{"id":1128,"date":"2018-04-23T18:31:54","date_gmt":"2018-04-23T10:31:54","guid":{"rendered":"https:\/\/damogame.cn:15443\/wordpress\/?p=1128"},"modified":"2018-12-27T17:04:53","modified_gmt":"2018-12-27T09:04:53","slug":"%e6%b7%b1%e5%ba%a6-%e6%9c%ba%e5%99%a8%e5%ad%a6%e4%b9%a0%e6%95%b2%e9%97%a8%e7%a0%96%ef%bc%9a%e4%bb%bb%e4%bd%95%e4%ba%ba%e9%83%bd%e8%83%bd%e7%9c%8b%e6%87%82%e7%9a%84tensorflow%e4%bb%8b%e7%bb%8d","status":"publish","type":"post","link":"https:\/\/i007.cc\/wordpress\/archives\/1128","title":{"rendered":"\u6df1\u5ea6 | \u673a\u5668\u5b66\u4e60\u6572\u95e8\u7816\uff1a\u4efb\u4f55\u4eba\u90fd\u80fd\u770b\u61c2\u7684TensorFlow\u4ecb\u7ecd"},"content":{"rendered":"

\u6df1\u5ea6 | \u673a\u5668\u5b66\u4e60\u6572\u95e8\u7816\uff1a\u4efb\u4f55\u4eba\u90fd\u80fd\u770b\u61c2\u7684TensorFlow\u4ecb\u7ecd<\/h2>\n
2016-08-21<\/em>\u00a0\u673a\u5668\u4e4b\u5fc3<\/a><\/div>\n
\n
\n
\n
\n
\n
\n
\n
\n
\n
\n
\n
\u9009\u81ea kdnuggets<\/p>\n
\u4f5c\u8005\uff1aSoon Hin Khor<\/strong><\/p>\n

\u673a\u5668\u4e4b\u5fc3\u7f16\u8bd1<\/strong><\/p>\n

\u53c2\u4e0e\uff1aRick\u3001\u5434\u6500\u3001\u674e\u4e9a\u6d32<\/strong><\/p>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n

\u672c\u6587\u662f\u65e5\u672c\u4e1c\u4eac TensorFlow \u805a\u4f1a\u8054\u5408\u7ec4\u7ec7\u8005 Hin Khor \u6240\u5199\u7684 TensorFlow \u7cfb\u5217\u4ecb\u7ecd\u6587\u7ae0\u7684\u524d\u4e24\u90e8\u5206\uff0c\u81ea\u79f0\u7ed9\u51fa\u4e86\u5173\u4e8e TensorFlow \u7684 gentlest \u7684\u4ecb\u7ecd\u3002\u8fd9\u4e24\u90e8\u5206\u8c08\u5230\u5355\u4e00\u7279\u5f81\u95ee\u9898\u7684\u7ebf\u6027\u56de\u5f52\u95ee\u9898\u4ee5\u53ca\u8bad\u7ec3\uff08training\uff09\u7684\u542b\u4e49\uff0c\u673a\u5668\u4e4b\u5fc3\u5c06\u7ee7\u7eed\u5173\u6ce8\u672c\u7cfb\u5217\u6587\u7ae0\u7684\u540e\u7eed\u66f4\u65b0\u3002<\/em><\/p><\/blockquote>\n

\u7b2c\u4e00\u90e8\u5206<\/strong><\/p>\n

\u5f15\u8a00<\/strong><\/p>\n

\u6211\u4eec\u8981\u89e3\u51b3\u7684\u662f\u4e00\u4e2a\u8fc7\u4e8e\u7b80\u5355\u4e14\u4e0d\u73b0\u5b9e\u7684\u95ee\u9898\uff0c\u4f46\u5176\u597d\u7684\u4e00\u9762\u662f\u4fbf\u4e8e\u6211\u4eec\u4e86\u89e3\u673a\u5668\u5b66\u4e60\u548c TensorFlow \u7684\u6982\u5ff5\u3002\u6211\u4eec\u8981\u9884\u6d4b\u4e00\u4e2a\u57fa\u4e8e\u5355\u4e00\u7279\u5f81\uff08\u623f\u95f4\u9762\u79ef\/\u5e73\u65b9\u7c73\uff09\u7684\u5355\u6807\u91cf\u8f93\u51fa\uff08\u623f\u4ef7\/\u7f8e\u5143\uff09\u3002\u8fd9\u6837\u505a\u6d88\u9664\u4e86\u5904\u7406\u591a\u7ef4\u6570\u636e\u7684\u9700\u8981\uff0c\u4f7f\u6211\u4eec\u80fd\u591f\u5728 TensorFlow \u4e2d\u53ea\u4e13\u6ce8\u4e8e\u786e\u5b9a\u3001\u5b9e\u73b0\u4ee5\u53ca\u8bad\u7ec3\u6a21\u578b\u3002<\/p>\n

