{"id":38,"date":"2026-09-28T22:58:30","date_gmt":"2026-09-28T14:58:30","guid":{"rendered":"http:\/\/luqimu.xyz\/?p=38"},"modified":"2026-10-01T14:49:51","modified_gmt":"2026-10-01T06:49:51","slug":"%e5%8d%b7%e7%a7%af%e7%a5%9e%e7%bb%8f%e7%bd%91%e7%bb%9c%e5%ad%a6%e4%b9%a0","status":"publish","type":"post","link":"https:\/\/luqimu.xyz\/index.php\/2026\/09\/28\/%e5%8d%b7%e7%a7%af%e7%a5%9e%e7%bb%8f%e7%bd%91%e7%bb%9c%e5%ad%a6%e4%b9%a0\/","title":{"rendered":"\u5377\u79ef\u795e\u7ecf\u7f51\u7edc\u5b66\u4e60"},"content":{"rendered":"\n

\u4e00\u3001CNN \u5230\u5e95\u89e3\u51b3\u4ec0\u4e48\u95ee\u9898<\/h2>\n\n\n\n

\u4f20\u7edf\u5168\u8fde\u63a5\u7f51\u7edc\u5904\u7406\u56fe\u50cf\u6709\u4e09\u4e2a\u81f4\u547d\u95ee\u9898\uff1a<\/p>\n\n\n\n

\u95ee\u9898<\/th>\u8bf4\u660e<\/th>CNN \u7684\u89e3\u6cd5<\/th><\/tr><\/thead>
\u53c2\u6570\u91cf\u7206\u70b8<\/td>\u4e00\u5f20 224\u00d7224\u00d73 \u7684\u56fe\uff0c\u63a5 1000 \u4e2a\u9690\u5c42\u795e\u7ecf\u5143\u5c31\u662f 1.5 \u4ebf\u53c2\u6570<\/td>\u5c40\u90e8\u8fde\u63a5<\/strong><\/td><\/tr>
\u4e22\u5931\u7a7a\u95f4\u7ed3\u6784<\/td>\u628a\u56fe\u50cf\u5c55\u5e73\u6210\u4e00\u7ef4\u5411\u91cf\uff0c\u90bb\u57df\u5173\u7cfb\u88ab\u7834\u574f<\/td>\u4fdd\u7559 2D \u7279\u5f81\u56fe<\/strong>\uff0c\u5377\u79ef\u6838\u5728\u7a7a\u95f4\u4e0a\u6ed1\u52a8<\/td><\/tr>
\u5e73\u79fb\u4e0d\u654f\u611f<\/td>\u732b\u5f80\u5de6\u632a 10 \u50cf\u7d20\uff0c\u7f51\u7edc\u5c31\u5f53\u65b0\u6837\u672c\u91cd\u5b66\u4e00\u904d<\/td>\u6743\u503c\u5171\u4eab + \u6c60\u5316<\/strong> \u2192 \u5e73\u79fb\u7b49\u53d8\u6027\/\u4e0d\u53d8\u6027<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n

\u4e09\u6761\u6838\u5fc3<\/strong>\uff1a<\/p>\n\n\n\n

    \n
  1. \u5c40\u90e8\u6027<\/strong>\uff1a\u76f8\u90bb\u50cf\u7d20\u9ad8\u5ea6\u76f8\u5173\uff0c\u8fdc\u5904\u7684\u53ef\u5148\u4e0d\u5173\u5fc3\u3002<\/li>\n\n\n\n
  2. \u5e73\u79fb\u7b49\u53d8\u6027<\/strong>\uff1a\u540c\u4e00\u7279\u5f81\u51fa\u73b0\u5728\u4efb\u4f55\u4f4d\u7f6e\uff0c\u90fd\u7528\u540c\u4e00\u4e2a\u5377\u79ef\u6838\u53bb\u68c0\u6d4b\u3002<\/li>\n\n\n\n
  3. \u5c42\u6b21\u6027<\/strong>\uff1a\u6d45\u5c42\u5b66\u8fb9\u7f18\/\u7eb9\u7406\uff0c\u4e2d\u5c42\u5b66\u90e8\u4ef6\uff0c\u6df1\u5c42\u5b66\u6574\u4f53\u8bed\u4e49<\/li>\n<\/ol>\n\n\n\n
    \"\"<\/div><\/figure>\n\n\n\n
    \n\n\n\n

    \u4e8c.\u6838\u5fc3\u67b6\u6784<\/h2>\n\n\n\n
    \"\"<\/div><\/figure>\n\n\n\n
    \"\"<\/div><\/figure>\n\n\n\n

    \u4e09\u3001\u6838\u5fc3\u7ec4\u4ef6<\/h2>\n\n\n\n

    3.1 \u5377\u79ef\u5c42<\/h3>\n\n\n\n
    \"\"<\/div><\/figure>\n\n\n\n

