{"id":286,"date":"2024-10-23T17:52:41","date_gmt":"2024-10-23T09:52:41","guid":{"rendered":"https:\/\/eve2333.top\/?p=286"},"modified":"2024-10-23T18:10:43","modified_gmt":"2024-10-23T10:10:43","slug":"vgg%e5%8e%9f%e7%90%86%e4%b8%8e%e5%ae%9e%e6%88%98","status":"publish","type":"post","link":"https:\/\/eve2333.top\/?p=286","title":{"rendered":"VGG\u539f\u7406\u4e0e\u5b9e\u6218"},"content":{"rendered":"
\n
\"\"<\/figure>\n<\/div>\n\n\n

VGGNet\u662f\u725b\u6d25\u5927\u5b66\u8ba1\u7b97\u673a\u89c6\u89c9\u7ec4(Visual Geometry Group\uff09\u548c\u8c37\u6b4c DeepMind \u4e00\u8d77\u7814\u7a76\u51fa\u6765\u7684\u6df1\u5ea6\u5377\u79ef\u795e\u7ecf\u7f51\u7edc\uff0c\u56e0\u800c\u51a0\u540d\u4e3a VGG\u3002VGG\u662f\u4e00\u79cd\u88ab\u5e7f\u6cdb\u4f7f\u7528\u7684\u5377\u79ef|\u795e\u7ecf\u7f51\u7edc\u7ed3\u6784\uff0c\u5176\u5728\u57282014\u5e74\u7684 ImageNet \u5927\u89c4\u6a21\u89c6\u89c9\u8bc6\u522b\u6311\u6218\uff08ILSVRC \u20142014\uff09\u4e2d\u83b7\u5f97\u4e86\u4e9a\u519b\uff0c\u4e0d\u662fVGG\u4e0d\u591f\u5f3a\uff0c\u800c\u662f\u5bf9\u624b\u592a\u5f3a\uff0c\u56e0\u4e3a\u5f53\u5e74\u83b7\u5f97\u51a0\u519b\u7684\u662fGoogLeNet\u3002
\u901a\u5e38\u4eba\u4eec\u8bf4\u7684VGG\u662f\u6307VGG\u201416\uff0813\u5c42\u5377\u79ef\u5c42+3\u5c42\u5168\u8fde\u63a5\u5c42\uff09\u3002\u867d\u7136\u5176\u5c48\u5c45\u4e9a\u519b\uff0c\u4f46\u662f\u7531\u4e8e\u5176\u89c4\u5f8b\u7684\u8bbe\u8ba1\u3001\u7b80\u6d01\u53ef\u5806\u53e0\u7684\u5377\u79ef\u5757\uff0c\u4e14\u5728\u5176\u4ed6\u6570\u636e\u96c6\u4e0a\u90fd\u6709\u7740\u5f88\u597d\u7684\u8868\u73b0\u4ece\u800c\u88ab\u4eba\u4eec\u5e7f\u6cdb\u4f7f\u7528\uff0c\u4ece\u8fd9\u70b9\u4e0a\u8fd8\u662f\u8d85\u8fc7\u4e86GoogLeNet\u3002VGG\u548c\u4e4b\u524d\u7684AlexNet\u76f8\u6bd4\uff0c\u6df1\u5ea6\u66f4\u6df1\uff0c\u53c2\u6570\u66f4\u591a\uff081.38\u4ebf\uff09\uff0c\u6548\u679c\u548c\u53ef\u79fb\u690d\u6027\u66f4\u597d\u3002<\/p>\n\n\n\n

VGG\u7f51\u7edc\u7ed3\u6784<\/h2>\n\n\n\n
\"\"<\/figure>\n\n\n\n

\u5377\u79ef\u5c31\u662f\u6211\u4eec\u7684\u4e00\u4e2a\u7279\u5f81\u56fe\u554a\u5f80\u5f80\u90fd\u4f1a\u7f29\u5c0f \uff0c\u7136\u540e\u7684\u8bdd\u4f46\u5b83\u901a\u9053\u4e0d\u4f1a\u53d8.\u5377\u79ef\u4e00\u822c\u662f\u4f7f\u7528\u6211\u4eec\u7684\u901a\u9053C\u53d8\u5927,\u78c1\u5316\u4f46\u662f\u5b83\u7684\u901a\u9053\u5c31\u662f\u6211\u4eec\u90a3\u4e2aH\u548cW\u4e00\u822c\u90fd\u4f1a\u53d8\u5c0f.\u4e0b\u91c7\u6837\u7684\u610f\u601d\u5c31\u662f\u4f7f\u5206\u8fa8\u7387\u53d8\u5c0f<\/p>\n\n\n\n

    \n
  1. vgg\u2014block\u5185\u7684\u5377\u79ef\u5c42\u90fd\u662f\u540c\u7ed3\u6784\u7684\u610f\u5473\u7740\u8f93\u5165\u548c\u8f93\u51fa\u7684\u5c3a\u5bf8\u4e00\u6837\uff0c\u4e14\u5377\u79ef\u5c42\u53ef\u4ee5\u5806\u53e0\u590d\u7528\uff0c\u5176\u4e2d\u7684\u5b9e\u73b0\u662f\u901a\u8fc7\u7edf\u4e00\u7684size\u4e3a3\u00d73\u7684kernel size + stride1 + padding1\u5b9e\u73b0\u3002<\/li>\n\n\n\n
  2. maxpool\u5c42\u5c06\u524d\u4e00\u5c42\uff08vgg\u2014block\u5c42\uff09\u7684\u7279\u5f81\u7f29\u51cf\u4e00\u534a \u4f7f\u5f97\u5c3a\u5bf8\u7f29\u51cf\u7684\u5f88\u89c4\u6574\uff0c\u4ece224\u2014112\u201456\u201428\u201414\u20147\u3002\u5176\u4e2d\u662f\u901a\u8fc7pool size2 + stride2\u5b9e\u73b0\u3002<\/li>\n\n\n\n
  3. \u6df1\u5ea6\u8f83\u6df1\uff0c\u53c2\u6570\u91cf\u591f\u5927\u00b7\u8f83\u6df1\u7684\u7f51\u7edc\u5c42\u6570\u4f7f\u5f97\u8bad\u7ec3\u5f97\u5230\u7684\u6a21\u578b\u5206\u7c7b\u6548\u679c\u4f18\u79c0\uff0c\u4f46\u662f\u8f83\u5927\u7684\u53c2\u6570\u5bf9\u8bad\u7ec3\u548c\u6a21\u578b\u4fdd\u5b58\u63d0\u51fa\u4e86\u66f4\u5927\u7684\u8d44\u6e90\u8981\u6c42\u3002(\u56e0\u4e3a\u5230\u540e\u9762\u7684RESNET\u7684\u65f6\u5019,\u4f60\u4f1a\u53d1\u73b0\u5b83\u89e3\u51b3\u5c31\u662f\u6211\u4eec\u6a21\u578b\u8f83\u6df1,\u4f7f\u6211\u4eec\u4f7f\u6211\u4eec\u7684\u6548\u679c\u4e0d\u597d\u7684\u95ee\u9898)<\/li>\n\n\n\n
  4. \u8f83\u5c0f\u7684filter size\/kernel size **\u8fd9\u91cc\u5168\u5c40\u7684kernel size\u90fd \u4e3a3\u00d73\uff0c\u76f8\u6bd4\u4ee5\u524d\u7684\u7f51\u7edc\u6a21\u578b\u6765\u8bf4\uff0c\u5c3a\u5bf8\u8db3\u591f\u5c0f\u3002<\/li>\n\n\n\n
  5. \u4f60\u4f1a\u53d1\u73b0\u6211\u4eec\u7684\u9009\u96c6\u90fd\u662f3\u00d73,3\u00d73\u6709\u4ec0\u4e48\u597d\u5904\u5462,\u53c2\u6570\u5c11\u5bf9\u5427\u9632\u6b62\u8fc7\u6ee4<\/li>\n<\/ol>\n\n\n\n

     \u53c2\u6570\u8be6\u89e3<\/h3>\n\n\n\n

    \u7b2c1\u4e2avgg block\u5c42\uff1a
    <\/strong>\uff081\uff09\u8f93\u5165\u4e3a224\u00d7224\u00d73\uff0c\u5377\u79ef\u6838\u6570\u91cf\u4e3a64\u4e2a\uff1b\u5377\u79ef\u6838\u7684\u5c3a\u5bf8\u5927\u5c0f\u4e3a3\u00d73\u00d73\uff1b\u6b65\u5e45\u4e3a1\uff08stride=1\uff09\uff0c\u586b\u5145\u4e3a1(padding=1);\u5377\u79ef\u540e\u5f97\u5230shape\u4e3a 224\u00d7224\u00d764\u7684\u7279\u5f81\u56fe\u8f93\u51fa\u3002
    \uff082\uff09\u8f93\u5165\u4e3a224\u00d7224\u00d764\uff0c\u5377\u79ef\u6838\u6570\u91cf\u4e3a64\u4e2a\uff1b\u5377\u79ef\u6838\u7684\u5c3a\u5bf8\u5927\u5c0f\u4e3a3\u00d73\u00d764\u6b65\u5e45\u4e3a1\uff08stride=1\uff09\u586b\u5145\u4e3a1(padding=1);\u5377\u79ef\u540e\u5f97\u5230shape\u4e3a 224\u00d7224\u00d764\u7684\u7279\u5f81\u56fe\u8f93\u51fa\u3002
    \uff083\uff09\u8f93\u5165\u4e3a224\u00d7224\u00d764\uff0c\u6c60\u5316\u6838\u4e3a2\u00d72\uff0c\u6b65\u5e45\u4e3a2\uff08stride=2\uff09\u540e\u5f97\u5230\u5c3a\u5bf8\u4e3a112\u00d7112\u00d764\u7684\u6c60\u5316\u5c42\u7684\u7279\u5f81\u56fe\u8f93\u51fa\u3002<\/p>\n\n\n\n

    \u7b2c2\u4e2avgg block\u5c42<\/strong>\uff1a
    (1)\u8f93\u5165\u4e3a112x112x64,\u5377\u79ef\u6838\u6570\u91cf\u4e3a128\u4e2a;\u5377\u79ef\u6838\u7684\u5c3a\u5bf8\u5927\u5c0f\u4e3a3x3x64\uff1b\u6b65\u5e45\u4e3a1 (stride =1),\u586b\u5145\u4e3a1(padding\u4e8c1)\uff1b\u5377\u79ef\u540e\u5f97\u5230shape\u4e3a112x112x128\u7684\u7279\u5f81\u56fe\u8f93\u51fa\u3002
    (2)\u8f93\u5165\u4e3a112x112x128,\u5377\u79ef\u6838\u6570\u91cf\u4e3a128\u4e2a;\u5377\u79ef\u6838\u7684\u5c3a\u5bf8\u5927\u5c0f\u4e3a3x3x128\uff1b\u6b65\u5e45\u4e3a1 (stride =1),\u586b\u5145\u4e3a1(padding=l)\uff1b\u5377\u79ef\u540e\u5f97\u5230shape\u4e3a112x112x128\u7684\u7279\u5f81\u56fe\u8f93\u51fa\u3002
    (3)\u8f93\u5165\u4e3a112x112x128,\u6c60\u5316\u6838\u4e3a2x2,\u6b65\u5e45\u4e3a2 (stride = 2)\u540e\u5f97\u5230\u5c3a\u5bf8\u4e3a56x5\u540e128\u7684\u6c60\u5316\u5c42\u7684\u7279\u5f81\u56fe\u8f93\u51fa<\/p>\n\n\n\n