\u673a\u5668\u5b66\u4e60\u7b80\u4ecb<\/strong><\/p>\n

\u6211\u4eec\u4ece\u4e00\u7ec4\u6536\u96c6\u5230\u7684\u6570\u636e\u70b9\u5f00\u59cb\uff08\u89c1\u4e0b\u56fe\uff09\uff0c\u6bcf\u4e2a\u6570\u636e\u70b9\u4ee3\u8868\u4e24\u4e2a\u503c\u4e4b\u95f4\u7684\u5173\u7cfb\u2014\u2014\u8f93\u51fa\uff08\u623f\u4ef7\uff09\u4e0e\u5f71\u54cd\u56e0\u7d20\uff08\u623f\u5b50\u9762\u79ef\uff09\u3002<\/p>\n

<\/p>\n

\u7136\u800c\u6211\u4eec\u65e0\u6cd5\u9884\u6d4b\u6ca1\u6709\u6570\u636e\u70b9\u7684\u7279\u5f81\u7684\u503c\uff08\u89c1\u4e0b\u56fe\uff09\u3002<\/p>\n

<\/p>\n

\u6211\u4eec\u53ef\u4ee5\u4f7f\u7528\u673a\u5668\u5b66\u4e60\u6765\u6316\u6398\u5b83\u4eec\u4e4b\u95f4\u7684\u5173\u7cfb\uff08\u89c1\u4e0b\u56fe\u7684\u300c\u6700\u4f73\u62df\u5408\u9884\u6d4b\u66f2\u7ebf\u300d\uff09\uff0c\u5373\u7ed9\u5b9a\u4e00\u4e2a\u4e0d\u5c5e\u4e8e\u6570\u636e\u70b9\u7684\u7279\u5f81\u503c\uff0c\u6211\u4eec\u53ef\u4ee5\u51c6\u786e\u5730\u9884\u6d4b\u51fa\u8f93\u51fa\uff08\u7279\u5f81\u503c\u548c\u9884\u6d4b\u7ebf\u7684\u4ea4\u70b9\uff09\u3002<\/p>\n

<\/p>\n

\u6b65\u9aa4\u4e00\uff1a\u9009\u62e9\u4e00\u4e2a\u6a21\u578b<\/strong><\/p>\n

1.\u6a21\u578b\u79cd\u7c7b<\/strong><\/p>\n

\u4e3a\u4e86\u4f7f\u7528\u673a\u5668\u5b66\u4e60\u6765\u505a\u9884\u6d4b\uff0c\u6211\u4eec\u9700\u8981\u9009\u62e9\u4e00\u4e2a\u80fd\u591f\u62df\u5408\u6536\u96c6\u5230\u7684\u6570\u636e\u7684\u6700\u4f73\u6a21\u578b\u3002<\/p>\n

\u6211\u4eec\u53ef\u4ee5\u9009\u62e9\u4e00\u4e2a\u7ebf\u6027\uff08\u76f4\u7ebf\uff09\u6a21\u578b\uff0c\u5e76\u901a\u8fc7\u6539\u53d8\u5176\u9661\u5ea6\/\u68af\u5ea6\u548c\u4f4d\u7f6e\u5bf9\u5176\u8fdb\u884c\u8c03\u6574\uff0c\u4ece\u800c\u5339\u914d\u6570\u636e\u70b9\u3002<\/p>\n

<\/p>\n

\u6211\u4eec\u4e5f\u53ef\u4ee5\u9009\u62e9\u4e00\u4e2a\u6307\u6570\uff08\u66f2\u7ebf\uff09\u6a21\u578b\uff0c\u5e76\u901a\u8fc7\u6539\u53d8\u5176\u66f2\u7387\uff08curvature\uff09\u548c\u4f4d\u7f6e\u5bf9\u5176\u8fdb\u884c\u8c03\u6574\uff0c\u4ece\u800c\u5339\u914d\u540c\u4e00\u6570\u636e\u70b9\u96c6\u3002<\/p>\n

