\u5206\u7c7b\u4efb\u52a1\u548c\u5927\u6a21\u578b\u8bad\u7ec3<\/li>\n<\/ul>\n<h4>9. BatchNorm \u4e3a\u4ec0\u4e48\u6709\u6b63\u5219\u6548\u679c&#xff1f;<\/h4>\n<ul>\n<li>\u8bad\u7ec3\u65f6\u6279\u6b21\u5185\u7684\u5747\u503c\u65b9\u5dee\u5f15\u5165\u4e86\u566a\u58f0&#xff0c;\u76f8\u5f53\u4e8e\u7ed9\u7279\u5f81\u6dfb\u52a0\u4e86\u6270\u52a8<\/li>\n<li>\u8f7b\u5fae\u7684\u6b63\u5219\u6548\u679c&#xff0c;\u53ef\u4e00\u5b9a\u7a0b\u5ea6\u964d\u4f4e\u8fc7\u62df\u5408<\/li>\n<\/ul>\n<hr \/>\n<h3>\u4e5d\u3001\u635f\u5931\u51fd\u6570\u5168\u89e3\u6790<\/h3>\n<h4>1. \u5206\u7c7b\u4efb\u52a1\u635f\u5931<\/h4>\n<h5>\u4ea4\u53c9\u71b5\u635f\u5931&#xff08;Cross Entropy Loss&#xff09;<\/h5>\n<ul>\n<li>\u4e8c\u5206\u7c7b&#xff1a;\\\\(Loss &#061; -[y\\\\log\\\\hat{y} &#043; (1-y)\\\\log(1-\\\\hat{y})]\\\\)<\/li>\n<li>\u591a\u5206\u7c7b&#xff1a;\\\\(Loss &#061; -\\\\sum_{i&#061;1}^C y_i \\\\log\\\\hat{y}_i\\\\)<\/li>\n<li>\u672c\u8d28&#xff1a;\u8861\u91cf\u4e24\u4e2a\u6982\u7387\u5206\u5e03\u7684\u5dee\u5f02&#xff0c;\u6700\u5c0f\u5316\u4ea4\u53c9\u71b5\u7b49\u4ef7\u4e8e\u6700\u5927\u5316\u4f3c\u7136<\/li>\n<li>\u642d\u914d&#xff1a;\u8f93\u51fa\u5c42\u7528 Softmax&#xff08;\u591a\u5206\u7c7b&#xff09;\u6216 Sigmoid&#xff08;\u4e8c\u5206\u7c7b&#xff09;<\/li>\n<\/ul>\n<h5>Focal Loss<\/h5>\n<ul>\n<li>\u516c\u5f0f&#xff1a;\\\\(FL &#061; -\\\\alpha_t (1-p_t)^\\\\gamma \\\\log(p_t)\\\\)<\/li>\n<li>\u89e3\u51b3\u95ee\u9898&#xff1a;\u6b63\u8d1f\u6837\u672c\u4e0d\u5747\u8861\u3001\u96be\u6613\u6837\u672c\u4e0d\u5747\u8861<\/li>\n<li>\u6838\u5fc3&#xff1a;\u964d\u4f4e\u6613\u5206\u6837\u672c\u7684\u6743\u91cd&#xff0c;\u805a\u7126\u96be\u5206\u6837\u672c<\/li>\n<li>\u5e38\u7528\u573a\u666f&#xff1a;\u76ee\u6807\u68c0\u6d4b\u3001\u957f\u5c3e\u5206\u5e03\u5206\u7c7b<\/li>\n<\/ul>\n<h5>Dice Loss<\/h5>\n<ul>\n<li>\u516c\u5f0f&#xff1a;\\\\(DiceLoss &#061; 1 &#8211; \\\\frac{2|A\\\\cap B|}{|A|&#043;|B|}\\\\)<\/li>\n<li>\u672c\u8d28&#xff1a;\u8861\u91cf\u9884\u6d4b\u4e0e\u771f\u5b9e\u7684\u91cd\u53e0\u7a0b\u5ea6<\/li>\n<li>\u9002\u7528&#xff1a;\u56fe\u50cf\u5206\u5272\u3001\u6837\u672c\u6781\u5ea6\u4e0d\u5747\u8861\u7684\u4efb\u52a1<\/li>\n<li>\u7f3a\u70b9&#xff1a;\u8bad\u7ec3\u4e0d\u7a33\u5b9a&#xff0c;\u6613\u9707\u8361<\/li>\n<\/ul>\n<h4>2. \u56de\u5f52\u4efb\u52a1\u635f\u5931<\/h4>\n<h5>MSE&#xff08;\u5747\u65b9\u8bef\u5dee&#xff09;<\/h5>\n<ul>\n<li>\u516c\u5f0f&#xff1a;\\\\(MSE &#061; \\\\frac{1}{N}\\\\sum(y_i-\\\\hat{y}_i)^2\\\\)<\/li>\n<li>\u7279\u70b9&#xff1a;\u5bf9\u5927\u8bef\u5dee\u60e9\u7f5a\u91cd&#xff0c;\u5bf9\u5f02\u5e38\u503c\u654f\u611f<\/li>\n<\/ul>\n<h5>MAE&#xff08;\u5e73\u5747\u7edd\u5bf9\u8bef\u5dee&#xff09;<\/h5>\n<ul>\n<li>\u516c\u5f0f&#xff1a;\\\\(MAE &#061; \\\\frac{1}{N}\\\\sum|y_i-\\\\hat{y}_i|\\\\)<\/li>\n<li>\u7279\u70b9&#xff1a;\u5bf9\u5f02\u5e38\u503c\u9c81\u68d2&#xff0c;\u4f46\u5bfc\u6570\u4e0d\u8fde\u7eed&#xff0c;\u6536\u655b\u6162<\/li>\n<\/ul>\n<h5>Smooth L1 Loss<\/h5>\n<ul>\n<li>\u516c\u5f0f&#xff1a;\u8bef\u5dee\u5c0f\u65f6\u7528 L2&#xff0c;\u8bef\u5dee\u5927\u65f6\u7528 L1<\/li>\n<li>\u4f18\u52bf&#xff1a;\u517c\u5177 MSE \u7684\u7a33\u5b9a\u68af\u5ea6\u548c MAE \u7684\u6297\u5f02\u5e38\u503c\u80fd\u529b<\/li>\n<li>\u7ecf\u5178\u5e94\u7528&#xff1a;Faster R-CNN \u7684\u8fb9\u6846\u56de\u5f52<\/li>\n<\/ul>\n<h4>3. \u5bf9\u6bd4\u5b66\u4e60\u635f\u5931<\/h4>\n<h5>\u4e09\u5143\u7ec4\u635f\u5931&#xff08;Triplet Loss&#xff09;<\/h5>\n<ul>\n<li>\u7ec4\u6210&#xff1a;\u951a\u70b9\u6837\u672c\u3001\u6b63\u6837\u672c\u3001\u8d1f\u6837\u672c<\/li>\n<li>\u76ee\u6807&#xff1a;\u951a\u70b9\u4e0e\u6b63\u6837\u672c\u8ddd\u79bb &lt; \u951a\u70b9\u4e0e\u8d1f\u6837\u672c\u8ddd\u79bb&#xff08;\u95f4\u9694 margin&#xff09;<\/li>\n<li>\u5e94\u7528&#xff1a;\u4eba\u8138\u8bc6\u522b\u3001\u884c\u4eba\u91cd\u8bc6\u522b\u3001\u7279\u5f81\u68c0\u7d22<\/li>\n<\/ul>\n<h5>InfoNCE Loss<\/h5>\n<ul>\n<li>\u5bf9\u6bd4\u5b66\u4e60\u6838\u5fc3\u635f\u5931&#xff0c;\u5c06\u6b63\u6837\u672c\u5bf9\u4e0e\u8d1f\u6837\u672c\u5bf9\u505a\u5206\u7c7b<\/li>\n<li>\u5e94\u7528&#xff1a;SimCLR\u3001MoCo \u7b49\u81ea\u76d1\u7763\u5bf9\u6bd4\u5b66\u4e60<\/li>\n<\/ul>\n<hr \/>\n<h3>\u5341\u3001\u6a21\u578b\u8bad\u7ec3\u6838\u5fc3\u6280\u5de7\u4e0e\u5e38\u89c1\u95ee\u9898<\/h3>\n<h4>1. \u6743\u91cd\u521d\u59cb\u5316<\/h4>\n<h5>\u4e3a\u4ec0\u4e48\u4e0d\u80fd\u5168\u96f6\u521d\u59cb\u5316&#xff1f;<\/h5>\n<ul>\n<li>\u5168\u96f6\u521d\u59cb\u5316\u4f1a\u5bfc\u81f4\u540c\u4e00\u5c42\u6240\u6709\u795e\u7ecf\u5143\u8f93\u51fa\u76f8\u540c&#xff0c;\u53cd\u5411\u4f20\u64ad\u68af\u5ea6\u76f8\u540c&#xff0c;\u53c2\u6570\u66f4\u65b0\u5b8c\u5168\u4e00\u81f4&#xff0c;\u7f51\u7edc\u9000\u5316\u4e3a\u5355\u5c42\u3002<\/li>\n<\/ul>\n<h5>Xavier \u521d\u59cb\u5316&#xff08;Glorot \u521d\u59cb\u5316&#xff09;<\/h5>\n<ul>\n<li>\u601d\u60f3&#xff1a;\u4fdd\u6301\u524d\u5411\u4f20\u64ad\u548c\u53cd\u5411\u4f20\u64ad\u65f6&#xff0c;\u5404\u5c42\u8f93\u51fa\u7684\u65b9\u5dee\u4e00\u81f4<\/li>\n<li>\u9002\u7528&#xff1a;tanh\u3001sigmoid \u7b49\u5bf9\u79f0\u6fc0\u6d3b\u51fd\u6570<\/li>\n<\/ul>\n<h5>He \u521d\u59cb\u5316&#xff08;MSRA \u521d\u59cb\u5316&#xff09;<\/h5>\n<ul>\n<li>\u601d\u60f3&#xff1a;\u9488\u5bf9 ReLU \u8bbe\u8ba1&#xff0c;\u8003\u8651\u8d1f\u533a\u95f4\u5931\u6d3b\u5e26\u6765\u7684\u65b9\u5dee\u51cf\u534a<\/li>\n<li>\u9002\u7528&#xff1a;ReLU \u7cfb\u5217\u6fc0\u6d3b\u51fd\u6570&#xff0c;CNN \u4e2d\u6700\u5e38\u7528<\/li>\n<\/ul>\n<h4>2. \u6279\u6b21\u5927\u5c0f&#xff08;Batch Size&#xff09;\u5982\u4f55\u9009\u62e9&#xff1f;<\/h4>\n<ul>\n<li>\u592a\u5927&#xff1a;\u663e\u5b58\u5360\u7528\u9ad8&#xff1b;\u6536\u655b\u6162&#xff0c;\u5bb9\u6613\u9677\u5165\u978d\u70b9&#xff1b;\u6cdb\u5316\u80fd\u529b\u53ef\u80fd\u4e0b\u964d<\/li>\n<li>\u592a\u5c0f&#xff1a;\u68af\u5ea6\u4f30\u8ba1\u566a\u58f0\u5927&#xff0c;\u8bad\u7ec3\u9707\u8361&#xff1b;\u96be\u4ee5\u5229\u7528 BatchNorm<\/li>\n<li>\u7ecf\u9a8c&#xff1a;\u5728\u663e\u5b58\u5141\u8bb8\u8303\u56f4\u5185\u9009\u9002\u4e2d\u5927\u5c0f&#xff0c;\u642d\u914d\u5b66\u4e60\u7387\u7ebf\u6027\u7f29\u653e\u89c4\u5219&#xff08;batch size \u7ffb\u500d&#xff0c;\u5b66\u4e60\u7387\u7ffb\u500d&#xff09;<\/li>\n<\/ul>\n<h4>3. \u68af\u5ea6\u88c1\u526a&#xff08;Gradient Clipping&#xff09;<\/h4>\n<ul>\n<li>\u4f5c\u7528&#xff1a;\u9650\u5236\u68af\u5ea6\u8303\u6570&#xff0c;\u9632\u6b62\u68af\u5ea6\u7206\u70b8<\/li>\n<li>\u4e24\u79cd\u65b9\u5f0f&#xff1a;\n<li>\u6309\u503c\u88c1\u526a&#xff1a;\u9650\u5236\u68af\u5ea6\u6bcf\u4e2a\u5206\u91cf\u7684\u4e0a\u4e0b\u754c<\/li>\n<li>\u6309\u8303\u6570\u88c1\u526a&#xff1a;\u9650\u5236\u68af\u5ea6\u6574\u4f53 L2 \u8303\u6570&#xff0c;\u66f4\u5e38\u7528<\/li>\n<\/li>\n<li>\u5e38\u7528\u4e8e RNN\u3001\u751f\u6210\u5bf9\u6297\u7f51\u7edc\u7b49\u6613\u68af\u5ea6\u7206\u70b8\u7684\u573a\u666f<\/li>\n<\/ul>\n<h4>4. \u6df7\u5408\u7cbe\u5ea6\u8bad\u7ec3&#xff08;Mixed Precision Training&#xff09;<\/h4>\n<ul>\n<li>\u539f\u7406&#xff1a;\u524d\u5411\u4f20\u64ad\u7528 FP16 \u534a\u7cbe\u5ea6&#xff0c;\u53cd\u5411\u4f20\u64ad\u66f4\u65b0\u7528 FP32 \u5168\u7cbe\u5ea6<\/li>\n<li>\u4f18\u52bf&#xff1a;\u663e\u5b58\u5360\u7528\u51cf\u534a\u3001\u8bad\u7ec3\u901f\u5ea6\u63d0\u5347\u3001\u8ba1\u7b97\u541e\u5410\u91cf\u66f4\u9ad8<\/li>\n<li>\u5173\u952e&#xff1a;\u635f\u5931\u7f29\u653e&#xff08;Loss Scaling&#xff09;&#xff0c;\u9632\u6b62 FP16 \u4e0b\u68af\u5ea6\u8fc7\u5c0f\u4e0b\u6ea2\u4e3a 0<\/li>\n<\/ul>\n<h4>5. \u68af\u5ea6\u7d2f\u79ef&#xff08;Gradient Accumulation&#xff09;<\/h4>\n<ul>\n<li>\u539f\u7406&#xff1a;\u591a\u6b65\u524d\u5411\u4f20\u64ad\u540e\u518d\u66f4\u65b0\u4e00\u6b21\u53c2\u6570&#xff0c;\u7b49\u6548\u4e8e\u6269\u5927 batch size<\/li>\n<li>\u9002\u7528\u573a\u666f&#xff1a;\u663e\u5b58\u4e0d\u8db3&#xff0c;\u65e0\u6cd5\u4f7f\u7528\u5927 batch size<\/li>\n<\/ul>\n<h4>6. \u68af\u5ea6\u68c0\u67e5\u70b9&#xff08;Gradient Checkpointing&#xff09;<\/h4>\n<ul>\n<li>\u539f\u7406&#xff1a;\u727a\u7272\u8ba1\u7b97\u65f6\u95f4\u6362\u53d6\u663e\u5b58&#xff0c;\u53ea\u4fdd\u5b58\u90e8\u5206\u4e2d\u95f4\u6fc0\u6d3b\u503c&#xff0c;\u53cd\u5411\u4f20\u64ad\u65f6\u91cd\u65b0\u8ba1\u7b97<\/li>\n<li>\u9002\u7528&#xff1a;\u8bad\u7ec3\u8d85\u5927\u6a21\u578b&#xff0c;\u663e\u5b58\u4e0d\u8db3\u65f6<\/li>\n<\/ul>\n<h4>7. \u8bad\u7ec3\u4e0d\u6536\u655b\u7684\u5e38\u89c1\u539f\u56e0\u6392\u67e5<\/h4>\n<li>\u6570\u636e\u95ee\u9898&#xff1a;\u6807\u7b7e\u9519\u8bef\u3001\u6570\u636e\u5f52\u4e00\u5316\u9519\u8bef\u3001\u6b63\u8d1f\u6837\u672c\u6781\u7aef\u4e0d\u5747\u8861<\/li>\n<li>\u53c2\u6570\u95ee\u9898&#xff1a;\u5b66\u4e60\u7387\u8fc7\u5927 \/ \u8fc7\u5c0f\u3001\u6743\u91cd\u521d\u59cb\u5316\u9519\u8bef<\/li>\n<li>\u6a21\u578b\u95ee\u9898&#xff1a;\u7f51\u7edc\u7ed3\u6784\u9519\u8bef\u3001\u68af\u5ea6\u6d88\u5931 \/ \u7206\u70b8<\/li>\n<li>\u635f\u5931\u51fd\u6570&#xff1a;\u635f\u5931\u51fd\u6570\u9009\u62e9\u4e0d\u5f53\u3001\u8ba1\u7b97\u9519\u8bef<\/li>\n<li>\u6b63\u5219\u5316&#xff1a;\u6b63\u5219\u5316\u5f3a\u5ea6\u8fc7\u5927<\/li>\n<hr \/>\n<h3>\u5341\u4e00\u3001\u7ecf\u5178\u7f51\u7edc\u67b6\u6784\u4e0e\u521b\u65b0\u70b9<\/h3>\n<h4>1. LeNet-5<\/h4>\n<ul>\n<li>\u5730\u4f4d&#xff1a;\u7b2c\u4e00\u4e2a\u73b0\u4ee3 CNN&#xff0c;\u7528\u4e8e\u624b\u5199\u6570\u5b57\u8bc6\u522b<\/li>\n<li>\u7ed3\u6784&#xff1a;\u5377\u79ef\u2192\u6c60\u5316\u2192\u5377\u79ef\u2192\u6c60\u5316\u2192\u5168\u8fde\u63a5\u2192\u8f93\u51fa<\/li>\n<li>\u5960\u5b9a\u4e86 CNN \u7684\u57fa\u672c\u8303\u5f0f<\/li>\n<\/ul>\n<h4>2. AlexNet<\/h4>\n<ul>\n<li>\u91cc\u7a0b\u7891\u610f\u4e49&#xff1a;2012 \u5e74 ImageNet \u51a0\u519b&#xff0c;\u5f00\u542f\u6df1\u5ea6\u5b66\u4e60\u65f6\u4ee3<\/li>\n<li>\u521b\u65b0\u70b9&#xff1a;\n<li>\u9996\u6b21\u4f7f\u7528 ReLU \u6fc0\u6d3b\u51fd\u6570&#xff0c;\u7f13\u89e3\u68af\u5ea6\u6d88\u5931<\/li>\n<li>\u4f7f\u7528 Dropout \u9632\u6b62\u8fc7\u62df\u5408<\/li>\n<li>\u6570\u636e\u589e\u5f3a<\/li>\n<li>GPU \u5e76\u884c\u8bad\u7ec3<\/li>\n<li>\u91cd\u53e0\u6c60\u5316<\/li>\n<\/li>\n<\/ul>\n<h4>3. VGGNet<\/h4>\n<ul>\n<li>\u6838\u5fc3\u601d\u60f3&#xff1a;\u5806\u53e0\u5c0f\u5377\u79ef\u6838\u66ff\u4ee3\u5927\u5377\u79ef\u6838<\/li>\n<li>\u7279\u70b9&#xff1a;\u7ed3\u6784\u89c4\u6574&#xff0c;\u5168\u90e8\u4f7f\u7528 3&#215;3 \u5377\u79ef\u548c 2&#215;2 \u6c60\u5316<\/li>\n<li>\u610f\u4e49&#xff1a;\u8bc1\u660e\u4e86\u52a0\u6df1\u7f51\u7edc\u53ef\u4ee5\u6709\u6548\u63d0\u5347\u6027\u80fd<\/li>\n<\/ul>\n<h4>4. GoogLeNet \/ Inception \u7cfb\u5217<\/h4>\n<ul>\n<li>\u6838\u5fc3&#xff1a;Inception \u6a21\u5757&#xff0c;\u591a\u5c3a\u5bf8\u5377\u79ef\u6838\u5e76\u884c&#xff08;1&#215;1\u30013&#215;3\u30015&#215;5&#xff09;&#xff0c;\u6355\u6349\u591a\u5c3a\u5ea6\u7279\u5f81<\/li>\n<li>\u521b\u65b0&#xff1a;\u5927\u91cf\u4f7f\u7528 1&#215;1 \u5377\u79ef\u964d\u7ef4&#xff0c;\u63a7\u5236\u53c2\u6570\u91cf<\/li>\n<li>\u8f85\u52a9\u5206\u7c7b\u5668&#xff1a;\u7f13\u89e3\u6df1\u5c42\u68af\u5ea6\u6d88\u5931<\/li>\n<\/ul>\n<h4>5. ResNet&#xff08;\u6b8b\u5dee\u7f51\u7edc&#xff09;<\/h4>\n<ul>\n<li>\u91cc\u7a0b\u7891&#xff1a;2015 \u5e74 ImageNet \u51a0\u519b&#xff0c;\u9996\u6b21\u8ba9\u8d85\u6df1\u7f51\u7edc&#xff08;152 \u5c42&#xff09;\u53ef\u8bad\u7ec3<\/li>\n<li>\u6838\u5fc3&#xff1a;\u6b8b\u5dee\u8fde\u63a5&#xff08;Shortcut&#xff09;&#xff0c;\u8ba9\u7f51\u7edc\u5b66\u4e60\u6b8b\u5dee\u6620\u5c04\u800c\u975e\u76f4\u63a5\u6620\u5c04<\/li>\n<li>\u89e3\u51b3\u95ee\u9898&#xff1a;\u6df1\u5c42\u7f51\u7edc\u7684\u9000\u5316\u95ee\u9898&#xff08;\u5c42\u6570\u52a0\u6df1&#xff0c;\u6548\u679c\u53cd\u800c\u4e0b\u964d&#xff09;<\/li>\n<li>\u57fa\u7840\u5355\u5143&#xff1a;BasicBlock&#xff08;\u4e24\u5c42 3&#215;3 \u5377\u79ef&#xff09;\u3001Bottleneck&#xff08;1&#215;1&#043;3&#215;3&#043;1&#215;1&#xff0c;\u964d\u7ef4\u5347\u7ef4&#xff0c;\u51cf\u5c11\u8ba1\u7b97\u91cf&#xff09;<\/li>\n<\/ul>\n<h4>6. DenseNet<\/h4>\n<ul>\n<li>\u6838\u5fc3&#xff1a;\u5bc6\u96c6\u8fde\u63a5&#xff0c;\u6bcf\u4e00\u5c42\u7684\u8f93\u5165\u90fd\u5305\u542b\u524d\u9762\u6240\u6709\u5c42\u7684\u8f93\u51fa<\/li>\n<li>\u4f18\u52bf&#xff1a;\n<li>\u7279\u5f81\u590d\u7528&#xff0c;\u4fe1\u606f\u6d41\u901a\u66f4\u987a\u7545<\/li>\n<li>\u7f13\u89e3\u68af\u5ea6\u6d88\u5931<\/li>\n<li>\u53c2\u6570\u91cf\u66f4\u5c11<\/li>\n<\/li>\n<li>\u7f3a\u70b9&#xff1a;\u663e\u5b58\u5360\u7528\u9ad8<\/li>\n<\/ul>\n<h4>7. MobileNet \u7cfb\u5217<\/h4>\n<ul>\n<li>\u5b9a\u4f4d&#xff1a;\u79fb\u52a8\u7aef\u8f7b\u91cf\u5316\u7f51\u7edc<\/li>\n<li>\u6838\u5fc3&#xff1a;\u6df1\u5ea6\u53ef\u5206\u79bb\u5377\u79ef<\/li>\n<li>\u8fdb\u9636&#xff1a;\u5bbd\u5ea6\u56e0\u5b50\u3001\u5206\u8fa8\u7387\u56e0\u5b50&#xff0c;\u7075\u6d3b\u63a7\u5236\u6a21\u578b\u5927\u5c0f<\/li>\n<\/ul>\n<h4>8. EfficientNet<\/h4>\n<ul>\n<li>\u6838\u5fc3\u601d\u60f3&#xff1a;\u590d\u5408\u7f29\u653e&#xff0c;\u540c\u65f6\u7f29\u653e\u7f51\u7edc\u6df1\u5ea6\u3001\u5bbd\u5ea6\u3001\u5206\u8fa8\u7387<\/li>\n<li>\u4f18\u52bf&#xff1a;\u5728\u8ba1\u7b97\u91cf\u548c\u7cbe\u5ea6\u4e4b\u95f4\u8fbe\u5230\u6700\u4f18\u5e73\u8861<\/li>\n<\/ul>\n<h4>9. ViT&#xff08;Vision Transformer&#xff09;<\/h4>\n<ul>\n<li>\u91cc\u7a0b\u7891&#xff1a;\u5c06 Transformer \u7eaf\u8fc1\u79fb\u5230\u56fe\u50cf\u4efb\u52a1<\/li>\n<li>\u505a\u6cd5&#xff1a;\u5c06\u56fe\u50cf\u5207\u5206\u6210 patch&#xff0c;\u5c55\u5e73\u6210\u5e8f\u5217&#xff0c;\u9001\u5165 Transformer Encoder<\/li>\n<li>\u610f\u4e49&#xff1a;\u8bc1\u660e\u4e86 Transformer \u5728\u89c6\u89c9\u9886\u57df\u7684\u6f5c\u529b&#xff0c;\u5f00\u542f CV \u5927\u6a21\u578b\u65f6\u4ee3<\/li>\n<\/ul>\n<h4>10. Swin Transformer<\/h4>\n<ul>\n<li>\u6838\u5fc3&#xff1a;\u6ed1\u52a8\u7a97\u53e3\u81ea\u6ce8\u610f\u529b &#043; \u5c42\u7ea7\u5316\u7ed3\u6784<\/li>\n<li>\u89e3\u51b3 ViT \u7684\u95ee\u9898&#xff1a;\u8ba1\u7b97\u91cf\u8fc7\u5927\u3001\u591a\u5c3a\u5ea6\u7279\u5f81\u4e0d\u8db3<\/li>\n<li>\u5730\u4f4d&#xff1a;\u89c6\u89c9 Transformer \u7684\u6807\u6746&#xff0c;\u5e7f\u6cdb\u7528\u4e8e\u68c0\u6d4b\u3001\u5206\u5272\u7b49\u4e0b\u6e38\u4efb\u52a1<\/li>\n<\/ul>\n<hr \/>\n<h3>\u5341\u4e8c\u3001\u6a21\u578b\u538b\u7f29\u4e0e\u8f7b\u91cf\u5316\u90e8\u7f72<\/h3>\n<h4>1. \u6a21\u578b\u538b\u7f29\u56db\u5927\u65b9\u5411<\/h4>\n<h5>&#xff08;1&#xff09;\u6a21\u578b\u526a\u679d&#xff08;Pruning&#xff09;<\/h5>\n<ul>\n<li>\u539f\u7406&#xff1a;\u79fb\u9664\u7f51\u7edc\u4e2d\u4e0d\u91cd\u8981\u7684\u6743\u91cd \/ \u901a\u9053 \/ \u5c42<\/li>\n<li>\u5206\u7c7b&#xff1a;\n<ul>\n<li>\u975e\u7ed3\u6784\u5316\u526a\u679d&#xff1a;\u526a\u5355\u4e2a\u6743\u91cd&#xff0c;\u7a00\u758f\u5ea6\u9ad8\u4f46\u9700\u7279\u6b8a\u786c\u4ef6\u652f\u6301<\/li>\n<li>\u7ed3\u6784\u5316\u526a\u679d&#xff1a;\u526a\u6574\u4e2a\u901a\u9053 \/ \u5c42&#xff0c;\u76f4\u63a5\u51cf\u5c0f\u6a21\u578b\u5c3a\u5bf8&#xff0c;\u901a\u7528\u786c\u4ef6\u53ef\u90e8\u7f72<\/li>\n<\/ul>\n<\/li>\n<li>\u6d41\u7a0b&#xff1a;\u8bad\u7ec3\u5927\u6a21\u578b\u2192\u8bc4\u4f30\u91cd\u8981\u6027\u2192\u526a\u679d\u2192\u5fae\u8c03\u6062\u590d\u7cbe\u5ea6<\/li>\n<\/ul>\n<h5>&#xff08;2&#xff09;\u91cf\u5316&#xff08;Quantization&#xff09;<\/h5>\n<ul>\n<li>\u539f\u7406&#xff1a;\u5c06\u9ad8\u7cbe\u5ea6\u6d6e\u70b9&#xff08;FP32&#xff09;\u8f6c\u4e3a\u4f4e\u7cbe\u5ea6&#xff08;FP16\u3001INT8\u3001INT4&#xff09;<\/li>\n<li>\u4f18\u52bf&#xff1a;\u51cf\u5c0f\u6a21\u578b\u4f53\u79ef\u3001\u63d0\u5347\u63a8\u7406\u901f\u5ea6\u3001\u964d\u4f4e\u663e\u5b58\u5360\u7528<\/li>\n<li>\u5206\u7c7b&#xff1a;\n<ul>\n<li>\u8bad\u7ec3\u540e\u91cf\u5316&#xff08;PTQ&#xff09;&#xff1a;\u5c11\u91cf\u6821\u51c6\u6570\u636e\u5373\u53ef&#xff0c;\u7b80\u5355\u5feb\u901f<\/li>\n<li>\u91cf\u5316\u611f\u77e5\u8bad\u7ec3&#xff08;QAT&#xff09;&#xff1a;\u8bad\u7ec3\u4e2d\u6a21\u62df\u91cf\u5316\u8bef\u5dee&#xff0c;\u7cbe\u5ea6\u66f4\u9ad8<\/li>\n<\/ul>\n<\/li>\n<li>\u6ce8\u610f&#xff1a;\u6781\u4f4e\u6bd4\u7279\u91cf\u5316\u53ef\u80fd\u5e26\u6765\u7cbe\u5ea6\u635f\u5931<\/li>\n<\/ul>\n<h5>&#xff08;3&#xff09;\u77e5\u8bc6\u84b8\u998f&#xff08;Knowledge Distillation&#xff09;<\/h5>\n<ul>\n<li>\u539f\u7406&#xff1a;\u5927\u6a21\u578b&#xff08;\u6559\u5e08&#xff09;\u6307\u5bfc\u5c0f\u6a21\u578b&#xff08;\u5b66\u751f&#xff09;\u8bad\u7ec3&#xff0c;\u5b66\u751f\u5b66\u4e60\u6559\u5e08\u7684\u8f6f\u6807\u7b7e<\/li>\n<li>\u4f18\u52bf&#xff1a;\u5c0f\u6a21\u578b\u83b7\u5f97\u63a5\u8fd1\u5927\u6a21\u578b\u7684\u7cbe\u5ea6<\/li>\n<li>\u8fdb\u9636&#xff1a;\u7279\u5f81\u84b8\u998f\u3001\u6ce8\u610f\u529b\u84b8\u998f\u7b49<\/li>\n<\/ul>\n<h5>&#xff08;4&#xff09;\u8f7b\u91cf\u5316\u7f51\u7edc\u8bbe\u8ba1<\/h5>\n<ul>\n<li>\u4ee3\u8868&#xff1a;MobileNet\u3001ShuffleNet\u3001SqueezeNet\u3001GhostNet<\/li>\n<li>\u601d\u8def&#xff1a;\u8bbe\u8ba1\u66f4\u9ad8\u6548\u7684\u5377\u79ef\u7ed3\u6784\u3001\u5206\u7ec4\u5377\u79ef\u3001\u901a\u9053\u6df7\u6d17\u7b49<\/li>\n<\/ul>\n<h4>2. \u63a8\u7406\u52a0\u901f\u6846\u67b6<\/h4>\n<ul>\n<li>TensorRT&#xff1a;NVIDIA