    \u56db\u4e2a\u5173\u952e\u8d85\u53c2\u6570<\/strong><\/p>\n\n\n\n

    \u53c2\u6570<\/th>\u7b26\u53f7<\/th>\u4f5c\u7528<\/th><\/tr><\/thead>
    \u5377\u79ef\u6838\u5927\u5c0f<\/td>K\uff08kernel size\uff09<\/td>\u611f\u53d7\u91ce\u7684\u8fb9\u957f\uff0c\u5e38\u7528 1\/3\/5\/7<\/td><\/tr>
    \u6b65\u957f<\/td>S\uff08stride\uff09<\/td>\u6ed1\u52a8\u8de8\u5ea6\uff0cS>1 \u5373\u4e0b\u91c7\u6837<\/td><\/tr>
    \u586b\u5145<\/td>P\uff08padding\uff09<\/td>\u8fb9\u7f18\u8865 0\uff0c\u63a7\u5236\u8f93\u51fa\u5c3a\u5bf8\u3001\u4fdd\u62a4\u8fb9\u7f18\u4fe1\u606f<\/td><\/tr>
    \u8f93\u51fa\u901a\u9053\u6570<\/td>C_out<\/td>\u5373\u5377\u79ef\u6838\u4e2a\u6570\uff0c\u6bcf\u4e2a\u6838\u5b66\u4e00\u79cd\u6a21\u5f0f<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n

    \u8f93\u51fa\u5c3a\u5bf8\u516c\u5f0f<\/strong><\/p>\n\n\n\n

    H_out = floor( (H_in + 2P - K) \/ S ) + 1
    W_out = floor( (W_in + 2P - K) \/ S ) + 1<\/pre>\n\n\n\n
    \u7a7a\u6d1e\u5377\u79ef<\/strong><\/p>\n\n\n\n
    H_out = floor( (H_in + 2P - d*(K-1) - 1) \/ S ) + 1<\/pre>\n\n\n\n
    d \u4e3a\u81a8\u80c0\u7387\u3002\u4e0d\u589e\u52a0\u53c2\u6570\u5373\u53ef\u6269\u5927\u611f\u53d7\u91ce\uff0c\u4ee3\u4ef7\u662f\u91c7\u6837\u7a00\u758f\uff0c\u5e38\u7528\u4e8e\u8bed\u4e49\u5206\u5272\u548c\u5b9e\u65f6\u68c0\u6d4b\u3002<\/p>\n\n\n\n
    \u53c2\u6570\u91cf\u4e0e\u8ba1\u7b97\u91cf\uff08\u5355\u4e2a\u5377\u79ef\u5c42\uff09<\/strong><\/p>\n\n\n\n
    \u53c2\u6570\u91cf  Params = (K \u00d7 K \u00d7 C_in + 1) \u00d7 C_out
    \u8ba1\u7b97\u91cf  FLOPs \u2248 2 \u00d7 H_out \u00d7 W_out \u00d7 C_out \u00d7 K \u00d7 K \u00d7 C_in<\/pre>\n\n\n\n
    \u6ce8\u610f<\/strong>\uff1a\u5377\u79ef\u5c42\u7684\u53c2\u6570\u91cf\u4e0e\u8f93\u5165\u56fe\u50cf\u5c3a\u5bf8\u65e0\u5173<\/strong>\uff0c\u53ea\u4e0e K\u3001C_in\u3001C_out \u6709\u5173\u3002\u8fd9\u662f CNN \u76f8\u5bf9 FC \u7684\u6700\u5927\u4f18\u52bf\u3002<\/p>\n\n\n\n
    \u51e0\u79cd\u91cd\u8981\u7684\u5377\u79ef\u53d8\u4f53<\/strong><\/p>\n\n\n\n
    \u53d8\u4f53<\/th>\u6838\u5fc3\u601d\u60f3<\/th>\u5178\u578b\u7528\u9014<\/th><\/tr><\/thead>
    1\u00d71 \u5377\u79ef<\/strong><\/td>\u53ea\u5728\u901a\u9053\u7ef4\u505a\u7ebf\u6027\u7ec4\u5408\uff0c\u8de8\u901a\u9053\u4fe1\u606f\u878d\u5408<\/td>\u5347\u964d\u7ef4\u3001\u74f6\u9888\u7ed3\u6784\u3001\u52a0\u975e\u7ebf\u6027<\/td><\/tr>
    \u5206\u7ec4\u5377\u79ef<\/strong><\/td>\u901a\u9053\u5206\u6210 g \u7ec4\uff0c\u7ec4\u5185\u72ec\u7acb\u5377\u79ef\uff0c\u53c2\u6570\u964d\u4e3a 1\/g<\/td>AlexNet\uff08\u53cc GPU\uff09\u3001ResNeXt<\/td><\/tr>