<\/p>\n

2.\u6210\u672c\u51fd\u6570<\/strong><\/p>\n

\u4e3a\u4e86\u6bd4\u8f83\u54ea\u4e2a\u6a21\u578b\u62df\u5408\u5f97\u66f4\u4e25\u5bc6\uff0c\u6570\u5b66\u4e0a\u6211\u4eec\u5c06\u6700\u4f73\u62df\u5408\u5b9a\u4e49\u4e3a\u4e00\u4e2a\u9700\u8981\u88ab\u6700\u5c0f\u5316\u7684\u6210\u672c\u51fd\u6570\u3002 \u6210\u672c\u51fd\u6570\u7684\u4e00\u4e2a\u7b80\u5355\u6837\u4f8b\u662f\u6bcf\u4e2a\u6570\u636e\u70b9\u6240\u4ee3\u8868\u7684\u5b9e\u9645\u8f93\u51fa\u4e0e\u9884\u6d4b\u8f93\u51fa\u4e4b\u95f4\u504f\u5dee\u7684\u7edd\u5bf9\u503c\u603b\u548c\uff08\u5b9e\u9645\u7ed3\u679c\u5230\u6700\u4f73\u62df\u5408\u66f2\u7ebf\u7684\u5782\u76f4\u6295\u5f71\uff09\u3002\u7528\u56fe\u8868\u8868\u793a\uff0c\u6210\u672c\u51fd\u6570\u88ab\u63cf\u8ff0\u4e3a\u4e0b\u8868\u4e2d\u84dd\u8272\u7ebf\u6bb5\u7684\u957f\u5ea6\u548c\u3002<\/p>\n

<\/p>\n

\u6ce8\u610f\uff1a\u66f4\u51c6\u786e\u5730\u8bf4\uff0c\u6210\u672c\u51fd\u6570\u5f80\u5f80\u662f\u5b9e\u9645\u8f93\u51fa\u548c\u9884\u6d4b\u8f93\u51fa\u4e4b\u95f4\u7684\u65b9\u5dee\uff0c\u56e0\u4e3a\u5dee\u503c\u6709\u65f6\u662f\u8d1f\u6570\uff1b\u8fd9\u4e5f\u79f0\u4e3a\u6700\u5c0f\u4e8c\u4e58\u6cd5\u3002<\/p>\n

3.\u7ebf\u6027\u6a21\u578b\u7b80\u4ecb<\/strong><\/p>\n

\u79c9\u6301\u7b80\u6d01\u7cbe\u795e\uff0c\u6211\u4eec\u5c06\u4f7f\u7528\u7ebf\u6027\u6a21\u578b\u6765\u5bf9\u6570\u636e\u70b9\u8fdb\u884c\u5efa\u6a21\u3002\u7ebf\u6027\u6a21\u578b\u7684\u6570\u5b66\u8868\u793a\u662f\uff1a<\/p>\n

\u00a0y = W.x + b<\/strong><\/em><\/p>\n

\u00a0 \u00a0Where:<\/strong><\/em><\/p>\n

\u00a0 \u00a0x: house size, in sqm<\/strong><\/em><\/p>\n

\u00a0 \u00a0y: predicted house price, in $<\/strong><\/em><\/p><\/blockquote>\n

\u4e3a\u4e86\u8c03\u6574\u6a21\u578b\u6765\u66f4\u597d\u5730\u62df\u5408\u6570\u636e\u70b9\uff0c\u6211\u4eec\u53ef\u4ee5\u8fd9\u6837\u505a\uff1a<\/p>\n

    \n
  • \u8c03\u6574 W \u6765\u6539\u53d8\u7ebf\u6027\u6a21\u578b\u7684\u68af\u5ea6<\/li>\n<\/ul>\n

    <\/p>\n

      \n
    • \u8c03\u6574 b \u6765\u6539\u53d8\u7ebf\u6027\u6a21\u578b\u7684\u4f4d\u7f6e<\/li>\n<\/ul>\n

      <\/p>\n

      \u901a\u8fc7\u4f7f\u7528\u8bb8\u591a\u4e2a W\u3001b \u7684\u503c\uff0c\u6700\u7ec8\u6211\u4eec\u53ef\u4ee5\u627e\u5230\u4e00\u4e2a\u6700\u4f73\u62df\u5408\u7ebf\u6027\u6a21\u578b\uff0c\u80fd\u591f\u5c06\u6210\u672c\u51fd\u6570\u964d\u5230\u6700\u5c0f\u3002<\/p>\n