GPU \u63a8\u7406\u52a0\u901f&#xff0c;\u652f\u6301\u91cf\u5316\u3001\u7b97\u5b50\u878d\u5408\u3001\u5185\u6838\u4f18\u5316<\/li>\n<li>ONNX Runtime&#xff1a;\u8de8\u5e73\u53f0\u63a8\u7406\u5f15\u64ce<\/li>\n<li>NCNN \/ MNN&#xff1a;\u79fb\u52a8\u7aef\u63a8\u7406\u6846\u67b6<\/li>\n<li>OpenVINO&#xff1a;Intel \u5e73\u53f0\u63a8\u7406\u4f18\u5316<\/li>\n<\/ul>\n<hr \/>\n<h3>\u5341\u4e09\u3001\u751f\u6210\u5f0f\u6a21\u578b\u57fa\u7840<\/h3>\n<h4>1. GAN&#xff08;\u751f\u6210\u5bf9\u6297\u7f51\u7edc&#xff09;<\/h4>\n<ul>\n<li>\u7ec4\u6210&#xff1a;\u751f\u6210\u5668 &#043; \u5224\u522b\u5668<\/li>\n<li>\u601d\u60f3&#xff1a;\u4e8c\u4eba\u96f6\u548c\u535a\u5f08&#xff0c;\u751f\u6210\u5668\u5c3d\u91cf\u751f\u6210\u903c\u771f\u6837\u672c&#xff0c;\u5224\u522b\u5668\u5c3d\u91cf\u5206\u8fa8\u771f\u5047<\/li>\n<li>\u635f\u5931&#xff1a;\u6700\u5c0f\u6700\u5927\u535a\u5f08<\/li>\n<li>\u95ee\u9898&#xff1a;\u8bad\u7ec3\u4e0d\u7a33\u5b9a\u3001\u6a21\u5f0f\u5d29\u584c\u3001\u68af\u5ea6\u6d88\u5931<\/li>\n<\/ul>\n<h4>2. VAE&#xff08;\u53d8\u5206\u81ea\u7f16\u7801\u5668&#xff09;<\/h4>\n<ul>\n<li>\u601d\u60f3&#xff1a;\u663e\u5f0f\u5efa\u6a21\u6570\u636e\u5206\u5e03&#xff0c;\u57fa\u4e8e\u53d8\u5206\u63a8\u65ad<\/li>\n<li>\u7ec4\u6210&#xff1a;\u7f16\u7801\u5668&#xff08;\u63a8\u65ad\u9690\u53d8\u91cf\u5206\u5e03&#xff09;&#043; \u89e3\u7801\u5668&#xff08;\u4ece\u9690\u53d8\u91cf\u751f\u6210\u6837\u672c&#xff09;<\/li>\n<li>\u7279\u70b9&#xff1a;\u8bad\u7ec3\u7a33\u5b9a&#xff0c;\u6709\u663e\u5f0f\u5206\u5e03&#xff1b;\u751f\u6210\u6837\u672c\u6e05\u6670\u5ea6\u901a\u5e38\u4f4e\u4e8e GAN<\/li>\n<\/ul>\n<h4>3. \u6269\u6563\u6a21\u578b&#xff08;Diffusion Model&#xff09;<\/h4>\n<ul>\n<li>\u539f\u7406&#xff1a;\u524d\u5411\u9010\u6b65\u52a0\u566a&#xff0c;\u53cd\u5411\u9010\u6b65\u53bb\u566a&#xff0c;\u5b66\u4e60\u4ece\u566a\u58f0\u8fd8\u539f\u6570\u636e<\/li>\n<li>\u4f18\u52bf&#xff1a;\u751f\u6210\u8d28\u91cf\u9ad8\u3001\u8bad\u7ec3\u7a33\u5b9a\u3001\u6a21\u5f0f\u8986\u76d6\u5168\u9762<\/li>\n<li>\u73b0\u72b6&#xff1a;\u5f53\u524d AIGC \u7684\u4e3b\u6d41\u6280\u672f&#xff08;Stable Diffusion\u3001DALL\u30fbE \u7b49&#xff09;<\/li>\n<li>\u7f3a\u70b9&#xff1a;\u63a8\u7406\u901f\u5ea6\u6162&#xff0c;\u9700\u8981\u591a\u6b65\u53bb\u566a<\/li>\n<\/ul>\n<hr \/>\n<h3>\u5341\u56db\u3001\u8bc4\u4f30\u6307\u6807\u4e0e\u6a21\u578b\u8bc4\u4ef7<\/h3>\n<h4>1. \u5206\u7c7b\u4efb\u52a1\u6307\u6807<\/h4>\n<ul>\n<li>\u51c6\u786e\u7387&#xff08;Accuracy&#xff09;&#xff1a;\u9884\u6d4b\u6b63\u786e\u7684\u6837\u672c\u5360\u603b\u6837\u672c\u6bd4\u4f8b&#xff1b;\u6837\u672c\u5747\u8861\u65f6\u9002\u7528<\/li>\n<li>\u7cbe\u786e\u7387&#xff08;Precision&#xff09;&#xff1a;\u9884\u6d4b\u4e3a\u6b63\u7684\u6837\u672c\u4e2d&#xff0c;\u771f\u6b63\u4e3a\u6b63\u7684\u6bd4\u4f8b&#xff1b;\u5173\u6ce8\u8bef\u62a5<\/li>\n<li>\u53ec\u56de\u7387&#xff08;Recall&#xff09;&#xff1a;\u771f\u5b9e\u4e3a\u6b63\u7684\u6837\u672c\u4e2d&#xff0c;\u88ab\u9884\u6d4b\u4e3a\u6b63\u7684\u6bd4\u4f8b&#xff1b;\u5173\u6ce8\u6f0f\u62a5<\/li>\n<li>F1 \u503c&#xff1a;\u7cbe\u786e\u7387\u548c\u53ec\u56de\u7387\u7684\u8c03\u548c\u5e73\u5747&#xff0c;\u7efc\u5408\u8bc4\u4ef7<\/li>\n<li>\u6df7\u6dc6\u77e9\u9635&#xff1a;\u76f4\u89c2\u5c55\u793a\u5404\u7c7b\u522b\u9884\u6d4b\u5bf9\u9519\u60c5\u51b5<\/li>\n<li>ROC-AUC&#xff1a;\u8861\u91cf\u6a21\u578b\u6574\u4f53\u6392\u5e8f\u80fd\u529b&#xff0c;\u4e0d\u53d7\u9608\u503c\u5f71\u54cd&#xff0c;\u4e0d\u5747\u8861\u6837\u672c\u9002\u7528<\/li>\n<\/ul>\n<h4>2. \u76ee\u6807\u68c0\u6d4b\u6307\u6807<\/h4>\n<ul>\n<li>mAP&#xff08;\u5e73\u5747\u7cbe\u5ea6\u5747\u503c&#xff09;&#xff1a;\u5404\u7c7b\u522b AP \u7684\u5e73\u5747\u503c&#xff0c;\u6838\u5fc3\u6307\u6807<\/li>\n<li>IoU&#xff08;\u4ea4\u5e76\u6bd4&#xff09;&#xff1a;\u9884\u6d4b\u6846\u4e0e\u771f\u5b9e\u6846\u7684\u91cd\u53e0\u7a0b\u5ea6<\/li>\n<li>FPS&#xff1a;\u6bcf\u79d2\u63a8\u7406\u5e27\u6570&#xff0c;\u8861\u91cf\u901f\u5ea6<\/li>\n<\/ul>\n<h4>3. \u5206\u5272\u4efb\u52a1\u6307\u6807<\/h4>\n<ul>\n<li>mIoU&#xff1a;\u5404\u7c7b\u522b\u4ea4\u5e76\u6bd4\u7684\u5747\u503c&#xff0c;\u6838\u5fc3\u6307\u6807<\/li>\n<li>Dice \u7cfb\u6570&#xff1a;\u4e0e IoU \u6b63\u76f8\u5173&#xff0c;\u533b\u5b66\u5206\u5272\u5e38\u7528<\/li>\n<\/ul>\n<h4>4. \u56de\u5f52\u4efb\u52a1\u6307\u6807<\/h4>\n<ul>\n<li>MAE\u3001MSE\u3001RMSE<\/li>\n<li>R\u00b2&#xff08;\u51b3\u5b9a\u7cfb\u6570&#xff09;&#xff1a;\u8861\u91cf\u6a21\u578b\u89e3\u91ca\u65b9\u5dee\u7684\u6bd4\u4f8b<\/li>\n<\/ul>\n<hr \/>\n<h3>\u5341\u4e94\u3001\u6df1\u5ea6\u5b66\u4e60\u6846\u67b6\u4e0e\u5de5\u7a0b\u5b9e\u8df5<\/h3>\n<h4>1. PyTorch vs TensorFlow<\/h4>\n<p>\u8868\u683c<\/p>\n<table>\n<tr>\u7ef4\u5ea6PyTorchTensorFlow<\/tr>\n<tbody>\n<tr>\n<td>\u7f16\u7a0b\u8303\u5f0f<\/td>\n<td>\u52a8\u6001\u56fe&#xff0c;\u547d\u4ee4\u5f0f&#xff0c;\u8c03\u8bd5\u65b9\u4fbf<\/td>\n<td>\u65e9\u671f\u9759\u6001\u56fe&#xff0c;\u73b0\u5728\u4e5f\u652f\u6301\u52a8\u6001\u56fe<\/td>\n<\/tr>\n<tr>\n<td>\u6613\u7528\u6027<\/td>\n<td>\u7b80\u5355\u76f4\u89c2&#xff0c;Python \u98ce\u683c\u5f3a<\/td>\n<td>\u8bed\u6cd5\u76f8\u5bf9\u590d\u6742&#xff0c;\u5b66\u4e60\u66f2\u7ebf\u9661<\/td>\n<\/tr>\n<tr>\n<td>\u5b66\u672f\u754c<\/td>\n<td>\u7edd\u5bf9\u4e3b\u6d41<\/td>\n<td>\u5360\u6bd4\u4f4e<\/td>\n<\/tr>\n<tr>\n<td>\u5de5\u4e1a\u90e8\u7f72<\/td>\n<td>\u9700\u989d\u5916\u5de5\u5177&#xff08;TorchScript\u3001ONNX&#xff09;<\/td>\n<td>TensorRT\u3001TFServing \u751f\u6001\u5b8c\u5584<\/td>\n<\/tr>\n<tr>\n<td>\u8c03\u8bd5<\/td>\n<td>\u53ef\u76f4\u63a5\u6253\u65ad\u70b9&#xff0c;\u50cf\u666e\u901a Python \u4ee3\u7801<\/td>\n<td>\u9759\u6001\u56fe\u8c03\u8bd5\u9ebb\u70e6<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<h4>2. PyTorch \u4e2d nn.Module \u7684\u4f5c\u7528<\/h4>\n<ul>\n<li>\u5c01\u88c5\u7f51\u7edc\u5c42\u3001\u53c2\u6570\u7ba1\u7406\u3001\u8bbe\u5907\u8fc1\u79fb\u3001\u4fdd\u5b58\u52a0\u8f7d\u3001\u5b50\u6a21\u5757\u7ba1\u7406<\/li>\n<li>forward \u65b9\u6cd5\u5b9a\u4e49\u524d\u5411\u4f20\u64ad\u903b\u8f91<\/li>\n<\/ul>\n<h4>3. model.eval () \u548c torch.no_grad () \u7684\u533a\u522b<\/h4>\n<ul>\n<li>model.eval()&#xff1a;\u5207\u6362\u5230\u8bc4\u4f30\u6a21\u5f0f&#xff0c;\u5173\u95ed Dropout\u3001BatchNorm \u4f7f\u7528\u79fb\u52a8\u5e73\u5747\u5747\u503c\u65b9\u5dee<\/li>\n<li>torch.no_grad()&#xff1a;\u5173\u95ed\u68af\u5ea6\u8ba1\u7b97&#xff0c;\u8282\u7701\u663e\u5b58\u548c\u8ba1\u7b97&#xff0c;\u4e0d\u5f71\u54cd\u6a21\u578b\u72b6\u6001<\/li>\n<li>\u63a8\u7406\u65f6\u4e24\u8005\u901a\u5e38\u4e00\u8d77\u4f7f\u7528<\/li>\n<\/ul>\n<h4>4. DataParallel vs DistributedDataParallel<\/h4>\n<ul>\n<li>DataParallel&#xff1a;\u5355\u673a\u591a\u5361&#xff0c;\u6570\u636e\u5e76\u884c&#xff0c;\u4e3b\u5361\u6c47\u603b\u68af\u5ea6&#xff0c;\u8d1f\u8f7d\u4e0d\u5747\u8861<\/li>\n<li>DistributedDataParallel&#xff08;DDP&#xff09;&#xff1a;\u591a\u673a\u591a\u5361 \/ \u5355\u673a\u591a\u5361&#xff0c;\u6bcf\u5f20\u5361\u72ec\u7acb\u8fdb\u7a0b&#xff0c;\u68af\u5ea6\u5168\u89c4\u7ea6&#xff0c;\u901f\u5ea6\u66f4\u5feb\u3001\u8d1f\u8f7d\u5747\u8861&#xff0c;\u662f\u4e3b\u6d41\u65b9\u6848<\/li>\n<\/ul>\n<hr \/>\n<h4>1. \u53cd\u5411\u4f20\u64ad\u4e0e\u68af\u5ea6\u95ee\u9898<\/h4>\n<p>\u53cd\u5411\u4f20\u64ad\u672c\u8d28&#xff1a;\u57fa\u4e8e\u94fe\u5f0f\u6cd5\u5219&#xff0c;\u4ece\u8f93\u51fa\u5c42\u5411\u8f93\u5165\u5c42\u9010\u5c42\u8ba1\u7b97\u635f\u5931\u5bf9\u6bcf\u4e2a\u53c2\u6570\u7684\u68af\u5ea6&#xff0c;\u7528\u4e8e\u68af\u5ea6\u4e0b\u964d\u66f4\u65b0\u6743\u91cd\u3002\u6838\u5fc3\u662f\u6784\u5efa\u8ba1\u7b97\u56fe&#xff0c;\u6b63\u5411\u4f20\u64ad\u8ba1\u7b97\u4e2d\u95f4\u503c&#xff0c;\u53cd\u5411\u4f20\u64ad\u9010\u5c42\u56de\u4f20\u68af\u5ea6\u3002<\/p>\n<p>\u68af\u5ea6\u6d88\u5931\u4e0e\u68af\u5ea6\u7206\u70b8<\/p>\n<ul>\n<li>\u6210\u56e0&#xff1a;\u6df1\u5c42\u7f51\u7edc\u4e2d&#xff0c;\u6fc0\u6d3b\u51fd\u6570\u5bfc\u6570\u8fde\u4e58\u540e\u6307\u6570\u7ea7\u8870\u51cf&#xff08;\u5c0f\u4e8e 1&#xff09;\u6216\u6307\u6570\u7ea7\u589e\u957f&#xff08;\u5927\u4e8e 1&#xff09;\u3002Sigmoid \u5bfc\u6570\u6700\u5927\u4ec5 0.25&#xff0c;\u591a\u5c42\u540e\u68af\u5ea6\u8d8b\u8fd1\u4e8e 0&#xff1b;\u6743\u91cd\u521d\u59cb\u5316\u8fc7\u5927\u5219\u4f1a\u5bfc\u81f4\u68af\u5ea6\u7206\u70b8\u3002<\/li>\n<li>\u68af\u5ea6\u6d88\u5931\u89e3\u51b3\u65b9\u6848&#xff1a;ReLU \u7cfb\u5217\u6fc0\u6d3b\u51fd\u6570\u3001\u6b8b\u5dee\u8fde\u63a5\u3001BatchNorm\u3001\u9884\u8bad\u7ec3 &#043; \u5fae\u8c03\u3001\u5408\u9002\u7684\u6743\u91cd\u521d\u59cb\u5316\u3002<\/li>\n<li>\u68af\u5ea6\u7206\u70b8\u89e3\u51b3\u65b9\u6848&#xff1a;\u68af\u5ea6\u88c1\u526a&#xff08;Gradient Clipping&#xff09;\u3001\u6743\u91cd\u521d\u59cb\u5316&#xff08;Xavier\/He \u521d\u59cb\u5316&#xff09;\u3001BatchNorm\u3001L2 \u6b63\u5219\u5316\u3002<\/li>\n<\/ul>\n<h4>2. \u6fc0\u6d3b\u51fd\u6570<\/h4>\n<p>\u8868\u683c<\/p>\n<table>\n<tr>\u6fc0\u6d3b\u51fd\u6570\u516c\u5f0f\u4f18\u70b9\u7f3a\u70b9\u9002\u7528\u573a\u666f<\/tr>\n<tbody>\n<tr>\n<td>Sigmoid<\/td>\n<td>\\\\(\\\\sigma(x)&#061;\\\\frac{1}{1&#043;e^{-x}}\\\\)<\/td>\n<td>\u8f93\u51fa 0~1&#xff0c;\u53ef\u89e3\u91ca\u4e3a\u6982\u7387<\/td>\n<td>\u68af\u5ea6\u6d88\u5931\u3001\u8f93\u51fa\u975e\u96f6\u5747\u503c\u3001\u6307\u6570\u8fd0\u7b97\u6162<\/td>\n<td>\u65e9\u671f\u7f51\u7edc\u3001\u95e8\u63a7\u5355\u5143<\/td>\n<\/tr>\n<tr>\n<td>Tanh<\/td>\n<td>\\\\(\\\\tanh(x)&#061;\\\\frac{e^x-e^{-x}}{e^x&#043;e^{-x}}\\\\)<\/td>\n<td>\u96f6\u5747\u503c&#xff0c;\u6536\u655b\u6bd4 Sigmoid \u5feb<\/td>\n<td>\u4ecd\u6709\u68af\u5ea6\u6d88\u5931\u3001\u6307\u6570\u8fd0\u7b97<\/td>\n<td>\u5faa\u73af\u795e\u7ecf\u7f51\u7edc<\/td>\n<\/tr>\n<tr>\n<td>ReLU<\/td>\n<td>\\\\(\\\\text{ReLU}(x)&#061;\\\\max(0,x)\\\\)<\/td>\n<td>\u6b63\u533a\u95f4\u65e0\u68af\u5ea6\u6d88\u5931\u3001\u8ba1\u7b97\u6781\u5feb\u3001\u7a00\u758f\u6fc0\u6d3b<\/td>\n<td>Dead ReLU&#xff08;\u8d1f\u533a\u95f4\u795e\u7ecf\u5143\u6c38\u4e45\u5931\u6d3b&#xff09;\u3001\u8f93\u51fa\u975e\u96f6\u5747\u503c<\/td>\n<td>\u7edd\u5927\u591a\u6570 CNN<\/td>\n<\/tr>\n<tr>\n<td>Leaky ReLU<\/td>\n<td>\\\\(\\\\max(\\\\alpha x,x),\\\\alpha\\\\approx0.01\\\\)<\/td>\n<td>\u89e3\u51b3 Dead ReLU \u95ee\u9898<\/td>\n<td>\u03b1 \u4e3a\u56fa\u5b9a\u8d85\u53c2&#xff0c;\u9700\u624b\u52a8\u8c03\u4f18<\/td>\n<td>\u5bf9\u8d1f\u68af\u5ea6\u6709\u9700\u6c42\u7684\u573a\u666f<\/td>\n<\/tr>\n<tr>\n<td>GELU<\/td>\n<td>\\\\(x\\\\cdot\\\\Phi(x)\\\\)&#xff0c;\u03a6 \u4e3a\u9ad8\u65af\u7d2f\u79ef\u5206\u5e03<\/td>\n<td>\u5e73\u6ed1\u975e\u9971\u548c\u3001\u7ed3\u5408\u968f\u673a\u6b63\u5219\u601d\u60f3\u3001\u6027\u80fd\u4f18\u5f02<\/td>\n<td>\u8ba1\u7b97\u7a0d\u590d\u6742<\/td>\n<td>Transformer\u3001ViT\u3001BERT<\/td>\n<\/tr>\n<tr>\n<td>Swish<\/td>\n<td>\\\\(x\\\\cdot\\\\sigma(\\\\beta x)\\\\)<\/td>\n<td>\u5e73\u6ed1\u6fc0\u6d3b\u3001\u81ea\u9002\u5e94\u95e8\u63a7<\/td>\n<td>\u03b2 \u9700\u5b66\u4e60<\/td>\n<td>\u8f7b\u91cf\u5316\u7f51\u7edc\u3001EfficientNet<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<h4>3. \u635f\u5931\u51fd\u6570<\/h4>\n<h5>\u5206\u7c7b\u635f\u5931<\/h5>\n<ul>\n<li>\u4ea4\u53c9\u71b5\u635f\u5931&#xff1a;\u591a\u5206\u7c7b\u6700\u5e38\u7528&#xff0c;\u8861\u91cf\u4e24\u4e2a\u5206\u5e03\u7684\u5dee\u5f02\u3002\u516c\u5f0f&#xff1a;\\\\(L&#061;-\\\\sum y_i\\\\log p_i\\\\)\u3002\u76f8\u6bd4 MSE&#xff0c;\u5206\u7c7b\u4efb\u52a1\u4e0b\u68af\u5ea6\u66f4\u5927\u3001\u6536\u655b\u66f4\u5feb&#xff0c;\u4e14\u4e3a\u51f8\u4f18\u5316\u95ee\u9898\u3002<\/li>\n<li>\u6807\u7b7e\u5e73\u6ed1&#xff1a;\u5c06\u786c\u6807\u7b7e\u8f6c\u5316\u4e3a\u8f6f\u6807\u7b7e \\\\(y&#039;&#061;y(1-\\\\epsilon)&#043;\\\\epsilon\/K\\\\)&#xff0c;\u9632\u6b62\u6a21\u578b\u8fc7\u5ea6\u81ea\u4fe1&#xff0c;\u63d0\u5347\u6cdb\u5316\u80fd\u529b\u3002<\/li>\n<li>Focal Loss&#xff1a;\u89e3\u51b3\u6b63\u8d1f\u6837\u672c\u3001\u96be\u6613\u6837\u672c\u4e0d\u5747\u8861\u3002\u516c\u5f0f&#xff1a;\\\\(FL&#061;-\\\\alpha_t(1-p_t)^\\\\gamma\\\\log(p_t)\\\\)\u3002\\\\(\\\\alpha\\\\)\u5e73\u8861\u6b63\u8d1f\u6837\u672c\u6743\u91cd&#xff0c;\\\\(\\\\gamma\\\\)\u964d\u4f4e\u6613\u5206\u7c7b\u6837\u672c\u6743\u91cd&#xff0c;\u805a\u7126\u96be\u6837\u672c\u3002<\/li>\n<\/ul>\n<h5>\u56de\u5f52\u635f\u5931<\/h5>\n<ul>\n<li>MSE\/L2 \u635f\u5931&#xff1a;\u5bf9\u5f02\u5e38\u503c\u654f\u611f&#xff0c;\u68af\u5ea6\u968f\u8bef\u5dee\u51cf\u5c0f\u800c\u51cf\u5c0f&#xff0c;\u6536\u655b\u5e73\u6ed1\u3002<\/li>\n<li>MAE\/L1 \u635f\u5931&#xff1a;\u5bf9\u5f02\u5e38\u503c\u9c81\u68d2&#xff0c;\u4f46\u96f6\u70b9\u68af\u5ea6\u4e0d\u8fde\u7eed&#xff0c;\u6536\u655b\u4e0d\u7a33\u5b9a\u3002<\/li>\n<li>Smooth L1&#xff1a;\\\\(|x|&lt;1\\\\)\u65f6\u4e3a\\\\(0.5x^2\\\\)&#xff0c;\u5426\u5219\u4e3a\\\\(|x|-0.5\\\\)\u3002\u7ed3\u5408 L1 \u4e0e L2 \u4f18\u70b9&#xff0c;\u9c81\u68d2\u4e14\u68af\u5ea6\u5e73\u6ed1&#xff0c;\u68c0\u6d4b\u6846\u56de\u5f52\u7ecf\u5178\u9009\u62e9\u3002<\/li>\n<\/ul>\n<h5>\u8fb9\u754c\u6846\u56de\u5f52\u635f\u5931<\/h5>\n<ul>\n<li>IoU Loss&#xff1a;\\\\(L&#061;1-\\\\text{IoU}\\\\)&#xff0c;\u76f4\u63a5\u4f18\u5316\u91cd\u53e0\u5ea6\u3002\u7f3a\u70b9&#xff1a;\u6846\u4e0d\u91cd\u53e0\u65f6\u65e0\u68af\u5ea6&#xff0c;\u65e0\u6cd5\u4f18\u5316\u3002<\/li>\n<li>GIoU Loss&#xff1a;\u5f15\u5165\u6700\u5c0f\u5916\u63a5\u6846\u60e9\u7f5a&#xff0c;\u89e3\u51b3\u65e0\u91cd\u53e0\u68af\u5ea6\u95ee\u9898&#xff0c;\u4f46\u6536\u655b\u6162\u3002<\/li>\n<li>DIoU Loss&#xff1a;\u52a0\u5165\u4e2d\u5fc3\u70b9\u8ddd\u79bb\u60e9\u7f5a&#xff0c;\u6536\u655b\u901f\u5ea6\u663e\u8457\u63d0\u5347\u3002<\/li>\n<li>CIoU Loss&#xff1a;\u518d\u52a0\u5165\u957f\u5bbd\u6bd4\u4e00\u81f4\u6027\u60e9\u7f5a&#xff0c;\u51e0\u4f55\u7ea6\u675f\u66f4\u5168\u9762&#xff0c;YOLO \u7cfb\u5217\u4e3b\u6d41\u9009\u62e9\u3002<\/li>\n<\/ul>\n<h4>4. \u4f18\u5316\u5668\u4e0e\u5b66\u4e60\u7387\u7b56\u7565<\/h4>\n<h5>\u7ecf\u5178\u4f18\u5316\u5668<\/h5>\n<ul>\n<li>SGD&#043;Momentum&#xff1a;\u79ef\u7d2f\u5386\u53f2\u68af\u5ea6\u5f62\u6210\u60ef\u6027&#xff0c;\u8df3\u51fa\u5c40\u90e8\u6700\u4f18\u4e0e\u978d\u70b9\u3002\u6cdb\u5316\u6027\u597d&#xff0c;\u4f46\u6536\u655b\u6162&#xff0c;\u5bf9\u5b66\u4e60\u7387\u654f\u611f\u3002<\/li>\n<li>RMSProp&#xff1a;\u6307\u6570\u79fb\u52a8\u5e73\u5747\u68af\u5ea6\u5e73\u65b9&#xff0c;\u81ea\u9002\u5e94\u8c03\u6574\u6bcf\u4e2a\u53c2\u6570\u7684\u5b66\u4e60\u7387&#xff0c;\u89e3\u51b3 AdaGrad \u5b66\u4e60\u7387\u5355\u8c03\u9012\u51cf\u95ee\u9898\u3002<\/li>\n<li>Adam&#xff1a;\u52a8\u91cf &#043; RMSProp&#xff0c;\u4e00\u9636\u77e9\u4e8c\u9636\u77e9\u6307\u6570\u79fb\u52a8\u5e73\u5747 &#043; \u504f\u5dee\u4fee\u6b63\u3002\u6536\u655b\u5feb\u3001\u81ea\u9002\u5e94\u5f3a&#xff0c;\u4f46\u6cdb\u5316\u6027\u5f31\u4e8e SGD&#xff0c;\u6613\u8fc7\u62df\u5408\u3002<\/li>\n<li>AdamW&#xff1a;\u89e3\u8026\u6743\u91cd\u8870\u51cf\u4e0e\u68af\u5ea6\u66f4\u65b0&#xff0c;\u5c06\u6743\u91cd\u8870\u51cf\u76f4\u63a5\u4f5c\u7528\u4e8e\u6743\u91cd\u672c\u8eab\u3002Transformer\u3001ViT \u7b49\u73b0\u4ee3\u6a21\u578b\u7684\u6807\u51c6\u9009\u62e9\u3002<\/li>\n<\/ul>\n<h5>\u5b66\u4e60\u7387\u8c03\u5ea6<\/h5>\n<ul>\n<li>Warmup&#xff1a;\u8bad\u7ec3\u521d\u671f\u7528\u5c0f\u5b66\u4e60\u7387\u9884\u70ed&#xff0c;\u907f\u514d\u68af\u5ea6\u9707\u8361\u5bfc\u81f4\u6a21\u578b\u5d29\u584c\u3002<\/li>\n<li>\u4f59\u5f26\u9000\u706b&#xff1a;\u5b66\u4e60\u7387\u6309\u4f59\u5f26\u51fd\u6570\u5468\u671f\u6027\u8870\u51cf&#xff0c;\u914d\u5408\u91cd\u542f\u53ef\u8df3\u51fa\u5c40\u90e8\u6700\u4f18\u3002<\/li>\n<li>StepLR \/ \u591a\u6b65\u8870\u51cf&#xff1a;\u56fa\u5b9a\u8f6e\u6570\u4e58\u4ee5\u8870\u51cf\u7cfb\u6570&#xff0c;\u7b80\u5355\u6613\u8c03\u3002<\/li>\n<li>ReduceLROnPlateau&#xff1a;\u76d1\u63a7\u9a8c\u8bc1\u96c6\u6307\u6807&#xff0c;\u4e0d\u518d\u4e0b\u964d\u65f6\u964d\u4f4e\u5b66\u4e60\u7387\u3002<\/li>\n<\/ul>\n<h4>5. \u5f52\u4e00\u5316\u65b9\u6cd5<\/h4>\n<p>\u6838\u5fc3\u4f5c\u7528&#xff1a;\u7f13\u89e3\u5185\u90e8\u534f\u53d8\u91cf\u504f\u79fb&#xff08;ICS&#xff09;&#xff0c;\u5c06\u8f93\u5165\u5206\u5e03\u62c9\u5230\u6fc0\u6d3b\u51fd\u6570\u975e\u9971\u548c\u533a&#xff0c;\u52a0\u901f\u6536\u655b\u3001\u964d\u4f4e\u521d\u59cb\u5316\u654f\u611f\u5ea6\u3002<\/p>\n<p>\u8868\u683c<\/p>\n<table>\n<tr>\u65b9\u6cd5\u5f52\u4e00\u5316\u7ef4\u5ea6\u4f9d\u8d56 Batch Size\u6838\u5fc3\u7279\u70b9\u9002\u7528\u573a\u666f<\/tr>\n<tbody>\n<tr>\n<td>BatchNorm (BN)<\/td>\n<td>[N, H, W]&#xff0c;\u9010\u901a\u9053<\/td>\n<td>\u662f<\/td>\n<td>\u6279\u6b21\u5185\u7edf\u8ba1&#xff0c;\u6709\u8f7b\u5fae\u6b63\u5219\u6548\u679c<\/td>\n<td>\u5927 Batch \u7684 CNN \u5206\u7c7b<\/td>\n<\/tr>\n<tr>\n<td>LayerNorm (LN)<\/td>\n<td>[C, H, W]&#xff0c;\u9010\u6837\u672c<\/td>\n<td>\u5426<\/td>\n<td>\u6837\u672c\u5185\u7edf\u8ba1&#xff0c;\u4e0d\u53d7\u6279\u6b21\u5f71\u54cd<\/td>\n<td>Transformer\u3001RNN<\/td>\n<\/tr>\n<tr>\n<td>InstanceNorm (IN)<\/td>\n<td>[H, W]&#xff0c;\u9010\u6837\u672c\u9010\u901a\u9053<\/td>\n<td>\u5426<\/td>\n<td>\u5173\u6ce8\u7eb9\u7406\u98ce\u683c&#xff0c;\u6d88\u9664\u4e2a\u4f53\u4eae\u5ea6\u5dee\u5f02<\/td>\n<td>\u98ce\u683c\u8fc1\u79fb\u3001\u56fe\u50cf\u751f\u6210<\/td>\n<\/tr>\n<tr>\n<td>GroupNorm (GN)<\/td>\n<td>\u901a\u9053\u5206\u7ec4\u540e\u7ec4\u5185\u5f52\u4e00\u5316<\/td>\n<td>\u5426<\/td>\n<td>\u5c0f\u6279\u6b21\u4e0b\u6027\u80fd\u7a33\u5b9a<\/td>\n<td>\u68c0\u6d4b\u3001\u5206\u5272\u7b49\u5927 Batch \u53d7\u9650\u573a\u666f<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>BN \u8bad\u7ec3\u4e0e\u63a8\u7406\u5dee\u5f02<\/p>\n<ul>\n<li>\u8bad\u7ec3&#xff1a;\u7528\u5f53\u524d Batch \u7684\u5747\u503c\u3001\u65b9\u5dee\u505a\u5f52\u4e00\u5316&#xff0c;\u540c\u65f6\u901a\u8fc7\u6ed1\u52a8\u5e73\u5747\u66f4\u65b0\u5168\u5c40\u7edf\u8ba1\u91cf\u3002<\/li>\n<li>\u63a8\u7406&#xff1a;\u7528\u8bad\u7ec3\u9636\u6bb5\u79ef\u7d2f\u7684\u5168\u5c40\u5747\u503c\u3001\u65b9\u5dee&#xff0c;\u4fdd\u8bc1\u8f93\u51fa\u786e\u5b9a\u6027\u3002<\/li>\n<li>\u53ef\u5b66\u4e60\u53c2\u6570\\\\(\\\\gamma\\\\)&#xff08;\u7f29\u653e&#xff09;\u548c\\\\(\\\\beta\\\\)&#xff08;\u504f\u79fb&#xff09;&#xff1a;\u6062\u590d\u7f51\u7edc\u7684\u8868\u8fbe\u80fd\u529b&#xff0c;\u907f\u514d\u5f52\u4e00\u5316\u7834\u574f\u7279\u5f81\u3002<\/li>\n<\/ul>\n<h4>6. \u6b63\u5219\u5316\u4e0e\u8fc7\u62df\u5408<\/h4>\n<p>\u8fc7\u62df\u5408\u672c\u8d28&#xff1a;\u6a21\u578b\u5b66\u4e60\u5230\u8bad\u7ec3\u96c6\u7684\u566a\u58f0\u800c\u975e\u901a\u7528\u89c4\u5f8b&#xff0c;\u8bad\u7ec3\u8bef\u5dee\u8fdc\u5c0f\u4e8e\u6d4b\u8bd5\u8bef\u5dee\u3002<\/p>\n<p>\u89e3\u51b3\u65b9\u6848<\/p>\n<li>\u6570\u636e\u5c42\u9762&#xff1a;\u6570\u636e\u589e\u5f3a\u3001\u6269\u5145\u6570\u636e\u96c6\u3001\u6570\u636e\u6e05\u6d17\u53bb\u566a\u3002<\/li>\n<li>\u6a21\u578b\u5c42\u9762&#xff1a;\u964d\u4f4e\u6a21\u578b\u590d\u6742\u5ea6\u3001\u52a0\u5165\u6b63\u5219\u9879\u3001Dropout\u3001BN\u3002<\/li>\n<li>\u8bad\u7ec3\u5c42\u9762&#xff1a;\u65e9\u505c&#xff08;Early Stopping&#xff09;\u3001\u6743\u91cd\u8870\u51cf\u3001\u5b66\u4e60\u7387\u8c03\u4f18\u3002<\/li>\n<li>\u96c6\u6210\u5c42\u9762&#xff1a;\u591a\u6a21\u578b\u878d\u5408\u3001\u6295\u7968\u5e73\u5747\u3002<\/li>\n<p>L1 vs L2 \u6b63\u5219\u5316<\/p>\n<ul>\n<li>L1&#xff08;Lasso&#xff09;&#xff1a;\u635f\u5931\u52a0\\\\(\\\\lambda\\\\|w\\\\|_1\\\\)&#xff0c;\u4ea7\u751f\u7a00\u758f\u89e3&#xff0c;\u53ef\u7528\u4e8e\u7279\u5f81\u9009\u62e9&#xff1b;\u96f6\u70b9\u4e0d\u53ef\u5bfc\u3002<\/li>\n<li>L2&#xff08;Ridge&#xff09;&#xff1a;\u635f\u5931\u52a0\\\\(\\\\lambda\\\\|w\\\\|_2^2\\\\)&#xff0c;\u6743\u91cd\u5e73\u6ed1\u8870\u51cf&#xff0c;\u9632\u6b62\u8fc7\u62df\u5408&#xff1b;\u5904\u5904\u53ef\u5bfc&#xff0c;\u66f4\u5e38\u7528\u3002<\/li>\n<li>\u672c\u8d28\u5dee\u5f02&#xff1a;L1 \u68af\u5ea6\u4e3a\u5e38\u6570&#xff0c;\u6743\u91cd\u6613\u88ab\u538b\u5230 0&#xff1b;L2 \u68af\u5ea6\u4e0e\u6743\u91cd\u6210\u6b63\u6bd4&#xff0c;\u6743\u91cd\u8d8b\u8fd1\u4e8e 0 \u4f46\u4e0d\u4e3a 0\u3002<\/li>\n<\/ul>\n<p>Dropout<\/p>\n<ul>\n<li>\u8bad\u7ec3&#xff1a;\u4ee5\u6982\u7387p\u968f\u673a\u5931\u6d3b\u795e\u7ecf\u5143&#xff0c;\u8f93\u51fa\u7f29\u653e\\\\(1\/(1-p)\\\\)\u4fdd\u8bc1\u671f\u671b\u4e0d\u53d8\u3002<\/li>\n<li>\u63a8\u7406&#xff1a;\u6240\u6709\u795e\u7ecf\u5143\u6fc0\u6d3b&#xff0c;\u6743\u91cd\u65e0\u9700\u989d\u5916\u7f29\u653e&#xff08;\u8bad\u7ec3\u65f6\u5df2\u505a\u53cd\u5411\u7f29\u653e&#xff09;\u3002<\/li>\n<li>\u4f5c\u7528&#xff1a;\u7b49\u6548\u4e8e\u96c6\u6210\u5927\u91cf\u5b50\u7f51\u7edc&#xff0c;\u6253\u7834\u795e\u7ecf\u5143\u5171\u9002\u5e94&#xff0c;\u7f13\u89e3\u8fc7\u62df\u5408\u3002<\/li>\n<li>\u53d8\u4f53&#xff1a;DropPath&#xff08;\u968f\u673a\u4e22\u5f03\u6b8b\u5dee\u8def\u5f84&#xff09;\u3001DropBlock&#xff08;\u968f\u673a\u4e22\u5f03\u7279\u5f81\u5757\u533a\u57df&#xff0c;\u9002\u914d CNN&#xff09;\u3002<\/li>\n<\/ul>\n<h3>\u4e8c\u3001\u5377\u79ef\u795e\u7ecf\u7f51\u7edc&#xff08;CNN&#xff09;\u57fa\u7840<\/h3>\n<h4>1. \u5377\u79ef\u6838\u5fc3\u8ba1\u7b97<\/h4>\n<ul>\n<li>\u8f93\u51fa\u5c3a\u5bf8\u516c\u5f0f&#xff1a;\\\\(O &#061; \\\\frac{I &#8211; K &#043; 2P}{S} &#043; 1\\\\) I\u8f93\u5165\u5c3a\u5bf8&#xff0c;K\u5377\u79ef\u6838\u5927\u5c0f&#xff0c;P\u586b\u5145&#xff0c;S\u6b65\u957f<\/li>\n<li>\u53c2\u6570\u91cf&#xff1a;\\\\(K\\\\times K\\\\times C_{in}\\\\times C_{out} &#043; C_{out}\\\\)&#xff08;\u504f\u7f6e&#xff09;<\/li>\n<li>\u8ba1\u7b97\u91cf&#xff08;FLOPs&#xff09;&#xff1a;\\\\(O\\\\times O\\\\times K\\\\times K\\\\times C_{in}\\\\times C_{out}\\\\)<\/li>\n<\/ul>\n<p>Padding \u4f5c\u7528&#xff1a;\u4fdd\u6301\u7279\u5f81\u56fe\u5c3a\u5bf8\u3001\u4fdd\u7559\u8fb9\u7f18\u4fe1\u606f&#xff1b;Stride \u4f5c\u7528&#xff1a;\u4e0b\u91c7\u6837\u3001\u6269\u5927\u611f\u53d7\u91ce\u3001\u51cf\u5c11\u8ba1\u7b97\u91cf\u3002<\/p>\n<h4>2. \u7ecf\u5178\u5377\u79ef\u53d8\u4f53<\/h4>\n<h5>1\u00d71 \u9010\u70b9\u5377\u79ef<\/h5>\n<p>\u6838\u5fc3\u4f5c\u7528&#xff1a;<\/p>\n<ul>\n<li>\u901a\u9053\u5347\u7ef4 \/ \u964d\u7ef4&#xff0c;\u5927\u5e45\u51cf\u5c11\u53c2\u6570\u91cf\u4e0e\u8ba1\u7b97\u91cf<\/li>\n<li>\u8de8\u901a\u9053\u4fe1\u606f\u4ea4\u4e92\u4e0e\u878d\u5408<\/li>\n<li>\u914d\u5408\u6fc0\u6d3b\u51fd\u6570\u589e\u52a0\u975e\u7ebf\u6027<\/li>\n<li>\u66ff\u4ee3\u5168\u8fde\u63a5\u5c42&#xff08;\u5168\u5c40\u5e73\u5747\u6c60\u5316 &#043; 1\u00d71 \u5377\u79ef&#xff09;<\/li>\n<\/ul>\n<h5>\u6df1\u5ea6\u53ef\u5206\u79bb\u5377\u79ef<\/h5>\n<p>\u5206\u4e3a\u4e24\u6b65&#xff1a;<\/p>\n<li>\u6df1\u5ea6\u5377\u79ef&#xff1a;\u6bcf\u4e2a\u8f93\u5165\u901a\u9053\u72ec\u7acb\u5377\u79ef&#xff0c;\u901a\u9053\u6570\u4e0d\u53d8<\/li>\n<li>\u9010\u70b9\u5377\u79ef&#xff1a;1\u00d71 \u5377\u79ef\u878d\u5408\u901a\u9053\u4fe1\u606f<\/li>\n<ul>\n<li>\u53c2\u6570\u91cf\u4e3a\u666e\u901a\u5377\u79ef\u7684\\\\(\\\\frac{1}{K^2}&#043;\\\\frac{1}{C_{out}}\\\\)&#xff0c;\u8f7b\u91cf\u5316\u7f51\u7edc\u6838\u5fc3&#xff08;MobileNet \u7cfb\u5217&#xff09;\u3002<\/li>\n<\/ul>\n<h5>\u7a7a\u6d1e\u5377\u79ef&#xff08;\u81a8\u80c0\u5377\u79ef&#xff09;<\/h5>\n<p>\u5377\u79ef\u6838\u5185\u63d2\u5165\u7a7a\u6d1e&#xff0c;\u76f8\u540c\u53c2\u6570\u91cf\u4e0b\u6269\u5927\u611f\u53d7\u91ce&#xff0c;\u4e0d\u964d\u4f4e\u7279\u5f81\u56fe\u5206\u8fa8\u7387\u3002<\/p>\n<ul>\n<li>\u95ee\u9898&#xff1a;\u7f51\u683c\u6548\u5e94&#xff0c;\u91c7\u6837\u4e0d\u8fde\u7eed\u4e22\u5931\u5c40\u90e8\u4fe1\u606f\u3002<\/li>\n<li>\u89e3\u51b3&#xff1a;\u6df7\u5408\u81a8\u80c0\u5377\u79ef&#xff08;HDC&#xff09;&#xff0c;\u7ea7\u8054\u4e0d\u540c\u81a8\u80c0\u7387\u4e14\u65e0\u516c\u7ea6\u6570\u3002<\/li>\n<li>\u4ee3\u8868&#xff1a;DeepLab \u7cfb\u5217\u8bed\u4e49\u5206\u5272\u3002<\/li>\n<\/ul>\n<h5>\u5206\u7ec4\u5377\u79ef<\/h5>\n<p>\u8f93\u5165\u901a\u9053\u5206\u4e3a G \u7ec4&#xff0c;\u6bcf\u7ec4\u72ec\u7acb\u5377\u79ef\u540e\u62fc\u63a5\u3002\u53c2\u6570\u91cf\u964d\u4f4e\u4e3a\u539f\u6765\u7684\\\\(1\/G\\\\)&#xff0c;\u589e\u52a0\u7279\u5f81\u591a\u6837\u6027\u3002<\/p>\n<ul>\n<li>\u8fdb\u9636&#xff1a;\u901a\u9053\u6df7\u6d17&#xff08;Channel Shuffle&#xff09;&#xff0c;\u89e3\u51b3\u5206\u7ec4\u540e\u901a\u9053\u95f4\u65e0\u4fe1\u606f\u4ea4\u4e92\u7684\u95ee\u9898&#xff08;ShuffleNet&#xff09;\u3002<\/li>\n<\/ul>\n<h5>\u8f6c\u7f6e\u5377\u79ef&#xff08;\u53cd\u5377\u79ef&#xff09;<\/h5>\n<p>\u901a\u8fc7\u8865\u96f6 &#043; \u666e\u901a\u5377\u79ef\u5b9e\u73b0\u4e0a\u91c7\u6837&#xff0c;\u5e76\u975e\u5377\u79ef\u7684\u6570\u5b66\u9006\u8fd0\u7b97\u3002<\/p>\n<ul>\n<li>\u8f93\u51fa\u5c3a\u5bf8&#xff1a;\\\\(O &#061; (I-1)\\\\times S &#8211; 2P &#043; K &#043; \\\\text{OutputPad}\\\\)<\/li>\n<li>\u95ee\u9898&#xff1a;\u68cb\u76d8\u6548\u5e94&#xff0c;\u8f93\u51fa\u5b58\u5728\u7f51\u683c\u4f2a\u5f71\u3002<\/li>\n<li>\u89e3\u51b3&#xff1a;\u5377\u79ef\u6838\u5c3a\u5bf8\u53ef\u88ab\u6b65\u957f\u6574\u9664&#xff0c;\u6216\u5148\u7528\u53cc\u7ebf\u6027\u63d2\u503c\u4e0a\u91c7\u6837\u518d\u505a\u666e\u901a\u5377\u79ef\u3002<\/li>\n<\/ul>\n<h4>3. \u611f\u53d7\u91ce<\/h4>\n<p>\u5b9a\u4e49&#xff1a;\u7279\u5f81\u56fe\u4e0a\u4e00\u4e2a\u50cf\u7d20\u5bf9\u5e94\u539f\u59cb\u8f93\u5165\u56fe\u50cf\u7684\u533a\u57df\u5927\u5c0f\u3002<\/p>\n<ul>\n<li>\u524d\u5411\u8ba1\u7b97\u516c\u5f0f&#xff1a; \\\\(RF_0 &#061; 1\\\\) \\\\(RF_i &#061; RF_{i-1} &#043; (K_i &#8211; 1) \\\\times \\\\prod_{j&#061;0}^{i-1} S_j\\\\)<\/li>\n<li>\u6709\u6548\u611f\u53d7\u91ce&#xff1a;\u5b9e\u9645\u5bf9\u8f93\u51fa\u6709\u8d21\u732e\u7684\u533a\u57df\u5c0f\u4e8e\u7406\u8bba\u611f\u53d7\u91ce&#xff0c;\u4e2d\u5fc3\u6743\u91cd\u9ad8\u3001\u8fb9\u7f18\u6743\u91cd\u4f4e\u3002<\/li>\n<\/ul>\n<h4>4. \u7ecf\u5178 CNN Backbone \u5168\u89e3\u6790<\/h4>\n<h5>AlexNet&#xff08;2012&#xff09;<\/h5>\n<p>\u6df1\u5ea6\u5b66\u4e60\u91cc\u7a0b\u7891&#xff0c;ImageNet \u51a0\u519b\u3002<\/p>\n<ul>\n<li>\u521b\u65b0&#xff1a;ReLU \u6fc0\u6d3b\u3001Dropout\u3001\u6570\u636e\u589e\u5f3a\u3001\u53cc GPU \u5206\u7ec4\u5377\u79ef\u3001\u91cd\u53e0\u6c60\u5316\u3002<\/li>\n<\/ul>\n<h5>VGGNet&#xff08;2014&#xff09;<\/h5>\n<ul>\n<li>\u6838\u5fc3&#xff1a;\u5806\u53e0 3\u00d73 \u5c0f\u5377\u79ef\u6838\u66ff\u4ee3\u5927\u5377\u79ef\u6838\u30022 \u4e2a 3\u00d73 \u611f\u53d7\u91ce\u7b49\u4ef7\u4e8e 1 \u4e2a 5\u00d75&#xff0c;\u53c2\u6570\u91cf\u66f4\u5c11\u3001\u975e\u7ebf\u6027\u66f4\u591a\u3002<\/li>\n<li>\u7ed3\u6784\u89c4\u6574&#xff0c;VGG16\/19 \u6700\u5e38\u7528&#xff1b;\u7f3a\u70b9&#xff1a;\u5168\u8fde\u63a5\u5c42\u53c2\u6570\u91cf\u5de8\u5927\u3002<\/li>\n<\/ul>\n<h5>GoogLeNet\/Inception v1&#xff08;2014&#xff09;<\/h5>\n<ul>\n<li>Inception \u6a21\u5757&#xff1a;\u5e76\u884c 1\u00d71\u30013\u00d73\u30015\u00d75 \u5377\u79ef &#043; \u6c60\u5316&#xff0c;\u591a\u5c3a\u5ea6\u7279\u5f81\u878d\u5408\u3002<\/li>\n<li>1\u00d71 \u5377\u79ef\u964d\u7ef4\u538b\u7f29\u8ba1\u7b97\u91cf&#xff1b;\u5168\u5c40\u5e73\u5747\u6c60\u5316\u66ff\u4ee3\u5168\u8fde\u63a5\u5c42&#xff1b;\u8f85\u52a9\u5206\u7c7b\u5668\u7f13\u89e3\u68af\u5ea6\u6d88\u5931\u3002<\/li>\n<\/ul>\n<h5>ResNet&#xff08;2015&#xff09;<\/h5>\n<p>\u89e3\u51b3\u6df1\u5c42\u7f51\u7edc\u9000\u5316\u95ee\u9898&#xff0c;\u4f55\u607a\u660e\u4ee3\u8868\u4f5c\u3002<\/p>\n<ul>\n<li>\u6838\u5fc3&#xff1a;\u6b8b\u5dee\u8fde\u63a5 \\\\(H(x) &#061; F(x) &#043; x\\\\)&#xff0c;\u5b66\u4e60\u6b8b\u5dee\u6620\u5c04\u6bd4\u76f4\u63a5\u5b66\u4e60\u6052\u7b49\u6620\u5c04\u66f4\u5bb9\u6613\u3002<\/li>\n<li>\u53cd\u5411\u4f20\u64ad\u65f6\u68af\u5ea6\u53ef\u901a\u8fc7 shortcut \u76f4\u63a5\u56de\u4f20&#xff0c;\u4ece\u6839\u672c\u4e0a\u7f13\u89e3\u68af\u5ea6\u6d88\u5931\u3002<\/li>\n<li>\u4e24\u79cd\u6b8b\u5dee\u5757&#xff1a;\n<ul>\n<li>BasicBlock&#xff1a;\u4e24\u4e2a 3\u00d73 \u5377\u79ef&#xff0c;\u7528\u4e8e ResNet18\/34<\/li>\n<li>Bottleneck&#xff1a;1\u00d71 \u964d\u7ef4 &#043; 3\u00d73 \u5377\u79ef &#043; 1\u00d71 \u5347\u7ef4&#xff0c;\u51cf\u5c11\u8ba1\u7b97\u91cf&#xff0c;\u7528\u4e8e ResNet50\/101\/152<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<h5>MobileNet \u7cfb\u5217<\/h5>\n<ul>\n<li>v1&#xff1a;\u6df1\u5ea6\u53ef\u5206\u79bb\u5377\u79ef&#xff0c;\u5bbd\u5ea6\u56e0\u5b50\u3001\u5206\u8fa8\u7387\u56e0\u5b50\u63a7\u5236\u6a21\u578b\u5927\u5c0f\u3002<\/li>\n<li>v2&#xff1a;\u5012\u6b8b\u5dee\u7ed3\u6784&#xff08;\u5347\u7ef4 &#8211; \u5377\u79ef &#8211; \u964d\u7ef4&#xff09;&#043; Linear Bottleneck&#xff08;\u5c3e\u90e8\u4e0d\u7528 ReLU&#xff0c;\u4fdd\u62a4\u4f4e\u7ef4\u7279\u5f81&#xff09;\u3002<\/li>\n<li>v3&#xff1a;NAS \u641c\u7d22\u67b6\u6784&#xff0c;h-swish \u6fc0\u6d3b&#xff0c;SE \u6ce8\u610f\u529b\u6a21\u5757\u3002<\/li>\n<\/ul>\n<h5>EfficientNet<\/h5>\n<p>\u590d\u5408\u7f29\u653e\u7b56\u7565&#xff1a;\u540c\u65f6\u5747\u8861\u7f29\u653e\u7f51\u7edc\u6df1\u5ea6\u3001\u5bbd\u5ea6\u3001\u8f93\u5165\u5206\u8fa8\u7387&#xff0c;\u7528 NAS \u641c\u7d22\u6700\u4f18\u6bd4\u4f8b&#xff0c;\u7cbe\u5ea6\u4e0e\u6548\u7387\u5e73\u8861\u6781\u4f73\u3002<\/p>\n<h3>\u4e09\u3001\u76ee\u6807\u68c0\u6d4b\u7b97\u6cd5<\/h3>\n<h4>1. \u6838\u5fc3\u6982\u5ff5\u4e0e\u8bc4\u4ef7\u6307\u6807<\/h4>\n<ul>\n<li>Anchor&#xff1a;\u9884\u8bbe\u7684\u591a\u5c3a\u5ea6\u3001\u591a\u5bbd\u9ad8\u6bd4\u57fa\u51c6\u6846&#xff0c;\u4f5c\u4e3a\u6846\u56de\u5f52\u7684\u521d\u59cb\u53c2\u8003\u3002<\/li>\n<li>\u6b63\u8d1f\u6837\u672c&#xff1a;\u4e0e GT \u6846 IoU \u5927\u4e8e\u9608\u503c\u4e3a\u6b63\u6837\u672c&#xff0c;\u5c0f\u4e8e\u9608\u503c\u4e3a\u8d1f\u6837\u672c\u3002<\/li>\n<li>mAP&#xff1a;\u6240\u6709\u7c7b\u522b AP&#xff08;PR \u66f2\u7ebf\u4e0b\u9762\u79ef&#xff09;\u7684\u5e73\u5747\u503c\u3002COCO \u6807\u51c6\u4e3a IoU 0.5~0.95 \u6b65\u957f 0.05 \u7684\u5e73\u5747 AP&#xff0c;\u66f4\u4e25\u683c\u3002<\/li>\n<li>FPS&#xff1a;\u6bcf\u79d2\u5904\u7406\u5e27\u6570&#xff0c;\u8861\u91cf\u63a8\u7406\u901f\u5ea6\u3002<\/li>\n<\/ul>\n<h4>2. \u4e24\u9636\u6bb5\u68c0\u6d4b\u7b97\u6cd5\u6f14\u8fdb<\/h4>\n<h5>R-CNN<\/h5>\n<ul>\n<li>\u6d41\u7a0b&#xff1a;\u9009\u62e9\u6027\u641c\u7d22\u751f\u6210 2000 \u5019\u9009\u6846\u2192\u7f29\u653e\u56fa\u5b9a\u5c3a\u5bf8\u2192\u9010\u4e2a\u8fc7 CNN \u63d0\u7279\u5f81\u2192SVM \u5206\u7c7b &#043; \u56de\u5f52\u5fae\u8c03\u3002<\/li>\n<li>\u7f3a\u70b9&#xff1a;\u91cd\u590d\u8ba1\u7b97\u3001\u901f\u5ea6\u6781\u6162\u3001\u591a\u9636\u6bb5\u8bad\u7ec3\u3002<\/li>\n<\/ul>\n<h5>Fast R-CNN<\/h5>\n<ul>\n<li>\u6539\u8fdb&#xff1a;\u6574\u5f20\u56fe\u8fc7 Backbone \u5171\u4eab\u5377\u79ef\u8ba1\u7b97&#xff1b;RoI Pooling \u5c06\u5019\u9009\u6846\u7279\u5f81\u6c60\u5316\u4e3a\u56fa\u5b9a\u5c3a\u5bf8&#xff1b;\u7aef\u5230\u7aef\u591a\u4efb\u52a1\u8bad\u7ec3\u3002<\/li>\n<li>\u74f6\u9888&#xff1a;\u5019\u9009\u6846\u4ecd\u7531\u9009\u62e9\u6027\u641c\u7d22\u751f\u6210\u3002<\/li>\n<\/ul>\n<h5>Faster R-CNN<\/h5>\n<p>\u5168\u5377\u79ef\u7aef\u5230\u7aef\u68c0\u6d4b&#xff0c;\u7528 RPN \u66ff\u4ee3\u9009\u62e9\u6027\u641c\u7d22\u3002<\/p>\n<ul>\n<li>\u5b8c\u6574\u6d41\u7a0b&#xff1a;\n<li>Backbone \u63d0\u53d6\u6574\u56fe\u7279\u5f81\u56fe<\/li>\n<li>RPN \u533a\u57df\u5efa\u8bae\u7f51\u7edc&#xff1a;\u6bcf\u4e2a\u7279\u5f81\u70b9\u751f\u6210 9 \u4e2a Anchor&#xff0c;\u8f93\u51fa\u524d\u666f \/ \u80cc\u666f\u5206\u7c7b &#043; \u6846\u504f\u79fb&#xff0c;NMS \u540e\u8f93\u51fa\u5019\u9009\u6846<\/li>\n<li>RoI Pooling&#xff1a;\u5019\u9009\u6846\u5bf9\u5e94\u7279\u5f81\u533a\u57df\u6c60\u5316\u4e3a 7\u00d77 \u56fa\u5b9a\u5c3a\u5bf8<\/li>\n<li>\u5168\u8fde\u63a5\u5934\u8f93\u51fa\u7c7b\u522b\u5206\u7c7b &#043; \u6846\u7cbe\u4fee\u56de\u5f52<\/li>\n<\/li>\n<li>RPN \u6b63\u8d1f\u6837\u672c\u89c4\u5219&#xff1a;\u4e0e GT IoU \u6700\u5927\u7684 Anchor&#xff0c;\u6216 IoU&gt;0.7 \u4e3a\u6b63&#xff1b;IoU&lt;0.3 \u4e3a\u8d1f\u3002<\/li>\n<\/ul>\n<h5>Mask R-CNN<\/h5>\n<p>\u5b9e\u4f8b\u5206\u5272\u6807\u6746&#xff0c;\u5728 Faster R-CNN \u57fa\u7840\u4e0a&#xff1a;<\/p>\n<ul>\n<li>RoI Align \u66ff\u4ee3 RoI Pooling&#xff0c;\u53cc\u7ebf\u6027\u63d2\u503c\u907f\u514d\u4e24\u6b21\u91cf\u5316\u8bef\u5dee&#xff0c;\u4f4d\u7f6e\u7cbe\u5ea6\u5927\u5e45\u63d0\u5347\u3002<\/li>\n<li>\u65b0\u589e Mask \u5206\u5272\u5206\u652f&#xff0c;\u9010\u50cf\u7d20\u9884\u6d4b\u5b9e\u4f8b\u63a9\u7801\u3002<\/li>\n<\/ul>\n<h5>FPN \u7279\u5f81\u91d1\u5b57\u5854\u7f51\u7edc<\/h5>\n<ul>\n<li>\u7ed3\u6784&#xff1a;\u81ea\u4e0b\u800c\u4e0a&#xff08;Backbone \u524d\u5411&#xff09;&#043; \u81ea\u4e0a\u800c\u4e0b&#xff08;\u4e0a\u91c7\u6837&#xff09;&#043; \u6a2a\u5411\u8fde\u63a5\u3002<\/li>\n<li>\u4ef7\u503c&#xff1a;\u878d\u5408\u9ad8\u5c42\u8bed\u4e49\u4e0e\u4f4e\u5c42\u7ec6\u8282&#xff0c;\u591a\u5c3a\u5ea6\u7279\u5f81\u56fe\u5206\u522b\u68c0\u6d4b\u4e0d\u540c\u5927\u5c0f\u76ee\u6807&#xff0c;\u663e\u8457\u63d0\u5347\u5c0f\u76ee\u6807\u6027\u80fd\u3002<\/li>\n<\/ul>\n<h4>3. \u4e00\u9636\u6bb5\u68c0\u6d4b\u7b97\u6cd5\u6f14\u8fdb<\/h4>\n<h5>YOLO v1<\/h5>\n<ul>\n<li>\u601d\u60f3&#xff1a;\u56fe\u50cf\u5212\u5206\u4e3a S\u00d7S \u7f51\u683c&#xff0c;\u6bcf\u4e2a\u7f51\u683c\u9884\u6d4b B \u4e2a\u6846\u4e0e\u7c7b\u522b&#xff0c;\u7aef\u5230\u7aef\u4e00\u6b21\u8f93\u51fa\u3002<\/li>\n<li>\u4f18\u70b9&#xff1a;\u901f\u5ea6\u6781\u5feb&#xff0c;\u5168\u5c40\u611f\u53d7\u91ce&#xff0c;\u80cc\u666f\u8bef\u68c0\u5c11\u3002<\/li>\n<li>\u7f3a\u70b9&#xff1a;\u5c0f\u76ee\u6807\u5dee\u3001\u5b9a\u4f4d\u7c97\u7cd9\u3001\u6bcf\u4e2a\u7f51\u683c\u53ea\u80fd\u9884\u6d4b\u4e00\u7c7b\u3002<\/li>\n<\/ul>\n<h5>SSD<\/h5>\n<p>\u591a\u5c3a\u5ea6\u7279\u5f81\u56fe\u68c0\u6d4b&#xff0c;\u4f4e\u5c42\u7279\u5f81\u68c0\u6d4b\u5c0f\u76ee\u6807&#xff0c;\u9ad8\u5c42\u7279\u5f81\u68c0\u6d4b\u5927\u76ee\u6807&#xff1b;\u6bcf\u4e2a\u7279\u5f81\u70b9\u751f\u6210\u591a\u4e2a Default Box\u3002\u7cbe\u5ea6\u8d85\u8d8a YOLO v1&#xff0c;\u901f\u5ea6\u5feb\u4e8e\u4e24\u9636\u6bb5\u3002<\/p>\n<h5>RetinaNet<\/h5>\n<ul>\n<li>\u6838\u5fc3\u8d21\u732e&#xff1a;Focal