    \u7b2c3\u4e2avgg block\u5c42\uff1a<\/strong>
    (1)\u8f93\u5165\u4e3a56x56x128,\u5377\u79ef\u6838\u6570\u91cf\u4e3a256\u4e2a\uff1b\u5377\u79ef\u6838\u7684\u5c3a\u5bf8\u5927\u5c0f\u4e3a3x3x128\uff1b\u6b65\u5e45\u4e3a1 (stride=1).\u586b\u5145\u4e3a 1 (padding=l)\uff1b\u5377\u79ef\u540e\u5f97\u5230shape\u4e3a56x56x256\u7684\u7279\u5f81\u56fe\u8f93\u51fa\u3002
    (2)\u8f93\u5165\u4e3a56x56x256,\u5377\u79ef\u6838\u6570\u91cf\u4e3a256\u4e2a\uff1b\u5377\u79ef\u6838\u7684\u5c3a\u5bf8\u5927\u5c0f\u4e3a3x3x256\uff1b\u6b65\u5e45\u4e3a1 (stride=1),\u586b\u5145\u4e3a 1 (padding=l)\uff1b\u5377\u79ef\u540e\u5f97\u5230shape\u4e3a56x56x256\u7684\u7279\u5f81\u56fe\u8f93\u51fa\u3002
    (3)\u8f93\u5165\u4e3a56x56x256,\u5377\u79ef\u6838\u6570\u91cf\u4e3a256\u4e2a\uff1b\u5377\u79ef\u6838\u7684\u5c3a\u5bf8\u5927\u5c0f\u4e3a3x3x256\uff1b\u6b65\u5e45\u4e3a1 (stride = 1),\u586b\u5145\u4e3a 1 (padding=l)\uff1b\u5377\u79ef\u540e\u5f97\u5230shape\u4e3a56x56x256\u7684\u7279\u5f81\u56fe\u8f93\u51fa\u3002
    (4)\u8f93\u5165\u4e3a56x56x256,\u6c60\u5316\u6838\u4e3a2x2,\u6b65\u5e45\u4e3a2(stride=2)\u540e\u5f97\u5230\u5c3a\u5bf8\u4e3a28x28x256\u7684\u6c60\u5316\u5c42\u7684\u7279\u5f81\u56fe\u8f93\u51fa\u3002<\/p>\n\n\n\n

    \u7b2c4\u4e2avgg block\u5c42\uff1a<\/strong>
    (1)\u8f93\u5165\u4e3a28x28x256,\u5377\u79ef\u6838\u6570\u91cf\u4e3a512\u4e2a\uff1b\u5377\u79ef\u6838\u7684\u5c3a\u5bf8\u5927\u5c0f\u4e3a3x3x256\uff1b\u6b65\u5e45\u4e3a1 (stride = 1),\u586b\u5145\u4e3a1(padding=1)\uff1b\u5377\u79ef\u540e\u5f97\u5230shape\u4e3a28x28x512\u7684\u7279\u5f81\u56fe\u8f93\u51fa\u3002
    (2)\u8f93\u5165\u4e3a28x28x512,\u5377\u79ef\u6838\u6570\u91cf\u4e3a512\u4e2a\uff1b\u5377\u79ef\u6838\u7684\u5c3a\u5bf8\u5927\u5c0f\u4e3a3x3x512\uff1b\u6b65\u5e45\u4e3a1 (stride=1),\u586b\u5145\u4e3a1 (padding=l)\uff1b\u5377\u79ef\u540e\u5f97\u5230sh\u53e9e\u4e3a28x28x512\u7684\u7279\u5f81\u56fe\u8f93\u51fa\u3002
    (3)\u8f93\u5165\u4e3a28x28x512,\u5377\u79ef\u6838\u6570\u91cf\u4e3a512\u4e2a\uff1b\u5377\u79ef\u6838\u7684\u5c3a\u5bf8\u5927\u5c0f\u4e3a3x3x512\uff1b\u6b65\u5e45\u4e3a1 (stride = 1).\u586b\u5145\u4e3a1 (padding=l)\uff1b\u5377\u79ef\u540e\u5f97\u5230shape\u4e3a28x28x512\u7684\u7279\u5f81\u56fe\u8f93\u51fa\u3002
    ( 4 ) \u8f93\u5165\u4e3a28x28x512,\u6c60\u5316\u6838\u4e3a2x2,\u6b65\u5e45\u4e3a2(stride=2)\u540e\u5f97\u5230\u5c3a\u5bf8\u4e3a14x14x512\u7684\u6c60\u5316\u5c42\u7684\u7279\u5f81\u56fe\u8f93\u51fa<\/p>\n\n\n\n

    \u7b2c5\u4e2avgg block\u5c42\uff1a<\/strong>
    (1)\u8f93\u5165\u4e3a14x14x512,\u5377\u79ef\u6838\u6570\u91cf\u4e3a512\u4e2a\uff1b\u5377\u79ef\u6838\u7684\u5c3a\u5bf8\u5927\u5c0f\u4e3a3x3x512\uff1b\u6b65\u5e45\u4e3a 1 (stride = 1),\u586b\u5145\u4e3a1(padding=l)\uff1b\u5fe0\u79ef\u540e\u5f97\u5230shape\u4e3a14x14*512\u7684\u7279\u5f81\u56fe\u8f93\u51fa\u3002
    (2)\u8f93\u5165\u4e3a14x14x512,\u5377\u79ef\u6838\u6570\u91cf\u4e3a512\u4e2a\uff1b\u5377\u79ef\u6838\u7684\u5c3a\u5bf8\u5927\u5c0f\u4e3a3x3x512\uff1b\u6b65\u5e45\u4e3a 1 (stride \u4e8c 1),\u586b\u5145\u4e3a1(padding\u4e8c 1)\uff1b\u5957\u79ef\u540e\u5f97\u5230shape\u4e3a14x14x512\u7684\u7279\u5f81\u56fe\u8f93\u51fa\u3002
    (3)\u8f93\u5165\u4e3a14x14x512,\u5377\u79ef\u6838\u6570\u91cf\u4e3a512\u4e2a\uff1b\u5377\u79ef\u6838\u7684\u5c3a\u5bf8\u5927\u5c0f\u4e3a3x3x512\uff1b\u6b65\u5e45\u4e3a 1 (stride \u4e8c 1),\u586b\u5145\u4e3a 1(padding=l)\uff1b\u5377\u79ef\u540e\u5f97\u5230shape\u4e3a14x14x512\u7684\u7279\u5f81\u56fe\u8f93\u51fa\u3002
    (4)\u8f93\u5165\u4e3a14x14x512,\u6c60\u5316\u6838\u4e3a2x2,\u6b65\u5e45\u4e3a2 (stride\u4e8c2)\u540e\u5f97\u5230\u5c3a\u5bf8\u4e3a7x7x512\u7684\u6c60\u5316\u5c42\u7684\u7279\u5f81\u56fe\u8f93\u51fa\u3002\u8be5\u5c42\u540e\u9762\u8fd8\u9690\u85cf\u4e86flatten\u64cd\u4f5c\uff0c\u901a\u8fc7\u5c55\u5e73\u5f97\u52307x7x512=25088\u4e2a\u53c2\u6570\u540e\u4e0e\u4e4b\u540e\u7684\u5168\u8fde\u63a5\u5c42\u76f8\u8fde\u3002
    \u7b2c1~3\u5c42\u5168\u8fde\u63a5\u5c42\uff1a\u7b2c1\u301c3\u5c42\u795e\u7ecf\u5143\u4e2a\u6570\u5206\u522b\u4e3a4096, 4096,1000\uff0c\u5176\u4e2d\u524d\u4e24\u5c42\u5728\u4f7f\u7528relu\u540e\u8fd8\u4f7f\u7528\u4e86Dropout\u5bf9\u795e\u7ecf\u5143\u968f\u673a\u5931\u6d3b\uff0c\u6700\u540e\u4e00\u5c42\u5168\u8fde\u63a5\u5c42\u7528softmax\u8f93\u51fa1000\u4e2a\u5206\u7c7b\u3002<\/p>\n\n\n\n

    VGGNet\u901a\u8fc7\u5728\u4f20\u7edf\u5377\u79ef\u795e\u7ecf\u7f51\u7edc\u6a21\u578b\uff08AlexNet\uff09\u4e0a\u7684\u62d3\u5c55\uff0c\u53d1\u73b0\u9664\u4e86\u8f83\u4e3a\u590d\u6742\u7684\u6a21\u578b\u7ed3\u6784\u7684\u8bbe\u8ba1\uff08\u5982GoogLeNet\uff09\u5916\uff0c\u6df1\u5ea6\u5bf9\u4e8e\u63d0\u9ad8\u6a21\u578b\u51c6\u786e\u7387\u5f88\u91cd\u8981\uff0cVGG\u548c\u4e4b\u524d\u7684AlexNet\u76f8\u6bd4\uff0c\u6df1\u5ea6\u66f4\u6df1\uff0c\u53c2\u6570\u66f4\u591a\uff081.38\u4ebf\uff09\uff0c\u6548\u679c\u548c\u53ef\u79fb\u690d\u6027\u66f4\u597d\uff0c\u4e14\u6a21\u578b\u8bbe\u8ba1\u7684\u7b80\u6d01\u800c\u89c4\u5f8b\uff0c\u4ece\u800c\u88ab\u5e7f\u6cdb\u4f7f\u7528\u3002\u8fd8\u6709\u4e00\u4e9b\u7279\u70b9\u603b\u7ed3\u5982\u4e0b\uff1a
    1\u3001\u5c0f\u5c3a\u5bf8\u7684filter\uff083\u00d73\uff09\u4e0d\u4ec5\u4f7f\u53c2\u6570\u66f4\u5c11\uff0c\u6548\u679c\u4e5f\u5e76\u4e0d\u5f31\u4e8e\u5927\u5c3a\u5bf8filter\u59825\u00d75
    2\u3001\u5757\u7684\u4f7f\u7528\u5bfc\u81f4\u7f51\u7edc\u5b9a\u4e49\u7684\u975e\u5e38\u7b80\u6d01\u3002\u4f7f\u7528\u5757\u53ef\u4ee5\u6709\u6548\u5730\u8bbe\u8ba1\u590d\u6742\u7684\u7f51\u7edc\u3002
    3\u3001AlexNet\u4e2d\u7684\u5c40\u90e8\u54cd\u5e94\u5f52\u4e00\u5316\u4f5c\u7528\u4e0d\u5927<\/p>\n\n\n\n