      \u9664\u4e86\u968f\u673a\u5c1d\u8bd5\u4e0d\u540c\u7684\u503c\uff0c\u6709\u6ca1\u6709\u4e00\u4e2a\u66f4\u597d\u7684\u65b9\u6cd5\u6765\u5feb\u901f\u627e\u5230 W\u3001b \u7684\u503c\uff1f<\/p>\n

      4.\u68af\u5ea6\u4e0b\u964d<\/strong><\/p>\n

      \u5982\u679c\u4f60\u8bd5\u56fe\u4ece\u5c71\u4e0a\u4e0b\u964d\u5230\u6700\u4f4e\u70b9\uff0c\u4f60\u7684\u89c6\u89d2\u5c31\u662f\u8fd9\u4e2a\u6837\u5b50\u3002<\/p>\n

      <\/p>\n

      \u4e0b\u964d\u8d8b\u52bf\u5e76\u4e0d\u660e\u663e\uff01\u5176\u6700\u4f73\u65b9\u5f0f\u662f\u6267\u884c\u68af\u5ea6\u4e0b\u964d\uff1a<\/p>\n

        \n
      • \u5728\u5f53\u524d\u4f4d\u7f6e\u4ee5\u6700\u9661\u7684\u4e0b\u964d\u68af\u5ea6\u786e\u5b9a\u65b9\u5411<\/li>\n
      • \u5728\u8be5\u65b9\u5411\u4e0a\u91c7\u53d6\u6b65\u957f X<\/li>\n
      • \u91cd\u590d & \u5237\u65b0\uff1b\u8fd9\u5c31\u662f\u8bad\u7ec3\u8fc7\u7a0b<\/li>\n<\/ul>\n

        \u6700\u5c0f\u5316\u6210\u672c\u51fd\u6570\u662f\u7c7b\u4f3c\u7684\uff0c\u56e0\u4e3a\u6210\u672c\u51fd\u6570\u5c31\u50cf\u662f\u8d77\u4f0f\u7684\u5c71\uff0c\u6211\u4eec\u60f3\u8981\u627e\u5230\u5176\u4e2d\u7684\u6700\u4f4e\u70b9\uff0c\u6211\u4eec\u53ef\u4ee5\u901a\u8fc7\u68af\u5ea6\u4e0b\u964d\u7c7b\u4f3c\u5730\u5b9e\u73b0\u3002<\/p>\n

        <\/p>\n

        \u73b0\u5728\u6211\u4eec\u6709\u4e86\u7ebf\u6027\u6a21\u578b\u3001\u6210\u672c\u51fd\u6570\u548c\u68af\u5ea6\u4e0b\u964d\u7684\u6982\u5ff5\uff0c\u53ef\u4ee5\u5f00\u59cb\u4f7f\u7528 TensorFlow \u4e86\u3002<\/p>\n

        \u6b65\u9aa4\u4e8c\uff1a\u5728TensorFlow \u4e2d\u5efa\u7acb\u6a21\u578b<\/strong><\/p>\n

        1.TensorFlow \u4e2d\u7684\u7ebf\u6027\u6a21\u578b<\/strong><\/p>\n

        TensorFlow \u76842\u4e2a\u57fa\u672c\u7ec4\u4ef6\u662f\uff1a<\/p>\n

          \n
        • \u5360\u4f4d\u7b26\uff08Placeholder\uff09\uff1a\u8868\u793a\u6267\u884c\u68af\u5ea6\u4e0b\u964d\u65f6\u5c06\u5b9e\u9645\u6570\u636e\u503c\u8f93\u5165\u5230\u6a21\u578b\u4e2d\u7684\u4e00\u4e2a\u5165\u53e3\u70b9\u3002\u4f8b\u5982\u623f\u5b50\u9762\u79ef \u00a0(x) \u548c\u623f\u4ef7 (y_)\u3002<\/li>\n<\/ul>\n

          <\/p>\n

            \n
          • \u53d8\u91cf\uff1a\u8868\u793a\u6211\u4eec\u8bd5\u56fe\u5bfb\u627e\u7684\u80fd\u591f\u4f7f\u6210\u672c\u51fd\u6570\u964d\u5230\u6700\u5c0f\u7684\u300cgood\u300d\u503c\u7684\u53d8\u91cf\uff0c\u4f8b\u5982 W \u548c b\u3002<\/li>\n<\/ul>\n

            <\/p>\n

            \u7136\u540e TensorFlow \u4e2d\u7684\u7ebf\u6027\u6a21\u578b (y = W.x + b) \u5c31\u662f\uff1a<\/p>\n