    \u6df1\u5ea6\u53ef\u5206\u79bb\u5377\u79ef<\/strong><\/td>Depthwise(K\u00d7K) + Pointwise(1\u00d71)\uff0c\u53c2\u6570\u91cf\u964d\u4e3a\u7ea6 1\/K\u00b2<\/td>MobileNet\u3001Xception<\/td><\/tr>
    \u8f6c\u7f6e\u5377\u79ef<\/strong><\/td>\u53cd\u5411\u7684\u5377\u79ef\uff08\u4e0a\u91c7\u6837\uff09\uff0c\u901a\u8fc7\u5b66\u4e60\u5f97\u5230\u800c\u975e\u63d2\u503c<\/td>\u5206\u5272\u3001\u8d85\u5206\u3001GAN \u751f\u6210\u5668<\/td><\/tr>
    \u53ef\u53d8\u5f62\u5377\u79ef<\/strong><\/td>\u91c7\u6837\u70b9\u4f4d\u7f6e\u53ef\u5b66\u4e60\u504f\u79fb<\/td>\u76ee\u6807\u68c0\u6d4b\uff08DCN\uff09<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n
    3.2 \u6fc0\u6d3b\u51fd\u6570<\/h3>\n\n\n\n
    \u51fd\u6570<\/th>\u516c\u5f0f<\/th>\u7279\u70b9<\/th><\/tr><\/thead>
    Sigmoid<\/td>1\/(1+e\u207b\u02e3)<\/td>\u6613\u68af\u5ea6\u6d88\u5931\u3001\u8f93\u51fa\u975e\u96f6\u5747\u503c\uff0c\u51e0\u4e4e\u4e0d\u7528<\/td><\/tr>
    Tanh<\/td>(e\u02e3-e\u207b\u02e3)\/(e\u02e3+e\u207b\u02e3)<\/td>\u96f6\u5747\u503c\uff0c\u4ecd\u4f1a\u68af\u5ea6\u6d88\u5931<\/td><\/tr>
    ReLU<\/strong><\/td>max(0, x)<\/td>\u8ba1\u7b97\u6781\u5feb\u3001\u7f13\u89e3\u68af\u5ea6\u6d88\u5931\uff1b\u6709”\u6b7b\u4ea1 ReLU”\u95ee\u9898<\/td><\/tr>
    LeakyReLU<\/td>max(\u03b1x, x)<\/td>\u03b1 \u5e38\u53d6 0.01\uff0c\u4fee\u590d\u6b7b\u4ea1\u795e\u7ecf\u5143<\/td><\/tr>
    GELU<\/strong><\/td>x\u00b7\u03a6(x)<\/td>\u5e73\u6ed1\uff0cTransformer\/ConvNeXt \u4e3b\u6d41<\/td><\/tr>
    SiLU\/Swish<\/td>x\u00b7sigmoid(x)<\/td>MobileNetV3\u3001EfficientNet \u5e38\u7528<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n
    3.3 \u6c60\u5316\u5c42<\/h3>\n\n\n\n
    \"\"<\/div><\/figure>\n\n\n\n
    \u7c7b\u578b<\/th>\u505a\u6cd5<\/th>\u7279\u70b9<\/th><\/tr><\/thead>
    \u6700\u5927\u6c60\u5316<\/strong><\/td>\u53d6\u7a97\u53e3\u5185\u6700\u5927\u503c<\/td>\u4fdd\u7559\u6700\u5f3a\u54cd\u5e94\uff0c\u6700\u5e38\u7528<\/td><\/tr>
    \u5e73\u5747\u6c60\u5316<\/td>\u53d6\u7a97\u53e3\u5185\u5747\u503c<\/td>\u66f4\u5e73\u6ed1\uff0c\u65e9\u671f\u7f51\u7edc\u7528<\/td><\/tr>
    \u5168\u5c40\u5e73\u5747\u6c60\u5316 GAP<\/strong><\/td>\u5bf9\u6bcf\u4e2a\u901a\u9053\u6574\u56fe\u53d6\u5747\u503c<\/td>\u76f4\u63a5\u66ff\u4ee3 FC \u5c42\uff0c\u53c2\u6570\u91cf\u4e3a 0\uff0c\u6297\u8fc7\u62df\u5408<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n
    \u6c60\u5316\u65e0\u53ef\u5b66\u4e60\u53c2\u6570<\/strong>\uff0c\u4f5c\u7528\u662f\u964d\u4f4e\u5206\u8fa8\u7387\u3001\u6269\u5927\u611f\u53d7\u91ce\u3001\u63d0\u4f9b\u4e00\u5b9a\u5e73\u79fb\u4e0d\u53d8\u6027\u3002<\/p>\n\n\n\n
    \"\"<\/div><\/figure>\n\n\n\n