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

    VGG\u4ee3\u7801\u5b9e\u6218<\/h2>\n\n\n\n

    model.py<\/p>\n\n\n\n

    import torch\nfrom torch import nn\nfrom torchsummary import summary\n\nclass VGG16(nn.Module()):\n    def __init__(self):\n        super(VGG16, self).__init__()\n        self.block1 = nn.Sequential(\n            nn.Conv2d(in_channels=1, out_channels=64, kernel_size=3, padding=1),\n            nn.ReLU(),\n            nn.Conv2d(in_channels=64, out_channels=64, kernel_size=3, padding=1),\n            nn.ReLU(),\n            nn.MaxPool2d(kernel_size=2, stride=2)\n        )\n        self.block2 = nn.Sequential(\n            nn.Conv2d(in_channels=64, out_channels=128, kernel_size=3, padding=1),\n            nn.ReLU(),\n            nn.Conv2d(in_channels=128, out_channels=128, kernel_size=3, padding=1),\n            nn.ReLU(),\n            nn.MaxPool2d(kernel_size=2, stride=2),\n        )\n        self.block3 = nn.Sequential(\n            nn.Conv2d(in_channels=128, out_channels=256, kernel_size=3, padding=1),\n            nn.ReLU(),\n            nn.Conv2d(in_channels=256, out_channels=256, kernel_size=3, padding=1),\n            nn.ReLU(),\n            nn.Conv2d(in_channels=256, out_channels=256, kernel_size=3, padding=1),\n            nn.ReLU(),\n            nn.MaxPool2d(kernel_size=2, stride=2)\n        )\n        self.block4 = nn.Sequential(\n            nn.Conv2d(in_channels=256, out_channels=512, kernel_size=3, padding=1),\n            nn.ReLU(),\n            nn.Conv2d(in_channels=512, out_channels=512, kernel_size=3, padding=1),\n            nn.ReLU(),\n            nn.Conv2d(in_channels=512, out_channels=512, kernel_size=3, padding=1),\n            nn.ReLU(),\n            nn.MaxPool2d(kernel_size=2, stride=2)\n        )\n        self.block5 = nn.Sequential(\n            nn.Conv2d(in_channels=512, out_channels=512, kernel_size=3, padding=1),\n            nn.ReLU(),\n            nn.Conv2d(in_channels=512, out_channels=512, kernel_size=3, padding=1),\n            nn.ReLU(),\n            nn.Conv2d(in_channels=512, out_channels=512, kernel_size=3, padding=1),\n            nn.ReLU(),\n            nn.MaxPool2d(kernel_size=2, stride=2)\n        )\n        self.block6 = nn.Sequential(\n            nn.FLatten(),\n            nn.Linear(7 * 7 * 512, 4096),\n            nn.ReLU(),\n            nn.Linear(4096, 4096),\n            nn.ReLU(),\n            nn.Linear(4096, 10),\n        )\n\n    def forward(self, x):\n        x = self.block1(x)\n        x = self.block2(x)\n        x = self.block3(x)\n        x = self.block4(x)\n        x = self.block5(x)\n        x = self.block6(x)\n        return x\n\n\nif __name__ == '__main__':\n    device = torch.device(\"cuda\" if torch.cuda.is_available() else \"cpu\")\n    model = VGG16().to(device)\n    print(summary(model, (1, 224, 224)))\n<\/code><\/pre>\n\n\n\n
    model_train.py\u8fd9\u6b21\u8bad\u7ec3\u7684\u7ed3\u679c\u4e0d\u662f\u4f1a\u5f88\u597d,\u66b4\u9732\u95ee\u9898,\u4f7f\u75281060\u8bad\u7ec3\u4e86332min\u53735.5\u4e2a\u5c0f\u65f6 \u200b<\/p>\n\n\n\n
    \u63a5\u4e0b\u6765\u6211\u4eec\u6765\u770b\u4e00\u4e0b\u6a21\u578b\u7684\u4e00\u4e2a\u8bad\u7ec3\u7684\u4e00\u4e2a\u7ed3\u679c\u554a\uff0c\u5443\u524d\u9762\u4e5f\u8bb2\u4e86\uff0c\u8fd9\u6b21\u7684\u8bad\u7ec3\u7684\u7ed3\u679c\u5462\uff0c\u5b83\u4e0d\u4f1a\u5f88\u597d\uff0c\u6211\u4eec\u662f\u7528\u6765\u66b4\u9732\u6211\u4eec\u7684\u95ee\u9898\u4e86\uff0c\u7136\u540e\u7684\u8bdd\u6839\u636e\u6211\u4eec\u7684\u95ee\u9898\u5e72\u561b\u554a\uff0c\u5c31\u662f\u6765\u89e3\u51b3\u5bf9\u5e94\u7684\u95ee\u9898\u561b\u5bf9\u5427\uff0c\u9996\u5148\u6211\u4eec\u6765\u770b\u4e00\u4e0b\u5c31\u662f\u6574\u4e2a\u8bad\u7ec3\u7684\u4e00\u4e2a\u8fc7\u7a0b\u554a\uff0c\u8bad\u7ec3\u8fc7\u7a0b\u662f 332 \u5206\u949f\uff0c\u4e5f\u5c31\u662f\u5c06\u8fd1\u5927\u6982\u662f\u4e94\u4e2a\u534a\u5c0f\u65f6\u5de6\u53f3\u561b\u5bf9\u5427\uff0c\u4e94\u4e2a\u534a\u5c0f\u65f6\u5de6\u53f3\u65f6\u95f4\u8fd8\u662f\u5f88\u957f\u7684\u554a\uff0c20 \u8f6e\u4e94\u4e2a\u534a\u5c0f\u65f6\u8fd8\u662f\u5f88\u957f\u7684\uff0c\u7136\u540e\u7684\u8bdd\u6211\u4eec\u8fd9\u4e2a\uff0c\u4f60\u5176\u5b9e\u6211\u4eec\u4ece\u8fd9\u4e2a\u901a\u8fc7\u4ec0\u4e48\u6253\u5370\u51fa\u6765\u4e00\u4e2a\u8bad\u7ec3\u65e5\u5fd7\u53ef\u4ee5\u53d1\u73b0\u554a\uff0c\u57fa\u672c\u4e0a\u5b83\u7684\u4e00\u4e2a\u4ec0\u4e48\u5443\u7cbe\u5ea6\u90fd\u662f\u5728 0.9 \u5de6\u53f3\uff0c\u4e5f\u5c31\u662f\u5728\u4ec0\u4e48\u554a\uff0c0.09 \u5de6\u53f3\u4e5f\u5c31 10%\u5de6\u53f3\u561b\u662f\u5427\uff0c\u7136\u540e\u5b83\u7684 loss \u503c\u554a\uff0c\u4e5f\u57fa\u672c\u4e0a\u5927\u6982\u5927\u6982\u5728\u8fd9\u4e2a\u503c\u5de6\u53f3\u5427\uff0c\u57fa\u672c\u4e0a\u662f\u4e0d\u53d8\u7684\u4e00\u4e2a\u8fc7\u7a0b\uff0c\u800c\u901a\u8fc7\u6211\u4eec\u7684\u56fe\u5462\u4e5f\u66f4\u52a0\u76f4\u89c2\uff0c\u4ed6\u8fd9\u4e2a\u7cbe\u786e\u5ea6\u5462\u57fa\u672c\u4e0a\u5c31\u5728\u8fd9\u4e2a\u9644\u8fd1\uff0c\u7136\u540e\u4ed6\u90a3\u4e2a loss \u503c\u554a\u4e5f\u4e5f\u4e5f\u4e0d\u6536\u655b\uff0c\u4e5f\u5c31\u662f\u8fd9\u4e24\u4e2a loss \u503c\u548c\u6211\u4eec\u7cbe\u786e\u5ea6\u90fd\u4e0d\u6536\u655b\uff0c\u989d\u8fd9\u91cc\u6211\u63d2\u63d2\u4e00\u4e2a\uff0c\u5c31\u662f\u9898\u5916\u8bdd\uff0c\u5f53\u65f6\u6211\u5728\u5199\u8fd9\u6bb5\u4ee3\u7801\u7684\u65f6\u5019\uff0c\u6211\u80af\u5b9a\u4e5f\u505a\u4e86\u4e00\u4e9b\u6d4b\u8bd5\u561b\uff0c\u54ce\u5f53\u65f6\u4f1a\u51fa\u73b0\u4ec0\u4e48\u60c5\u51b5\u5462\uff0c\u6211\u8bad\u7ec3\u5b8c\u4e4b\u540e\u5462\uff0c\u8bf6\u4ed6\u51fa\u73b0\u8fd9\u4e2a\u95ee\u9898\u5443\uff0c\u6211\u6211\u6211\u68c0\u67e5\u4e00\u4e0b\u4e86\u6a21\u578b\u554a\uff0c\u53d1\u73b0\u7684\u6a21\u578b\u597d\u50cf\u662f\u5427\uff0c\u6ca1\u6709\u4ec0\u4e48\u95ee\u9898\u554a\u662f\u5427\uff0c\u6211\u5c31\u5f88\u82e6\u607c\uff0c\u540e\u9762\u5462\u5c31\u662f\u5e72\u561b\uff0c\u540e\u9762\u8fc7\u4e86\u51e0\u5929\u4e4b\u540e\uff0c\u6211\u540c\u6837\u4ee3\u7801\u6211\u4e5f\u6ca1\u6709\u6539\uff0c\u6211\u4e5f\u5f88\u5fd9\u561b\uff0c\u6211\u6ca1\u6709\u6539\uff0c\u7136\u540e\u53c8\u8fd0\u884c\u540c\u6837\u7684\u4ee3\u7801\uff0c\u6211\u53d1\u73b0\u8bf6\u4ed6\u53c8\u8fd0\u884c OK \u4e86\uff0c\u7136\u540e\u6211\u5c31\u4e00\u76f4\u601d\u8003\uff0c\u6211\u5230\u5e95\u662f\u4e0d\u662f\u6709\u4eba\u52a8\u4e86\u6211\u4ee3\u7801\uff0c\u8fd8\u662f\u6211\u81ea\u5df1\u5c31\u662f\u8ff7\u8ff7\u7cca\uff0c\u4e0d\u77e5\u9053\uff0c\u52a8\u4e86\u4ee3\u7801\u4e4b\u540e\u5462\uff0c\u7136\u540e\u8fd0\u884c\u8fd0\u884c\u4e00\u4e0b\u4ee3\u7801\u4e4b\u540e\u5462\uff0c\u7136\u540e\u5443\u5b83\u5c31\u53d8\u597d\u4e86\uff0c\u5bf9\u4e0d\u5bf9\uff0c\u5c31\u6bd4\u5982\u6211\u4eec\u7684\u7cbe\u5ea6\u554a\uff0c\u4e00\u76f4\u4e00\u76f4\u4e0a\u5347\uff0c\u53ef\u80fd\u6700\u540e\u5728\u5927\u6982\u5728 