            <\/p>\n

            2.TensorFlow \u4e2d\u7684\u6210\u672c\u51fd\u6570<\/strong><\/p>\n

            \u4e0e\u5c06\u6570\u636e\u70b9\u7684\u5b9e\u9645\u623f\u4ef7 (y_) \u8f93\u5165\u6a21\u578b\u7c7b\u4f3c\uff0c\u6211\u4eec\u521b\u5efa\u4e00\u4e2a\u5360\u4f4d\u7b26\u3002<\/p>\n

            <\/p>\n

            \u6210\u672c\u51fd\u6570\u7684\u6700\u5c0f\u65b9\u5dee\u5c31\u662f\uff1a<\/p>\n

            <\/p>\n

            3.\u6570\u636e<\/strong><\/p>\n

            \u7531\u4e8e\u6ca1\u6709\u623f\u4ef7(y_) \u548c\u623f\u5b50\u9762\u79ef (x) \u7684\u5b9e\u9645\u6570\u636e\u70b9\uff0c\u6211\u4eec\u5c31\u751f\u6210\u5b83\u4eec\u3002<\/p>\n

            <\/p>\n

            \u7b80\u5355\u8d77\u89c1\uff0c\u6211\u4eec\u5c06\u623f\u4ef7 (ys) \u8bbe\u7f6e\u6210\u6c38\u8fdc\u662f\u623f\u5b50\u9762\u79ef (xs) \u7684 2 \u500d\u3002<\/p>\n

            4.\u68af\u5ea6\u4e0b\u964d<\/strong><\/p>\n

            \u6709\u4e86\u7ebf\u6027\u6a21\u578b\u3001\u6210\u672c\u51fd\u6570\u548c\u6570\u636e\uff0c\u6211\u4eec\u5c31\u53ef\u4ee5\u5f00\u59cb\u6267\u884c\u68af\u5ea6\u4e0b\u964d\u4ece\u800c\u6700\u5c0f\u5316\u4ee3\u4ef7\u51fd\u6570\uff0c\u4ee5\u83b7\u5f97 W\u3001b \u7684\u300cgood\u300d\u503c\u3002<\/p>\n

            <\/p>\n

            0.00001 \u662f\u6211\u4eec\u6bcf\u6b21\u8fdb\u884c\u8bad\u7ec3\u65f6\u5728\u6700\u9661\u7684\u68af\u5ea6\u65b9\u5411\u4e0a\u6240\u91c7\u53d6\u7684\u300c\u6b65\u300d\u957f\uff1b\u5b83\u4e5f\u88ab\u79f0\u4f5c\u5b66\u4e60\u7387\uff08learning rate\uff09\u3002<\/p>\n

            \u6b65\u9aa4\u4e09\uff1a\u8bad\u7ec3\u6a21\u578b<\/strong><\/p>\n

            \u8bad\u7ec3\u5305\u542b\u4ee5\u9884\u5148\u786e\u5b9a\u597d\u7684\u6b21\u6570\u6267\u884c\u68af\u5ea6\u4e0b\u964d\uff0c\u6216\u8005\u662f\u76f4\u5230\u6210\u672c\u51fd\u6570\u4f4e\u4e8e\u67d0\u4e2a\u9884\u5148\u786e\u5b9a\u7684\u4e34\u754c\u503c\u4e3a\u6b62\u3002<\/p>\n

            1.TensorFlow \u7684\u602a\u5f02<\/strong><\/p>\n

            \u6240\u6709\u53d8\u91cf\u90fd\u9700\u8981\u5728\u8bad\u7ec3\u5f00\u59cb\u65f6\u8fdb\u884c\u521d\u59cb\u5316\uff0c\u5426\u5219\u5b83\u4eec\u53ef\u80fd\u4f1a\u5e26\u6709\u4e4b\u524d\u6267\u884c\u8fc7\u7a0b\u4e2d\u7684\u6b8b\u4f59\u503c\u3002<\/p>\n