    3.4 \u5f52\u4e00\u5316\u5c42<\/h3>\n\n\n\n
    \u7c7b\u578b<\/th>\u5f52\u4e00\u5316\u7ef4\u5ea6<\/th>\u9002\u7528\u573a\u666f<\/th><\/tr><\/thead>
    BN<\/strong>\uff08Batch Norm\uff09<\/td>\u5bf9 (N, H, W) \u6c42\u7edf\u8ba1\u91cf<\/td>CNN \u9ed8\u8ba4\u9009\u62e9\uff1b\u5c0f batch \u65f6\u6548\u679c\u5dee<\/td><\/tr>
    LN\uff08Layer Norm\uff09<\/td>\u5bf9 (C, H, W) \u6c42\u7edf\u8ba1\u91cf<\/td>Transformer\u3001RNN\u3001\u5c0f batch<\/td><\/tr>
    IN\uff08Instance Norm\uff09<\/td>\u5bf9 (H, W) \u6c42\u7edf\u8ba1\u91cf\uff0c\u6bcf\u6837\u672c\u6bcf\u901a\u9053\u72ec\u7acb<\/td>\u98ce\u683c\u8fc1\u79fb<\/td><\/tr>
    GN\uff08Group Norm\uff09<\/td>\u901a\u9053\u5206\u7ec4\u540e\u5f52\u4e00\u5316<\/td>\u68c0\u6d4b\/\u5206\u5272\u5c0f batch \u573a\u666f<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n
    BN \u7684\u4f5c\u7528<\/strong>\uff1a\u7f13\u89e3\u5185\u90e8\u534f\u53d8\u91cf\u504f\u79fb\u3001\u5141\u8bb8\u66f4\u5927\u5b66\u4e60\u7387\u3001\u52a0\u901f\u6536\u655b\u3001\u6709\u8f7b\u5fae\u6b63\u5219\u6548\u679c\u3002<\/p>\n\n\n\n
    3.5 Dropout \u4e0e\u6b63\u5219\u5316<\/h3>\n\n\n\n
      \n
    • Dropout<\/strong>\uff1a\u8bad\u7ec3\u65f6\u968f\u673a\u7f6e\u96f6\u90e8\u5206\u795e\u7ecf\u5143\uff08p \u5e38 0.5\uff0cCNN \u4e2d\u7528 0.1~0.3\uff09\uff0c\u63a8\u7406\u65f6\u5168\u90e8\u4f7f\u7528\u5e76\u6309\u6bd4\u4f8b\u7f29\u653e\u3002\u7b49\u4ef7\u4e8e\u96c6\u6210\u591a\u4e2a\u5b50\u7f51\u7edc\u3002\u73b0\u591a\u7528\u5728 FC \u5c42\u3002<\/li>\n\n\n\n
    • DropBlock<\/strong>\uff1aCNN \u4e13\u7528\uff0c\u968f\u673a\u4e22\u5f03\u8fde\u7eed\u533a\u57df\u800c\u975e\u5355\u70b9\uff0c\u56e0\u4e3a\u76f8\u90bb\u50cf\u7d20\u4fe1\u606f\u5197\u4f59\u3002<\/li>\n\n\n\n
    • \u6743\u91cd\u8870\u51cf L2<\/strong>\u3001\u6807\u7b7e\u5e73\u6ed1<\/strong>\u3001\u6570\u636e\u589e\u5f3a<\/strong>\u4e5f\u662f\u6807\u914d\u3002<\/li>\n<\/ul>\n\n\n\n
      \n\n\n\n
      \u56db\u3001\u611f\u53d7\u91ce<\/h2>\n\n\n\n
      \u5b9a\u4e49<\/strong>\uff1a\u8f93\u51fa\u7279\u5f81\u56fe\u4e0a\u4e00\u4e2a\u50cf\u7d20\uff0c\u5bf9\u5e94\u5230\u539f\u59cb\u8f93\u5165\u56fe\u50cf\u4e0a\u7684\u533a\u57df\u5927\u5c0f\u3002<\/p>\n\n\n\n
      \u9012\u63a8\u516c\u5f0f<\/strong>\uff1a<\/p>\n\n\n\n
      RF_l = RF_{l-1} + (K_l - 1) \u00d7 \u220f_{i=1}^{l-1} S_i<\/pre>\n\n\n\n
        \n
      • \u521d\u59cb RF_0 = 1<\/code><\/li>\n\n\n\n
      • S_i<\/code> \u662f\u7b2c i \u5c42\u7684\u6b65\u957f\uff0c\u220f S_i<\/code> \u662f\u5230\u7b2c l-1 \u5c42\u7684\u7d2f\u8ba1\u4e0b\u91c7\u6837\u500d\u6570\uff08stride product\uff09<\/li>\n<\/ul>\n\n\n\n