90%\u591a\u7684\u51c6\u786e\u5ea6\uff0c\u8fd8\u662f\u86ee\u86ee\u4e0d\u9519\u7684\uff0c\u5bf9\u4e0d\u5bf9\uff0c\u7136\u540e\u7684\u8bdd\u90a3\u4ee3\u7801\u6211\u4e5f\u6ca1\u52a8\uff0c\u8fc7\u51e0\u5929\u6211\u53c8\u53c8\u53c8\u91cd\u65b0\u8bad\u4e86\u4e00\u4e0b\uff0c\u4e4b\u540e\u53d1\u73b0\u53c8\u53c8\u4e0d\u884c\u4e86\uff0c\u7136\u540e\u4e0d\u65ad\u7684\u91cd\u590d\u53c8\u8bad\u53c8\u91cd\u590d\u4e0d\u8bad\u5443\uff0c\u5c31\u91cd\u590d\u8bad\uff0c\u7136\u540e\u6bcf\u6b21\u662f\u5427\u55ef\u7ed3\u679c\u90fd\u4e0d\u4e00\u6837\u554a\uff0c\u5c31\u662f\u6bd4\u5982\u8bf4\u8bad\u4e86\u4e94\u6b21\u4e4b\u540e\uff0c\u54ce\u5440\u5b83\u7684\u7cbe\u5ea6\u90fd\u662f\u8fd9\u6837\u5b50\u7684\uff0c\u4e0d\u6536\u655b\u662f\u5427\u554a\uff0c\u8fc7\u4e86\u4e0b\u5348\u7684\u65f6\u5019\u5462\uff0c\u6216\u8005\u5173\u5173\u673a\u91cd\u542f\u4e4b\u540e\u5462\uff0c\u53ef\u80fd\u8fc7\u4e86\u4e00\u4f1a\u513f\u4ed6\u53c8\u597d\u4e86\uff0c\u54ce\u90a3\u65f6\u5019\u6211\u5c31\u4e0d\u77e5\u9053\u662f\u4ec0\u4e48\u6837\u7684\u539f\u56e0\uff0c\u5c31\u5f88\u7591\u60d1\uff0c\u6211\u5f53\u65f6\u6000\u7591\u662f\u4e0d\u662f\u73af\u5883\u7684\u95ee\u9898\uff0c\u95ee\u54ea\u4e2a\u95ee\u9898\u554a\uff0c\u7ec8\u4e8e\u5728\u540e\u9762\u554a\uff0c\u6211\u7ec8\u4e8e\u77e5\u9053\u95ee\u9898\u51fa\u5728\u54ea\u91cc\u554a\uff0c\u8fd9\u91cc\u6211\u7ed9\u5927\u5bb6\u8bb2\u4e00\u4e0b\u554a\uff0c\u9996\u5148\u554a\u7ed9\u5927\u5bb6\u8bb2\u4e2a\u8fd9\u4e2a\u77e5\u8bc6\u70b9\u554a\uff0c\u9996\u5148\u6211\u4eec\u6a21\u578b\u642d\u5efa\uff0c\u6211\u8fd9\u91cc\u7ed9\u5927\u5bb6\u8bb2\u4e00\u4e0b\uff0c\u80af\u5b9a\u662f\u6ca1\u6709\u95ee\u9898\u7684\u554a\uff0c\u8fd9\u91cc\u6211\u53ef\u4ee5\u7ed9\u5927\u5bb6\u4fdd\u8bc1\uff0c\u8fd9\u91cc\u6a21\u578b\u642d\u5efa\u80af\u5b9a\u662f\u6ca1\u6709\u95ee\u9898\u7684\uff0c\u7136\u540e\u7684\u8bdd\u554a\u8fd9\u91cc\u7ed9\u5927\u5bb6\u8bb2\u8fd9\u4e2a\u77e5\u8bc6\u70b9\u554a\uff0c\u9996\u5148\u6211\u4eec\u8fd9\u4e2a VGG \u5bf9\u4e0d\u5bf9\uff0c\u5371\u673a\u662f\u5341\u516d\u5341\u516d\u4e2a\u5c42\uff0c\u5bf9\u4e0d\u5bf9\uff0c\u7136\u540e\u7684\u8bdd\u56e0\u4e3a\u6a21\u578b\u554a\uff0c\u5b83\u76f8\u5bf9\u4e8e\u6211\u4eec\u524d\u9762 NSNET \u548c n net\uff0c\u5b83\u76f8\u5bf9\u6765\u8bf4\u6bd4\u8f83\u6df1\uff0c\u58f0\u5bf9\u4e0d\u5bf9\uff0c\u90a3\u5c31\u610f\u5473\u7740\u4ec0\u4e48\u5462\uff0c\u56e0\u4e3a\u4f60\u8fd9\u91cc\u8f93\u5165\u4e00\u4e2a X \u662f\u5427\uff0c\u8f93\u5165 X\uff0c\u7136\u540e\u8fd9\u91cc\u6bd4\u5982\u8bf4\u901a\u8fc7\u6a21\u578b\uff0c\u901a\u8fc7\u6a21\u578b\uff0c\u901a\u8fc7\u6a21\u578b\u554a\uff0c\u6211\u4eec\u6253\u4e2a\u7701\u7565\u53f7\uff0c\u901a\u8fc7\u6a21\u578b\u5bf9\u4e0d\u5bf9\uff0c\u7136\u540e\u4e00\u76f4\u5230\u6700\u540e\u7684 Y \u5bf9\u554a\uff0c\u8fd9\u91cc\u80af\u5b9a\u662f\u6ca1\u95ee\u9898\u7684\uff0c\u5bf9\u4e0d\u5bf9\uff0c\u7136\u540e\u7684\u8bdd\u5462\u55ef\u5bf9\u7684\u8bdd\uff0c\u4f60\u6700\u597d\u8fd8\u80fd\u5f97\u51fa\u6211\u4eec\u7684 loss \u503c\uff0c\u4f60\u901a\u8fc7 loss \u503c\u4e0d\u65ad\u7684\u53bb\u53cd\u9988\u8fc7\u6765\u662f\u5427\uff0c\u53bb\u66f4\u65b0\u6211\u4eec\u7684\u4e00\u4e2a\u4ec0\u4e48 W \u7684\uff0c\u5bf9\u4e0d\u5bf9\uff0c\u7136\u540e\u6211\u4eec\u524d\u9762\u4e5f\u4e5f\u77e5\u9053\uff0c\u6211\u4eec W \u66f4\u65b0\u5e94\u8be5\u662f\u4ec0\u4e48\u6837\u5b50\u7684\uff0c\u5e94\u8be5\u662f\u8fd9\u6837\u7684\uff0c\u5bf9\u4e0d\u5bf9\u662f\u5427\uff0c\u8fd9\u91cc\u662f\u6211\u4eec\u7684\u4ec0\u4e48 loss \u503c\u7684\u51fd\u6570\uff0c\u7136\u540e\u5bf9\u5bf9\u4ec0\u4e48\u5bf9\u6211\u4eec\u7684 W \u662f\u4e00\u4e2a\u4ec0\u4e48\uff0c\u662f\u6c42\u5bfc\u7684\u8fc7\u7a0b\uff0c\u4e8b\u5b9e\u4e0a\u4f60\u4f60\u4f1a\u53d1\u73b0\u968f\u7740\u7f51\u7edc\u8d8a\u8d8a\u6765\u8d8a\u6df1\u554a\uff0c\u968f\u7740\u7f51\u7edc\u8d8a\u6765\u8d8a\u6df1\uff0c\u4f60\u5176\u5b9e\u4f60\u5f88\u5bb9\u6613\u51fa\u73b0\u4ec0\u4e48\uff0c\u524d\u9762\u6211\u4e5f\u8bb2\u8fc7\uff0c\u4e5f\u5bb9\u6613\u51fa\u73b0\u4ec0\u4e48\u68af\u5ea6\u6d88\u5931\u8fd9\u79cd\u95ee\u9898\uff0c\u5c31\u6d88\u5931\u6216\u8005\u68af\u5ea6\u7206\u70b8\uff0c\u8fd9\u4e2a\u95ee\u9898\u5bfc\u81f4\u4f60\u7684 W \u66f4\u65b0\u4e4b\u540e\uff0c\u5176\u5b9e\u4e8b\u5b9e\u4e0a\u6548\u679c\u4e0d\u4f1a\u5f88\u597d\uff0c\u6216\u8005\u76f4\u63a5\u5c31\u4e0d\u6536\u655b\u4e86\uff0c\u4e0d\u6536\u655b\u5bf9\u4e0d\u5bf9\u554a\uff0c\u4e0d\u6536\u655b\u554a\uff0c\u76f4\u63a5\u5c31\u4e0d\u6536\u655b\uff0c\u56e0\u4e3a\u4ec0\u4e48\u5462\uff0c\u56e0\u4e3a\u4f60\u8fd9\u91cc\u94fe\u5f0f\u6c42\u5bfc\u561b\uff0c\u4f60\u6bd4\u5982\u8bf4\u6253\u4e2a\u7b80\u5355\uff0c\u6bd4\u65b9\u4f60\u4f60\u4f60\u4f60\u7684\u4f60\u7684\u4e00\u4e2a loss \u503c\u5bf9\u4ec0\u4e48\uff0c\u5bf9\u6211\u4eec\u7684\u4e00\u4e2a\u4f60\u7684 loss \u662f\u5bf9\u6211\u4eec\u7684 W \u8fdb\u884c\u6c42\u5bfc\uff0c\u4e8b\u5b9e\u4e0a\u5b83\u662f\u8fd9\u6837\u5b50\u7684\uff0c\u4ed6\u53ef\u80fd\u5bf9 A \u51fd\u6570\u6c42\u5bfc\uff0cA \u51fd\u6570\u53c8\u662f B \u51fd\u6570\u7684\u5bfc\u6570\uff0c\u662f\u5427\u554a\uff0cB \u51fd\u6570\u53ef\u80fd\u662f C \u51fd\u6570\u7684\u5bfc\u6570\uff0cC \u51fd\u6570\u53ef\u80fd\u662f D \u51fd\u6570\u7684\u5bfc\u6570\uff0c\u5bf9\u4e0d\u5bf9\uff0c\u7136\u540e D \u51fd\u6570\uff0cD \u51fd\u6570\u53ef\u80fd\u624d\u662f W \u7684\u51fd\u6570\u7684\u5bfc\u6570\uff0c\u4e8b\u5b9e\u4e0a\u55ef\u8fd9\u4e2a\u8fd8\u662f\u5c11\u7684\uff0c\u53ef\u80fd\u5c31\u4e94\u5c42\uff0c\u4e8b\u5b9e\u4e0a\u4f60\u60f3\u8c61\u4e00\u4e0b\u5c31\u662f\u4e94\u5c42\u5341\u5c42\u662f\u5427\uff0c\u5443\u53ef\u80fd\u5341\u51e0\u5c42\u4ed6\u53ef\u80fd\u5c31\u662f\u8d8a\u6765\u8d8a\u6df1\u4e86\u561b\uff0c\u8d8a\u6765\u8d8a\u6df1\u7684\u8bdd\uff0c\u4f60\u4f1a\u53d1\u73b0\u4f60\u4f1a\u53d1\u73b0\uff0c\u5c31\u662f\u5982\u679c\u4f60\u4f60\u4f60\u4f60\u94fe\u5f0f\u6c42\u5bfc\u7684\u8bdd\uff0c\u4f60\u8fd9\u91cc\u5bfc\u6570\u5047\u8bbe\u662f\u5f88\u5c0f\u7684\u8bdd\uff0c\u4f60\u8fd9\u91cc\u4e5f\u5f88\u5c0f\uff0c\u4f60\u8fd9\u91cc\u4e5f\u5f88\u5c0f\uff0c\u4f60\u8fd9\u5c0f\u8fd9\u91cc\u4e5f\u5f88\u5c0f\u7684\u8bdd\u5c31\u4f1a\u5bfc\u81f4\u4ec0\u4e48\uff0c\u4f60\u8fd9\u91cc\u8fd9\u4e2a\u503c\u8d8b\u8fd1\u4e8e\u96f6\u561b\uff0c\u5c31\u95ee\u4f60\u8fd9\u91cc\u51cf\u8ddf\u6ca1\u51cf\u5b83\u4e0d\u662f\u4e00\u6837\u7684\u5417\uff0c\u4f60 W \u5c31\u4f1a\u4e0d\u66f4\u65b0\uff0c\u5bf9\u4e0d\u5bf9\uff0c\u7136\u540e\u7684\u8bdd\u7136\u540e\u7684\u8bdd\u5f80\u5f80\u5f80\u5f80\u4f1a\u51fa\u73b0\u4ec0\u4e48\u60c5\u51b5\u5462\uff0c\u5c31\u662f\u56e0\u4e3a\u4e00\u5f00\u59cb\u662f\u968f\u673a\u7684\u561b\uff0c\u5e94\u8be5\u662f\u968f\u673a\u7684\uff0c\u5bf9\u4e0d\u5bf9\uff0c\u4e00\u5f00\u59cb\u968f\u673a\u4f1a\u51fa\u73b0\u4ec0\u4e48\u8fc7\u5927\u6216\u8fc7\u5c0f\uff0c\u8fc7\u5927\u6216\u8fc7\u5c0f\u7684\u8fd9\u79cd\u60c5\u51b5\u662f\u5427\uff0c\u524d\u9762\u6211\u4eec\u5728\u8bb2\u6211\u4eec\u90a3\u4e2a\u4ec0\u4e48\u539f\u7406\u7684\u65f6\u5019\uff0c\u518d\u8bb2\u6df1\u5ea6\u5b66\u4e60\u90a3\u4e2a\u5443\u524d\u9762\u90a3\u4e2a\u57fa\u7840\u77e5\u8bc6\u7684\u65f6\u5019\uff0c\u6211\u4eec\u524d\u9762\u4e5f\u8bb2\u4e86\u561b\uff0c\u5c31\u662f\u6211\u4eec\u524d\u9762\u7528\u4e00\u4e2a\u6848\u4f8b\u53bb\u8ba1\u7b97\u4ec0\u4e48\uff0c\u68af\u5ea6\u4e0b\u964d\u6cd5\uff0c\u5c31\u662f\u53cd\u5411\u4f20\u64ad\u8fd9\u6837\u7684\u8fc7\u7a0b\u561b\uff0c\u5f53\u65f6\u7684\u8bdd\u6211\u4eec W \u554a\u662f\u5427\uff0c\u6211\u4eec\u4e00\u4e2a\u521d\u59cb\u5316\u7684\u503c\uff0c\u6bd4\u5982\u8bf4\u7b49\u4e8e\u4e00\uff0c\u90a3\u5047\u8bbe\u4f60\u521d\u59cb\u5316\u503c\u53ef\u80fd\u7b49\u4e8e\u4e00\u4e2a\u8fc7\u5927\u4e00\u4e2a\u503c\uff0c\u4f60\u4f60\u53bb\u6bd4\u5982\u8bf4\u4f60\u8fd9\u91cc\u554a\u8fc7\u5927\u503c\u662f\u5427\uff0c\u4e8b\u5b9e\u4e0a\u771f\u5b9e\u503c\u6700\u4f18\u7684\u503c\u53ef\u80fd\u662f\uff0c\u5047\u8bbe\u6211\u4eec