            <\/p>\n

            2.TensorFlow \u4f1a\u8bdd<\/strong><\/p>\n

            \u867d\u7136 TensorFlow \u662f\u4e00\u4e2a Python \u5e93\uff0cPython \u662f\u4e00\u79cd\u89e3\u91ca\u6027\u7684\u8bed\u8a00\uff0c\u4f46\u662f\u9ed8\u8ba4\u60c5\u51b5\u4e0b\u4e0d\u628a TensorFlow \u8fd0\u7b97\u7528\u4f5c\u89e3\u91ca\u6027\u80fd\u7684\u539f\u56e0\uff0c\u56e0\u6b64\u4e0d\u6267\u884c\u4e0a\u9762\u7684 init \u3002\u76f8\u53cd TensorFlow \u662f\u5728\u4e00\u4e2a\u4f1a\u8bdd\u4e2d\u8fdb\u884c\uff1b\u521b\u5efa\u4e00\u4e2a\u4f1a\u8bdd (sess) \u7136\u540e\u4f7f\u7528 sess.run() \u53bb\u6267\u884c\u3002<\/p>\n

            <\/p>\n

            \u7c7b\u4f3c\u5730\u6211\u4eec\u5728\u4e00\u4e2a\u5faa\u73af\u4e2d\u8c03\u7528 withinsess.run() \u6765\u6267\u884c\u4e0a\u9762\u7684 train_step\u3002<\/p>\n

            <\/p>\n

            \u4f60\u9700\u8981\u5c06\u7531 x, y_ \u6240\u7ec4\u6210\u7684\u5b9e\u9645\u6570\u636e\u8f93\u5165\u518d\u63d0\u4f9b\u7ed9\u8f93\u5165\uff0c\u56e0\u4e3a TensorFlow \u5c06 train_step \u5206\u89e3\u4e3a\u5b83\u7684\u4ece\u5c5e\u9879\uff1a<\/p>\n

            <\/p>\n

            \u4ece\u5c5e\u9879\u7684\u5e95\u90e8\u662f\u5360\u4f4d\u7b26 x\uff0cy_\uff1b\u800c\u4e14\u6b63\u5982\u6211\u4eec\u4e4b\u524d\u63d0\u5230\u7684\uff0ctf.placeholders \u662f\u7528\u6765\u8868\u793a\u6240\u8981\u63d0\u4f9b\u7684\u5b9e\u9645\u6570\u636e\u70b9\u503c\u623f\u4ef7 (y_) \u548c\u623f\u5b50\u9762\u79ef \u00a0(x) \u7684\u4f4d\u7f6e\u3002<\/p>\n

            \u7ed3\u679c<\/strong><\/p>\n

            \u5faa\u73af\u4e2d\u7684 print \u8bed\u53e5\u5c06\u663e\u793a TensorFlow \u5982\u4f55\u5728\u6bcf\u6b21\u8fed\u4ee3\u4e2d\u5b66\u4e60 W \u548c b \u7684\u300cgood\u300d\u503c\u3002<\/p>\n

            <\/p>\n

            \u5c0f\u7ed3<\/strong><\/p>\n

            \u6211\u4eec\u5df2\u7ecf\u4ee5\u6700\u7b80\u5355\u7684\u5f62\u5f0f\u5b66\u4e60\u4e86\u673a\u5668\u5b66\u4e60\uff1b\u4ece\u4e00\u4e2a\u5355\u4e00\u7279\u5f81\u9884\u6d4b\u7ed3\u679c\u3002\uff08\u4e3a\u7b80\u5355\u8d77\u89c1\uff09\u6211\u4eec\u9009\u62e9\u4e86\u4e00\u4e2a\u7ebf\u6027\u6a21\u578b\u6765\u62df\u5408\u6211\u4eec\u7684\u6570\u636e\u70b9\uff0c\u5b9a\u4e49\u4e00\u4e2a\u6210\u672c\u51fd\u6570\u6765\u8868\u793a\u6700\u4f73\u62df\u5408\uff0c\u5e76\u901a\u8fc7\u53cd\u590d\u8c03\u6574\u5176\u68af\u5ea6\u53d8\u91cf W \u4e0e\u4f4d\u7f6e\u53d8\u91cf b \u6765\u8bad\u7ec3\u6211\u4eec\u7684\u6a21\u578b\uff0c\u4f7f\u6210\u672c\u51fd\u6570\u964d\u5230\u6700\u5c0f\u3002<\/p>\n