        \u5982\u679c\u4f60\u77e5\u9053\u5f53\u524d\u5c42\u7684\u611f\u53d7\u91ce\uff0c\u60f3\u63a8\u524d\u4e00\u5c42\u7684\u611f\u53d7\u91ce\uff1a<\/p>\n\n\n\n
        RF(\u524d) = (RF(\u540e) – 1) \u00d7 S(\u524d) + K(\u524d)<\/p>\n\n\n\n
        \u5173\u952e\u7ed3\u8bba<\/strong><\/p>\n\n\n\n
          \n
        • \u5806\u53e0\u5c0f\u5377\u79ef\u6838 > \u7528\u4e00\u4e2a\u5927\u5377\u79ef\u6838\uff1a\u4e24\u4e2a 3\u00d73 \u7684\u611f\u53d7\u91ce\u7b49\u4e8e\u4e00\u4e2a 5\u00d75\uff0c\u4f46\u53c2\u6570\u4ece 25 \u964d\u5230 18\uff0c\u4e14\u591a\u4e86\u4e24\u6b21\u975e\u7ebf\u6027\u3002<\/li>\n\n\n\n
        • \u611f\u53d7\u91ce\u53ef\u4ee5\u6bd4\u8f93\u5165\u56fe\u50cf\u8fd8\u5927\uff08\u8bf4\u660e\u8986\u76d6\u4e86\u5168\u56fe\uff09\u3002<\/li>\n\n\n\n
        • \u6709\u4e9b\u50cf\u7d20\u7684\u6709\u6548\u611f\u53d7\u91ce<\/strong>\u8fdc\u5c0f\u4e8e\u7406\u8bba\u503c\uff0c\u5448\u9ad8\u65af\u5206\u5e03\uff0c\u4e2d\u5fc3\u6743\u91cd\u9ad8\u3002<\/li>\n<\/ul>\n\n\n\n
          \n\n\n\n
          \u4e94\u3001\u6b8b\u5dee\u8fde\u63a5<\/h2>\n\n\n\n
          ResNet \u6b8b\u5dee\u5757\u6838\u5fc3<\/strong><\/p>\n\n\n\n
          \u8f93\u51fa = F(x) + x       \uff08\u6052\u7b49\u6620\u5c04\u6377\u5f84\uff09<\/pre>\n\n\n\n
          \"\"<\/div><\/figure>\n\n\n\n
          \"\"<\/div><\/figure>\n\n\n\n
          \"\"<\/div><\/figure>\n\n\n\n
          \u4e3a\u4ec0\u4e48\u6709\u6548\uff1a\u53cd\u5411\u4f20\u64ad\u65f6\u68af\u5ea6\u53ef\u7ecf\u6377\u5f84\u65e0\u635f\u56de\u4f20\uff0c\u7f13\u89e3\u68af\u5ea6\u6d88\u5931\u5728\u6df1\u5c42\u7f51\u7edc\u7684\u7d2f\u79ef\u8870\u51cf\uff1b\u540c\u65f6\u628a\u5b66\u4e60\u76ee\u6807\u4ece”\u5b66\u5b8c\u6574\u6620\u5c04 H(x)”\u8f6c\u53d8\u4e3a”\u5b66\u6b8b\u5dee F(x)=H(x)-x”\uff0c\u4efb\u52a1\u66f4\u7b80\u5355\u3002<\/p>\n\n\n\n
          \"\"<\/div><\/figure>\n\n\n\n
          \"\"<\/div><\/figure>\n\n\n\n
          \u516d\u3001\u8bad\u7ec3\u4e0e\u8c03\u4f18\u8981\u70b9<\/h2>\n\n\n\n
          \u521d\u59cb\u5316<\/strong><\/p>\n\n\n\n
            \n
          • \u4e0d\u80fd\u5168\u96f6\uff08\u5bf9\u79f0\u6027\u65e0\u6cd5\u6253\u7834\uff09<\/li>\n\n\n\n
          • Xavier\/Glorot\uff1a\u9002\u5408 tanh\/sigmoid\uff0cVar = 2\/(fan_in + fan_out)<\/code><\/li>\n\n\n\n
          • He\/Kaiming<\/strong>\uff1a\u914d\u5408 ReLU\uff0cVar = 2\/fan_in<\/code>\uff0cCNN \u9ed8\u8ba4<\/li>\n<\/ul>\n\n\n\n
            \u4f18\u5316\u5668<\/strong><\/p>\n\n\n\n
              \n
            • SGD + Momentum\uff1a\u6cdb\u5316\u5e38\u66f4\u597d\uff0cCNN \u89c6\u89c9\u4efb\u52a1\u7ecf\u5178\u9009\u62e9<\/li>\n\n\n\n