W \u521d\u59cb\u503c\u662f 100 \u4e0a\uff0c\u6700\u4f18\u7684\u503c\u53ef\u80fd\u662f\u4ec0\u4e48\uff0c\u662f 0.1\uff0c\u6709\u53ef\u80fd\u5427\uff0c\u7136\u540e\u7684\u8bdd\u4f60\u8fd9\u91cc\u554a\u4f60\u6b63\u5e38\uff0c\u54ea\u6015\u4f60\u6b63\u5e38\u66f4\u65b0\u7684\u60c5\u51b5\u4e0b\u662f\u5427\uff0c\u6b63\u5e38\u66f4\u65b0\u7684\u60c5\u51b5\u4e0b\uff0c\u5b83\u4f1a\u5bfc\u81f4\u4ec0\u4e48\uff0c\u4f60\u6700\u540e\u4f60\u5176\u5b9e\u4f60\u66f4\u65b0\u5b8c\u4e4b\u540e\uff0c\u4f60\u53ef\u80fd\u518d\u4e58\u4e2a\u5b66\u4e60\u7387\u4e4b\u540e\uff0c\u4f60\u8fd9\u91cc\u76f8\u5f53\u4e8e\u51cf\u4e00\u4e2a 0.0001\uff0c\u5bf9\u4e0d\u5bf9\uff0c\u6709\u8fd9\u79cd\u53ef\u80fd\u5427\uff0c\u6709\u53ef\u80fd\u5427\uff0c\u4f60\u8bb2\u5b8c\u4e4b\u540e\u5462\uff0c\u4e8b\u5b9e\u4e0a\u4f60 0.001 \u548c\u4ec0\u4e48\u548c 99.9999 \u51e0\u4e4e\u4e0a\u6ca1\u4ec0\u4e48\u533a\u522b\u561b\uff0c\u4f46\u4f60\u6700\u4f18\u7684\u4e00\u4e2a\u4ec0\u4e48 W\uff0c\u5e94\u8be5\u662f\u5728\u8fd9\u4e2a\u533a\u57df\u561b\u5bf9\u5427\uff0c\u6240\u4ee5\u8bf4\u6700\u540e\u5bfc\u81f4\u4f60\u561b\u6a21\u578b\u5b83\u4e0d\u6536\u655b\uff0c\u4e8b\u5b9e\u4e0a\u6211\u4eec\u5e0c\u671b\u4ec0\u4e48\uff0c\u4e8b\u5b9e\u4e0a\u4f60\u5176\u5b9e\u4f60\u7684\u4e00\u4e2a\u7801\uff0c\u4f60\u968f\u673a\u521d\u59cb\u5c31\u662f\u4f60\u7684\u4f60 W1 \u5f00\u59cb\u968f\u673a\u968f\u673a\uff0c\u968f\u673a\u968f\u673a\u4ec0\u4e48\u8d4b\u503c\u7684\u65f6\u5019\uff0c\u4f60\u4e0d\u5e94\u8be5\u8fd9\u4e48\u8fd9\u4e48\u8fd9\u4e48\u8fd9\u4e48\u8fd9\u4e48\u8fd9\u4e48\u968f\u4fbf\uff0c\u5bf9\u4e0d\u5bf9\uff0c\u6240\u4ee5\u7684\u8bdd\u5c31\u662f\u6211\u4eec\u4f1a\u6709\u4ec0\u4e48\uff0c\u4f1a\u6709\u4e00\u4e2a\u5c31\u662f\u968f\u673a\u521d\u59cb\u5316\u7684\u4e00\u4e2a\u65b9\u6cd5\u662f\u5427\uff0c\u6765\u6765\u7ed9\u4f60\u4eec\u518d\u518d\u518d\u4e48\u5728\u8bad\u7ec3\u6211\u4eec\u6a21\u578b\u4e4b\u524d\uff0c\u6211\u4eec\u628a\u6211\u4eec\u628a\u6211\u4eec W \u6309\u7167\u4e00\u5b9a\u7684\u65b9\u5f0f\u662f\u5427\uff0c\u6309\u7167\u4e00\u5b9a\u7684\u65b9\u5f0f\u53bb\u521d\u59cb\u5316\uff0c\u6309\u7167\u4e00\u4e2a\u65b9\u6cd5\u53bb\u521d\u59cb\u5316\u54ce\uff0c\u8ba9\u4ed6\u4e0d\u81f3\u4e8e\u8fd9\u4e48\u79bb\u8c31\u5417\uff0c\u660e\u767d\u6211\u610f\u601d\u5427\uff0c\u800c\u4e14\u56e0\u4e3a W \u4f1a\u4e58\u4e0a\u4e58\u4e0a\u6211\u4eec\u5bf9\u5e94\u7684 X\uff0c\u5b83\u4f1a\u7ecf\u8fc7\u4ec0\u4e48\uff0c\u7ecf\u8fc7\u6211\u4eec\u7684\u6fc0\u6d3b\u51fd\u6570\uff0c\u6bd4\u5982\u516d\u6fc0\u6d3b\u51fd\u6570\u554a\uff0c\u6216\u8005\u4ec0\u4e48 SM \u7684\u6fc0\u6d3b\u51fd\u6570\u554a\uff0c\u4e8b\u5b9e\u4e0a\u5728\u6709\u65f6\u5019\uff0c\u8fd9\u516d\u6fc0\u6d3b\u51fd\u6570\u6216\u8005 single mode \u6fc0\u6d3b\u51fd\u6570\u554a\uff0c\u5b83\u5728\u67d0\u4e2a\u533a\u95f4\u5185\u5b83\u5728\u4ec0\u4e48\u5bfc\u6570\u662f\u600e\u4e48\u8bf4\u5462\uff0c\u6216\u8005\u8d8b\u5411\u4e8e\u8d8b\u5411\u4e8e\u96f6\u7684\uff0c\u5bf9\u4e0d\u5bf9\u662f\u5427\uff0c\u6216\u8005\u6bd4\u5982\u8bf4\u6211\u4eec LOL \u51fd\u6570\uff0c\u76f4\u63a5\u5c31\u662f\u8fd9\u6837\u4e00\u4e2a\u5206\u6bb5\u51fd\u6570\uff0c\u5728\u8fd9\u4e2a\u533a\u95f4\u5185\uff0c\u5728\u8fd9\u4e2a\u533a\u95f4\u5b83\u662f\u6ca1\u6709\u6ca1\u6709\u5bfc\u6570\u7684\uff0c\u5bf9\u4e0d\u5bf9\u662f\u5427\uff0c\u5982\u679c\u4f60\u7684\u503c\u6240\u4ee5\u5c31\u843d\u5728\u8fd9\u91cc\u7684\u8bdd\uff0c\u4f60\u5c31\u4e0d\u66f4\u65b0\u4e86\u5440\uff0c\u5bf9\u4e0d\u5bf9\uff0c\u6211\u4eec\u53ef\u80fd\u5e0c\u671b\u8fd9\u4e2a\u503c\u53ef\u80fd\u843d\u5728\u8fd9\u4e2a\u9644\u8fd1\uff0c\u5728\u67d0\u4e2a\u503c\u7684\u65f6\u5019\u5b83\u4e0d\u66f4\u65b0\u554a\uff0c\u518d\u6709\u7684\u53ea\u662f\u5b83\u66f4\u65b0\uff0c\u5bf9\u4e0d\u5bf9\uff0c\u8fd9\u6837\u6211\u4eec\u5e0c\u671b\u8fd9\u6837\u5b50\u7684\u5427\uff0c\u5982\u679c\u4f60\u8fd9\u4e2a W \u4e58\u4ee5 X \u7684\u503c\u5168\u90e8\u843d\u5728\u8fd9\u4e2a\u5730\u65b9\uff0c\u90a3\u90a3\u4f60\u5c31\u5168\u90e8\u53bb\u6309\u96f6\uff0c\u5b83\u76f4\u63a5\u4e0d\u66f4\u65b0\u4e86\u5440\uff0c\u90a3\u4f60 W \u51cf\u53bb\u6700\u540e\u51cf\u53bb\u4e86 W \u7b49\u4e8e\u4ec0\u4e48\uff0c\u7b49\u4e8e\u51cf\u53bb\u4e00\u4e2a\u51e0\u4e4e\u63a5\u8fd1\u4e8e\u96f6\u7684\u4e00\u4e2a\u6570\uff0c\u4ed6\u5c31\u4e0d\u66f4\u65b0\u4e86\u5417\uff0c\u90a3\u6700\u540e\u4f60\u7b49\u4e8e\u4ec0\u4e48\uff0c\u4f60\u7684 loss \u503c\uff0c\u4f60\u7684 loss \u662f\u80af\u5b9a\u662f\u4e0d\u4f1a\u4e0b\u964d\u7684\u5440\uff0c\u6700\u540e\u4f60\u7684\u7cbe\u786e\u5ea6\u4e5f\u4e5f\u4e0d\u4f1a\u4e0a\u5347\u7684\u5440\uff0c\u660e\u767d\u6211\u7684\u610f\u601d\u5427\uff0c\u6240\u4ee5\u5176\u5b9e\u5c31\u662f\u6709\u8fd9\u6837\u7684\u4e00\u4e2a\u65b9\u6cd5\u6765\u5e72\u561b\uff0c\u6765\u521d\u59cb\u5316\u6211\u4eec\u7684 W \u8fd9\u6837\u5b50\u600e\u4e48\u5e72\u561b\uff0c\u5c31\u4f7f\u6211\u4eec\u6700\u540e\u7684\u4e00\u4e2a\u66f4\u65b0\u7684\u901f\u5ea6\uff0c\u7b2c\u4e00\u4e2a\u53ef\u80fd\u4f1a\u53d8\u5feb\uff0c\u7b2c\u4e8c\u4e2a\u4ec0\u4e48\u4e0d\u81f3\u4e8e\u50cf\u521a\u521a\u90a3\u6837\u4e0d\u6536\u655b\u554a\uff0c\u6240\u4ee5\u7684\u8bdd\u5f53\u65f6\u554a\u5c31\u662f\u5728\u5199\u8fd9\u6bb5\u4ee3\u7801\u7684\u65f6\u5019\uff0c\u6211\u6ca1\u6709\u52a0\u4e0a\u4e86\uff0c\u52a0\u4e0a\u4e00\u4e2a\uff0c\u8fd9\u5c31\u662f\u6211\u4eec\u6743\u91cd\u521d\u59cb\u5316\u8fd9\u4e2a\u8fd9\u4e00\u9879\u5bfc\u81f4\u4ec0\u4e48\uff0c\u5bfc\u81f4\u6211\u4eec\u6a21\u578b\u5728\u8bad\u7ec3\u7684\u65f6\u5019\u8bf6\uff0c\u4ed6\u597d\u50cf\u4ec0\u4e48\u5c31\u662f\u4e00\u76f4\u4e0d\u6536\u655b\u662f\u5427\uff0c\u7136\u540e\u7684\u8bdd\u5c31\u53c8\u53c8\u524d\u9762\u4e5f\u8bb2\u4e86\uff0c\u90a3\u53ef\u80fd\u8fc7\u4e00\u6bb5\u65f6\u95f4\u4e4b\u540e\uff0c\u6211\u518d\u91cd\u65b0\u8fd0\u884c\u8fd9\u6bb5\u4ee3\u7801\u7684\u65f6\u5019\u662f\u5427\uff0c\u54ce\u53ef\u80fd\u5f53\u65f6\u4ed6\u7684\u56e0\u4e3a\u6743\u91cd\u662f\u90a3\u4e2a W \u662f\u968f\u673a\u7684\u561b\uff0cW \u968f\u673a\u7684\u8bdd\u53ef\u80fd\u6070\u597d\u8bf6\u662f\u5427\uff0c\u6070\u597d\u8fd9\u4e00\u7ec4 WA \u6bd4\u8f83\u63a5\u8fd1\u6211\u4eec\u771f\u5b9e\u503c\uff0c\u8fd9\u7ec4 W \u6bd4\u8f83\u63a5\u8fd1\u6211\u4eec\u7684\u771f\u5b9e\u771f\u5b9e W \u5bf9\u4e0d\u5bf9\uff0c\u54ce\u90a3\u53ef\u80fd\u66f4\u65b0\u5f88\u5feb\u5c31\u66f4\u65b0\u597d\u4e86\u662f\u5427\uff0c\u6240\u4ee5\u5c31\u4f1a\u5bfc\u81f4\u4ec0\u4e48\u5462\uff0c\u6bcf\u4e00\u6b21\u66f4\u65b0\u4e4b\u540e\u662f\u6bcf\u4e00\u6b21\u8fd0\u884c\uff0c\u53ef\u80fd\u8fc7\u4e94\u6b21\u8fd0\u884c\u5b83\u4e0d\u6536\u655b\uff0c\u53ef\u80fd\u7b2c\u516d\u6b21\u4ed6\u53c8\u6536\u655b\u4e86\uff0c\u6240\u4ee5\u5f53\u65f6\u5c31\u8ba9\u6211\u5f88\u82e6\u607c\u554a\uff0c\u6240\u4ee5\u554a\u6240\u4ee5\u5f53\u6211\u52a0\u4e86\u4ec0\u4e48\uff0c\u52a0\u4e0a\u6211\u4eec\u4e00\u4e2a\u4ec0\u4e48\u554a\uff0c\u6743\u91cd\u521d\u59cb\u5316\u7684\u4e00\u4e2a\u4ec0\u4e48\u65b9\u6cd5\u4e4b\u540e\uff0c\u521d\u59cb\u5316\u554a\uff0c\u521d\u59cb\u5316\u590d\u5236\u8fd9\u4e2a\u65b9\u6cd5\u4e4b\u540e\u5462\uff0c\u54ce\u57fa\u672c\u4e0a\u4f60\u5728\u4ec0\u4e48\u5728\u6211\u4eec\u90a3\u4e2a\u8fd0\u884c\u4e4b\u540e\u554a\uff0c\u4ed6\u7684\u4e00\u4e2a\u4ec0\u4e48\u5c31\u662f\u7ed3\u679c\u662f\u5f88\u7a33\u5b9a\u7684\u554a\uff0c\u57fa\u672c\u4e0a\u5c31\u662f\u5443\u5927\u6982\u5341\u8f6e\u4e8c\u8f6e\u4e4b\u540e\uff0c\u4ed6\u5c31\u57fa\u672c\u6536\u655b\u4e86\uff0c\u6240\u4ee5\u7684\u8bdd\u63a5\u4e0b\u6765\u6211\u4eec\u6765\u770b\u4e00\u4e0b\uff0c\u5c31\u662f\u65e2\u7136\u6211\u4eec\u6743\u91cd\u662f\u9700\u8981\u521d\u59cb\u5316\u7684\u5bf9\u5427\uff0c\u6211\u4eec\u6743\u91cd\u662f\u9700\u8981\u521d\u59cb\u5316\u7684\uff0c\u5443\uff0c\u90a3\u6211\u4eec\u7528\u4ee3\u7801\u600e\u4e48\u53bb\uff0c\u600e\u4e48\u53bb\u521d\u59cb\u5316\u6211\u4eec\u90a3\u4e2a\u6743\u91cd\u554a\uff0c\u8fd9\u91cc\u9762\u8fd8\u8bb2\u4e00\u4e0b\uff0c\u5176\u5b9e\u4f60\u4f1a\u53d1\u73b0\u6211\u4eec\u90a3\u4e2a