            \u7b2c\u4e8c\u90e8\u5206<\/strong><\/p>\n

            \u7b80\u5355\u56de\u987e<\/strong><\/p>\n

            \u5728\u4e0a\u4e00\u90e8\u5206\uff0c\u6211\u4eec\u4f7f\u7528 TensorFlow \u6784\u5efa\u5e76\u5b66\u4e60\u4e86\u4e00\u4e2a\u5e26\u6709\u5355\u4e00\u7279\u5f81\u7684\u7ebf\u6027\u56de\u5f52\u6a21\u578b\u2014\u2014\u7ed9\u5b9a\u4e00\u4e2a\u7279\u5f81\u503c\uff08\u623f\u5c4b\u9762\u79ef\/\u5e73\u65b9\u7c73\uff09\uff0c\u6211\u4eec\u53ef\u4ee5\u9884\u6d4b\u8f93\u51fa\uff08\u623f\u4ef7\/\u7f8e\u5143\uff09\u3002<\/p>\n

            \u4e0b\u9762\u662f\u4e00\u4e9b\u603b\u7ed3\uff1a<\/p>\n

              \n
            1. \u6211\u4eec\u6709\u4e00\u4e9b\u623f\u5c4b\u9762\u79ef\u548c\u623f\u4ef7\u7684\u6570\u636e\uff08\u7070\u8272\u70b9\uff09<\/li>\n
            2. \u6211\u4eec\u4f7f\u7528\u7ebf\u6027\u56de\u5f52\u5bf9\u8fd9\u4e9b\u6570\u636e\u8fdb\u884c\u4e86\u5efa\u6a21\uff08\u7ea2\u8272\u865a\u7ebf\uff09<\/li>\n
            3. \u6211\u4eec\u901a\u8fc7\u8bad\u7ec3\u8be5\u7ebf\u6027\u56de\u5f52\u6a21\u578b\u7684 W\uff08\u6743\u91cd\uff09\u548c b\uff08\u504f\u7f6e\uff09\u627e\u5230\u4e86\u6700\u5c0f\u5316\u300c\u6210\u672c\u300d\uff08\u7ad6\u76f4\u84dd\u8272\u5b9e\u7ebf\u7684\u957f\u5ea6\u603b\u548c\uff0c\u8fd9\u4e9b\u84dd\u7ebf\u4ee3\u8868\u4e86\u9884\u6d4b\u548c\u5b9e\u9645\u8f93\u51fa\u4e4b\u95f4\u7684\u5dee\u5f02\uff09\u7684\u300c\u6700\u597d\u300d\u6a21\u578b<\/li>\n
            4. \u7ed9\u5b9a\u4efb\u610f\u623f\u5c4b\u9762\u79ef\uff0c\u6211\u4eec\u53ef\u4ee5\u4f7f\u7528\u8be5\u7ebf\u6027\u6a21\u578b\u9884\u6d4b\u623f\u4ef7\uff08\u5e26\u7bad\u5934\u7684\u84dd\u8272\u865a\u7ebf\uff09<\/li>\n<\/ol>\n

              <\/p>\n

              \u4e00\u5f20\u56fe\u89e3\u91ca\u7ebf\u6027\u56de\u5f52<\/em><\/p>\n

              \u5728\u673a\u5668\u5b66\u4e60\u6587\u732e\u4e2d\uff0c\u6211\u4eec\u5e38\u5e38\u770b\u5230\u300c\u8bad\u7ec3\uff08training\uff09\u300d\u8fd9\u4e2a\u8bcd\u3002\u5728\u8fd9\u4e00\u90e8\u5206\uff0c\u6211\u4eec\u5c06\u5728 TensorFlow \u4e2d\u7406\u89e3\u300c\u8bad\u7ec3\u300d\u7684\u542b\u4e49\u3002<\/p>\n

              \u7ebf\u6027\u56de\u5f52\u5efa\u6a21<\/strong><\/p>\n

              Linear Model (in TF notation): y = tf.matmul(x,W) + b<\/strong><\/em><\/p><\/blockquote>\n

              \u7ebf\u6027\u56de\u5f52\u7684\u76ee\u6807\u662f\u5bfb\u627e W \u548c b\uff0c\u8fd9\u6837\u5bf9\u4e8e\u7ed9\u5b9a\u7684\u4efb\u610f\u7279\u5f81\u503c x\uff0c\u6211\u4eec\u53ef\u4ee5\u901a\u8fc7\u5c06 W\u3001b \u548c x \u7684\u503c\u4ee3\u5165\u5230\u6a21\u578b\u4e2d\u5f97\u5230\u9884\u6d4b y\u3002<\/p>\n