Loss&#xff0c;\u89e3\u51b3\u4e00\u9636\u6bb5\u6b63\u8d1f\u6837\u672c\u6781\u5ea6\u4e0d\u5747\u8861&#xff08;1:1000&#xff09;\u95ee\u9898&#xff0c;\u4e00\u9636\u6bb5\u7cbe\u5ea6\u9996\u6b21\u8ffd\u5e73\u4e24\u9636\u6bb5\u3002<\/li>\n<li>\u67b6\u6784&#xff1a;ResNet&#043;FPN &#043; \u5206\u7c7b\u56de\u5f52\u5934\u3002<\/li>\n<\/ul>\n<h5>YOLO \u7cfb\u5217\u6f14\u8fdb<\/h5>\n<ul>\n<li>v2&#xff1a;Darknet19\u3001BN\u3001Anchor \u673a\u5236\u3001\u591a\u5c3a\u5ea6\u8bad\u7ec3\u3002<\/li>\n<li>v3&#xff1a;Darknet53\u3001FPN \u591a\u5c3a\u5ea6\u9884\u6d4b\u3001Logistic \u5206\u7c7b\u652f\u6301\u591a\u6807\u7b7e\u3002<\/li>\n<li>v4&#xff1a;CSPDarknet53\u3001PANet \u7279\u5f81\u878d\u5408\u3001Mosaic \u589e\u5f3a\u3001CIoU Loss\u3001\u5927\u91cf\u5de5\u7a0b Trick\u3002<\/li>\n<li>v8&#xff1a;C2f \u6a21\u5757\u3001\u89e3\u8026\u68c0\u6d4b\u5934\u3001Anchor-Free\u3001\u4efb\u52a1\u5bf9\u9f50\u5b66\u4e60&#xff08;TAL&#xff09;\u3002<\/li>\n<\/ul>\n<h4>4. Anchor-Free \u68c0\u6d4b\u7b97\u6cd5<\/h4>\n<p>\u65e0\u9700\u9884\u8bbe Anchor&#xff0c;\u907f\u514d\u8d85\u53c2\u8c03\u4f18&#xff0c;\u964d\u4f4e\u8ba1\u7b97\u590d\u6742\u5ea6\u3002<\/p>\n<ul>\n<li>CenterNet&#xff1a;\u68c0\u6d4b\u76ee\u6807\u4e2d\u5fc3\u70b9&#xff0c;\u9884\u6d4b\u4e2d\u5fc3\u70ed\u529b\u56fe\u3001\u5bbd\u9ad8\u3001\u504f\u79fb\u91cf\u3002\u540e\u5904\u7406\u7b80\u5355&#xff0c;\u65e0\u9700 NMS\u3002<\/li>\n<li>FCOS&#xff1a;\u9010\u50cf\u7d20\u9884\u6d4b&#xff0c;\u6bcf\u4e2a\u50cf\u7d20\u8f93\u51fa\u5230\u56db\u6761\u8fb9\u7684\u8ddd\u79bb &#043; \u4e2d\u5fc3\u5ea6&#xff08;Center-ness&#xff09;\u6291\u5236\u4f4e\u8d28\u91cf\u6846\u3002<\/li>\n<li>YOLOX&#xff1a;Anchor-Free &#043; \u89e3\u8026\u5934 &#043; SimOTA \u6b63\u6837\u672c\u5339\u914d &#043; \u5f3a\u6570\u636e\u589e\u5f3a\u3002<\/li>\n<\/ul>\n<h4>5. \u540e\u5904\u7406&#xff1a;NMS \u53ca\u5176\u53d8\u4f53<\/h4>\n<p>\u6807\u51c6 NMS \u6d41\u7a0b&#xff1a;<\/p>\n<li>\u6309\u5206\u7c7b\u5f97\u5206\u964d\u5e8f\u6392\u5217<\/li>\n<li>\u53d6\u6700\u9ad8\u5206\u6846&#xff0c;\u5220\u9664 IoU \u5927\u4e8e\u9608\u503c\u7684\u5176\u4f59\u6846<\/li>\n<li>\u91cd\u590d\u76f4\u81f3\u6240\u6709\u6846\u5904\u7406\u5b8c\u6bd5<\/li>\n<p>\u53d8\u4f53<\/p>\n<ul>\n<li>Soft-NMS&#xff1a;\u4e0d\u76f4\u63a5\u5220\u9664\u9ad8 IoU \u6846&#xff0c;\u800c\u662f\u8870\u51cf\u5176\u5f97\u5206&#xff0c;\u9002\u5408\u5bc6\u96c6\u76ee\u6807\u573a\u666f\u3002<\/li>\n<li>DIoU-NMS&#xff1a;\u7528 DIoU \u66ff\u4ee3 IoU&#xff0c;\u8003\u8651\u4e2d\u5fc3\u70b9\u8ddd\u79bb&#xff0c;\u66f4\u7b26\u5408\u51e0\u4f55\u76f4\u89c9\u3002<\/li>\n<\/ul>\n<h3>\u56db\u3001\u8bed\u4e49\u5206\u5272\u4e0e\u5b9e\u4f8b\u5206\u5272<\/h3>\n<h4>1. \u4efb\u52a1\u5206\u7c7b<\/h4>\n<ul>\n<li>\u8bed\u4e49\u5206\u5272&#xff1a;\u9010\u50cf\u7d20\u5206\u7c7b&#xff0c;\u540c\u7c7b\u76ee\u6807\u4e0d\u533a\u5206\u4e2a\u4f53\u3002<\/li>\n<li>\u5b9e\u4f8b\u5206\u5272&#xff1a;\u9010\u50cf\u7d20\u5206\u7c7b &#043; \u5b9e\u4f8b\u533a\u5206&#xff0c;\u517c\u987e\u68c0\u6d4b\u4e0e\u5206\u5272\u3002<\/li>\n<li>\u5168\u666f\u5206\u5272&#xff1a;\u6240\u6709\u50cf\u7d20\u5747\u5206\u914d\u7c7b\u522b &#043; \u5b9e\u4f8b ID&#xff0c;\u6db5\u76d6\u524d\u666f\u4e0e\u80cc\u666f\u3002<\/li>\n<\/ul>\n<h4>2. \u7ecf\u5178\u5206\u5272\u7f51\u7edc<\/h4>\n<h5>FCN&#xff08;2015&#xff09;<\/h5>\n<p>\u5168\u5377\u79ef\u7f51\u7edc\u5f00\u5c71\u4e4b\u4f5c\u3002\u7528\u5377\u79ef\u66ff\u4ee3\u5168\u8fde\u63a5&#xff0c;\u652f\u6301\u4efb\u610f\u5c3a\u5bf8\u8f93\u5165&#xff1b;\u8f6c\u7f6e\u5377\u79ef\u4e0a\u91c7\u6837&#xff1b;\u8df3\u5c42\u8fde\u63a5\u878d\u5408\u9ad8\u4f4e\u5c42\u7279\u5f81\u3002\u7f3a\u70b9&#xff1a;\u4e0a\u91c7\u6837\u7c97\u7cd9&#xff0c;\u8fb9\u754c\u7cbe\u5ea6\u5dee\u3002<\/p>\n<h5>U-Net&#xff08;2015&#xff09;<\/h5>\n<p>\u533b\u5b66\u5f71\u50cf\u5206\u5272\u4e8b\u5b9e\u6807\u51c6\u3002<\/p>\n<ul>\n<li>\u7ed3\u6784&#xff1a;\u7f16\u7801\u5668&#xff08;\u4e0b\u91c7\u6837\u63d0\u7279\u5f81&#xff09;&#043; \u89e3\u7801\u5668&#xff08;\u4e0a\u91c7\u6837\u6062\u590d\u5c3a\u5bf8&#xff09;&#043; \u8df3\u8dc3\u8fde\u63a5&#xff08;\u62fc\u63a5\u5bf9\u5e94\u5c42\u7279\u5f81&#xff09;\u3002<\/li>\n<li>\u7279\u70b9&#xff1a;U \u578b\u5bf9\u79f0\u7ed3\u6784&#xff0c;\u591a\u5c3a\u5ea6\u7279\u5f81\u878d\u5408&#xff0c;\u5c0f\u6570\u636e\u96c6\u4e0a\u8868\u73b0\u4f18\u5f02\u3002<\/li>\n<\/ul>\n<h5>DeepLab \u7cfb\u5217<\/h5>\n<ul>\n<li>v1&#xff1a;\u7a7a\u6d1e\u5377\u79ef\u6269\u5927\u611f\u53d7\u91ce &#043; \u5168\u8fde\u63a5 CRF \u540e\u5904\u7406\u3002<\/li>\n<li>v2&#xff1a;ASPP \u7a7a\u6d1e\u7a7a\u95f4\u91d1\u5b57\u5854\u6c60\u5316&#xff0c;\u591a\u81a8\u80c0\u7387\u5e76\u884c\u6355\u6349\u591a\u5c3a\u5ea6\u4e0a\u4e0b\u6587\u3002<\/li>\n<li>v3&#043;&#xff1a;\u7f16\u7801\u5668 &#8211; \u89e3\u7801\u5668\u67b6\u6784&#xff0c;ASPP \u878d\u5408\u9ad8\u5c42\u8bed\u4e49&#xff0c;\u8df3\u8dc3\u8fde\u63a5\u8865\u5145\u4f4e\u5c42\u7ec6\u8282\u3002<\/li>\n<\/ul>\n<h5>PSPNet<\/h5>\n<p>\u91d1\u5b57\u5854\u6c60\u5316\u6a21\u5757&#xff08;PPM&#xff09;&#xff0c;\u591a\u5c3a\u5ea6\u6c60\u5316\u878d\u5408\u5168\u5c40\u4e0a\u4e0b\u6587\u4fe1\u606f&#xff0c;\u89e3\u51b3\u573a\u666f\u89e3\u6790\u7684\u591a\u5c3a\u5ea6\u95ee\u9898\u3002<\/p>\n<h5>HRNet<\/h5>\n<p>\u5168\u7a0b\u4fdd\u6301\u9ad8\u5206\u8fa8\u7387\u7279\u5f81&#xff0c;\u5e76\u884c\u591a\u5206\u8fa8\u7387\u5206\u652f\u53cd\u590d\u4ea4\u4e92\u878d\u5408\u3002\u4f4d\u7f6e\u7cbe\u5ea6\u6781\u9ad8&#xff0c;\u9002\u7528\u4e8e\u59ff\u6001\u4f30\u8ba1\u3001\u533b\u5b66\u5206\u5272\u3002<\/p>\n<h5>SegFormer<\/h5>\n<p>\u7eaf Transformer \u5206\u5272\u67b6\u6784&#xff0c;\u5206\u5c42 Transformer \u7f16\u7801\u5668 &#043; \u8f7b\u91cf MLP \u89e3\u7801\u5668&#xff0c;\u517c\u987e\u7cbe\u5ea6\u4e0e\u901f\u5ea6\u3002<\/p>\n<h4>3. \u4e0a\u91c7\u6837\u6280\u672f<\/h4>\n<li>\u63d2\u503c\u6cd5&#xff1a;\u6700\u8fd1\u90bb\u3001\u53cc\u7ebf\u6027\u3001\u53cc\u4e09\u6b21\u3002\u65e0\u53c2\u6570\u3001\u901f\u5ea6\u5feb&#xff0c;\u4f46\u4e0d\u53ef\u5b66\u4e60&#xff0c;\u7ec6\u8282\u6062\u590d\u5dee\u3002<\/li>\n<li>\u8f6c\u7f6e\u5377\u79ef&#xff1a;\u53ef\u5b66\u4e60\u4e0a\u91c7\u6837&#xff0c;\u53c2\u6570\u591a&#xff0c;\u6613\u51fa\u73b0\u68cb\u76d8\u6548\u5e94\u3002<\/li>\n<li>\u50cf\u7d20\u6d17\u724c&#xff08;Pixel Shuffle&#xff09;&#xff1a;\u901a\u9053\u7ef4\u5ea6\u91cd\u6392\u5230\u7a7a\u95f4\u7ef4\u5ea6&#xff0c;\u65e0\u68cb\u76d8\u6548\u5e94&#xff0c;\u8d85\u5206\u8fa8\u7387\u5e38\u7528\u3002<\/li>\n<li>\u53cd\u6c60\u5316&#xff1a;\u5229\u7528\u6c60\u5316\u7d22\u5f15\u6062\u590d\u4f4d\u7f6e&#xff0c;\u65e0\u53c2\u6570\u4f46\u4fe1\u606f\u4e22\u5931\u591a\u3002<\/li>\n<h4>4. \u5206\u5272\u635f\u5931\u51fd\u6570<\/h4>\n<ul>\n<li>\u4ea4\u53c9\u71b5&#xff1a;\u50cf\u7d20\u7ea7\u5206\u7c7b\u635f\u5931&#xff0c;\u7b80\u5355\u901a\u7528&#xff0c;\u4f46\u6837\u672c\u4e0d\u5747\u8861\u65f6\u6548\u679c\u5dee\u3002<\/li>\n<li>Dice Loss&#xff1a;\\\\(Dice&#061;\\\\frac{2|A\\\\cap B|}{|A|&#043;|B|}\\\\)&#xff0c;\u76f4\u63a5\u4f18\u5316\u91cd\u53e0\u5ea6&#xff0c;\u9002\u914d\u524d\u666f\u80cc\u666f\u4e0d\u5747\u8861\u573a\u666f&#xff08;\u533b\u5b66\u5f71\u50cf&#xff09;\u3002\u8bad\u7ec3\u4e0d\u7a33\u5b9a&#xff0c;\u5e38\u4e0e\u4ea4\u53c9\u71b5\u6df7\u5408\u4f7f\u7528\u3002<\/li>\n<li>Lov\u00e1sz-Softmax Loss&#xff1a;\u76f4\u63a5\u4f18\u5316 IoU \u6307\u6807&#xff0c;\u57fa\u4e8e\u5b50\u6a21\u635f\u5931&#xff0c;\u7cbe\u5ea6\u66f4\u9ad8\u4f46\u8ba1\u7b97\u7a0d\u590d\u6742\u3002<\/li>\n<li>Focal Loss&#xff1a;\u89e3\u51b3\u96be\u6613\u6837\u672c\u4e0d\u5747\u8861\u3002<\/li>\n<\/ul>\n<h3>\u4e94\u3001\u56fe\u50cf\u751f\u6210\u6a21\u578b<\/h3>\n<h4>1. GAN \u751f\u6210\u5bf9\u6297\u7f51\u7edc<\/h4>\n<p>\u6838\u5fc3\u601d\u60f3&#xff1a;\u751f\u6210\u5668 G \u4e0e\u5224\u522b\u5668 D \u535a\u5f08\u8bad\u7ec3\u3002\u751f\u6210\u5668\u8f93\u5165\u566a\u58f0\u751f\u6210\u5047\u56fe\u9a97\u8fc7\u5224\u522b\u5668&#xff1b;\u5224\u522b\u5668\u533a\u5206\u771f\u5047\u6837\u672c\u3002<\/p>\n<ul>\n<li>\u635f\u5931&#xff1a;\\\\(\\\\min_G\\\\max_D V(D,G) &#061; \\\\mathbb{E}_x[\\\\log D(x)] &#043; \\\\mathbb{E}_z[\\\\log(1-D(G(z)))]\\\\)<\/li>\n<\/ul>\n<p>\u8bad\u7ec3\u96be\u70b9<\/p>\n<li>\u6a21\u5f0f\u5d29\u6e83&#xff1a;\u751f\u6210\u5668\u53ea\u4ea7\u51fa\u5c11\u6570\u6a21\u5f0f&#xff0c;\u591a\u6837\u6027\u4e0d\u8db3\u3002<\/li>\n<li>\u8bad\u7ec3\u4e0d\u7a33\u5b9a&#xff1a;G \u4e0e D \u96be\u4ee5\u5e73\u8861&#xff0c;\u5224\u522b\u5668\u8fc7\u5f3a\u5bfc\u81f4\u751f\u6210\u5668\u68af\u5ea6\u6d88\u5931\u3002<\/li>\n<p>\u7ecf\u5178\u53d8\u4f53<\/p>\n<ul>\n<li>DCGAN&#xff1a;\u6df1\u5ea6\u5377\u79ef GAN&#xff0c;\u7528\u8f6c\u7f6e\u5377\u79ef &#043; \u6b65\u957f\u5377\u79ef &#043; BN&#xff0c;\u5927\u5e45\u63d0\u5347\u7a33\u5b9a\u6027\u3002<\/li>\n<li>WGAN&#xff1a;\u7528\u63a8\u571f\u673a\u8ddd\u79bb\u66ff\u4ee3 JS \u6563\u5ea6&#xff0c;\u5206\u5e03\u4e0d\u91cd\u53e0\u65f6\u4ecd\u6709\u68af\u5ea6&#xff0c;\u89e3\u51b3\u8bad\u7ec3\u4e0d\u7a33\u5b9a\u4e0e\u6a21\u5f0f\u5d29\u6e83&#xff1b;\u6743\u91cd\u88c1\u526a\u7ea6\u675f Lipschitz \u6761\u4ef6\u3002<\/li>\n<li>WGAN-GP&#xff1a;\u7528\u68af\u5ea6\u60e9\u7f5a\u66ff\u4ee3\u6743\u91cd\u88c1\u526a&#xff0c;\u6548\u679c\u66f4\u7a33\u5b9a\u3001\u751f\u6210\u8d28\u91cf\u66f4\u9ad8\u3002<\/li>\n<li>CycleGAN&#xff1a;\u65e0\u914d\u5bf9\u56fe\u50cf\u98ce\u683c\u8fc1\u79fb&#xff0c;\u5faa\u73af\u4e00\u81f4\u6027\u635f\u5931\u4fdd\u8bc1\u5185\u5bb9\u4e0d\u53d8\u3002<\/li>\n<li>StyleGAN&#xff1a;\u98ce\u683c\u89e3\u8026&#xff0c;\u53ef\u7cbe\u51c6\u63a7\u5236\u751f\u6210\u56fe\u50cf\u7684\u5c5e\u6027\u4e0e\u98ce\u683c\u3002<\/li>\n<\/ul>\n<h4>2. \u6269\u6563\u6a21\u578b<\/h4>\n<p>\u6838\u5fc3\u539f\u7406&#xff1a;\u524d\u5411\u9010\u6b65\u52a0\u566a&#xff0c;\u53cd\u5411\u9010\u6b65\u53bb\u566a\u751f\u6210\u56fe\u50cf\u3002<\/p>\n<li>\u524d\u5411\u6269\u6563&#xff1a;T \u6b65\u9010\u6b65\u5411\u56fe\u50cf\u52a0\u5165\u9ad8\u65af\u566a\u58f0&#xff0c;\u6700\u7ec8\u53d8\u4e3a\u7eaf\u566a\u58f0&#xff0c;\u53ef\u76f4\u63a5\u8ba1\u7b97\u4efb\u610f\u65f6\u523b\u7684\u566a\u58f0\u56fe\u3002<\/li>\n<li>\u53cd\u5411\u53bb\u566a&#xff1a;\u795e\u7ecf\u7f51\u7edc\u9884\u6d4b\u6bcf\u4e00\u6b65\u7684\u566a\u58f0&#xff0c;\u4ece\u7eaf\u566a\u58f0\u9010\u6b65\u8fd8\u539f\u6e05\u6670\u56fe\u50cf\u3002<\/li>\n<li>\u8bad\u7ec3\u76ee\u6807&#xff1a;\u6700\u5c0f\u5316\u9884\u6d4b\u566a\u58f0\u4e0e\u771f\u5b9e\u566a\u58f0\u7684 MSE\u3002<\/li>\n<p>\u4ee3\u8868\u5de5\u4f5c&#xff1a;DDPM\u3001DDIM\u3001Stable Diffusion\u3002<\/p>\n<ul>\n<li>\u4f18\u70b9&#xff1a;\u751f\u6210\u8d28\u91cf\u9ad8\u3001\u591a\u6837\u6027\u597d\u3001\u8bad\u7ec3\u7a33\u5b9a\u3001\u65e0\u6a21\u5f0f\u5d29\u6e83\u3002<\/li>\n<li>\u7f3a\u70b9&#xff1a;\u539f\u751f\u63a8\u7406\u901f\u5ea6\u6162&#xff0c;\u9700\u591a\u6b65\u91c7\u6837&#xff1b;\u5df2\u6709 DDIM\u3001DPM-Solver \u7b49\u52a0\u901f\u7b97\u6cd5\u3002<\/li>\n<li>Stable Diffusion&#xff1a;\u5728\u9690\u7a7a\u95f4\u505a\u6269\u6563&#xff0c;\u5927\u5e45\u964d\u4f4e\u8ba1\u7b97\u91cf&#xff0c;\u652f\u6301\u6587\u751f\u56fe\u3001\u56fe\u751f\u56fe\u7b49\u4e30\u5bcc\u5e94\u7528\u3002<\/li>\n<\/ul>\n<h4>3. VAE \u53d8\u5206\u81ea\u7f16\u7801\u5668<\/h4>\n<ul>\n<li>\u7ed3\u6784&#xff1a;\u7f16\u7801\u5668\u5c06\u8f93\u5165\u6620\u5c04\u4e3a\u9690\u7a7a\u95f4\u7684\u5747\u503c\u4e0e\u65b9\u5dee&#xff0c;\u91cd\u53c2\u6570\u5316\u6280\u5de7\u91c7\u6837\u9690\u53d8\u91cf&#xff0c;\u89e3\u7801\u5668\u8fd8\u539f\u56fe\u50cf\u3002<\/li>\n<li>\u635f\u5931&#xff1a;\u91cd\u6784\u635f\u5931 &#043; KL \u6563\u5ea6&#xff08;\u7ea6\u675f\u9690\u5206\u5e03\u63a5\u8fd1\u6807\u51c6\u6b63\u6001&#xff09;\u3002<\/li>\n<li>\u7279\u70b9&#xff1a;\u6709\u663e\u5f0f\u9690\u7a7a\u95f4\u3001\u53ef\u89e3\u91ca\u3001\u751f\u6210\u7a33\u5b9a&#xff1b;\u4f46\u751f\u6210\u56fe\u50cf\u504f\u6a21\u7cca&#xff0c;\u8d28\u91cf\u5f31\u4e8e GAN \u4e0e\u6269\u6563\u6a21\u578b\u3002<\/li>\n<\/ul>\n<h3>\u516d\u3001\u89c6\u89c9 Transformer<\/h3>\n<h4>1. Transformer \u6838\u5fc3\u7ec4\u4ef6<\/h4>\n<ul>\n<li>\u81ea\u6ce8\u610f\u529b&#xff1a;\\\\(Attention(Q,K,V) &#061; \\\\text{softmax}(\\\\frac{QK^T}{\\\\sqrt{d_k}})V\\\\) Q\/K\/V \u7531\u8f93\u5165\u7ebf\u6027\u6295\u5f71\u5f97\u5230&#xff1b;\\\\(\\\\sqrt{d_k}\\\\)\u7f29\u653e\u9632\u6b62\u70b9\u79ef\u8fc7\u5927\u5bfc\u81f4 softmax \u9971\u548c\u3002\u5168\u5c40\u611f\u53d7\u91ce&#xff0c;\u5efa\u6a21\u957f\u8ddd\u79bb\u4f9d\u8d56\u3002<\/li>\n<li>\u591a\u5934\u6ce8\u610f\u529b&#xff1a;\u5c06 QKV \u62c6\u5206\u4e3a\u591a\u4e2a\u5934\u72ec\u7acb\u8ba1\u7b97\u6ce8\u610f\u529b\u518d\u62fc\u63a5&#xff0c;\u6355\u6349\u4e0d\u540c\u5b50\u7a7a\u95f4\u7279\u5f81\u3002<\/li>\n<li>\u4f4d\u7f6e\u7f16\u7801&#xff1a;Transformer \u65e0\u4f4d\u7f6e\u611f\u77e5&#xff0c;\u9700\u6ce8\u5165\u4f4d\u7f6e\u4fe1\u606f\u3002\u5206\u4e3a\u6b63\u5f26\u4f59\u5f26\u56fa\u5b9a\u7f16\u7801\u3001\u53ef\u5b66\u4e60\u4f4d\u7f6e\u7f16\u7801\u3002<\/li>\n<li>FFN \u524d\u9988\u7f51\u7edc&#xff1a;\u4e24\u5c42\u7ebf\u6027\u5c42 &#043; GELU \u6fc0\u6d3b&#xff0c;\u5347\u7ef4\u518d\u964d\u7ef4\u3002<\/li>\n<li>\u57fa\u7840\u5355\u5143&#xff1a;\u591a\u5934\u6ce8\u610f\u529b &#043; FFN&#xff0c;\u6bcf\u5c42\u5747\u6709\u6b8b\u5dee\u8fde\u63a5 &#043; LayerNorm\u3002<\/li>\n<\/ul>\n<h4>2. ViT \u89c6\u89c9 Transformer<\/h4>\n<ul>\n<li>\u6d41\u7a0b&#xff1a;\u56fe\u50cf\u5212\u5206\u4e3a Patch\u2192\u7ebf\u6027\u6295\u5f71\u4e3a Patch Embedding\u2192\u52a0\u5165\u4f4d\u7f6e\u7f16\u7801\u4e0e Class Token\u2192Transformer \u7f16\u7801\u5668\u2192Class Token \u8f93\u51fa\u5206\u7c7b\u3002<\/li>\n<li>\u4e0e CNN \u5bf9\u6bd4\n<ul>\n<li>\u4f18\u70b9&#xff1a;\u5168\u5c40\u5efa\u6a21\u80fd\u529b\u5f3a\u3001\u957f\u8ddd\u79bb\u4f9d\u8d56\u4f18\u3001\u5927\u6570\u636e\u4e0b\u6027\u80fd\u4e0a\u9650\u9ad8\u3001\u8fc1\u79fb\u6cdb\u5316\u6027\u597d\u3002<\/li>\n<li>\u7f3a\u70b9&#xff1a;\u8ba1\u7b97\u91cf\u4e0e\u5e8f\u5217\u957f\u5ea6\u5e73\u65b9\u6210\u6b63\u6bd4\u3001\u5c0f\u6570\u636e\u96c6\u6548\u679c\u5f31\u4e8e CNN\u3001\u7f3a\u4e4f\u5c40\u90e8\u6027\u5f52\u7eb3\u504f\u7f6e\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<h4>3. \u89c6\u89c9 Transformer \u53d8\u4f53<\/h4>\n<ul>\n<li>DeiT&#xff1a;\u77e5\u8bc6\u84b8\u998f\u65b9\u6848&#xff0c;\u7528 CNN \u6559\u5e08\u6a21\u578b\u6307\u5bfc ViT \u5b66\u751f&#xff0c;\u5c0f\u6570\u636e\u96c6\u4e5f\u53ef\u8bad\u7ec3\u3002<\/li>\n<li>Swin Transformer&#xff1a;\u5206\u5c42 Transformer&#xff0c;\u6ed1\u52a8\u7a97\u53e3\u6ce8\u610f\u529b\u5c06\u8ba1\u7b97\u91cf\u964d\u4e3a\u7ebf\u6027&#xff1b;\u79fb\u4f4d\u7a97\u53e3\u5b9e\u73b0\u7a97\u53e3\u95f4\u4ea4\u4e92\u3002\u652f\u6301\u591a\u5c3a\u5ea6\u7279\u5f81&#xff0c;\u9002\u914d\u68c0\u6d4b\u3001\u5206\u5272\u7b49\u5bc6\u96c6\u9884\u6d4b\u4efb\u52a1\u3002<\/li>\n<li>MAE&#xff1a;\u63a9\u7801\u81ea\u7f16\u7801\u5668\u81ea\u76d1\u7763\u9884\u8bad\u7ec3&#xff0c;\u968f\u673a\u63a9\u7801\u5927\u90e8\u5206 Patch&#xff0c;\u91cd\u5efa\u50cf\u7d20&#xff0c;\u5b66\u4e60\u901a\u7528\u89c6\u89c9\u8868\u5f81\u3002<\/li>\n<\/ul>\n<h3>\u4e03\u3001\u6a21\u578b\u538b\u7f29\u4e0e\u90e8\u7f72\u4f18\u5316<\/h3>\n<h4>1. \u6a21\u578b\u8f7b\u91cf\u5316\u8bbe\u8ba1<\/h4>\n<ul>\n<li>\u6df1\u5ea6\u53ef\u5206\u79bb\u5377\u79ef\u3001\u5206\u7ec4\u5377\u79ef\u3001\u5012\u6b8b\u5dee\u7ed3\u6784<\/li>\n<li>\u795e\u7ecf\u67b6\u6784\u641c\u7d22&#xff08;NAS&#xff09;\u81ea\u52a8\u641c\u7d22\u9ad8\u6548\u7ed3\u6784<\/li>\n<\/ul>\n<h4>2. \u6a21\u578b\u538b\u7f29\u6280\u672f<\/h4>\n<h5>\u6a21\u578b\u526a\u679d<\/h5>\n<ul>\n<li>\u975e\u7ed3\u6784\u5316\u526a\u679d&#xff1a;\u79fb\u9664\u4e0d\u91cd\u8981\u7684\u6743\u91cd&#xff0c;\u7a00\u758f\u5ea6\u9ad8\u4f46\u9700\u7279\u6b8a\u786c\u4ef6\u652f\u6301\u3002<\/li>\n<li>\u7ed3\u6784\u5316\u526a\u679d&#xff1a;\u79fb\u9664\u6574\u4e2a\u901a\u9053 \/ \u6ee4\u6ce2\u5668&#xff0c;\u89c4\u5219\u7a00\u758f&#xff0c;\u901a\u7528\u786c\u4ef6\u5373\u53ef\u52a0\u901f\u3002<\/li>\n<li>\u6d41\u7a0b&#xff1a;\u8bad\u7ec3\u5927\u6a21\u578b\u2192\u8bc4\u4f30\u6743\u91cd\u91cd\u8981\u6027\u2192\u526a\u679d\u2192\u5fae\u8c03\u6062\u590d\u7cbe\u5ea6\u3002<\/li>\n<\/ul>\n<h5>\u6a21\u578b\u91cf\u5316<\/h5>\n<p>\u5c06\u6d6e\u70b9\u8fd0\u7b97\u8f6c\u4e3a\u5b9a\u70b9&#xff08;\u5982 INT8&#xff09;&#xff0c;\u51cf\u5c11\u663e\u5b58\u5360\u7528\u3001\u63d0\u5347\u63a8\u7406\u901f\u5ea6\u3002<\/p>\n<ul>\n<li>PTQ \u8bad\u7ec3\u540e\u91cf\u5316&#xff1a;\u8bad\u597d\u6a21\u578b\u76f4\u63a5\u91cf\u5316&#xff0c;\u5c11\u91cf\u6821\u51c6\u6570\u636e&#xff0c;\u7b80\u5355\u4f46\u7cbe\u5ea6\u635f\u5931\u8f83\u5927\u3002<\/li>\n<li>QAT \u91cf\u5316\u611f\u77e5\u8bad\u7ec3&#xff1a;\u8bad\u7ec3\u4e2d\u6a21\u62df\u91cf\u5316\u8bef\u5dee&#xff0c;\u7cbe\u5ea6\u635f\u5931\u5c0f&#xff0c;\u6548\u679c\u66f4\u4f18\u3002<\/li>\n<li>\u5206\u7c7b&#xff1a;\u6743\u91cd\u91cf\u5316\u3001\u6fc0\u6d3b\u91cf\u5316\u3001\u5168\u91cf\u5316&#xff1b;\u5bf9\u79f0\u91cf\u5316\u3001\u975e\u5bf9\u79f0\u91cf\u5316\u3002<\/li>\n<\/ul>\n<h5>\u77e5\u8bc6\u84b8\u998f<\/h5>\n<p>\u5927\u6a21\u578b&#xff08;\u6559\u5e08&#xff09;\u8f93\u51fa\u8f6f\u6807\u7b7e\u6307\u5bfc\u5c0f\u6a21\u578b&#xff08;\u5b66\u751f&#xff09;\u8bad\u7ec3&#xff0c;\u5b66\u751f\u7ee7\u627f\u6559\u5e08\u7684\u6cdb\u5316\u80fd\u529b\u3002<\/p>\n<ul>\n<li>\u635f\u5931&#xff1a;\u5b66\u751f\u4e0e\u771f\u5b9e\u6807\u7b7e\u7684\u786c\u635f\u5931 &#043; \u5b66\u751f\u4e0e\u6559\u5e08\u8f93\u51fa\u7684\u84b8\u998f\u635f\u5931&#xff08;\u6e29\u5ea6 T \u63a7\u5236\u5e73\u6ed1\u5ea6&#xff09;\u3002<\/li>\n<li>\u8fdb\u9636&#xff1a;\u7279\u5f81\u84b8\u998f\u3001\u6ce8\u610f\u529b\u84b8\u998f&#xff0c;\u4f20\u9012\u4e2d\u95f4\u5c42\u77e5\u8bc6\u3002<\/li>\n<\/ul>\n<h5>\u4f4e\u79e9\u5206\u89e3<\/h5>\n<p>\u5c06\u5927\u6743\u91cd\u77e9\u9635\u5206\u89e3\u4e3a\u4e24\u4e2a\u5c0f\u77e9\u9635\u76f8\u4e58&#xff0c;\u5927\u5e45\u51cf\u5c11\u53c2\u6570\u91cf\u4e0e\u8ba1\u7b97\u91cf\u3002<\/p>\n<h4>3. \u6a21\u578b\u90e8\u7f72\u4e0e\u63a8\u7406\u52a0\u901f<\/h4>\n<ul>\n<li>ONNX&#xff1a;\u901a\u7528\u6a21\u578b\u4ea4\u6362\u683c\u5f0f&#xff0c;\u8de8\u6846\u67b6\u8f6c\u6362\u7684\u4e2d\u95f4\u6807\u51c6\u3002<\/li>\n<li>TensorRT&#xff1a;NVIDIA \u63a8\u7406\u5f15\u64ce&#xff0c;\u7b97\u5b50\u878d\u5408\u3001\u91cf\u5316\u3001\u5185\u6838\u81ea\u52a8\u8c03\u4f18&#xff0c;GPU \u4e0a\u6570\u500d\u52a0\u901f\u3002<\/li>\n<li>\u6838\u5fc3\u52a0\u901f\u624b\u6bb5&#xff1a;\u7b97\u5b50\u878d\u5408&#xff08;Conv&#043;BN&#043;ReLU \u5408\u5e76&#xff09;\u3001\u663e\u5b58\u590d\u7528\u3001\u6df7\u5408\u7cbe\u5ea6\u63a8\u7406&#xff08;FP16\/BF16&#xff09;\u3001\u6279\u91cf\u63a8\u7406\u3002<\/li>\n<\/ul>\n<h3>\u516b\u3001\u4f20\u7edf\u6570\u5b57\u56fe\u50cf\u5904\u7406<\/h3>\n<h4>1. \u8272\u5f69\u7a7a\u95f4<\/h4>\n<ul>\n<li>RGB&#xff1a;\u52a0\u8272\u6a21\u578b&#xff0c;\u663e\u793a\u8bbe\u5907\u901a\u7528\u3002<\/li>\n<li>HSV&#xff1a;\u8272\u8c03\u3001\u9971\u548c\u5ea6\u3001\u660e\u5ea6&#xff0c;\u7b26\u5408\u4eba\u773c\u611f\u77e5&#xff0c;\u9002\u7528\u4e8e\u989c\u8272\u5206\u5272\u3002<\/li>\n<li>YUV\/YCbCr&#xff1a;\u4eae\u5ea6 &#043; \u8272\u5ea6&#xff0c;\u8272\u5ea6\u53ef\u964d\u91c7\u6837&#xff0c;\u89c6\u9891\u7f16\u7801\u901a\u7528\u3002<\/li>\n<li>\u7070\u5ea6\u8f6c\u6362&#xff1a;\\\\(Gray &#061; 0.299R &#043; 0.587G &#043; 0.114B\\\\)&#xff0c;\u4eba\u773c\u5bf9\u7eff\u8272\u6700\u654f\u611f\u3002<\/li>\n<\/ul>\n<h4>2. \u56fe\u50cf\u6ee4\u6ce2\u4e0e\u53bb\u566a<\/h4>\n<p>\u8868\u683c<\/p>\n<table>\n<tr>\u6ee4\u6ce2\u65b9\u6cd5\u539f\u7406\u7279\u70b9\u9002\u7528\u573a\u666f<\/tr>\n<tbody>\n<tr>\n<td>\u5747\u503c\u6ee4\u6ce2<\/td>\n<td>\u90bb\u57df\u50cf\u7d20\u5e73\u5747<\/td>\n<td>\u7b80\u5355\u3001\u8fb9\u7f18\u6a21\u7cca\u4e25\u91cd<\/td>\n<td>\u8f7b\u5ea6\u5747\u5300\u566a\u58f0<\/td>\n<\/tr>\n<tr>\n<td>\u9ad8\u65af\u6ee4\u6ce2<\/td>\n<td>\u9ad8\u65af\u52a0\u6743&#xff0c;\u4e2d\u5fc3\u6743\u91cd\u9ad8<\/td>\n<td>\u5e73\u6ed1\u6548\u679c\u597d&#xff0c;\u4fdd\u8fb9\u4f18\u4e8e\u5747\u503c<\/td>\n<td>\u9ad8\u65af\u566a\u58f0\u53bb\u566a\u3001\u9884\u5904\u7406<\/td>\n<\/tr>\n<tr>\n<td>\u4e2d\u503c\u6ee4\u6ce2<\/td>\n<td>\u90bb\u57df\u6392\u5e8f\u53d6\u4e2d\u503c&#xff0c;\u975e\u7ebf\u6027<\/td>\n<td>\u6912\u76d0\u566a\u58f0\u6548\u679c\u6781\u4f73&#xff0c;\u4fdd\u8fb9\u7f18<\/td>\n<td>\u6912\u76d0\u566a\u58f0\u53bb\u9664<\/td>\n<\/tr>\n<tr>\n<td>\u53cc\u8fb9\u6ee4\u6ce2<\/td>\n<td>\u7a7a\u95f4\u9ad8\u65af &#043; \u503c\u57df\u9ad8\u65af\u8054\u5408\u52a0\u6743<\/td>\n<td>\u4fdd\u8fb9\u53bb\u566a&#xff0c;\u8ba1\u7b97\u8f83\u6162<\/td>\n<td>\u4eba\u50cf\u7f8e\u989c\u3001\u7ec6\u8282\u4fdd\u7559<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<h4>3. \u8fb9\u7f18\u4e0e\u7279\u5f81\u68c0\u6d4b<\/h4>\n<h5>Canny \u8fb9\u7f18\u68c0\u6d4b&#xff08;\u5de5\u4e1a\u6807\u51c6&#xff09;<\/h5>\n<p>\u6b65\u9aa4&#xff1a;<\/p>\n<li>\u9ad8\u65af\u6a21\u7cca\u53bb\u566a<\/li>\n<li>Sobel \u7b97\u5b50\u8ba1\u7b97\u68af\u5ea6\u5e45\u503c\u4e0e\u65b9\u5411<\/li>\n<li>\u975e\u6781\u5927\u503c\u6291\u5236&#xff1a;\u6cbf\u68af\u5ea6\u65b9\u5411\u4fdd\u7559\u5c40\u90e8\u6700\u5927\u503c&#xff0c;\u7ec6\u5316\u8fb9\u7f18<\/li>\n<li>\u53cc\u9608\u503c\u68c0\u6d4b&#xff1a;\u5f3a\u8fb9\u7f18\u76f4\u63a5\u4fdd\u7559&#xff0c;\u5f31\u8fb9\u7f18\u4ec5\u4e0e\u5f3a\u8fb9\u7f18\u76f8\u8fde\u65f6\u4fdd\u7559<\/li>\n<li>\u8fb9\u7f18\u8fde\u63a5<\/li>\n<h5>SIFT \u5c3a\u5ea6\u4e0d\u53d8\u7279\u5f81\u53d8\u6362<\/h5>\n<ul>\n<li>\u7279\u70b9&#xff1a;\u5c3a\u5ea6\u3001\u65cb\u8f6c\u3001\u5149\u7167\u4e0d\u53d8\u6027&#xff0c;\u5339\u914d\u7cbe\u5ea6\u9ad8\u3002<\/li>\n<li>\u6b65\u9aa4&#xff1a;DoG \u5c3a\u5ea6\u7a7a\u95f4\u6781\u503c\u68c0\u6d4b\u2192\u5173\u952e\u70b9\u7cbe\u5b9a\u4f4d\u2192\u68af\u5ea6\u76f4\u65b9\u56fe\u8d4b\u4e3b\u65b9\u5411\u2192128 \u7ef4\u63cf\u8ff0\u5b50\u751f\u6210\u3002<\/li>\n<\/ul>\n<h5>Harris \u89d2\u70b9\u68c0\u6d4b<\/h5>\n<p>\u57fa\u4e8e\u7a97\u53e3\u79fb\u52a8\u7684\u7070\u5ea6\u53d8\u5316&#xff0c;\u5404\u65b9\u5411\u7070\u5ea6\u53d8\u5316\u5747\u5927\u4e3a\u89d2\u70b9\u3002\u6784\u9020 M \u77e9\u9635\u8ba1\u7b97\u54cd\u5e94\u503c&#xff0c;\u9608\u503c &#043; NMS \u8f93\u51fa\u89d2\u70b9\u3002<\/p>\n<h4>4. \u56fe\u50cf\u589e\u5f3a\u4e0e\u5f62\u6001\u5b66<\/h4>\n<h5>\u76f4\u65b9\u56fe\u5747\u8861\u5316<\/h5>\n<p>\u901a\u8fc7\u7d2f\u79ef\u5206\u5e03\u51fd\u6570\u6620\u5c04\u7070\u5ea6&#xff0c;\u5c06\u76f4\u65b9\u56fe\u62c9\u4f38\u4e3a\u5747\u5300\u5206\u5e03&#xff0c;\u63d0\u5347\u5bf9\u6bd4\u5ea6\u3002<\/p>\n<ul>\n<li>\u7f3a\u70b9&#xff1a;\u5168\u5c40\u589e\u5f3a\u6613\u653e\u5927\u566a\u58f0\u3001\u5c40\u90e8\u8fc7\u66dd \/ \u8fc7\u6697\u3002<\/li>\n<li>\u6539\u8fdb&#xff1a;CLAHE&#xff08;\u9650\u5236\u5bf9\u6bd4\u5ea6\u81ea\u9002\u5e94\u76f4\u65b9\u56fe\u5747\u8861&#xff09;&#xff0c;\u5206\u5757\u5904\u7406 &#043; \u5bf9\u6bd4\u5ea6\u9650\u5236\u3002<\/li>\n<\/ul>\n<h5>\u5f62\u6001\u5b66\u64cd\u4f5c&#xff08;\u4e8c\u503c\u56fe\u50cf&#xff09;<\/h5>\n<ul>\n<li>\u8150\u8680&#xff1a;\u53d6\u90bb\u57df\u6700\u5c0f\u503c&#xff0c;\u7f29\u5c0f\u76ee\u6807\u3001\u53bb\u9664\u5c0f\u4eae\u70b9\u3001\u65ad\u5f00\u7ec6\u8fde\u63a5\u3002<\/li>\n<li>\u81a8\u80c0&#xff1a;\u53d6\u90bb\u57df\u6700\u5927\u503c&#xff0c;\u6269\u5927\u76ee\u6807\u3001\u586b\u8865\u7a7a\u6d1e\u3001\u8fde\u63a5\u65ad\u88c2\u3002<\/li>\n<li>\u5f00\u8fd0\u7b97&#xff1a;\u5148\u8150\u8680\u540e\u81a8\u80c0&#xff0c;\u53bb\u566a\u4e14\u4e0d\u6539\u53d8\u76ee\u6807\u5927\u5c0f\u3002<\/li>\n<li>\u95ed\u8fd0\u7b97&#xff1a;\u5148\u81a8\u80c0\u540e\u8150\u8680&#xff0c;\u586b\u8865\u5c0f\u5b54\u3001\u8fde\u63a5\u76f8\u90bb\u533a\u57df\u3002<\/li>\n<li>\u9876\u5e3d&#xff1a;\u539f\u56fe\u51cf\u5f00\u8fd0\u7b97&#xff0c;\u63d0\u53d6\u4eae\u7ec6\u8282&#xff1b;\u9ed1\u5e3d&#xff1a;\u95ed\u8fd0\u7b97\u51cf\u539f\u56fe&#xff0c;\u63d0\u53d6\u6697\u7ec6\u8282\u3002<\/li>\n<\/ul>\n<h3>\u4e5d\u3001\u4f20\u7edf\u673a\u5668\u5b66\u4e60\u57fa\u7840<\/h3>\n<h4>1. \u504f\u5dee\u4e0e\u65b9\u5dee\u6743\u8861<\/h4>\n<ul>\n<li>\u504f\u5dee&#xff1a;\u6a21\u578b\u671f\u671b\u9884\u6d4b\u4e0e\u771f\u5b9e\u503c\u7684\u5dee\u8ddd&#xff0c;\u8861\u91cf\u62df\u5408\u80fd\u529b&#xff1b;\u504f\u5dee\u5927\u2192\u6b20\u62df\u5408\u3002<\/li>\n<li>\u65b9\u5dee&#xff1a;\u6a21\u578b\u9884\u6d4b\u7684\u6ce2\u52a8\u7a0b\u5ea6&#xff0c;\u8861\u91cf\u7a33\u5b9a\u6027&#xff1b;\u65b9\u5dee\u5927\u2192\u8fc7\u62df\u5408\u3002<\/li>\n<li>\u603b\u8bef\u5dee &#061; \u504f\u5dee \u00b2 &#043; \u65b9\u5dee &#043; \u4e0d\u53ef\u907f\u514d\u8bef\u5dee\u3002<\/li>\n<li>\u6b20\u62df\u5408\u89e3\u51b3&#xff1a;\u589e\u52a0\u6a21\u578b\u590d\u6742\u5ea6\u3001\u589e\u52a0\u7279\u5f81\u3001\u51cf\u5c0f\u6b63\u5219\u5316\u3002<\/li>\n<li>\u8fc7\u62df\u5408\u89e3\u51b3&#xff1a;\u589e\u52a0\u6570\u636e\u3001\u589e\u5f3a\u6b63\u5219\u3001\u964d\u4f4e\u6a21\u578b\u590d\u6742\u5ea6\u3001\u65e9\u505c\u3002<\/li>\n<\/ul>\n<h4>2. \u7ecf\u5178\u7b97\u6cd5\u6838\u5fc3<\/h4>\n<h5>\u51b3\u7b56\u6811<\/h5>\n<ul>\n<li>\u5206\u88c2\u51c6\u5219&#xff1a;ID3&#xff08;\u4fe1\u606f\u589e\u76ca&#xff09;\u3001C4.5&#xff08;\u4fe1\u606f\u589e\u76ca\u6bd4&#xff09;\u3001CART&#xff08;\u57fa\u5c3c\u7cfb\u6570&#xff09;\u3002<\/li>\n<li>\u526a\u679d&#xff1a;\u9884\u526a\u679d&#xff08;\u5206\u88c2\u524d\u5224\u65ad\u589e\u76ca&#xff09;\u3001\u540e\u526a\u679d&#xff08;\u751f\u6210\u540e\u56de\u6eaf\u526a\u679d&#xff09;&#xff0c;\u9632\u6b62\u8fc7\u62df\u5408\u3002<\/li>\n<\/ul>\n<h5>\u968f\u673a\u68ee\u6797<\/h5>\n<p>Bagging \u96c6\u6210 &#043; \u51b3\u7b56\u6811&#xff0c;\u6837\u672c Bootstrap \u91c7\u6837 &#043; \u7279\u5f81\u968f\u673a\u9009\u62e9\u3002<\/p>\n<ul>\n<li>\u4f18\u70b9&#xff1a;\u6297\u8fc7\u62df\u5408\u3001\u6cdb\u5316\u5f3a\u3001\u53ef\u5e76\u884c\u3001\u5904\u7406\u9ad8\u7ef4\u6570\u636e\u3002<\/li>\n<\/ul>\n<h5>XGBoost<\/h5>\n<p>GBDT \u5de5\u7a0b\u4f18\u5316\u6807\u6746\u3002<\/p>\n<ul>\n<li>\u6838\u5fc3\u6539\u8fdb&#xff1a;\u4e8c\u9636\u6cf0\u52d2\u5c55\u5f00\u635f\u5931\u3001\u663e\u5f0f\u6b63\u5219\u9879\u3001\u8fd1\u4f3c\u5206\u88c2\u7b97\u6cd5\u3001\u7f3a\u5931\u503c\u81ea\u52a8\u5904\u7406\u3001\u5217\u91c7\u6837\u3001\u7279\u5f81\u7ef4\u5ea6\u5e76\u884c\u3002<\/li>\n<\/ul>\n<h5>SVM \u652f\u6301\u5411\u91cf\u673a<\/h5>\n<ul>\n<li>\u6838\u5fc3&#xff1a;\u6700\u5927\u5316\u5206\u7c7b\u95f4\u9694\u7684\u8d85\u5e73\u9762&#xff0c;\u652f\u6301\u5411\u91cf\u51b3\u5b9a\u5206\u9694\u9762\u3002<\/li>\n<li>\u6838\u6280\u5de7&#xff1a;\u4f4e\u7ef4\u6620\u5c04\u9ad8\u7ef4\u5b9e\u73b0\u975e\u7ebf\u6027\u5206\u7c7b&#xff0c;\u5e38\u7528 RBF \u9ad8\u65af\u6838\u3002<\/li>\n<li>\u4f18\u70b9&#xff1a;\u5c0f\u6837\u672c\u6cdb\u5316\u597d\u3001\u7406\u8bba\u5b8c\u5907&#xff1b;\u7f3a\u70b9&#xff1a;\u5927\u6570\u636e\u6162\u3001\u8c03\u53c2\u590d\u6742\u3002<\/li>\n<\/ul>\n<h4>3. \u5206\u7c7b\u8bc4\u4ef7\u6307\u6807<\/h4>\n<ul>\n<li>\u7cbe\u786e\u7387 Precision&#xff1a;\u9884\u6d4b\u6b63\u4f8b\u4e2d\u771f\u9633\u6027\u6bd4\u4f8b&#xff0c;\u5173\u6ce8\u8bef\u68c0\u3002<\/li>\n<li>\u53ec\u56de\u7387 Recall&#xff1a;\u771f\u9633\u6027\u4e2d\u88ab\u68c0\u51fa\u6bd4\u4f8b&#xff0c;\u5173\u6ce8\u6f0f\u68c0\u3002<\/li>\n<li>F1&#xff1a;\u7cbe\u786e\u7387\u4e0e\u53ec\u56de\u7387\u8c03\u548c\u5e73\u5747&#xff0c;\u7efc\u5408\u6307\u6807\u3002<\/li>\n<li>ROC-AUC&#xff1a;\u6a2a\u8f74\u5047\u9633\u7387\u3001\u7eb5\u8f74\u771f\u9633\u7387&#xff0c;AUC \u8861\u91cf\u6a21\u578b\u6574\u4f53\u6392\u5e8f\u80fd\u529b&#xff0c;\u4e0d\u53d7\u9608\u503c\u5f71\u54cd\u3002<\/li>\n<li>PR \u66f2\u7ebf&#xff1a;\u6837\u672c\u4e0d\u5747\u8861\u573a\u666f\u4e0b\u6bd4 ROC \u66f4\u654f\u611f\u3002<\/li>\n<\/ul>\n<h3>\u5341\u3001\u9ad8\u9891\u624b\u6495\u4ee3\u7801&#xff08;PyTorch \u7248&#xff09;<\/h3>\n<h4>1. IoU \u8ba1\u7b97<\/h4>\n<p>python<\/p>\n<p>\u8fd0\u884c<\/p>\n<p>def compute_iou(box1, box2):<br \/>\n    # box\u683c\u5f0f: [x1, y1, x2, y2]<br \/>\n    x1 &#061; max(box1[0], box2[0])<br \/>\n    y1 &#061; max(box1[1], box2[1])<br \/>\n    x2 &#061; min(box1[2], box2[2])<br \/>\n    y2 &#061; min(box1[3], box2[3])<\/p>\n<p>    inter &#061; max(0, x2 &#8211; x1) * max(0, y2 &#8211; y1)<br \/>\n    area1 &#061; (box1[2]-box1[0]) * (box1[3]-box1[1])<br \/>\n    area2 &#061; (box2[2]-box2[0]) * (box2[3]-box2[1])<br \/>\n    union &#061; area1 &#043; area2 &#8211; inter<br \/>\n    return inter \/ union if union &gt; 0 else 0<\/p>\n<h4>2. NMS \u5b9e\u73b0<\/h4>\n<p>python<\/p>\n<p>\u8fd0\u884c<\/p>\n<p>def nms(boxes, scores, iou_threshold):<br \/>\n    order &#061; scores.argsort(descending&#061;True)<br \/>\n    keep &#061; []<br \/>\n    while len(order) &gt; 0:<br \/>\n        idx &#061; order[0]<br \/>\n        keep.append(idx.item())<br \/>\n        if len(order) &#061;&#061; 1:<br \/>\n            break<br \/>\n        rest &#061; order[1:]<br \/>\n        ious &#061; torch.tensor([compute_iou(boxes[idx], boxes[i]) for i in rest])<br \/>\n        order &#061; rest[ious &lt; iou_threshold]<br \/>\n    return keep<\/p>\n<h4>3. Focal Loss<\/h4>\n<p>python<\/p>\n<p>\u8fd0\u884c<\/p>\n<p>class FocalLoss(nn.Module):<br \/>\n    def __init__(self, alpha&#061;0.25, gamma&#061;2):<br \/>\n        super().__init__()<br \/>\n        self.alpha &#061; alpha<br \/>\n        self.gamma &#061; gamma<\/p>\n<p>    def forward(self, pred, target):<br \/>\n        ce &#061; F.cross_entropy(pred, target, reduction&#061;&#039;none&#039;)<br \/>\n        p &#061; torch.exp(-ce)<br \/>\n        focal_weight &#061; self.alpha * (1 &#8211; p) ** self.gamma<br \/>\n        return (focal_weight * ce).mean()<\/p>\n<h4>4. Dice Loss<\/h4>\n<p>python<\/p>\n<p>\u8fd0\u884c<\/p>\n<p>class DiceLoss(nn.Module):<br \/>\n    def __init__(self, smooth&#061;1e-6):<br \/>\n        super().__init__()<br \/>\n        self.smooth &#061; smooth<\/p>\n<p>    def forward(self, pred, target):<br \/>\n        pred &#061; torch.sigmoid(pred)<br \/>\n        pred &#061; pred.view(-1)<br \/>\n        target &#061; target.view(-1)<br \/>\n        intersection &#061; (pred * target).sum()<br \/>\n        dice &#061; (2. * intersection &#043; self.smooth) \/ (pred.sum() &#043; target.sum() &#043; self.smooth)<br \/>\n        return 1 &#8211; dice<\/p>\n<h4>5. ResNet BasicBlock<\/h4>\n<p>python<\/p>\n<p>\u8fd0\u884c<\/p>\n<p>class BasicBlock(nn.Module):<br \/>\n    expansion &#061; 1<br \/>\n    def __init__(self, in_channels, out_channels, stride&#061;1):<br \/>\n        super().__init__()<br \/>\n        self.conv1 &#061; nn.Conv2d(in_channels, out_channels, 3, stride, 1, bias&#061;False)<br \/>\n        self.bn1 &#061; nn.BatchNorm2d(out_channels)<br \/>\n        self.conv2 &#061; nn.Conv2d(out_channels, out_channels, 3, 1, 1, bias&#061;False)<br \/>\n        self.bn2 &#061; nn.BatchNorm2d(out_channels)<\/p>\n<p>        self.shortcut &#061; nn.Sequential()<br \/>\n        if stride !&#061; 1 or in_channels !&#061; out_channels:<br \/>\n            self.shortcut &#061; nn.Sequential(<br \/>\n                nn.Conv2d(in_channels, out_channels, 1, stride, bias&#061;False),<br \/>\n                nn.BatchNorm2d(out_channels)<br \/>\n            )<\/p>\n<p>    def forward(self, x):<br \/>\n        out &#061; F.relu(self.bn1(self.conv1(x)))<br \/>\n        out &#061; self.bn2(self.conv2(out))<br \/>\n        out &#043;&#061; self.shortcut(x)<br \/>\n        return F.relu(out)<\/p>\n<h4>6. \u591a\u5934\u6ce8\u610f\u529b<\/h4>\n<p>python<\/p>\n<p>\u8fd0\u884c<\/p>\n<p>class MultiHeadAttention(nn.Module):<br \/>\n    def __init__(self, dim, num_heads):<br \/>\n        super().