ALICENE \u548c\u6211\u4eec\u7684 net\uff0c\u4ed6\u5f53\u65f6\u662f\u6ca1\u6709\u521d\u59cb\u5316\u8fd9\u4e2a\u9009\u9879\u7684\uff0c\u5443\u6211\u601d\u8003\u4e86\u4e00\u4e0b\uff0c\u53ef\u80fd\u56e0\u4e3a\u8fd9\u4e2a\u8fd9\u4e2a\u6743\u91cd\u554a\uff0c\u8fd9\u4e2a\u8fd9\u4e2a\u8fd9\u4e2a\u6a21\u578b\u5b83\u4e0d\u591f\u90a3\u4e48\u6df1\uff0c\u6240\u4ee5\u7684\u8bdd\u5462\u53ef\u80fd\u5168\u90fd\u5728\u5728\u5728\u6211\u4eec\u4ec0\u4e48\u968f\u673a\u968f\u673a\uff0c\u4ec0\u4e48\u968f\u673a\u8fdb\u884c\u4ec0\u4e48\u554a\uff0c\u8fdb\u884c\u4e00\u4e2a\u590d\u5236\u4e4b\u540e\uff0c\u7136\u540e\u8fdb\u884c\u66f4\u65b0\uff0c\u5b83\u4e5f\u662f OK \u7684\uff0c\u660e\u767d\u6211\u610f\u601d\u5427\uff0c\u4e8b\u5b9e\u4e0a\u5728 alex net \u6211\u4e5f\u51fa\u73b0\u8fc7\u4ec0\u4e48\uff0c\u5c31\u662f\u5f53\u65f6\u6ca1\u6709\u6743\u91cd\u521d\u59cb\u5316\u662f\u5427\uff0c\u7136\u540e\u6211\u8fd0\u6c14\u8bf4\u54ce\u53d1\u73b0\u4ed6\u4e5f\u6ca1\u6536\u655b\uff0c\u6709\u7528\u8fd9\u79cd\u60c5\u51b5\uff0c\u4f46\u57fa\u672c\u4e0a\u6211\u57fa\u672c\u4e0a\u6bcf\u6b21\u8fd0\u884c\u8fd9\u4e2a ANNASNET \u7684\u65f6\u5019\uff0c\u5b83\u7684\u6548\u679c\u90fd\u662f\u8fd8\u662f OK \u7684\uff0c\u6b63\u5e38\u90fd\u4f1a\u90fd\u4f1a\u8bad\u7ec3\u7684\uff0c\u4f46\u662f\u5462\u968f\u7740\u7f51\u7edc\u8d8a\u6765\u8d8a\u6df1\uff0c\u6bd4\u5982\u8bf4\u6211\u4eec\u4ec0\u4e48 VG \u554a\uff0c\u540e\u9762\u8bb2\u7684 google\uff0c\u540e\u9762\u8bb2\u516c\u5f00\u7684\uff0c\u5b83\u90fd\u662f\u9700\u8981\u4ec0\u4e48\u6743\u91cd\u77e5\u8bc6\u5316\u7684\uff0c\u5426\u5219\u5f88\u5bb9\u6613\u51fa\u73b0\u4ec0\u4e48\uff0c\u5f88\u5bb9\u6613\u51fa\u73b0 A \u4f60\u4f60\u4f60\u4f60\u4f60\u4f60\u4f60\u8fd0\u884c\uff0c\u7136\u540e\u7684\u8bdd\u53ef\u80fd\u4e8c\u8f6e\u4ed6\u90fd\u4e0d\u6536\u655b\u554a\uff0closs \u503c\u4e5f\u4e0d\u4e0b\u964d\u662f\u5427\uff0c\u6211\u4eec\u90a3\u4e2a\u7cbe\u5ea6\u4e5f\u4e0d\u4e0a\u5347\u8fd9\u79cd\u60c5\u51b5\u554a\uff0c\u6240\u4ee5\u7684\u8bdd\u8fd9\u4e00\u70b9\u6211\u4eec\u5148\u8bb2\u4e00\u4e0b\uff0c\u8fd9\u4e2a\u5c31\u662f\u6211\u4eec\u4ec0\u4e48\u6743\u91cd\u521d\u59cb\u5316\uff0c\u600e\u4e48\u5728\u4ee3\u7801\u5f53\u4e2d\u52a0\u4e0a\u4e00\u4e2a\u6743\u91cd\u521d\u59cb\u5316\uff0c\u8fd9\u4e2a\u65b9\u6cd5\u4f7f\u6211\u4eec\u6743\u91cd\u554a\uff0c\u66f4\u7b26\u5408\u6211\u4eec\u771f\u5b9e\u7684\u5b9e\u9645\u60c5\u51b5\u554a\uff0c\u597d\u5427\uff0c\u63a5\u4e0b\u6765\u6211\u4eec\u6765\u5199\u5b9a\u4e49\u7684\u4ee3\u7801\u3002<\/p>\n\n\n\n
    \"\"<\/figure>\n\n\n\n
    VGG16\u4ee3\u7801<\/h1>\n\n\n\n
    model.py<\/p>\n\n\n\n
    import torch\nfrom torch import nn\nfrom torchsummary import summary\n\n\nclass VGG16(nn.Module):\n    def __init__(self):\n        super(VGG16, self).__init__()\n        self.block1 = nn.Sequential(\n            nn.Conv2d(in_channels=1, out_channels=64, kernel_size=3, padding=1),\n            nn.ReLU(),\n            nn.Conv2d(in_channels=64, out_channels=64, kernel_size=3, padding=1),\n            nn.ReLU(),\n            nn.MaxPool2d(kernel_size=2, stride=2)\n        )\n        self.block2 = nn.Sequential(\n            nn.Conv2d(in_channels=64, out_channels=128, kernel_size=3, padding=1),\n            nn.ReLU(),\n            nn.Conv2d(in_channels=128, out_channels=128, kernel_size=3, padding=1),\n            nn.ReLU(),\n            nn.MaxPool2d(kernel_size=2, stride=2)\n        )\n        self.block3 = nn.Sequential(\n            nn.Conv2d(in_channels=128, out_channels=256, kernel_size=3, padding=1),\n            nn.ReLU(),\n            nn.Conv2d(in_channels=256, out_channels=256, kernel_size=3, padding=1),\n            nn.ReLU(),\n            nn.Conv2d(in_channels=256, out_channels=256, kernel_size=3, padding=1),\n            nn.ReLU(),\n            nn.MaxPool2d(kernel_size=2, stride=2)\n        )\n        self.block4 = nn.Sequential(\n            nn.Conv2d(in_channels=256, out_channels=512, kernel_size=3, padding=1),\n            nn.ReLU(),\n            nn.Conv2d(in_channels=512, out_channels=512, kernel_size=3, padding=1),\n            nn.ReLU(),\n            nn.Conv2d(in_channels=512, out_channels=512, kernel_size=3, padding=1),\n            nn.ReLU(),\n            nn.MaxPool2d(kernel_size=2, stride=2)\n        )\n        self.block5 = nn.Sequential(\n            nn.Conv2d(in_channels=512, out_channels=512, kernel_size=3, padding=1),\n            nn.ReLU(),\n            nn.Conv2d(in_channels=512, out_channels=512, kernel_size=3, padding=1),\n            nn.ReLU(),\n            nn.Conv2d(in_channels=512, out_channels=512, kernel_size=3, padding=1),\n            nn.ReLU(),\n            nn.MaxPool2d(kernel_size=2, stride=2)\n        )\n\n        self.block6 = nn.Sequential(\n            nn.Flatten(),\n            nn.Linear(7*7*512, 256),\n            nn.ReLU(),\n            nn.Linear(256, 128),\n            nn.ReLU(),\n            nn.Linear(128, 10),\n        )\n\n        for m in self.modules():\n            if isinstance(m, nn.Conv2d):\n                nn.init.kaiming_normal_(m.weight, nonlinearity='relu')\n                if m.bias is not None:\n                    nn.init.constant_(m.bias, 0)\n            elif isinstance(m, nn.Linear):\n                nn.init.normal_(m.weight, 0, 0.01)\n                if m.bias is not None:\n                    nn.init.constant_(m.bias, 0)\n\n    def forward(self, x):\n        x = self.block1(x)\n        x = self.block2(x)\n        x = self.block3(x)\n        x = self.block4(x)\n        x = self.block5(x)\n        x = self.block6(x)\n        return x\n\n\n\nif __name__==\"__main__\":\n    device = torch.device(\"cuda\" if torch.cuda.is_available() else \"cpu\")\n    model = VGG16().to(device)\n    print(summary(model, (1, 224, 224)))<\/code><\/pre>\n\n\n\n
    model_test.py<\/p>\n\n\n\n