            • Adam \/ AdamW\uff1a\u6536\u655b\u5feb\uff0c\u5c0f\u6570\u636e\u96c6\/Transformer \u53cb\u597d\uff1bAdamW \u4fee\u6b63\u4e86\u6743\u91cd\u8870\u51cf\u7684\u5b9e\u73b0<\/li>\n<\/ul>\n\n\n\n
              \u5b66\u4e60\u7387\u8c03\u5ea6<\/strong><\/p>\n\n\n\n
                \n
              • Warmup + Cosine Annealing \/ Step Decay \/ OneCycle<\/li>\n\n\n\n
              • \u5927 batch \u8981\u76f8\u5e94\u653e\u5927\u5b66\u4e60\u7387\uff08\u7ebf\u6027\u7f29\u653e\u89c4\u5219\uff09<\/li>\n<\/ul>\n\n\n\n
                \u6570\u636e\u589e\u5f3a<\/strong><\/p>\n\n\n\n
                  \n
                • \u57fa\u7840\uff1a\u968f\u673a\u88c1\u526a\u3001\u6c34\u5e73\u7ffb\u8f6c\u3001\u8272\u5f69\u6296\u52a8<\/li>\n\n\n\n
                • \u8fdb\u9636\uff1aMixup\u3001CutMix\u3001RandAugment\u3001AutoAugment<\/li>\n\n\n\n
                • \u6ce8\u610f\uff1a\u6c34\u5e73\u7ffb\u8f6c\u5bf9\u6570\u5b57\/\u6587\u5b57\u8bc6\u522b\u7b49\u6709\u8bed\u4e49\u7684\u573a\u666f\u53ef\u80fd\u6709\u5bb3<\/li>\n<\/ul>\n\n\n\n
                  \u8fc1\u79fb\u5b66\u4e60<\/strong><\/p>\n\n\n\n
                    \n
                  • \u51bb\u7ed3\u4e3b\u5e72\u53ea\u8bad\u5206\u7c7b\u5934\uff08\u5c0f\u6570\u636e\uff09\u2192 \u9010\u6b65\u89e3\u51bb\u5fae\u8c03 \u2192 \u5168\u7f51\u7edc\u5fae\u8c03 + \u5c0f\u5b66\u4e60\u7387<\/li>\n<\/ul>\n\n\n\n
                    \n\n\n\n
                    \u4e03\u3001PyTorch \u6700\u5c0f\u5b9e\u73b0\u6a21\u677f<\/h2>\n\n\n\n
                    import torch
                    import torch.nn as nn
                    \u200b
                    class SmallCNN(nn.Module):
                        def __init__(self, num_classes=10):
                            super().__init__()
                            self.features = nn.Sequential(
                                nn.Conv2d(3, 32, kernel_size=3, padding=1, bias=False),
                                nn.BatchNorm2d(32),
                                nn.ReLU(inplace=True),
                                nn.MaxPool2d(2),                    # 224 -> 112
                    \u200b
                                nn.Conv2d(32, 64, kernel_size=3, padding=1, bias=False),
                                nn.BatchNorm2d(64),
                                nn.ReLU(inplace=True),
                                nn.MaxPool2d(2),                    # 112 -> 56
                    \u200b
                                nn.Conv2d(64, 128, kernel_size=3, padding=1, bias=False),
                                nn.BatchNorm2d(128),