__init__()<br \/>\n        self.num_heads &#061; num_heads<br \/>\n        self.head_dim &#061; dim \/\/ num_heads<br \/>\n        self.scale &#061; self.head_dim ** -0.5<\/p>\n<p>        self.qkv &#061; nn.Linear(dim, dim * 3)<br \/>\n        self.proj &#061; nn.Linear(dim, dim)<\/p>\n<p>    def forward(self, x):<br \/>\n        B, N, C &#061; x.shape<br \/>\n        qkv &#061; self.qkv(x).reshape(B, N, 3, self.num_heads, self.head_dim).permute(2, 0, 3, 1, 4)<br \/>\n        q, k, v &#061; qkv.unbind(0)<\/p>\n<p>        attn &#061; (q &#064; k.transpose(-2, -1)) * self.scale<br \/>\n        attn &#061; attn.softmax(dim&#061;-1)<\/p>\n<p>        out &#061; (attn &#064; v).transpose(1, 2).reshape(B, N, C)<br \/>\n        return self.proj(out)<\/p>\n<h3>\u5341\u4e00\u3001\u9762\u8bd5\u9ad8\u9891\u5f00\u653e\u6027\u95ee\u9898<\/h3>\n<li>\n<p>\u5c0f\u76ee\u6807\u68c0\u6d4b\u96be\u70b9\u4e0e\u89e3\u51b3\u65b9\u6848 \u96be\u70b9&#xff1a;\u50cf\u7d20\u5c11\u3001\u7279\u5f81\u5f31\u3001\u6807\u6ce8\u8bef\u5dee\u5927\u3001\u6837\u672c\u4e0d\u5747\u8861\u3002 \u65b9\u6848&#xff1a;\u9ad8\u5206\u8fa8\u7387\u8f93\u5165\u3001\u591a\u5c3a\u5ea6\u7279\u5f81\u878d\u5408&#xff08;FPN&#xff09;\u3001\u6570\u636e\u589e\u5f3a&#xff08;Mosaic\u3001\u590d\u5236\u7c98\u8d34&#xff09;\u3001Anchor \u4f18\u5316\u3001\u9488\u5bf9\u5c0f\u76ee\u6807\u7684\u635f\u5931\u52a0\u6743\u3001\u589e\u52a0\u4f4e\u5c42\u7279\u5f81\u76d1\u7763\u3002<\/p>\n<\/li>\n<li>\n<p>\u6837\u672c\u4e0d\u5747\u8861\u5904\u7406\u65b9\u6cd5 \u6570\u636e\u5c42\u9762&#xff1a;\u91cd\u91c7\u6837&#xff08;\u8fc7\u91c7\u6837\u3001\u6b20\u91c7\u6837&#xff09;\u3001\u6570\u636e\u589e\u5f3a\u3001\u5408\u6210\u6837\u672c\u3002 \u7b97\u6cd5\u5c42\u9762&#xff1a;\u7c7b\u522b\u52a0\u6743\u635f\u5931\u3001Focal Loss\u3001\u96be\u4f8b\u6316\u6398&#xff08;OHEM&#xff09;\u3001\u751f\u6210\u5f0f\u8865\u5145\u6837\u672c\u3002<\/p>\n<\/li>\n<li>\n<p>CNN \u4e0e Transformer \u9009\u578b<\/p>\n<ul>\n<li>\u6570\u636e\u91cf\u5c0f\u3001\u8fb9\u7f18\u8bbe\u5907\u90e8\u7f72\u3001\u5b9e\u65f6\u6027\u8981\u6c42\u9ad8&#xff1a;\u4f18\u5148 CNN\u3002<\/li>\n<li>\u6570\u636e\u91cf\u5927\u3001\u901a\u7528\u89c6\u89c9\u4efb\u52a1\u3001\u9700\u8981\u5f3a\u5168\u5c40\u5efa\u6a21&#xff1a;\u4f18\u5148 Transformer\u3002<\/li>\n<li>\u5de5\u4e1a\u843d\u5730\u4e3b\u6d41&#xff1a;CNN Backbone \u4ecd\u662f\u68c0\u6d4b\u5206\u5272\u4e3b\u529b&#xff0c;Transformer \u9010\u6b65\u6e17\u900f\u9ad8\u7aef\u573a\u666f\u3002<\/li>\n<\/ul>\n<\/li>\n<li>\n<p>\u6a21\u578b\u90e8\u7f72\u901f\u5ea6\u4f18\u5316\u601d\u8def \u6a21\u578b\u4fa7&#xff1a;\u8f7b\u91cf\u5316\u67b6\u6784\u3001\u526a\u679d\u3001\u91cf\u5316\u3001\u77e5\u8bc6\u84b8\u998f\u3002 \u63a8\u7406\u4fa7&#xff1a;\u7b97\u5b50\u878d\u5408\u3001TensorRT\/OpenVINO \u52a0\u901f\u3001\u6df7\u5408\u7cbe\u5ea6\u3001\u6279\u91cf\u63a8\u7406\u3001\u663e\u5b58\u4f18\u5316\u3002 \u5de5\u7a0b\u4fa7&#xff1a;\u591a\u7ebf\u7a0b\u3001\u6d41\u6c34\u7ebf\u3001\u6a21\u578b\u5e76\u884c \/ \u6570\u636e\u5e76\u884c\u3002<\/p>\n<\/li>\n<p>\u9700\u8981\u6211\u8865\u5145Transformer \u68c0\u6d4b\u67b6\u6784&#xff08;DETR \u7cfb\u5217&#xff09;\u3001\u81ea\u76d1\u7763\u5b66\u4e60\u3001\u591a\u6a21\u6001\u5927\u6a21\u578b\u76f8\u5173\u7684\u516b\u80a1\u5185\u5bb9\u5417&#xff1f;<\/p>\n<p>\u6df1\u5ea6\u5b66\u4e60\u66f4\u591a\u516b\u80a1<\/p>\n<p>\u5df2\u5b8c\u6210\u601d\u8003<\/p>\n<h2>\u6df1\u5ea6\u5b66\u4e60\u8fdb\u9636\u516b\u80a1\u5927\u5168&#xff08;\u6df1\u5ea6\u62d3\u5c55\u7248&#xff09;<\/h2>\n<p>\u4ee5\u4e0b\u8865\u5145\u6df1\u5ea6\u5b66\u4e60\u9886\u57df\u66f4\u7ec6\u5206\u3001\u66f4\u9ad8\u9891\u7684\u9762\u8bd5\u6df1\u5ea6\u8003\u70b9&#xff0c;\u8986\u76d6\u4f18\u5316\u5668\u539f\u7406\u3001\u5f52\u4e00\u5316\u4e0e\u521d\u59cb\u5316\u3001Transformer \u8fdb\u9636\u3001\u81ea\u76d1\u7763\u5b66\u4e60\u3001\u5927\u6a21\u578b\u5fae\u8c03\u3001\u5206\u5e03\u5f0f\u8bad\u7ec3\u3001\u8bad\u7ec3\u6392\u969c\u3001\u7406\u8bba\u57fa\u7840\u7b49\u6838\u5fc3\u516b\u80a1\u65b9\u5411&#xff0c;\u5168\u90e8\u8d34\u5408\u7b97\u6cd5\u5c97\u9762\u8bd5\u63d0\u95ee\u903b\u8f91\u3002<\/p>\n<h3>\u5341\u4e8c\u3001\u4f18\u5316\u5668\u4e0e\u68af\u5ea6\u4e0b\u964d\u6df1\u5ea6\u8fdb\u9636<\/h3>\n<h4>1. \u68af\u5ea6\u4e0b\u964d\u5bb6\u65cf\u5b8c\u6574\u8c31\u7cfb<\/h4>\n<ul>\n<li>BGD \u6279\u91cf\u68af\u5ea6\u4e0b\u964d&#xff1a;\u6bcf\u8f6e\u7528\u5168\u90e8\u6837\u672c\u8ba1\u7b97\u68af\u5ea6\u540e\u66f4\u65b0\u3002\u6536\u655b\u7a33\u5b9a\u3001\u8fed\u4ee3\u6b21\u6570\u5c11&#xff1b;\u4f46\u8ba1\u7b97\u91cf\u5927&#xff0c;\u65e0\u6cd5\u5728\u7ebf\u66f4\u65b0&#xff0c;\u5927\u6570\u636e\u96c6\u4e0d\u53ef\u884c\u3002<\/li>\n<li>SGD \u968f\u673a\u68af\u5ea6\u4e0b\u964d&#xff1a;\u6bcf\u8f6e\u53ea\u7528\u5355\u4e2a\u6837\u672c\u8ba1\u7b97\u68af\u5ea6\u3002\u8fed\u4ee3\u6ce2\u52a8\u5927&#xff0c;\u6709\u9690\u5f0f\u6b63\u5219\u6548\u679c&#xff0c;\u6cdb\u5316\u6027\u66f4\u597d&#xff1b;\u8ba1\u7b97\u6781\u5feb&#xff0c;\u652f\u6301\u5728\u7ebf\u5b66\u4e60\u3002<\/li>\n<li>MBGD \u5c0f\u6279\u91cf\u68af\u5ea6\u4e0b\u964d&#xff1a;\u5de5\u4e1a\u754c\u6807\u51c6&#xff0c;\u6298\u4e2d\u4e24\u8005\u4f18\u70b9&#xff0c;\u6bcf\u6279\u7528\u56fa\u5b9a\u6570\u91cf\u6837\u672c\u66f4\u65b0&#xff0c;\u517c\u987e\u7a33\u5b9a\u6027\u4e0e\u6548\u7387\u3002<\/li>\n<li>\u52a8\u91cf Momentum&#xff1a;\u5f15\u5165\u6307\u6570\u52a0\u6743\u5e73\u5747\u7684\u5386\u53f2\u68af\u5ea6&#xff0c;\u5f62\u6210\u66f4\u65b0\u60ef\u6027\u3002\u516c\u5f0f&#xff1a;\\\\(v_t &#061; \\\\beta v_{t-1} &#043; (1-\\\\beta)g_t,\\\\ w &#061; w &#8211; \\\\eta v_t\\\\)\u3002\n<ul>\n<li>\u4f5c\u7528&#xff1a;\u52a0\u901f\u6536\u655b\u3001\u51b2\u8fc7\u5c40\u90e8\u6700\u4f18\u4e0e\u978d\u70b9\u3001\u6291\u5236\u68af\u5ea6\u9707\u8361\u3002<\/li>\n<\/ul>\n<\/li>\n<li>Nesterov \u52a8\u91cf&#xff1a;\u5148\u6309\u60ef\u6027\u65b9\u5411\u524d\u8fdb\u4e00\u6b65&#xff0c;\u518d\u5728\u65b0\u4f4d\u7f6e\u8ba1\u7b97\u68af\u5ea6\u4fee\u6b63\u65b9\u5411\u3002\u6bd4\u666e\u901a\u52a8\u91cf\u66f4\u7075\u654f&#xff0c;\u80fd\u63d0\u524d\u51cf\u901f&#xff0c;\u6536\u655b\u66f4\u5e73\u7a33\u3002<\/li>\n<\/ul>\n<h4>2. Adam \u7cfb\u4f18\u5316\u5668\u6df1\u5ea6\u8fa8\u6790<\/h4>\n<ul>\n<li>Adam \u5b8c\u6574\u673a\u5236&#xff1a;\u540c\u65f6\u7ef4\u62a4\u68af\u5ea6\u7684\u4e00\u9636\u77e9&#xff08;\u52a8\u91cf&#xff0c;\u5e73\u6ed1\u68af\u5ea6\u65b9\u5411&#xff09;\u548c\u4e8c\u9636\u77e9&#xff08;\u68af\u5ea6\u5e73\u65b9&#xff0c;\u81ea\u9002\u5e94\u5b66\u4e60\u7387&#xff09;&#xff0c;\u5e76\u52a0\u5165\u504f\u5dee\u4fee\u6b63\u89e3\u51b3\u521d\u59cb\u9636\u6bb5\u7edf\u8ba1\u91cf\u6709\u504f\u7684\u95ee\u9898\u3002<\/li>\n<li>Adam \u6cdb\u5316\u6027\u5f31\u4e8e SGD \u7684\u539f\u56e0&#xff08;\u9762\u8bd5\u9ad8\u9891&#xff09;\n<li>\u81ea\u9002\u5e94\u5b66\u4e60\u7387\u5bfc\u81f4\u53c2\u6570\u66f4\u65b0\u65b9\u5411\u66f4\u6fc0\u8fdb&#xff0c;\u6613\u6536\u655b\u5230\u5c16\u9510\u6781\u5c0f\u503c&#xff0c;\u5bf9\u6570\u636e\u6270\u52a8\u5bb9\u5fcd\u5ea6\u4f4e&#xff0c;\u6cdb\u5316\u5dee&#xff1b;SGD \u6052\u5b9a\u5b66\u4e60\u7387\u66f4\u6613\u627e\u5230\u5e73\u5766\u6781\u5c0f\u503c\u3002<\/li>\n<li>\u4e8c\u9636\u77e9\u6301\u7eed\u7d2f\u79ef\u4f1a\u5bfc\u81f4\u8bad\u7ec3\u540e\u671f\u5b66\u4e60\u7387\u8fc7\u5c0f&#xff0c;\u6a21\u578b\u8fc7\u65e9\u6536\u655b\u5230\u6b21\u4f18\u70b9\u3002<\/li>\n<li>\u539f\u59cb Adam \u4e2d L2 \u6b63\u5219\u4e0e\u81ea\u9002\u5e94\u5b66\u4e60\u7387\u8026\u5408&#xff0c;\u6b63\u5219\u6548\u679c\u88ab\u524a\u5f31\u3002<\/li>\n<\/li>\n<li>AdamW vs Adam\n<ul>\n<li>\u6838\u5fc3\u5dee\u5f02&#xff1a;\u89e3\u8026\u6743\u91cd\u8870\u51cf\u3002\u539f\u59cb Adam \u5c06\u6743\u91cd\u8870\u51cf\u5e76\u5165\u68af\u5ea6\u8ba1\u7b97&#xff0c;\u81ea\u9002\u5e94\u5b66\u4e60\u7387\u4f1a\u7f29\u653e\u68af\u5ea6&#xff0c;\u5bfc\u81f4\u6b63\u5219\u5f3a\u5ea6\u4e0d\u7a33\u5b9a&#xff1b;AdamW \u5c06\u6743\u91cd\u8870\u51cf\u76f4\u63a5\u4f5c\u7528\u4e8e\u6743\u91cd\u672c\u8eab&#xff0c;\u4e0e\u68af\u5ea6\u66f4\u65b0\u5b8c\u5168\u89e3\u8026\u3002<\/li>\n<li>\u5730\u4f4d&#xff1a;Transformer\u3001ViT\u3001\u5927\u6a21\u578b\u7684\u6807\u51c6\u4f18\u5316\u5668&#xff0c;\u6b63\u5219\u66f4\u7a33\u5b9a&#xff0c;\u6cdb\u5316\u663e\u8457\u4f18\u4e8e\u539f\u751f Adam\u3002<\/li>\n<\/ul>\n<\/li>\n<li>LAMB \u4f18\u5316\u5668&#xff1a;\u9010\u5c42\u81ea\u9002\u5e94\u77e9\u4f18\u5316&#xff0c;\u4e3a\u6bcf\u4e00\u5c42\u5355\u72ec\u8ba1\u7b97\u5b66\u4e60\u7387\u7f29\u653e\u56e0\u5b50&#xff0c;\u9002\u914d\u8d85\u5927 Batch \u9884\u8bad\u7ec3\u573a\u666f&#xff0c;\u89e3\u51b3\u5927 Batch \u4e0b Adam \u8bad\u7ec3\u4e0d\u7a33\u5b9a\u7684\u95ee\u9898\u3002<\/li>\n<\/ul>\n<h4>3. \u6743\u91cd\u8870\u51cf vs L2 \u6b63\u5219\u5316<\/h4>\n<ul>\n<li>SGD \u573a\u666f&#xff1a;\u4e24\u8005\u6570\u5b66\u4e0a\u8fd1\u4f3c\u7b49\u4ef7&#xff0c;L2 \u6b63\u5219\u7684\u68af\u5ea6\u9879\u7b49\u4ef7\u4e8e\u6743\u91cd\u8870\u51cf\u3002<\/li>\n<li>\u81ea\u9002\u5e94\u4f18\u5316\u5668\u573a\u666f&#xff08;Adam\/RMSProp&#xff09;&#xff1a;\u4e24\u8005\u4e0d\u7b49\u4ef7\u3002L2 \u6b63\u5219\u7684\u68af\u5ea6\u4f1a\u88ab\u81ea\u9002\u5e94\u5b66\u4e60\u7387\u7f29\u653e&#xff0c;\u6b63\u5219\u6548\u679c\u88ab\u524a\u5f31&#xff1b;\u6743\u91cd\u8870\u51cf\u76f4\u63a5\u4f5c\u7528\u4e8e\u6743\u91cd&#xff0c;\u4e0d\u53d7\u5b66\u4e60\u7387\u5f71\u54cd&#xff0c;\u6b63\u5219\u5f3a\u5ea6\u66f4\u7a33\u5b9a\u3002<\/li>\n<li>\u5de5\u7a0b\u7ed3\u8bba&#xff1a;\u6df1\u5ea6\u5b66\u4e60\u4e2d\u4f18\u5148\u4f7f\u7528\u6743\u91cd\u8870\u51cf&#xff0c;\u800c\u975e\u5728\u635f\u5931\u4e2d\u52a0 L2 \u6b63\u5219\u9879\u3002<\/li>\n<\/ul>\n<h4>4. \u4e8c\u9636\u4f18\u5316\u65b9\u6cd5\u4e3a\u4ec0\u4e48\u5de5\u4e1a\u754c\u4e0d\u5e38\u7528<\/h4>\n<p>\u4ee3\u8868\u65b9\u6cd5&#xff1a;\u725b\u987f\u6cd5\u3001\u62df\u725b\u987f\u6cd5&#xff08;L-BFGS&#xff09;\u3002<\/p>\n<ul>\n<li>\u6838\u5fc3\u74f6\u9888&#xff1a;\u6d77\u68ee\u77e9\u9635\u8ba1\u7b97\u4e0e\u5b58\u50a8\u6210\u672c\u6781\u9ad8&#xff0c;\u53c2\u6570\u91cf\u767e\u4e07\u7ea7\u4ee5\u4e0a\u5b8c\u5168\u4e0d\u53ef\u884c\u3002<\/li>\n<li>\u975e\u51f8\u95ee\u9898&#xff1a;\u6df1\u5ea6\u5b66\u4e60\u635f\u5931\u66f2\u9762\u975e\u51f8&#xff0c;\u6d77\u68ee\u77e9\u9635\u53ef\u80fd\u4e0d\u6b63\u5b9a&#xff0c;\u725b\u987f\u65b9\u5411\u4e0d\u4e00\u5b9a\u662f\u4e0b\u964d\u65b9\u5411\u3002<\/li>\n<li>\u6027\u4ef7\u6bd4\u4f4e&#xff1a;\u4e00\u9636\u65b9\u6cd5\u914d\u5408\u52a8\u91cf\u3001\u81ea\u9002\u5e94\u5b66\u4e60\u7387\u3001\u5b66\u4e60\u7387\u8c03\u5ea6&#xff0c;\u5728\u5927\u6570\u636e\u4e0b\u6548\u679c\u8db3\u591f\u597d&#xff0c;\u5de5\u7a0b\u6210\u672c\u8fdc\u4f4e\u4e8e\u4e8c\u9636\u65b9\u6cd5\u3002<\/li>\n<\/ul>\n<h3>\u5341\u4e09\u3001\u5f52\u4e00\u5316\u4e0e\u521d\u59cb\u5316\u6df1\u5ea6\u89e3\u6790<\/h3>\n<h4>1. BatchNorm \u6df1\u5c42\u7406\u89e3<\/h4>\n<ul>\n<li>BN \u52a0\u901f\u6536\u655b\u7684\u672c\u8d28\n<li>\u7ecf\u5178\u89e3\u91ca&#xff1a;\u7f13\u89e3\u5185\u90e8\u534f\u53d8\u91cf\u504f\u79fb&#xff08;ICS&#xff09;&#xff0c;\u8ba9\u6bcf\u5c42\u8f93\u5165\u5206\u5e03\u7a33\u5b9a&#xff0c;\u65e0\u9700\u53cd\u590d\u9002\u5e94\u8f93\u5165\u5206\u5e03\u53d8\u5316\u3002<\/li>\n<li>\u540e\u7eed\u7814\u7a76&#xff1a;\u5e73\u6ed1\u635f\u5931\u66f2\u9762&#xff0c;\u964d\u4f4e\u68af\u5ea6\u7684 Lipschitz \u5e38\u6570&#xff0c;\u8ba9\u68af\u5ea6\u66f4\u7a33\u5b9a&#xff0c;\u5927\u5e45\u964d\u4f4e\u5bf9\u521d\u59cb\u5316\u548c\u5b66\u4e60\u7387\u7684\u654f\u611f\u5ea6&#xff1b;\u6279\u6b21\u7edf\u8ba1\u5f15\u5165\u566a\u58f0&#xff0c;\u5177\u5907\u8f7b\u5fae\u6b63\u5219\u6548\u679c\u3002<\/li>\n<\/li>\n<li>BN \u7684\u6838\u5fc3\u7f3a\u70b9\n<li>\u5f3a\u4f9d\u8d56 Batch Size&#xff0c;\u5c0f Batch \u4e0b\u7edf\u8ba1\u91cf\u504f\u5dee\u5927&#xff0c;\u6027\u80fd\u66b4\u8dcc\u3002<\/li>\n<li>\u8bad\u7ec3\u4e0e\u63a8\u7406\u884c\u4e3a\u4e0d\u4e00\u81f4&#xff0c;\u63a8\u7406\u9700\u4f7f\u7528\u8bad\u7ec3\u671f\u7d2f\u79ef\u7684\u6ed1\u52a8\u5e73\u5747\u7edf\u8ba1\u91cf\u3002<\/li>\n<li>\u5bf9\u5e8f\u5217\u6570\u636e\u3001\u50cf\u7d20\u7ea7\u5bc6\u96c6\u9884\u6d4b\u4efb\u52a1\u9002\u914d\u6027\u5dee\u3002<\/li>\n<li>\u5f15\u5165\u6279\u6b21\u95f4\u566a\u58f0&#xff0c;\u5728\u751f\u6210\u6a21\u578b\u4e2d\u4f1a\u7834\u574f\u751f\u6210\u8d28\u91cf\u3002<\/li>\n<\/li>\n<li>\u63a8\u7406\u7528\u6ed1\u52a8\u5e73\u5747\u7684\u539f\u56e0&#xff1a;\u4fdd\u8bc1\u8f93\u51fa\u786e\u5b9a\u6027&#xff0c;\u540c\u4e00\u6837\u672c\u4e0d\u53d7\u63a8\u7406\u6279\u6b21\u5f71\u54cd&#xff1b;\u540c\u65f6\u8ba9\u63a8\u7406\u5206\u5e03\u4e0e\u8bad\u7ec3\u671f\u671f\u671b\u5206\u5e03\u5bf9\u9f50\u3002<\/li>\n<\/ul>\n<h4>2. \u5f52\u4e00\u5316\u5bb6\u65cf\u9009\u578b\u8003\u70b9<\/h4>\n<ul>\n<li>\u4e3a\u4ec0\u4e48 Transformer \u7528 LayerNorm \u4e0d\u7528 BN&#xff1f;\n<li>Transformer \u8f93\u5165\u662f\u53d8\u957f\u5e8f\u5217&#xff0c;\u5e8f\u5217\u957f\u5ea6\u4e0d\u56fa\u5b9a&#xff0c;BN \u7684\u901a\u9053\u7ef4\u5ea6\u7edf\u8ba1\u65e0\u610f\u4e49\u3002<\/li>\n<li>\u5e8f\u5217\u4e2d\u4e0d\u540c token \u8bed\u4e49\u5dee\u5f02\u5927&#xff0c;BN \u7684\u6279\u6b21\u5e73\u5747\u4f1a\u62b9\u5e73\u4e2a\u4f53\u7279\u5f81\u3002<\/li>\n<li>LN \u9010\u6837\u672c\u5f52\u4e00\u5316&#xff0c;\u4e0d\u53d7 Batch Size\u3001\u5e8f\u5217\u957f\u5ea6\u5f71\u54cd&#xff0c;\u9002\u914d\u6027\u66f4\u5f3a\u3002<\/li>\n<\/li>\n<li>RMSNorm&#xff1a;LayerNorm \u7684\u8f7b\u91cf\u5316\u7248\u672c&#xff0c;\u53ea\u5f52\u4e00\u5316\u65b9\u5dee\u3001\u4e0d\u51cf\u5747\u503c&#xff0c;\u8ba1\u7b97\u91cf\u66f4\u4f4e&#xff0c;\u6548\u679c\u76f8\u5f53&#xff0c;\u662f\u5927\u6a21\u578b&#xff08;LLaMA \u7b49&#xff09;\u7684\u4e3b\u6d41\u9009\u62e9\u3002<\/li>\n<li>InstanceNorm vs BN&#xff1a;IN \u9010\u6837\u672c\u9010\u901a\u9053\u5f52\u4e00\u5316&#xff0c;\u6d88\u9664\u56fe\u50cf\u4e2a\u4f53\u4eae\u5ea6\u3001\u5bf9\u6bd4\u5ea6\u5dee\u5f02&#xff0c;\u4fdd\u7559\u7eb9\u7406\u98ce\u683c&#xff0c;\u9002\u7528\u4e8e\u98ce\u683c\u8fc1\u79fb\u3001\u56fe\u50cf\u751f\u6210&#xff1b;BN \u5173\u6ce8\u6279\u6b21\u6574\u4f53\u5206\u5e03&#xff0c;\u9002\u914d\u5206\u7c7b\u4efb\u52a1\u3002<\/li>\n<\/ul>\n<h4>3. \u6743\u91cd\u521d\u59cb\u5316\u6838\u5fc3\u539f\u7406<\/h4>\n<ul>\n<li>\u5168\u96f6\u521d\u59cb\u5316\u4e3a\u4ec0\u4e48\u5931\u6548&#xff1f;&#xff1a;\u6240\u6709\u795e\u7ecf\u5143\u8f93\u51fa\u5b8c\u5168\u76f8\u540c&#xff0c;\u53cd\u5411\u4f20\u64ad\u68af\u5ea6\u4e00\u81f4&#xff0c;\u53c2\u6570\u66f4\u65b0\u6c38\u8fdc\u5bf9\u79f0&#xff0c;\u7f51\u7edc\u9000\u5316\u4e3a\u5355\u5c42\u795e\u7ecf\u5143&#xff0c;\u65e0\u6cd5\u5b66\u4e60\u5dee\u5f02\u5316\u7279\u5f81\u3002<\/li>\n<li>Xavier \u521d\u59cb\u5316&#xff08;Glorot&#xff09;\n<ul>\n<li>\u5047\u8bbe&#xff1a;\u6fc0\u6d3b\u51fd\u6570\u8fd1\u4f3c\u7ebf\u6027&#xff0c;\u4fdd\u8bc1\u524d\u5411\u4f20\u64ad\u4e0e\u53cd\u5411\u4f20\u64ad\u65f6\u6bcf\u5c42\u7279\u5f81\u7684\u65b9\u5dee\u4e00\u81f4&#xff0c;\u907f\u514d\u68af\u5ea6\u6d88\u5931 \/ \u7206\u70b8\u3002<\/li>\n<li>\u7ed3\u8bba&#xff1a;\u6743\u91cd\u670d\u4ece\u5747\u5300\u5206\u5e03 \\\\(U(-\\\\sqrt{6\/(n_{in}&#043;n_{out})},\\\\sqrt{6\/(n_{in}&#043;n_{out})})\\\\)\u3002<\/li>\n<li>\u9002\u914d&#xff1a;Sigmoid\u3001Tanh \u7b49\u5bf9\u79f0\u6fc0\u6d3b\u51fd\u6570\u3002<\/li>\n<\/ul>\n<\/li>\n<li>He \u521d\u59cb\u5316&#xff08;MSRA&#xff09;\n<ul>\n<li>\u9488\u5bf9 ReLU \u8bbe\u8ba1&#xff1a;ReLU \u4f1a\u8ba9\u7ea6\u4e00\u534a\u795e\u7ecf\u5143\u5931\u6d3b&#xff0c;\u8f93\u51fa\u65b9\u5dee\u51cf\u534a\u3002<\/li>\n<li>\u7ed3\u8bba&#xff1a;\u6743\u91cd\u65b9\u5dee\u4e3a \\\\(2\/n_{in}\\\\)&#xff0c;\u670d\u4ece\u6b63\u6001\u5206\u5e03 \\\\(N(0,\\\\sqrt{2\/n_{in}})\\\\)\u3002<\/li>\n<li>\u9002\u914d&#xff1a;ReLU \u7cfb\u5217\u6fc0\u6d3b\u51fd\u6570&#xff0c;CNN \u7684\u6807\u51c6\u521d\u59cb\u5316\u65b9\u6848\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<h3>\u5341\u56db\u3001\u6ce8\u610f\u529b\u673a\u5236\u4e0e Transformer \u8fdb\u9636\u516b\u80a1<\/h3>\n<h4>1. \u81ea\u6ce8\u610f\u529b\u6838\u5fc3\u8003\u70b9<\/h4>\n<ul>\n<li>\u4e3a\u4ec0\u4e48\u9664\u4ee5\\\\(\\\\sqrt{d_k}\\\\)&#xff1f; \u70b9\u79ef\u7ed3\u679c\u7684\u65b9\u5dee\u4e3a\\\\(d_k\\\\)&#xff0c;\u7ef4\u5ea6\u8f83\u5927\u65f6\u70b9\u79ef\u503c\u6ce2\u52a8\u6781\u5927&#xff0c;\u4f1a\u5bfc\u81f4 softmax \u8fdb\u5165\u9971\u548c\u533a&#xff0c;\u68af\u5ea6\u8d8b\u8fd1\u4e8e 0\u3002\u9664\u4ee5\\\\(\\\\sqrt{d_k}\\\\)\u540e\u70b9\u79ef\u65b9\u5dee\u5f52\u4e00\u5316\u4e3a 1&#xff0c;softmax \u5206\u5e03\u66f4\u5e73\u7f13&#xff0c;\u68af\u5ea6\u7a33\u5b9a\u3002<\/li>\n<li>\u81ea\u6ce8\u610f\u529b\u590d\u6742\u5ea6&#xff1a;\u5e8f\u5217\u957f\u5ea6N\u3001\u7279\u5f81\u7ef4\u5ea6d&#xff0c;\u65f6\u95f4 \/ \u7a7a\u95f4\u590d\u6742\u5ea6\u5747\u4e3a\\\\(O(N^2d)\\\\)\u3002\u5e8f\u5217\u957f\u5ea6\u662f\u6027\u80fd\u74f6\u9888&#xff0c;\u56e0\u6b64\u884d\u751f\u51fa\u6ed1\u52a8\u7a97\u53e3\u6ce8\u610f\u529b\u3001\u7a00\u758f\u6ce8\u610f\u529b\u3001\u7ebf\u6027\u6ce8\u610f\u529b\u7b49\u4f18\u5316\u65b9\u5411\u3002<\/li>\n<li>\u591a\u5934\u6ce8\u610f\u529b\u7684\u4f5c\u7528\n<li>\u591a\u5b50\u7a7a\u95f4\u5efa\u6a21&#xff1a;\u4e0d\u540c\u5934\u5b66\u4e60\u4e0d\u540c\u6a21\u5f0f\u7684\u6ce8\u610f\u529b&#xff08;\u5c40\u90e8\u7eb9\u7406\u3001\u5168\u5c40\u8bed\u4e49\u3001\u8fb9\u7f18\u8f6e\u5ed3\u7b49&#xff09;&#xff0c;\u7279\u5f81\u8868\u8fbe\u66f4\u4e30\u5bcc\u3002<\/li>\n<li>\u8ba1\u7b97\u6548\u7387&#xff1a;\u62c6\u5206\u591a\u5934\u540e\u5355\u5934\u7ef4\u5ea6\u964d\u4f4e&#xff0c;\u603b\u8ba1\u7b97\u91cf\u4e0e\u5355\u5934\u76f8\u5f53&#xff0c;\u4f46\u8868\u8fbe\u80fd\u529b\u663e\u8457\u63d0\u5347\u3002<\/li>\n<\/li>\n<\/ul>\n<h4>2. \u4f4d\u7f6e\u7f16\u7801\u5168\u89e3\u6790<\/h4>\n<p>\u8868\u683c<\/p>\n<table>\n<tr>\u4f4d\u7f6e\u7f16\u7801\u7c7b\u578b\u539f\u7406\u4f18\u70b9\u7f3a\u70b9\u4ee3\u8868\u6a21\u578b<\/tr>\n<tbody>\n<tr>\n<td>\u6b63\u5f26\u4f59\u5f26\u7f16\u7801<\/td>\n<td>\u4e0d\u540c\u9891\u7387\u6b63\u4f59\u5f26\u51fd\u6570&#xff0c;\u56fa\u5b9a\u4e0d\u53ef\u5b66\u4e60<\/td>\n<td>\u652f\u6301\u957f\u5ea6\u5916\u63a8&#xff0c;\u65e0\u989d\u5916\u53c2\u6570\u91cf<\/td>\n<td>\u62df\u5408\u80fd\u529b\u6709\u9650<\/td>\n<td>\u539f\u751f Transformer<\/td>\n<\/tr>\n<tr>\n<td>\u53ef\u5b66\u4e60\u4f4d\u7f6e\u7f16\u7801<\/td>\n<td>\u76f4\u63a5\u521d\u59cb\u5316\u53ef\u5b66\u4e60\u53c2\u6570\u5411\u91cf<\/td>\n<td>\u62df\u5408\u80fd\u529b\u5f3a&#xff0c;\u6548\u679c\u597d<\/td>\n<td>\u65e0\u6cd5\u5916\u63a8\u8d85\u957f\u5e8f\u5217<\/td>\n<td>ViT\u3001BERT<\/td>\n<\/tr>\n<tr>\n<td>\u76f8\u5bf9\u4f4d\u7f6e\u7f16\u7801<\/td>\n<td>\u7f16\u7801 token \u95f4\u7684\u76f8\u5bf9\u8ddd\u79bb<\/td>\n<td>\u7b26\u5408\u89c6\u89c9 \/ \u8bed\u8a00\u76f4\u89c9&#xff0c;\u6cdb\u5316\u6027\u597d<\/td>\n<td>\u5b9e\u73b0\u7a0d\u590d\u6742<\/td>\n<td>Swin Transformer<\/td>\n<\/tr>\n<tr>\n<td>RoPE \u65cb\u8f6c\u4f4d\u7f6e\u7f16\u7801<\/td>\n<td>\u5bf9 Q\/K \u65bd\u52a0\u65cb\u8f6c\u77e9\u9635\u6ce8\u5165\u4f4d\u7f6e\u4fe1\u606f<\/td>\n<td>\u4fdd\u7559\u76f8\u5bf9\u4f4d\u7f6e&#xff0c;\u652f\u6301\u957f\u5ea6\u5916\u63a8<\/td>\n<td>\u9700\u4fee\u6539\u6ce8\u610f\u529b\u8ba1\u7b97<\/td>\n<td>LLaMA\u3001\u5927\u6a21\u578b\u4e3b\u6d41<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<h4>3. Pre-LN vs Post-LN<\/h4>\n<ul>\n<li>Post-LN&#xff1a;\u539f\u751f Transformer \u7ed3\u6784&#xff0c;\u987a\u5e8f\u4e3a\u300cAttention\/FFN \u2192 \u6b8b\u5dee\u76f8\u52a0 \u2192 LayerNorm\u300d\u3002\n<ul>\n<li>\u95ee\u9898&#xff1a;\u8bad\u7ec3\u521d\u671f\u5e95\u5c42\u68af\u5ea6\u5927\u3001\u9876\u5c42\u68af\u5ea6\u5c0f&#xff0c;\u6df1\u5c42\u7f51\u7edc\u8bad\u7ec3\u6781\u4e0d\u7a33\u5b9a&#xff0c;\u5fc5\u987b\u914d\u5408\u5b66\u4e60\u7387 Warmup\u3002<\/li>\n<\/ul>\n<\/li>\n<li>Pre-LN&#xff1a;\u987a\u5e8f\u4e3a\u300cLayerNorm \u2192 Attention\/FFN \u2192 \u6b8b\u5dee\u76f8\u52a0\u300d&#xff0c;ViT\u3001GPT\u3001\u5927\u6a21\u578b\u5747\u91c7\u7528\u6b64\u7ed3\u6784\u3002\n<ul>\n<li>\u4f18\u70b9&#xff1a;\u6bcf\u5c42\u8f93\u5165\u90fd\u7ecf\u8fc7\u5f52\u4e00\u5316&#xff0c;\u68af\u5ea6\u5c3a\u5ea6\u59cb\u7ec8\u7a33\u5b9a&#xff0c;\u6df1\u5c42\u7f51\u7edc\u6613\u8bad\u7ec3&#xff0c;\u5bf9 Warmup \u4f9d\u8d56\u4f4e\u3002<\/li>\n<li>\u7f3a\u70b9&#xff1a;\u6700\u7ec8\u8f93\u51fa\u65e0\u5f52\u4e00\u5316&#xff0c;\u9876\u5c42\u7279\u5f81\u8868\u8fbe\u529b\u7565\u6709\u635f\u5931\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<h4>4. Transformer \u6df1\u5c42\u68af\u5ea6\u95ee\u9898<\/h4>\n<ul>\n<li>\u4e3a\u4ec0\u4e48\u51e0\u5341\u4e0a\u767e\u5c42\u7684 Transformer \u4e0d\u4f1a\u4e25\u91cd\u68af\u5ea6\u6d88\u5931&#xff1f;\n<li>\u6b8b\u5dee\u8fde\u63a5&#xff1a;\u68af\u5ea6\u53ef\u901a\u8fc7 shortcut \u76f4\u63a5\u56de\u4f20&#xff0c;\u4ece\u6839\u672c\u4e0a\u7f29\u77ed\u68af\u5ea6\u4f20\u64ad\u8def\u5f84\u3002<\/li>\n<li>Pre-LN \u5f52\u4e00\u5316&#xff1a;\u4fdd\u8bc1\u6bcf\u5c42\u8f93\u5165\u5c3a\u5ea6\u7a33\u5b9a&#xff0c;\u68af\u5ea6\u4e0d\u4f1a\u6307\u6570\u7ea7\u8870\u51cf\u3002<\/li>\n<li>\u81ea\u6ce8\u610f\u529b\u5168\u5c40\u8fde\u63a5&#xff1a;\u6bcf\u4e2a token \u76f4\u63a5\u4e0e\u6240\u6709 token \u4ea4\u4e92&#xff0c;\u68af\u5ea6\u4f20\u64ad\u8def\u5f84\u77ed\u3002<\/li>\n<\/li>\n<\/ul>\n<h3>\u5341\u4e94\u3001\u81ea\u76d1\u7763\u5b66\u4e60&#xff08;SSL&#xff09;\u516b\u80a1<\/h3>\n<h4>1. \u81ea\u76d1\u7763\u6838\u5fc3\u601d\u60f3\u4e0e\u4ef7\u503c<\/h4>\n<ul>\n<li>\u65e0\u9700\u4eba\u5de5\u6807\u6ce8&#xff0c;\u4ece\u6570\u636e\u672c\u8eab\u6784\u9020\u76d1\u7763\u4fe1\u53f7\u9884\u8bad\u7ec3&#xff0c;\u5b66\u4e60\u901a\u7528\u89c6\u89c9\u8868\u5f81&#xff0c;\u4e0b\u6e38\u4efb\u52a1\u5fae\u8c03\u5373\u53ef\u9002\u914d\u3002<\/li>\n<li>\u6838\u5fc3\u4ef7\u503c&#xff1a;\u53ef\u5229\u7528\u6d77\u91cf\u65e0\u6807\u6ce8\u6570\u636e&#xff1b;\u9884\u8bad\u7ec3\u5f97\u5230\u7684\u7279\u5f81\u8fc1\u79fb\u6027\u3001\u6cdb\u5316\u6027\u66f4\u5f3a&#xff1b;\u5c0f\u6837\u672c\u3001\u5c11\u6837\u672c\u573a\u666f\u4e0b\u8fdc\u4f18\u4e8e\u76d1\u7763\u9884\u8bad\u7ec3\u3002<\/li>\n<\/ul>\n<h4>2. \u5bf9\u6bd4\u5b66\u4e60\u6838\u5fc3\u539f\u7406<\/h4>\n<ul>\n<li>\u6838\u5fc3\u76ee\u6807&#xff1a;\u8ba9\u6b63\u6837\u672c\u5bf9\u7684\u7279\u5f81\u8ddd\u79bb\u62c9\u8fd1&#xff0c;\u8d1f\u6837\u672c\u5bf9\u7684\u7279\u5f81\u8ddd\u79bb\u63a8\u8fdc&#xff0c;\u5b66\u4e60\u5177\u6709\u5224\u522b\u6027\u7684\u8868\u5f81\u3002<\/li>\n<li>InfoNCE \u635f\u5931&#xff1a; \\\\(L &#061; -\\\\log \\\\frac{\\\\exp(\\\\text{sim}(z_i,z_j)\/\\\\tau)}{\\\\sum_{k&#061;1}^{2N} \\\\exp(\\\\text{sim}(z_i,z_k)\/\\\\tau)}\\\\)<\/li>\n<li>\u6e29\u5ea6\u7cfb\u6570\\\\(\\\\tau\\\\)\u7684\u4f5c\u7528&#xff1a;\u63a7\u5236\u5206\u5e03\u5c16\u9510\u7a0b\u5ea6\u3002\\\\(\\\\tau\\\\)\u8d8a\u5c0f&#xff0c;\u5bf9\u96be\u8d1f\u6837\u672c\u7684\u60e9\u7f5a\u8d8a\u5f3a&#xff0c;\u533a\u5206\u5ea6\u8d8a\u9ad8&#xff0c;\u4f46\u6613\u8fc7\u62df\u5408&#xff1b;\\\\(\\\\tau\\\\)\u8d8a\u5927&#xff0c;\u5206\u5e03\u8d8a\u5e73\u6ed1&#xff0c;\u533a\u5206\u5ea6\u5f31\u3002\u89c6\u89c9\u4efb\u52a1\u5e38\u7528 0.05~0.1\u3002<\/li>\n<\/ul>\n<h4>3. \u7ecf\u5178\u5bf9\u6bd4\u5b66\u4e60\u7b97\u6cd5<\/h4>\n<ul>\n<li>SimCLR&#xff1a;\u540c\u4e00\u5f20\u56fe\u505a\u4e24\u6b21\u4e0d\u540c\u589e\u5f3a\u5f97\u5230\u4e24\u4e2a\u89c6\u56fe&#xff0c;\u4e92\u4e3a\u6b63\u6837\u672c&#xff0c;Batch \u5185\u5176\u4f59\u6240\u6709\u6837\u672c\u4e3a\u8d1f\u6837\u672c\u3002\n<ul>\n<li>\u6838\u5fc3\u7ed3\u8bba&#xff1a;\u5f3a\u6570\u636e\u589e\u5f3a\u662f\u6027\u80fd\u5173\u952e&#xff1b;\u5927 Batch \u63d0\u4f9b\u5145\u8db3\u8d1f\u6837\u672c&#xff1b;MLP \u6295\u5f71\u5934\u5927\u5e45\u63d0\u5347\u8868\u5f81\u8d28\u91cf\u3002<\/li>\n<\/ul>\n<\/li>\n<li>MoCo \u7cfb\u5217&#xff1a;\u52a8\u91cf\u7f16\u7801\u5668 &#043; \u961f\u5217\u5b57\u5178&#xff0c;\u7ef4\u62a4\u4e00\u4e2a\u5927\u5bb9\u91cf\u8d1f\u6837\u672c\u961f\u5217&#xff0c;\u65e0\u9700\u5927 Batch \u5373\u53ef\u83b7\u5f97\u5927\u91cf\u8d1f\u6837\u672c&#xff0c;\u5de5\u7a0b\u843d\u5730\u6027\u66f4\u5f3a\u3002<\/li>\n<li>BYOL&#xff1a;\u65e0\u8d1f\u6837\u672c\u5bf9\u6bd4\u5b66\u4e60&#xff0c;\u4ec5\u901a\u8fc7\u6b63\u6837\u672c\u4e00\u81f4\u6027\u5b66\u4e60&#xff0c;\u4f9d\u9760\u52a8\u91cf\u7f16\u7801\u5668 &#043; \u505c\u6b62\u68af\u5ea6\u907f\u514d\u6a21\u578b\u5d29\u584c&#xff08;\u6240\u6709\u6837\u672c\u7279\u5f81\u8d8b\u540c&#xff09;\u3002<\/li>\n<li>DINO&#xff1a;\u81ea\u76d1\u7763\u84b8\u998f\u65b9\u6848&#xff0c;\u5b66\u751f\u7f51\u7edc\u62df\u5408\u52a8\u91cf\u6559\u5e08\u7684\u8f93\u51fa&#xff0c;\u65e0\u9700\u8d1f\u6837\u672c&#xff0c;\u5728 ViT \u4e0a\u6548\u679c\u7a81\u51fa&#xff0c;\u6ce8\u610f\u529b\u5929\u7136\u5bf9\u5e94\u7269\u4f53\u8bed\u4e49\u533a\u57df\u3002<\/li>\n<\/ul>\n<h4>4. \u63a9\u7801\u81ea\u7f16\u7801\u5668&#xff08;MAE&#xff09;<\/h4>\n<ul>\n<li>\u6838\u5fc3\u673a\u5236&#xff1a;\u968f\u673a\u63a9\u7801\u56fe\u50cf 75% \u7684 Patch&#xff0c;\u8ba9\u6a21\u578b\u91cd\u5efa\u88ab\u63a9\u7801\u533a\u57df\u7684\u50cf\u7d20\u3002\u7f16\u7801\u5668\u4ec5\u5904\u7406\u53ef\u89c1 Patch&#xff0c;\u89e3\u7801\u5668\u8f7b\u91cf\u8d1f\u8d23\u91cd\u5efa\u3002<\/li>\n<li>\u4e0e\u5bf9\u6bd4\u5b66\u4e60\u7684\u5dee\u5f02&#xff1a;\u5bf9\u6bd4\u5b66\u4e60\u5b66\u4e60\u5224\u522b\u6027\u8868\u5f81&#xff0c;\u9002\u914d\u5206\u7c7b\u3001\u68c0\u6d4b&#xff1b;MAE \u5b66\u4e60\u91cd\u5efa\u6027\u8868\u5f81&#xff0c;\u5bf9\u56fe\u50cf\u7ed3\u6784\u3001\u7ec6\u8282\u7406\u89e3\u66f4\u6df1&#xff0c;\u9002\u914d\u5206\u5272\u3001\u751f\u6210\u3001\u4f4e\u5c42\u6b21\u89c6\u89c9\u4efb\u52a1\u3002<\/li>\n<\/ul>\n<h3>\u5341\u516d\u3001\u5927\u6a21\u578b\u4e0e\u591a\u6a21\u6001\u6838\u5fc3\u516b\u80a1<\/h3>\n<h4>1. \u5927\u6a21\u578b\u5fae\u8c03\u6280\u672f\u5168\u5bf9\u6bd4<\/h4>\n<p>\u8868\u683c<\/p>\n<table>\n<tr>\u5fae\u8c03\u65b9\u6cd5\u539f\u7406\u53c2\u6570\u91cf\u663e\u5b58\u5360\u7528\u63a8\u7406\u5f00\u9500\u6548\u679c<\/tr>\n<tbody>\n<tr>\n<td>\u5168\u53c2\u5fae\u8c03<\/td>\n<td>\u66f4\u65b0\u6a21\u578b\u5168\u90e8\u53c2\u6570<\/td>\n<td>100%<\/td>\n<td>\u6781\u9ad8<\/td>\n<td>\u65e0\u989d\u5916\u5f00\u9500<\/td>\n<td>\u6700\u597d<\/td>\n<\/tr>\n<tr>\n<td>LoRA<\/td>\n<td>\u51bb\u7ed3\u539f\u6743\u91cd&#xff0c;\u8bad\u7ec3\u4f4e\u79e9\u65c1\u8def\u77e9\u9635<\/td>\n<td>\u6781\u4f4e&#xff08;\u79e9 r \u8fdc\u5c0f\u4e8e\u539f\u7ef4\u5ea6&#xff09;<\/td>\n<td>\u4f4e<\/td>\n<td>\u53ef\u5408\u5e76\u6743\u91cd&#xff0c;\u65e0\u989d\u5916\u5ef6\u8fdf<\/td>\n<td>\u63a5\u8fd1\u5168\u53c2<\/td>\n<\/tr>\n<tr>\n<td>QLoRA<\/td>\n<td>\u57fa\u5ea7\u6a21\u578b\u91cf\u5316\u5230 4bit&#043;LoRA \u5fae\u8c03<\/td>\n<td>\u6781\u4f4e<\/td>\n<td>\u6781\u4f4e<\/td>\n<td>\u6709\u91cf\u5316\u7cbe\u5ea6\u635f\u5931<\/td>\n<td>\u63a5\u8fd1\u5168\u53c2<\/td>\n<\/tr>\n<tr>\n<td>Prefix Tuning<\/td>\n<td>\u4ec5\u4f18\u5316\u524d\u7f00 Prompt \u5411\u91cf<\/td>\n<td>\u6781\u5c11<\/td>\n<td>\u4f4e<\/td>\n<td>\u6709\u989d\u5916\u5e8f\u5217\u957f\u5ea6\u5f00\u9500<\/td>\n<td>\u5f31\u4e8e LoRA<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<ul>\n<li>LoRA \u6838\u5fc3\u7ec6\u8282&#xff1a;\u901a\u5e38\u4ec5\u5fae\u8c03\u6ce8\u610f\u529b\u7684 Q\u3001V \u77e9\u9635\u3002\u539f\u56e0&#xff1a;\u6ce8\u610f\u529b\u7684\u67e5\u8be2\u4e0e\u503c\u4ea4\u4e92\u662f\u4fe1\u606f\u5904\u7406\u7684\u6838\u5fc3&#xff0c;\u5fae\u8c03\u8fd9\u4e24\u4e2a\u77e9\u9635\u6548\u679c\u8db3\u591f\u597d&#xff0c;\u540c\u65f6\u53c2\u6570\u91cf\u6700\u5c0f\u3002<\/li>\n<\/ul>\n<h4>2. \u89c6\u89c9\u591a\u6a21\u6001\u6838\u5fc3\u6a21\u578b<\/h4>\n<ul>\n<li>CLIP\n<ul>\n<li>\u539f\u7406&#xff1a;\u6d77\u91cf\u56fe\u6587\u5bf9\u5bf9\u6bd4\u9884\u8bad\u7ec3&#xff0c;\u56fe\u50cf\u7f16\u7801\u5668\u4e0e\u6587\u672c\u7f16\u7801\u5668\u5206\u522b\u7f16\u7801&#xff0c;\u8ba9\u5339\u914d\u7684\u56fe\u6587\u7279\u5f81\u62c9\u8fd1\u3001\u4e0d\u5339\u914d\u7684\u63a8\u8fdc&#xff0c;\u5b9e\u73b0\u8de8\u6a21\u6001\u5bf9\u9f50\u3002<\/li>\n<li>\u80fd\u529b&#xff1a;\u96f6\u6837\u672c\u56fe\u50cf\u5206\u7c7b\u3001\u8de8\u6a21\u6001\u68c0\u7d22\u3001\u8fc1\u79fb\u6027\u6781\u5f3a\u3002<\/li>\n<li>\u5c40\u9650&#xff1a;\u7ec6\u7c92\u5ea6\u5206\u7c7b\u80fd\u529b\u5f31\u3001\u7a7a\u95f4\u4f4d\u7f6e\u7406\u89e3\u5dee\u3001\u5bf9\u8bad\u7ec3\u6570\u636e\u5206\u5e03\u654f\u611f\u3002<\/li>\n<\/ul>\n<\/li>\n<li>BLIP \u7cfb\u5217&#xff1a;\u5728 CLIP \u5bf9\u6bd4\u9884\u8bad\u7ec3\u57fa\u7840\u4e0a\u52a0\u5165\u751f\u6210\u5f0f\u9884\u8bad\u7ec3&#xff0c;\u652f\u6301\u56fe\u6587\u68c0\u7d22\u3001\u56fe\u50cf\u63cf\u8ff0\u3001\u89c6\u89c9\u95ee\u7b54\u7b49\u591a\u7c7b\u4efb\u52a1&#xff0c;\u591a\u6a21\u6001\u7406\u89e3\u80fd\u529b\u66f4\u5168\u9762\u3002<\/li>\n<\/ul>\n<h4>3. \u5927\u6a21\u578b\u9ad8\u9891\u6982\u5ff5<\/h4>\n<ul>\n<li>\u6d8c\u73b0\u80fd\u529b&#xff1a;\u6a21\u578b\u89c4\u6a21\u8fbe\u5230\u7279\u5b9a\u9608\u503c\u540e&#xff0c;\u6027\u80fd\u51fa\u73b0\u975e\u7ebf\u6027\u8dc3\u5347&#xff0c;\u5c0f\u6a21\u578b\u4e0d\u5177\u5907\u7684\u80fd\u529b\u7a81\u7136\u51fa\u73b0\u3002<\/li>\n<li>\u5e7b\u89c9&#xff1a;\u6a21\u578b\u8f93\u51fa\u770b\u4f3c\u5408\u7406\u4f46\u4e0e\u4e8b\u5b9e\u4e0d\u7b26\u7684\u5185\u5bb9&#xff0c;\u591a\u6a21\u6001\u573a\u666f\u4e0b\u8868\u73b0\u4e3a\u865a\u6784\u7269\u4f53\u3001\u63cf\u8ff0\u4e0e\u56fe\u50cf\u5185\u5bb9\u4e0d\u7b26\u3002<\/li>\n<li>\u5bf9\u9f50&#xff1a;\u901a\u8fc7\u6307\u4ee4\u5fae\u8c03\u3001\u4eba\u7c7b\u53cd\u9988\u5f3a\u5316\u5b66\u4e60\u7b49\u65b9\u5f0f&#xff0c;\u8ba9\u6a21\u578b\u8f93\u51fa\u7b26\u5408\u4eba\u7c7b\u8ba4\u77e5\u3001\u4ef7\u503c\u89c2\u4e0e\u4e8b\u5b9e\u6807\u51c6\u3002<\/li>\n<\/ul>\n<h3>\u5341\u4e03\u3001\u5206\u5e03\u5f0f\u8bad\u7ec3\u4e0e\u6df7\u5408\u7cbe\u5ea6<\/h3>\n<h4>1. \u5e76\u884c\u8bad\u7ec3\u5206\u7c7b\u4e0e\u9009\u578b<\/h4>\n<ul>\n<li>\u6570\u636e\u5e76\u884c&#xff08;DDP&#xff09;&#xff1a;\u6bcf\u5f20\u5361\u590d\u5236\u5b8c\u6574\u6a21\u578b&#xff0c;\u6570\u636e\u62c6\u5206\u5230\u5404\u5361&#xff0c;\u72ec\u7acb\u524d\u5411\u53cd\u5411&#xff0c;\u540c\u6b65\u68af\u5ea6\u540e\u7edf\u4e00\u66f4\u65b0\u3002\n<ul>\n<li>\u9002\u7528&#xff1a;\u6a21\u578b\u5355\u5361\u53ef\u5bb9\u7eb3&#xff0c;\u8ffd\u6c42\u8bad\u7ec3\u63d0\u901f\u3002\u5de5\u4e1a\u754c\u6700\u5e38\u7528\u65b9\u6848&#xff0c;\u73af\u5f0f\u68af\u5ea6\u540c\u6b65\u6548\u7387\u9ad8\u3002<\/li>\n<\/ul>\n<\/li>\n<li>\u6a21\u578b\u5e76\u884c&#xff1a;\u6a21\u578b\u62c6\u5206\u5230\u591a\u5f20\u5361&#xff0c;\u6bcf\u5f20\u5361\u5b58\u50a8\u90e8\u5206\u53c2\u6570\u3002\n<ul>\n<li>\u5f20\u91cf\u5e76\u884c&#xff1a;\u6309\u77e9\u9635\u7ef4\u5ea6\u62c6\u5206\u5355\u5c42&#xff0c;\u8ba1\u7b97\u65f6\u8de8\u5361\u901a\u4fe1&#xff0c;\u901a\u4fe1\u5f00\u9500\u5927\u3002<\/li>\n<li>\u6d41\u6c34\u7ebf\u5e76\u884c&#xff1a;\u6309\u7f51\u7edc\u5c42\u62c6\u5206&#xff0c;\u6d41\u6c34\u7ebf\u5f0f\u6267\u884c&#xff0c;\u901a\u4fe1\u91cf\u5c0f\u4f46\u5b58\u5728\u6d41\u6c34\u7ebf\u6c14\u6ce1\u3002<\/li>\n<\/ul>\n<\/li>\n<li>3D \u5e76\u884c&#xff1a;\u6570\u636e\u5e76\u884c &#043; \u5f20\u91cf\u5e76\u884c &#043; \u6d41\u6c34\u7ebf\u5e76\u884c\u7ed3\u5408&#xff0c;\u9002\u914d\u8d85\u5927\u6a21\u578b\u9884\u8bad\u7ec3\u3002<\/li>\n<\/ul>\n<h4>2. \u6df7\u5408\u7cbe\u5ea6\u8bad\u7ec3<\/h4>\n<ul>\n<li>\u6838\u5fc3\u673a\u5236&#xff1a;\u6743\u91cd\u3001\u68af\u5ea6\u4e3b\u5907\u4efd\u7528 FP32 \u4fdd\u8bc1\u7cbe\u5ea6&#xff0c;\u524d\u5411\u53cd\u5411\u8ba1\u7b97\u7528 FP16\/BF16 \u63d0\u901f\u3001\u7701\u663e\u5b58\u3002<\/li>\n<li>\u63d0\u901f\u539f\u7406&#xff1a;FP16 \u4f4d\u5bbd\u51cf\u534a&#xff0c;\u663e\u5b58\u5360\u7528\u51cf\u534a&#xff1b;GPU \u5f20\u91cf\u6838\u5fc3\u5bf9 FP16 \u7684\u8ba1\u7b97\u541e\u5410\u91cf\u662f FP32 \u7684\u6570\u500d\u3002<\/li>\n<li>\u635f\u5931\u7f29\u653e&#xff08;Loss Scaling&#xff09;\n<ul>\n<li>\u95ee\u9898&#xff1a;FP16 \u52a8\u6001\u8303\u56f4\u5c0f&#xff0c;\u5c0f\u68af\u5ea6\u4f1a\u4e0b\u6ea2\u4e3a 0&#xff0c;\u5bfc\u81f4\u68af\u5ea6\u6d88\u5931\u3002<\/li>\n<li>\u89e3\u51b3&#xff1a;\u635f\u5931\u4e58\u4ee5\u7f29\u653e\u56e0\u5b50&#xff0c;\u53cd\u5411\u4f20\u64ad\u68af\u5ea6\u540c\u6b65\u653e\u5927&#xff0c;\u907f\u514d\u4e0b\u6ea2&#xff1b;\u66f4\u65b0\u6743\u91cd\u65f6\u8fd8\u539f\u7f29\u653e&#xff0c;\u4e0d\u5f71\u54cd\u7cbe\u5ea6\u3002<\/li>\n<\/ul>\n<\/li>\n<li>BF16 vs FP16&#xff1a;BF16 \u52a8\u6001\u8303\u56f4\u4e0e FP32 \u4e00\u81f4&#xff0c;\u4e0d\u4f1a\u6ea2\u51fa&#xff0c;\u65e0\u9700\u635f\u5931\u7f29\u653e&#xff1b;\u4f46\u5c3e\u6570\u7cbe\u5ea6\u66f4\u4f4e\u3002\u73b0\u4ee3\u6df1\u5ea6\u5b66\u4e60\u8bad\u7ec3\u4f18\u5148\u9009 BF16\u3002<\/li>\n<\/ul>\n<h4>3. \u68af\u5ea6\u7d2f\u79ef<\/h4>\n<ul>\n<li>\u539f\u7406&#xff1a;\u591a\u4e2a Step \u4e0d\u66f4\u65b0\u53c2\u6570&#xff0c;\u7d2f\u79ef\u68af\u5ea6\u540e\u7edf\u4e00\u66f4\u65b0\u3002<\/li>\n<li>\u4f5c\u7528&#xff1a;\u663e\u5b58\u4e0d\u8db3\u65f6\u6a21\u62df\u5927 Batch \u6548\u679c\u3002<\/li>\n<li>\u6ce8\u610f&#xff1a;BN \u7684\u6279\u6b21\u7edf\u8ba1\u91cf\u6309\u5b9e\u9645 Batch \u8ba1\u7b97&#xff0c;\u68af\u5ea6\u7d2f\u79ef\u4e0d\u4f1a\u6539\u53d8 BN \u7684\u6709\u6548 Batch \u5927\u5c0f&#xff0c;\u56e0\u6b64\u65e0\u6cd5\u5b8c\u5168\u7b49\u4ef7\u4e8e\u771f\u5b9e\u5927 Batch\u3002<\/li>\n<\/ul>\n<h3>\u5341\u516b\u3001\u8bad\u7ec3\u8c03\u4f18\u5b9e\u6218\u6392\u969c\u516b\u80a1<\/h3>\n<h4>1. Loss \u4e0d\u6536\u655b\u6392\u67e5\u601d\u8def<\/h4>\n<p>\u6309\u4f18\u5148\u7ea7\u4ece\u9ad8\u5230\u4f4e\u6392\u67e5&#xff1a;<\/p>\n<li>\u6570\u636e\u5c42&#xff1a;\u6807\u7b7e\u9519\u8bef\u3001\u9884\u5904\u7406\u5f52\u4e00\u5316\u5f02\u5e38\u3001\u6570\u636e\u96c6\u672c\u8eab\u4e0d\u53ef\u5206\u3001\u7c7b\u522b\u6781\u5ea6\u4e0d\u5747\u8861\u3002<\/li>\n<li>\u4ee3\u7801\u5c42&#xff1a;\u635f\u5931\u51fd\u6570\u8ba1\u7b97\u9519\u8bef\u3001\u68af\u5ea6\u672a\u56de\u4f20\u3001\u6a21\u578b\u53c2\u6570\u88ab\u51bb\u7ed3\u3001\u5f52\u4e00\u5316\u5c42\u8bad\u7ec3 \/ \u63a8\u7406\u6a21\u5f0f\u6df7\u6dc6\u3001\u6570\u636e\u52a0\u8f7d Bug\u3002<\/li>\n<li>\u8d85\u53c2\u5c42&#xff1a;\u5b66\u4e60\u7387\u8fc7\u5927\u5bfc\u81f4 Loss \u7206\u70b8\u3001\u5b66\u4e60\u7387\u8fc7\u5c0f\u5bfc\u81f4 Loss \u505c\u6ede\u3001\u6b63\u5219\u5f3a\u5ea6\u8fc7\u9ad8\u3001\u4f18\u5316\u5668\u9009\u578b\u4e0d\u5f53\u3002<\/li>\n<li>\u6a21\u578b\u5c42&#xff1a;\u7f51\u7edc\u7ed3\u6784\u9519\u8bef\u5bfc\u81f4\u68af\u5ea6\u4f20\u64ad\u4e2d\u65ad\u3001\u6a21\u578b\u5bb9\u91cf\u4e0d\u8db3\u4ee5\u62df\u5408\u4efb\u52a1\u3002<\/li>\n<h4>2. Nan\/Inf Loss \u5e38\u89c1\u539f\u56e0<\/h4>\n<li>\u5b66\u4e60\u7387\u8fc7\u5927&#xff0c;\u68af\u5ea6\u7206\u70b8&#xff0c;\u53c2\u6570\u66f4\u65b0\u5e45\u5ea6\u8d85\u51fa\u6570\u503c\u8303\u56f4\u3002<\/li>\n<li>\u635f\u5931\u8ba1\u7b97\u5f02\u5e38&#xff1a;\u9664\u96f6\u9519\u8bef\u3001log (0)\u3001\u4ea4\u53c9\u71b5\u4e2d\u9884\u6d4b\u6982\u7387\u4e3a 0\u3002<\/li>\n<li>\u8f93\u5165\u6570\u636e\u672c\u8eab\u5305\u542b Nan \u503c\u3002<\/li>\n<li>BN \u5c42\u65b9\u5dee\u4e3a 0&#xff0c;\u51fa\u73b0\u9664\u96f6\u9519\u8bef\u3002<\/li>\n<li>\u6df7\u5408\u7cbe\u5ea6\u4e0b\u6570\u503c\u4e0a\u6ea2 \/ \u4e0b\u6ea2&#xff0c;\u635f\u5931\u7f29\u653e\u53c2\u6570\u4e0d\u5f53\u3002<\/li>\n<h4>3. Loss \u9707\u8361\u7684\u539f\u56e0<\/h4>\n<li>\u5b66\u4e60\u7387\u8fc7\u9ad8&#xff0c;\u53c2\u6570\u5728\u6700\u4f18\u70b9\u9644\u8fd1\u6765\u56de\u8df3\u52a8\u3002<\/li>\n<li>Batch Size \u8fc7\u5c0f&#xff0c;\u6279\u6b21\u566a\u58f0\u5927&#xff0c;\u68af\u5ea6\u65b9\u5411\u6ce2\u52a8\u5267\u70c8\u3002<\/li>\n<li>\u6570\u636e\u672c\u8eab\u566a\u58f0\u5927\u3001\u6807\u6ce8\u8d28\u91cf\u5dee\u3002<\/li>\n<li>\u6b63\u5219\u5f3a\u5ea6\u4e0d\u8db3&#xff0c;\u6a21\u578b\u5728\u6837\u672c\u95f4\u53cd\u590d\u62df\u5408\u3002<\/li>\n<li>\u5b66\u4e60\u7387\u8c03\u5ea6\u4e0d\u5408\u7406&#xff0c;\u8bad\u7ec3\u540e\u671f\u672a\u53ca\u65f6\u8870\u51cf\u3002<\/li>\n<h4>4. \u9a8c\u8bc1\u96c6\u7cbe\u5ea6\u9ad8\u4e8e\u8bad\u7ec3\u96c6\u7684\u539f\u56e0<\/h4>\n<li>\u6700\u5e38\u89c1\u539f\u56e0&#xff1a;Dropout\u3001BN \u7b49\u6b63\u5219\u5316\u5c42\u8bad\u7ec3\u65f6\u5f15\u5165\u566a\u58f0 \/ \u6270\u52a8&#xff0c;\u63a8\u7406\u65f6\u5173\u95ed&#xff0c;\u6a21\u578b\u8868\u73b0\u66f4\u7a33\u5b9a\u3002<\/li>\n<li>\u9a8c\u8bc1\u96c6\u6570\u636e\u5206\u5e03\u66f4\u7b80\u5355\u3001\u96be\u5ea6\u4f4e\u4e8e\u8bad\u7ec3\u96c6\u3002<\/li>\n<li>\u8bc4\u4f30\u65f6\u673a\u95ee\u9898&#xff1a;\u8bad\u7ec3\u96c6\u8bc4\u4f30\u65f6\u6a21\u578b\u5904\u4e8e\u8bad\u7ec3\u4e2d\u9014&#xff0c;\u9a8c\u8bc1\u96c6\u5728\u4e00\u8f6e\u8bad\u7ec3\u7ed3\u675f\u540e\u8bc4\u4f30\u3002<\/li>\n<h4>5. Batch Size \u7684\u5f71\u54cd\u4e0e\u9009\u578b<\/h4>\n<ul>\n<li>\u5927 Batch&#xff1a;\u68af\u5ea6\u4f30\u8ba1\u51c6\u786e&#xff0c;\u6536\u655b\u5e73\u7a33&#xff1b;\u8bad\u7ec3\u5e76\u884c\u6548\u7387\u9ad8&#xff1b;\u4f46\u6cdb\u5316\u6027\u901a\u5e38\u66f4\u5dee&#xff0c;\u6613\u6536\u655b\u5230\u5c16\u9510\u6781\u5c0f\u503c&#xff1b;\u663e\u5b58\u9700\u6c42\u9ad8\u3002<\/li>\n<li>\u5c0f Batch&#xff1a;\u68af\u5ea6\u5e26\u566a\u58f0&#xff0c;\u5177\u5907\u9690\u5f0f\u6b63\u5219\u6548\u679c&#xff0c;\u6cdb\u5316\u66f4\u597d&#xff1b;\u4f46\u6536\u655b\u9707\u8361&#xff0c;\u8fed\u4ee3\u6b65\u6570\u591a&#xff1b;\u663e\u5b58\u9700\u6c42\u4f4e\u3002<\/li>\n<li>\u5927 Batch \u6cdb\u5316\u5dee\u7684\u89e3\u51b3\u65b9\u6848&#xff1a;\u5b66\u4e60\u7387\u7ebf\u6027\u7f29\u653e\u3001\u52a0\u957f Warmup\u3001\u589e\u5f3a\u6743\u91cd\u8870\u51cf\u4e0e\u6807\u7b7e\u5e73\u6ed1\u3002<\/li>\n<\/ul>\n<h3>\u5341\u4e5d\u3001\u6df1\u5ea6\u5b66\u4e60\u7406\u8bba\u57fa\u7840\u516b\u80a1<\/h3>\n<h4>1. \u504f\u5dee &#8211; \u65b9\u5dee\u5206\u89e3<\/h4>\n<ul>\n<li>\u6cdb\u5316\u8bef\u5dee &#061; \u504f\u5dee \u00b2 &#043; \u65b9\u5dee &#043; \u4e0d\u53ef\u907f\u514d\u566a\u58f0\u3002<\/li>\n<li>\u504f\u5dee&#xff1a;\u6a21\u578b\u671f\u671b\u9884\u6d4b\u4e0e\u771f\u5b9e\u503c\u7684\u5dee\u8ddd&#xff0c;\u8861\u91cf\u6a21\u578b\u62df\u5408\u80fd\u529b\u3002\u504f\u5dee\u5927\u2192\u6b20\u62df\u5408\u3002<\/li>\n<li>\u65b9\u5dee&#xff1a;\u6a21\u578b\u5728\u4e0d\u540c\u8bad\u7ec3\u96c6\u4e0a\u9884\u6d4b\u7ed3\u679c\u7684\u6ce2\u52a8&#xff0c;\u8861\u91cf\u7a33\u5b9a\u6027\u3002\u65b9\u5dee\u5927\u2192\u8fc7\u62df\u5408\u3002<\/li>\n<li>\u53d8\u5316\u89c4\u5f8b&#xff1a;\u6a21\u578b\u590d\u6742\u5ea6\u63d0\u5347&#xff0c;\u504f\u5dee\u6301\u7eed\u51cf\u5c0f&#xff0c;\u65b9\u5dee\u6301\u7eed\u589e\u5927&#xff0c;\u603b\u8bef\u5dee\u5448\u5148\u964d\u540e\u5347\u7684 U \u578b\u3002<\/li>\n<\/ul>\n<h4>2. \u53cc\u4e0b\u964d\u73b0\u8c61<\/h4>\n<ul>\n<li>\u4f20\u7edf\u7edf\u8ba1\u5b66\u4e60&#xff1a;\u53c2\u6570\u91cf\u8d85\u8fc7\u62df\u5408\u9608\u503c\u540e&#xff0c;\u8fc7\u62df\u5408\u52a0\u5267&#xff0c;\u6cdb\u5316\u8bef\u5dee\u4e0a\u5347&#xff08;U \u578b\u66f2\u7ebf&#xff09;\u3002<\/li>\n<li>\u6df1\u5ea6\u5b66\u4e60&#xff1a;\u53c2\u6570\u91cf\u8d85\u8fc7\u67d0\u4e00\u9608\u503c\u540e&#xff0c;\u6cdb\u5316\u8bef\u5dee\u4f1a\u518d\u6b21\u4e0b\u964d&#xff0c;\u5f62\u6210 \u201c\u53cc\u4e0b\u964d\u201d \u66f2\u7ebf\u3002<\/li>\n<li>\u6838\u5fc3\u539f\u56e0&#xff1a;\u8d85\u53c2\u6570\u5316\u6a21\u578b\u7684\u89e3\u7a7a\u95f4\u66f4\u5927&#xff0c;\u68af\u5ea6\u4e0b\u964d\u503e\u5411\u4e8e\u627e\u5230\u8303\u6570\u66f4\u5c0f\u3001\u66f4\u5e73\u5766\u7684\u89e3&#xff0c;\u6cdb\u5316\u80fd\u529b\u53cd\u800c\u63d0\u5347&#xff1b;\u5927\u6a21\u578b\u81ea\u8eab\u7684\u5f52\u7eb3\u504f\u7f6e\u4e5f\u66f4\u9002\u914d\u6570\u636e\u89c4\u5f8b\u3002<\/li>\n<\/ul>\n<h4>3. \u4e3a\u4ec0\u4e48\u8d85\u53c2\u6570\u6a21\u578b\u4e0d\u4f1a\u4e25\u91cd\u8fc7\u62df\u5408<\/h4>\n<li>\u9690\u5f0f\u6b63\u5219&#xff1a;SGD \u7684\u68af\u5ea6\u566a\u58f0\u3001BN \u7684\u6279\u6b21\u566a\u58f0\u3001\u6743\u91cd\u8870\u51cf\u7b49\u90fd\u6709\u6b63\u5219\u6548\u679c\u3002<\/li>\n<li>\u6570\u636e\u7ea6\u675f&#xff1a;\u6d77\u91cf\u8bad\u7ec3\u6570\u636e\u5bf9\u6a21\u578b\u5f62\u6210\u5f3a\u7ea6\u675f&#xff0c;\u9650\u5236\u4e86\u8fc7\u62df\u5408\u7a7a\u95f4\u3002<\/li>\n<li>\u5f52\u7eb3\u504f\u7f6e&#xff1a;CNN \u7684\u5c40\u90e8\u6027\u3001\u5e73\u79fb\u4e0d\u53d8\u6027&#xff0c;Transformer \u7684\u5168\u5c40\u5efa\u6a21\u80fd\u529b&#xff0c;\u90fd\u7b26\u5408\u771f\u5b9e\u6570\u636e\u7684\u5185\u5728\u89c4\u5f8b\u3002<\/li>\n<li>\u4f18\u5316\u504f\u597d&#xff1a;\u68af\u5ea6\u4e0b\u964d\u9690\u5f0f\u503e\u5411\u4e8e\u627e\u5230\u8303\u6570\u5c0f\u3001\u5e73\u6ed1\u5ea6\u9ad8\u7684\u89e3&#xff0c;\u8fd9\u7c7b\u89e3\u6cdb\u5316\u6027\u66f4\u597d\u3002<\/li>\n<h4>4. \u5c40\u90e8\u6700\u4f18\u4e0e\u978d\u70b9<\/h4>\n<p>\u9ad8\u7ef4\u53c2\u6570\u7a7a\u95f4\u4e2d&#xff0c;\u771f\u6b63\u7684\u5c40\u90e8\u6700\u4f18\u70b9\u6570\u91cf\u5f88\u5c11&#xff0c;\u5927\u91cf\u68af\u5ea6\u4e3a 0 \u7684\u70b9\u662f\u978d\u70b9&#xff08;\u90e8\u5206\u65b9\u5411\u662f\u6781\u5927\u503c\u3001\u90e8\u5206\u65b9\u5411\u662f\u6781\u5c0f\u503c&#xff09;\u3002SGD \u7684\u566a\u58f0\u3001\u52a8\u91cf\u673a\u5236\u53ef\u4ee5\u6709\u6548\u8df3\u51fa\u978d\u70b9&#xff0c;\u56e0\u6b64\u6df1\u5ea6\u5b66\u4e60\u4e2d\u5c40\u90e8\u6700\u4f18\u5e76\u975e\u4e3b\u8981\u95ee\u9898\u3002<\/p>\n<h3>\u4e8c\u5341\u3001\u8865\u5145\u9ad8\u9891\u624b\u6495\u4ee3\u7801<\/h3>\n<h4>1. BatchNorm \u524d\u5411\u5b9e\u73b0<\/h4>\n<p>python<\/p>\n<p>\u8fd0\u884c<\/p>\n<p>def batchnorm_forward(x, gamma, beta, eps&#061;1e-5):<br \/>\n    # x shape: [N, C, H, W]<br \/>\n    mean &#061; x.mean(dim&#061;(0, 2, 3), keepdim&#061;True)<br \/>\n    var &#061; x.var(dim&#061;(0, 2, 3), keepdim&#061;True, unbiased&#061;False)<br \/>\n    x_hat &#061; (x &#8211; mean) \/ torch.sqrt(var &#043; eps)<br \/>\n    return gamma * x_hat &#043; beta<\/p>\n<h4>2. \u6570\u503c\u7a33\u5b9a\u7684\u4ea4\u53c9\u71b5\u635f\u5931<\/h4>\n<p>python<\/p>\n<p>\u8fd0\u884c<\/p>\n<p>def cross_entropy_loss(pred, target):<br \/>\n    # pred: [N, C], target: [N]<br \/>\n    pred_max &#061; pred.max(dim&#061;1, keepdim&#061;True)[0]<br \/>\n    exp_pred &#061; torch.exp(pred &#8211; pred_max)  # \u51cf\u6700\u5927\u503c\u9632\u6ea2\u51fa<br \/>\n    softmax &#061; exp_pred \/ exp_pred.sum(dim&#061;1, keepdim&#061;True)<br \/>\n    p &#061; softmax[torch.arange(len(target)), target]<br \/>\n    return -torch.log(p &#043; 1e-8).mean()<\/p>\n<h4>3. \u5bf9\u6bd4\u5b66\u4e60 InfoNCE \u635f\u5931<\/h4>\n<p>python<\/p>\n<p>\u8fd0\u884c<\/p>\n<p>def info_nce_loss(feat1, feat2, temperature&#061;0.07):<br \/>\n    # feat1, feat2: [N, D] \u4e24\u4e2a\u589e\u5f3a\u89c6\u56fe\u7684\u7279\u5f81<br \/>\n    features &#061; torch.cat([feat1, feat2], dim&#061;0)<br \/>\n    batch_size &#061; feat1.shape[0]<br \/>\n    # \u6b63\u6837\u672c\u6807\u7b7e&#xff1a;\u7b2ci\u4e2a\u5bf9\u5e94\u7b2ci&#043;batch_size\u4e2a<br \/>\n    labels &#061; torch.arange(2 * batch_size).to(features.device)<br \/>\n    labels &#061; (labels &#043; batch_size) % (2 * batch_size)<\/p>\n<p>    sim_matrix &#061; F.cosine_similarity(features.unsqueeze(1), features.unsqueeze(0), dim&#061;-1)<br \/>\n    sim_matrix &#061; sim_matrix \/ temperature<br \/>\n    return F.cross_entropy(sim_matrix, labels)<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u4e00\u3001\u6df1\u5ea6\u5b66\u4e60\u57fa\u7840\u6838\u5fc3\u6982\u5ff51. \u6df1\u5ea6\u5b66\u4e60 vs \u673a\u5668\u5b66\u4e60 vs \u4eba\u5de5\u667a\u80fd\u7684\u5173\u7cfb\u4eba\u5de5\u667a\u80fd&#xff08;AI&#xff09;&#xff1a;\u6700\u5927\u8303\u7574&#xff0c;\u76ee\u6807\u662f\u8ba9\u673a\u5668\u5177\u5907\u4eba\u7c7b\u667a\u80fd\u80fd\u529b&#xff0c;\u6db5\u76d6\u673a\u5668\u5b66\u4e60\u3001\u4e13\u5bb6\u7cfb\u7edf\u3001\u77e5\u8bc6\u56fe\u8c31\u7b49\u3002\u673a\u5668\u5b66\u4e60&#xff08;ML&#xff09;&#xff1a;AI \u7684\u5b50\u96c6&#xff0c;\u901a\u8fc7\u6570\u636e\u548c\u7b97\u6cd5\u8ba9\u673a\u5668\u81ea\u52a8\u5b66\u4e60\u89c4\u5f8b&#xff0c;\u65e0\u9700\u4eba\u5de5\u786c\u7f16\u7801\u89c4\u5219&#xff0c;\u5305\u542b\u4f20\u7edf\u673a\u5668\u5b66\u4e60&#xff08;SVM\u3001\u51b3\u7b56\u6811\u3001LR \u7b49&amp;#xff09<\/p>\n","protected":false},"author":2,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[50,86],"topic":[],"class_list":["post-88202","post","type-post","status-publish","format-standard","hentry","category-server","tag-50","tag-86"],"yoast_head":"<!-- This site is optimized with the Yoast SEO plugin v20.3 - https:\/\/yoast.com\/wordpress\/plugins\/seo\/ -->\n<title>\u6df1\u5ea6\u5b66\u4e60\u516b\u80a1\uff0c\uff08\u56fe\u50cf\uff09 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3<\/title>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/www.wsisp.com\/helps\/88202.html\" \/>\n<meta property=\"og:locale\" content=\"zh_CN\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"\u6df1\u5ea6\u5b66\u4e60\u516b\u80a1\uff0c\uff08\u56fe\u50cf\uff09 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"og:description\" content=\"\u4e00\u3001\u6df1\u5ea6\u5b66\u4e60\u57fa\u7840\u6838\u5fc3\u6982\u5ff51. \u6df1\u5ea6\u5b66\u4e60 vs \u673a\u5668\u5b66\u4e60 vs \u4eba\u5de5\u667a\u80fd\u7684\u5173\u7cfb\u4eba\u5de5\u667a\u80fd&#xff08;AI&#xff09;&#xff1a;\u6700\u5927\u8303\u7574&#xff0c;\u76ee\u6807\u662f\u8ba9\u673a\u5668\u5177\u5907\u4eba\u7c7b\u667a\u80fd\u80fd\u529b&#xff0c;\u6db5\u76d6\u673a\u5668\u5b66\u4e60\u3001\u4e13\u5bb6\u7cfb\u7edf\u3001\u77e5\u8bc6\u56fe\u8c31\u7b49\u3002\u673a\u5668\u5b66\u4e60&#xff08;ML&#xff09;&#xff1a;AI \u7684\u5b50\u96c6&#xff0c;\u901a\u8fc7\u6570\u636e\u548c\u7b97\u6cd5\u8ba9\u673a\u5668\u81ea\u52a8\u5b66\u4e60\u89c4\u5f8b&#xff0c;\u65e0\u9700\u4eba\u5de5\u786c\u7f16\u7801\u89c4\u5219&#xff0c;\u5305\u542b\u4f20\u7edf\u673a\u5668\u5b66\u4e60&#xff08;SVM\u3001\u51b3\u7b56\u6811\u3001LR \u7b49&amp;#xff09\" \/>\n<meta property=\"og:url\" content=\"https:\/\/www.wsisp.com\/helps\/88202.html\" \/>\n<meta property=\"og:site_name\" content=\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"article:published_time\" content=\"2026-07-31T09:56:44+00:00\" \/>\n<meta name=\"author\" content=\"admin\" \/>\n<meta name=\"twitter:card\" content=\"summary_large_image\" \/>\n<meta name=\"twitter:label1\" content=\"\u4f5c\u8005\" \/>\n\t<meta name=\"twitter:data1\" content=\"admin\" \/>\n\t<meta name=\"twitter:label2\" content=\"\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4\" \/>\n\t<meta name=\"twitter:data2\" content=\"27 \u5206\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"WebPage\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/88202.html\",\"url\":\"https:\/\/www.wsisp.com\/helps\/88202.html\",\"name\":\"\u6df1\u5ea6\u5b66\u4e60\u516b\u80a1\uff0c\uff08\u56fe\u50cf\uff09 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"isPartOf\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\"},\"datePublished\":\"2026-07-31T09:56:44+00:00\",\"dateModified\":\"2026-07-31T09:56:44+00:00\",\"author\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\"},\"breadcrumb\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/88202.html#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/www.wsisp.com\/helps\/88202.html\"]}]},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/88202.html#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\/\/www.wsisp.com\/helps\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"\u6df1\u5ea6\u5b66\u4e60\u516b\u80a1\uff0c\uff08\u56fe\u50cf\uff09\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\",\"url\":\"https:\/\/www.wsisp.com\/helps\/\",\"name\":\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"description\":\"\u9999\u6e2f\u670d\u52a1\u5668_\u9999\u6e2f\u4e91\u670d\u52a1\u5668\u8d44\u8baf_\u670d\u52a1\u5668\u5e2e\u52a9\u6587\u6863_\u670d\u52a1\u5668\u6559\u7a0b\",\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\/\/www.wsisp.com\/helps\/?s={search_term_string}\"},\"query-input\":\"required name=search_term_string\"}],\"inLanguage\":\"zh-Hans\"},{\"@type\":\"Person\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\",\"name\":\"admin\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/image\/\",\"url\":\"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery\",\"contentUrl\":\"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery\",\"caption\":\"admin\"},\"sameAs\":[\"http:\/\/wp.wsisp.com\"],\"url\":\"https:\/\/www.wsisp.com\/helps\/author\/admin\"}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"\u6df1\u5ea6\u5b66\u4e60\u516b\u80a1\uff0c\uff08\u56fe\u50cf\uff09 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/www.wsisp.com\/helps\/88202.html","og_locale":"zh_CN","og_type":"article","og_title":"\u6df1\u5ea6\u5b66\u4e60\u516b\u80a1\uff0c\uff08\u56fe\u50cf\uff09 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","og_description":"\u4e00\u3001\u6df1\u5ea6\u5b66\u4e60\u57fa\u7840\u6838\u5fc3\u6982\u5ff51. \u6df1\u5ea6\u5b66\u4e60 vs \u673a\u5668\u5b66\u4e60 vs \u4eba\u5de5\u667a\u80fd\u7684\u5173\u7cfb\u4eba\u5de5\u667a\u80fd&#xff08;AI&#xff09;&#xff1a;\u6700\u5927\u8303\u7574&#xff0c;\u76ee\u6807\u662f\u8ba9\u673a\u5668\u5177\u5907\u4eba\u7c7b\u667a\u80fd\u80fd\u529b&#xff0c;\u6db5\u76d6\u673a\u5668\u5b66\u4e60\u3001\u4e13\u5bb6\u7cfb\u7edf\u3001\u77e5\u8bc6\u56fe\u8c31\u7b49\u3002\u673a\u5668\u5b66\u4e60&#xff08;ML&#xff09;&#xff1a;AI \u7684\u5b50\u96c6&#xff0c;\u901a\u8fc7\u6570\u636e\u548c\u7b97\u6cd5\u8ba9\u673a\u5668\u81ea\u52a8\u5b66\u4e60\u89c4\u5f8b&#xff0c;\u65e0\u9700\u4eba\u5de5\u786c\u7f16\u7801\u89c4\u5219&#xff0c;\u5305\u542b\u4f20\u7edf\u673a\u5668\u5b66\u4e60&#xff08;SVM\u3001\u51b3\u7b56\u6811\u3001LR \u7b49&amp;#xff09","og_url":"https:\/\/www.wsisp.com\/helps\/88202.html","og_site_name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","article_published_time":"2026-07-31T09:56:44+00:00","author":"admin","twitter_card":"summary_large_image","twitter_misc":{"\u4f5c\u8005":"admin","\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4":"27 \u5206"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/www.wsisp.com\/helps\/88202.html","url":"https:\/\/www.wsisp.com\/helps\/88202.html","name":"\u6df1\u5ea6\u5b66\u4e60\u516b\u80a1\uff0c\uff08\u56fe\u50cf\uff09 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","isPartOf":{"@id":"https:\/\/www.wsisp.com\/helps\/#website"},"datePublished":"2026-07-31T09:56:44+00:00","dateModified":"2026-07-31T09:56:44+00:00","author":{"@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41"},"breadcrumb":{"@id":"https:\/\/www.wsisp.com\/helps\/88202.html#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/www.wsisp.com\/helps\/88202.html"]}]},{"@type":"BreadcrumbList","@id":"https:\/\/www.wsisp.com\/helps\/88202.html#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/www.wsisp.com\/helps"},{"@type":"ListItem","position":2,"name":"\u6df1\u5ea6\u5b66\u4e60\u516b\u80a1\uff0c\uff08\u56fe\u50cf\uff09"}]},{"@type":"WebSite","@id":"https:\/\/www.wsisp.com\/helps\/#website","url":"https:\/\/www.wsisp.com\/helps\/","name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","description":"\u9999\u6e2f\u670d\u52a1\u5668_\u9999\u6e2f\u4e91\u670d\u52a1\u5668\u8d44\u8baf_\u670d\u52a1\u5668\u5e2e\u52a9\u6587\u6863_\u670d\u52a1\u5668\u6559\u7a0b","potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/www.wsisp.com\/helps\/?s={search_term_string}"},"query-input":"required name=search_term_string"}],"inLanguage":"zh-Hans"},{"@type":"Person","@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41","name":"admin","image":{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/image\/","url":"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery","contentUrl":"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery","caption":"admin"},"sameAs":["http:\/\/wp.wsisp.com"],"url":"https:\/\/www.wsisp.com\/helps\/author\/admin"}]}},"_links":{"self":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/88202","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/comments?post=88202"}],"version-history":[{"count":0,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/88202\/revisions"}],"wp:attachment":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/media?parent=88202"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/categories?post=88202"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/tags?post=88202"},{"taxonomy":"topic","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/topic?post=88202"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}