    import torch\nimport torch.utils.data as Data\nfrom torchvision import transforms\nfrom torchvision.datasets import FashionMNIST\nfrom model import VGG16\n\n\n\ndef test_data_process():\n    test_data = FashionMNIST(root='.\/data',\n                              train=False,\n                              transform=transforms.Compose([transforms.Resize(size=224), transforms.ToTensor()]),\n                              download=True)\n\n    test_dataloader = Data.DataLoader(dataset=test_data,\n                                       batch_size=1,\n                                       shuffle=True,\n                                       num_workers=0)\n    return test_dataloader\n\n\ndef test_model_process(model, test_dataloader):\n    # \u8bbe\u5b9a\u6d4b\u8bd5\u6240\u7528\u5230\u7684\u8bbe\u5907\uff0c\u6709GPU\u7528GPU\u6ca1\u6709GPU\u7528CPU\n    device = \"cuda\" if torch.cuda.is_available() else 'cpu'\n\n    # \u8bb2\u6a21\u578b\u653e\u5165\u5230\u8bad\u7ec3\u8bbe\u5907\u4e2d\n    model = model.to(device)\n\n    # \u521d\u59cb\u5316\u53c2\u6570\n    test_corrects = 0.0\n    test_num = 0\n\n    # \u53ea\u8fdb\u884c\u524d\u5411\u4f20\u64ad\u8ba1\u7b97\uff0c\u4e0d\u8ba1\u7b97\u68af\u5ea6\uff0c\u4ece\u800c\u8282\u7701\u5185\u5b58\uff0c\u52a0\u5feb\u8fd0\u884c\u901f\u5ea6\n    with torch.no_grad():\n        for test_data_x, test_data_y in test_dataloader:\n            # \u5c06\u7279\u5f81\u653e\u5165\u5230\u6d4b\u8bd5\u8bbe\u5907\u4e2d\n            test_data_x = test_data_x.to(device)\n            # \u5c06\u6807\u7b7e\u653e\u5165\u5230\u6d4b\u8bd5\u8bbe\u5907\u4e2d\n            test_data_y = test_data_y.to(device)\n            # \u8bbe\u7f6e\u6a21\u578b\u4e3a\u8bc4\u4f30\u6a21\u5f0f\n            model.eval()\n            # \u524d\u5411\u4f20\u64ad\u8fc7\u7a0b\uff0c\u8f93\u5165\u4e3a\u6d4b\u8bd5\u6570\u636e\u96c6\uff0c\u8f93\u51fa\u4e3a\u5bf9\u6bcf\u4e2a\u6837\u672c\u7684\u9884\u6d4b\u503c\n            output= model(test_data_x)\n            # \u67e5\u627e\u6bcf\u4e00\u884c\u4e2d\u6700\u5927\u503c\u5bf9\u5e94\u7684\u884c\u6807\n            pre_lab = torch.argmax(output, dim=1)\n            # \u5982\u679c\u9884\u6d4b\u6b63\u786e\uff0c\u5219\u51c6\u786e\u5ea6test_corrects\u52a01\n            test_corrects += torch.sum(pre_lab == test_data_y.data)\n            # \u5c06\u6240\u6709\u7684\u6d4b\u8bd5\u6837\u672c\u8fdb\u884c\u7d2f\u52a0\n            test_num += test_data_x.size(0)\n\n    # \u8ba1\u7b97\u6d4b\u8bd5\u51c6\u786e\u7387\n    test_acc = test_corrects.double().item() \/ test_num\n    print(\"\u6d4b\u8bd5\u7684\u51c6\u786e\u7387\u4e3a\uff1a\", test_acc)\n\n\nif __name__ == \"__main__\":\n    # \u52a0\u8f7d\u6a21\u578b\n    model = VGG16()\n    model.load_state_dict(torch.load('best_model.pth'))\n    # # \u5229\u7528\u73b0\u6709\u7684\u6a21\u578b\u8fdb\u884c\u6a21\u578b\u7684\u6d4b\u8bd5\n    test_dataloader = test_data_process()\n    test_model_process(model, test_dataloader)\n\n\n    # \u8bbe\u5b9a\u6d4b\u8bd5\u6240\u7528\u5230\u7684\u8bbe\u5907\uff0c\u6709GPU\u7528GPU\u6ca1\u6709GPU\u7528CPU\n    device = \"cuda\" if torch.cuda.is_available() else 'cpu'\n    model = model.to(device)\n\n    classes = ['T-shirt\/top', 'Trouser', 'Pullover', 'Dress', 'Coat', 'Sandal', 'Shirt', 'Sneaker', 'Bag', 'Ankle boot']\n    with torch.no_grad():\n        for b_x, b_y in test_dataloader:\n            b_x = b_x.to(device)\n            b_y = b_y.to(device)\n\n            # \u8bbe\u7f6e\u6a21\u578b\u4e3a\u9a8c\u8bc1\u6a21\u578b\n            model.eval()\n            output = model(b_x)\n            pre_lab = torch.argmax(output, dim=1)\n            result = pre_lab.item()\n            label = b_y.item()\n            print(\"\u9884\u6d4b\u503c\uff1a\",  classes[result], \"------\", \"\u771f\u5b9e\u503c\uff1a\", classes[label])<\/code><\/pre>\n\n\n\n
    model_train.py<\/p>\n\n\n\n
    import copy\nimport time\n\nimport torch\nfrom torchvision.datasets import FashionMNIST\nfrom torchvision import transforms\nimport torch.utils.data as Data\nimport numpy as np\nimport matplotlib.pyplot as plt\nfrom model import VGG16\nimport torch.nn as nn\nimport pandas as pd\n\n\ndef train_val_data_process():\n    train_data = FashionMNIST(root='.\/data',\n                              train=True,\n                              transform=transforms.Compose([transforms.Resize(size=224), transforms.ToTensor()]),\n                              download=True)\n\n    train_data, val_data = Data.random_split(train_data, [round(0.8*len(train_data)), round(0.2*len(train_data))])\n    train_dataloader = Data.DataLoader(dataset=train_data,\n                                       batch_size=28,\n                                       shuffle=True,\n                                       num_workers=2)\n\n    val_dataloader = Data.DataLoader(dataset=val_data,\n                                       batch_size=28,\n                                       shuffle=True,\n                                       num_workers=2)\n\n    return train_dataloader, val_dataloader\n\n\ndef train_model_process(model, train_dataloader, val_dataloader, num_epochs):\n    # \u8bbe\u5b9a\u8bad\u7ec3\u6240\u7528\u5230\u7684\u8bbe\u5907\uff0c\u6709GPU\u7528GPU\u6ca1\u6709GPU\u7528CPU\n    device = torch.device(\"cuda\" if torch.cuda.is_available() else \"cpu\")\n    # \u4f7f\u7528Adam\u4f18\u5316\u5668\uff0c\u5b66\u4e60\u7387\u4e3a0.001\n    optimizer = torch.optim.Adam(model.parameters(), lr=0.001)\n    # \u635f\u5931\u51fd\u6570\u4e3a\u4ea4\u53c9\u71b5\u51fd\u6570\n    criterion = nn.CrossEntropyLoss()\n    # \u5c06\u6a21\u578b\u653e\u5165\u5230\u8bad\u7ec3\u8bbe\u5907\u4e2d\n    model = model.to(device)\n    # \u590d\u5236\u5f53\u524d\u6a21\u578b\u7684\u53c2\u6570\n    best_model_wts = copy.deepcopy(model.state_dict())\n\n    # \u521d\u59cb\u5316\u53c2\u6570\n    # \u6700\u9ad8\u51c6\u786e\u5ea6\n    best_acc = 0.0\n    # \u8bad\u7ec3\u96c6\u635f\u5931\u5217\u8868\n    train_loss_all = []\n    # \u9a8c\u8bc1\u96c6\u635f\u5931\u5217\u8868\n    val_loss_all = []\n    # \u8bad\u7ec3\u96c6\u51c6\u786e\u5ea6\u5217\u8868\n    train_acc_all = []\n    # \u9a8c\u8bc1\u96c6\u51c6\u786e\u5ea6\u5217\u8868\n    val_acc_all = []\n    # \u5f53\u524d\u65f6\u95f4\n    since = time.time()\n\n    for epoch in range(num_epochs):\n        print(\"Epoch {}\/{}\".format(epoch, num_epochs-1))\n        print(\"-\"*10)\n\n        # \u521d\u59cb\u5316\u53c2\u6570\n        # \u8bad\u7ec3\u96c6\u635f\u5931\u51fd\u6570\n        train_loss = 0.0\n        # \u8bad\u7ec3\u96c6\u51c6\u786e\u5ea6\n        train_corrects = 0\n        # \u9a8c\u8bc1\u96c6\u635f\u5931\u51fd\u6570\n        val_loss = 0.0\n        # \u9a8c\u8bc1\u96c6\u51c6\u786e\u5ea6\n        val_corrects = 0\n        # \u8bad\u7ec3\u96c6\u6837\u672c\u6570\u91cf\n        train_num = 0\n        # \u9a8c\u8bc1\u96c6\u6837\u672c\u6570\u91cf\n        val_num = 0\n\n        # \u5bf9\u6bcf\u4e00\u4e2amini-batch\u8bad\u7ec3\u548c\u8ba1\u7b97\n        for step, (b_x, b_y) in enumerate(train_dataloader):\n            # \u5c06\u7279\u5f81\u653e\u5165\u5230\u8bad\u7ec3\u8bbe\u5907\u4e2d\n            b_x = b_x.to(device)\n            # \u5c06\u6807\u7b7e\u653e\u5165\u5230\u8bad\u7ec3\u8bbe\u5907\u4e2d\n            b_y = b_y.to(device)\n            # \u8bbe\u7f6e\u6a21\u578b\u4e3a\u8bad\u7ec3\u6a21\u5f0f\n            model.train()\n\n            # \u524d\u5411\u4f20\u64ad\u8fc7\u7a0b\uff0c\u8f93\u5165\u4e3a\u4e00\u4e2abatch\uff0c\u8f93\u51fa\u4e3a\u4e00\u4e2abatch\u4e2d\u5bf9\u5e94\u7684\u9884\u6d4b\n            output = model(b_x)\n            # \u67e5\u627e\u6bcf\u4e00\u884c\u4e2d\u6700\u5927\u503c\u5bf9\u5e94\u7684\u884c\u6807\n            pre_lab = torch.argmax(output, dim=1)\n            # \u8ba1\u7b97\u6bcf\u4e00\u4e2abatch\u7684\u635f\u5931\u51fd\u6570\n            loss = criterion(output, b_y)\n\n            # \u5c06\u68af\u5ea6\u521d\u59cb\u5316\u4e3a0\n            optimizer.zero_grad()\n            # \u53cd\u5411\u4f20\u64ad\u8ba1\u7b97\n            loss.backward()\n            # \u6839\u636e\u7f51\u7edc\u53cd\u5411\u4f20\u64ad\u7684\u68af\u5ea6\u4fe1\u606f\u6765\u66f4\u65b0\u7f51\u7edc\u7684\u53c2\u6570\uff0c\u4ee5\u8d77\u5230\u964d\u4f4eloss\u51fd\u6570\u8ba1\u7b97\u503c\u7684\u4f5c\u7528\n            optimizer.step()\n            # \u5bf9\u635f\u5931\u51fd\u6570\u8fdb\u884c\u7d2f\u52a0\n            train_loss += loss.item() * b_x.size(0)\n            # \u5982\u679c\u9884\u6d4b\u6b63\u786e\uff0c\u5219\u51c6\u786e\u5ea6train_corrects\u52a01\n            train_corrects += torch.sum(pre_lab == b_y.data)\n            # \u5f53\u524d\u7528\u4e8e\u8bad\u7ec3\u7684\u6837\u672c\u6570\u91cf\n            train_num += b_x.size(0)\n        for step, (b_x, b_y) in enumerate(val_dataloader):\n            # \u5c06\u7279\u5f81\u653e\u5165\u5230\u9a8c\u8bc1\u8bbe\u5907\u4e2d\n            b_x = b_x.to(device)\n            # \u5c06\u6807\u7b7e\u653e\u5165\u5230\u9a8c\u8bc1\u8bbe\u5907\u4e2d\n            b_y = b_y.to(device)\n            # \u8bbe\u7f6e\u6a21\u578b\u4e3a\u8bc4\u4f30\u6a21\u5f0f\n            model.eval()\n            # \u524d\u5411\u4f20\u64ad\u8fc7\u7a0b\uff0c\u8f93\u5165\u4e3a\u4e00\u4e2abatch\uff0c\u8f93\u51fa\u4e3a\u4e00\u4e2abatch\u4e2d\u5bf9\u5e94\u7684\u9884\u6d4b\n            output = model(b_x)\n            # \u67e5\u627e\u6bcf\u4e00\u884c\u4e2d\u6700\u5927\u503c\u5bf9\u5e94\u7684\u884c\u6807\n            pre_lab = torch.argmax(output, dim=1)\n            # \u8ba1\u7b97\u6bcf\u4e00\u4e2abatch\u7684\u635f\u5931\u51fd\u6570\n            loss = criterion(output, b_y)\n\n\n            # \u5bf9\u635f\u5931\u51fd\u6570\u8fdb\u884c\u7d2f\u52a0\n            val_loss += loss.item() * b_x.size(0)\n            # \u5982\u679c\u9884\u6d4b\u6b63\u786e\uff0c\u5219\u51c6\u786e\u5ea6train_corrects\u52a01\n            val_corrects += torch.sum(pre_lab == b_y.data)\n            # \u5f53\u524d\u7528\u4e8e\u9a8c\u8bc1\u7684\u6837\u672c\u6570\u91cf\n            val_num += b_x.size(0)\n\n        # \u8ba1\u7b97\u5e76\u4fdd\u5b58\u6bcf\u4e00\u6b21\u8fed\u4ee3\u7684loss\u503c\u548c\u51c6\u786e\u7387\n        # \u8ba1\u7b97\u5e76\u4fdd\u5b58\u8bad\u7ec3\u96c6\u7684loss\u503c\n        train_loss_all.append(train_loss \/ train_num)\n        # \u8ba1\u7b97\u5e76\u4fdd\u5b58\u8bad\u7ec3\u96c6\u7684\u51c6\u786e\u7387\n        train_acc_all.append(train_corrects.double().item() \/ train_num)\n\n        # \u8ba1\u7b97\u5e76\u4fdd\u5b58\u9a8c\u8bc1\u96c6\u7684loss\u503c\n        val_loss_all.append(val_loss \/ val_num)\n        # \u8ba1\u7b97\u5e76\u4fdd\u5b58\u9a8c\u8bc1\u96c6\u7684\u51c6\u786e\u7387\n        val_acc_all.append(val_corrects.double().item() \/ val_num)\n\n        print(\"{} train loss:{:.4f} train acc: {:.4f}\".format(epoch, train_loss_all[-1], train_acc_all[-1]))\n        print(\"{} val loss:{:.4f} val acc: {:.4f}\".format(epoch, val_loss_all[-1], val_acc_all[-1]))\n\n        if val_acc_all[-1] > best_acc:\n            # \u4fdd\u5b58\u5f53\u524d\u6700\u9ad8\u51c6\u786e\u5ea6\n            best_acc = val_acc_all[-1]\n            # \u4fdd\u5b58\u5f53\u524d\u6700\u9ad8\u51c6\u786e\u5ea6\u7684\u6a21\u578b\u53c2\u6570\n            best_model_wts = copy.deepcopy(model.state_dict())\n\n        # \u8ba1\u7b97\u8bad\u7ec3\u548c\u9a8c\u8bc1\u7684\u8017\u65f6\n        time_use = time.time() - since\n        print(\"\u8bad\u7ec3\u548c\u9a8c\u8bc1\u8017\u8d39\u7684\u65f6\u95f4{:.0f}m{:.0f}s\".format(time_use\/\/60, time_use%60))\n\n    # \u9009\u62e9\u6700\u4f18\u53c2\u6570\uff0c\u4fdd\u5b58\u6700\u4f18\u53c2\u6570\u7684\u6a21\u578b\n    model.load_state_dict(best_model_wts)\n    # torch.save(model.load_state_dict(best_model_wts), \"C:\/Users\/86159\/Desktop\/LeNet\/best_model.pth\")\n    torch.save(best_model_wts, \"C:\/Users\/86159\/Desktop\/VGG16\/best_model.pth\")\n\n\n    train_process = pd.DataFrame(data={\"epoch\":range(num_epochs),\n                                       \"train_loss_all\":train_loss_all,\n                                       \"val_loss_all\":val_loss_all,\n                                       \"train_acc_all\":train_acc_all,\n                                       \"val_acc_all\":val_acc_all,})\n\n    return train_process\n\n\ndef matplot_acc_loss(train_process):\n    # \u663e\u793a\u6bcf\u4e00\u6b21\u8fed\u4ee3\u540e\u7684\u8bad\u7ec3\u96c6\u548c\u9a8c\u8bc1\u96c6\u7684\u635f\u5931\u51fd\u6570\u548c\u51c6\u786e\u7387\n    plt.figure(figsize=(12, 4))\n    plt.subplot(1, 2, 1)\n    plt.plot(train_process['epoch'], train_process.train_loss_all, \"ro-\", label=\"Train loss\")\n    plt.plot(train_process['epoch'], train_process.val_loss_all, \"bs-\", label=\"Val loss\")\n    plt.legend()\n    plt.xlabel(\"epoch\")\n    plt.ylabel(\"Loss\")\n    plt.subplot(1, 2, 2)\n    plt.plot(train_process['epoch'], train_process.train_acc_all, \"ro-\", label=\"Train acc\")\n    plt.plot(train_process['epoch'], train_process.val_acc_all, \"bs-\", label=\"Val acc\")\n    plt.xlabel(\"epoch\")\n    plt.ylabel(\"acc\")\n    plt.legend()\n    plt.show()\n\n\nif __name__ == '__main__':\n    # \u52a0\u8f7d\u9700\u8981\u7684\u6a21\u578b\n    VGG16 = VGG16()\n    # \u52a0\u8f7d\u6570\u636e\u96c6\n    train_data, val_data = train_val_data_process()\n    # \u5229\u7528\u73b0\u6709\u7684\u6a21\u578b\u8fdb\u884c\u6a21\u578b\u7684\u8bad\u7ec3\n    train_process = train_model_process(VGG16, train_data, val_data, num_epochs=20)\n    matplot_acc_loss(train_process)<\/code><\/pre>\n","protected":false},"excerpt":{"rendered":"
    VGGNet\u662f\u725b\u6d25\u5927\u5b66\u8ba1\u7b97\u673a\u89c6\u89c9\u7ec4(Visual Geometry Group\uff09\u548c\u8c37\u6b4c DeepMind \u4e00\u8d77\u7814\u7a76\u51fa\u6765\u7684\u6df1\u5ea6\u5377\u79ef\u795e …<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"emotion":"","emotion_color":"","title_style":"","license":"","footnotes":""},"categories":[2],"tags":[9],"class_list":["post-286","post","type-post","status-publish","format-standard","hentry","category-2","tag-9"],"_links":{"self":[{"href":"https:\/\/eve2333.top\/index.php?rest_route=\/wp\/v2\/posts\/286","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/eve2333.top\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/eve2333.top\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/eve2333.top\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/eve2333.top\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=286"}],"version-history":[{"count":0,"href":"https:\/\/eve2333.top\/index.php?rest_route=\/wp\/v2\/posts\/286\/revisions"}],"wp:attachment":[{"href":"https:\/\/eve2333.top\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=286"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/eve2333.top\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=286"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/eve2333.top\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=286"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}