                                nn.ReLU(inplace=True),
                            )
                            self.pool = nn.AdaptiveAvgPool2d(1)     # \u5168\u5c40\u5e73\u5747\u6c60\u5316
                            self.classifier = nn.Sequential(
                                nn.Flatten(),
                                nn.Dropout(0.2),
                                nn.Linear(128, num_classes),
                            )
                    \u200b
                        def forward(self, x):
                            x = self.features(x)
                            x = self.pool(x)
                            return self.classifier(x)
                    \u200b
                    if __name__ == \"__main__\":
                        model = SmallCNN()
                        x = torch.randn(8, 3, 224, 224)
                        print(model(x).shape)                       # torch.Size([8, 10])
                    \u200b
                        total = sum(p.numel() for p in model.parameters())
                        print(f\"\u53c2\u6570\u91cf: {total\/1e6:.3f}M\")<\/pre>\n","protected":false},"excerpt":{"rendered":"
                    \u4e00\u3001CNN \u5230\u5e95\u89e3\u51b3\u4ec0\u4e48\u95ee\u9898 \u4f20\u7edf\u5168\u8fde\u63a5\u7f51\u7edc\u5904\u7406\u56fe\u50cf\u6709\u4e09\u4e2a\u81f4\u547d\u95ee\u9898\uff1a \u95ee\u9898 \u8bf4\u660e CNN \u7684\u89e3\u6cd5 \u53c2\u6570\u91cf\u7206\u70b8  […]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[6],"tags":[],"class_list":["post-38","post","type-post","status-publish","format-standard","hentry","category-6"],"_links":{"self":[{"href":"https:\/\/luqimu.xyz\/index.php\/wp-json\/wp\/v2\/posts\/38","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/luqimu.xyz\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/luqimu.xyz\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/luqimu.xyz\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/luqimu.xyz\/index.php\/wp-json\/wp\/v2\/comments?post=38"}],"version-history":[{"count":2,"href":"https:\/\/luqimu.xyz\/index.php\/wp-json\/wp\/v2\/posts\/38\/revisions"}],"predecessor-version":[{"id":104,"href":"https:\/\/luqimu.xyz\/index.php\/wp-json\/wp\/v2\/posts\/38\/revisions\/104"}],"wp:attachment":[{"href":"https:\/\/luqimu.xyz\/index.php\/wp-json\/wp\/v2\/media?parent=38"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/luqimu.xyz\/index.php\/wp-json\/wp\/v2\/categories?post=38"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/luqimu.xyz\/index.php\/wp-json\/wp\/v2\/tags?post=38"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}