              \u4f46\u662f\u4e3a\u4e86\u627e\u5230\u80fd\u51c6\u786e\u505a\u51fa\u9884\u6d4b\u7684 W \u548c b \u7684\u503c\uff0c\u6211\u4eec\u9700\u8981\u4f7f\u7528\u53ef\u7528\u7684\u6570\u636e\uff08\u8bb8\u591a\u5b9e\u9645\u7279\u5f81 x \u548c\u5b9e\u9645\u8f93\u51fa y_ \u7684\u914d\u5bf9\uff0c\u6ce8\u610f\u4e0b\u5212\u7ebf\uff09\u6765\u300c\u8bad\u7ec3\u300d\u8be5\u6a21\u578b\u3002<\/p>\n

              \u89e3\u91ca\u300c\u8bad\u7ec3\u300d<\/strong><\/p>\n

              \u4e3a\u4e86\u627e\u5230\u6700\u4f73\u7684 W \u548c b \u503c\uff0c\u6211\u4eec\u53ef\u4ee5\u4ece\u4efb\u610f\u7684 W \u548c b \u503c\u5f00\u59cb\u3002\u6211\u4eec\u4e5f\u9700\u8981\u5b9a\u4e49\u4e00\u4e2a\u6210\u672c\u51fd\u6570\uff0c\u8be5\u51fd\u6570\u53ef\u4ee5\u8861\u91cf\u5bf9\u4e8e\u4e00\u4e2a\u7ed9\u5b9a\u7279\u5f81\u503c x \u9884\u6d4b\u8f93\u51fa y \u548c\u5b9e\u9645\u8f93\u51fa y_ \u4e4b\u95f4\u5dee\u5f02\u3002\u4e3a\u4e86\u7b80\u5355\u8d77\u89c1\uff0c\u6211\u4eec\u4f7f\u7528\u6700\u7b80\u5355\u7684\u6700\u5c0f\u5747\u65b9\u8bef\u5dee\uff08MSE\uff1aminimum squared error\uff09\u4f5c\u4e3a\u6211\u4eec\u7684\u6210\u672c\u51fd\u6570\u3002<\/p>\n

              Cost function (in TF notation): tf.reduce_mean(tf.square(y_ – y))<\/strong><\/em><\/p><\/blockquote>\n

              \u901a\u8fc7\u6700\u5c0f\u5316\u6210\u672c\u51fd\u6570\uff0c\u6211\u4eec\u53ef\u4ee5\u5f97\u5230\u5f88\u597d\u7684 W \u548c b \u503c\u3002<\/p>\n

              \u6211\u4eec\u7684\u8bad\u7ec3\u4ee3\u7801\u5b9e\u9645\u4e0a\u975e\u5e38\u7b80\u5355\uff0c\u5e76\u4e14\u7528 [A, B, C, D] \u8fdb\u884c\u4e86\u6ce8\u91ca\uff0c\u540e\u9762\u6211\u4eec\u8fd8\u4f1a\u8c08\u5230\u8fd9\u4e9b\u4ee3\u7801\u3002\u5b8c\u6574\u4ee3\u7801\u8bf7\u8bbf\u95ee\uff1ahttps:\/\/github.com\/nethsix\/gentle_tensorflow\/blob\/master\/code\/linear_regression_one_feature_using_mini_batch_with_tensorboard.py<\/p>\n

              # … (\u7701\u7565) \u53d8\u91cf\/\u5e38\u91cf\u58f0\u660e …<\/strong><\/em><\/p>\n

              # [A] TensorFlow\u56fe<\/strong><\/em><\/p>\n

              y = tf.matmul(x,W) + b<\/strong><\/em><\/p>\n

              cost = tf.reduce_mean(tf.square(y_-y))<\/strong><\/em><\/p>\n

              # [B] \u7528\u56fa\u5b9a\u300c\u5b66\u4e60\u7387\uff08learn_rate\uff09\u300d\u8bad\u7ec3<\/strong><\/em><\/p>\n

              learn_rate = 0.1<\/strong><\/em><\/p>\n

              train_step =<\/strong><\/em><\/p>\n

              \u00a0 \u00a0 tf.train.GradientDescentOptimizer(learn_rate).minimize(cost)<\/strong><\/em><\/p>\n

              for i in range(steps):<\/strong><\/em><\/p>\n

              \u00a0 # [C] \u51c6\u5907\u6570\u636e\u70b9<\/strong><\/em><\/p>\n

              \u00a0 # … (\u7701\u7565) \u51c6\u5907\u4f5c\u4e3ax\u548cy\u7684\u6570\u636e\u70b9\u7684\u4ee3\u7801 …<\/strong><\/em><\/p>\n

              \u00a0 # [D] \u5728\u6bcf\u4e2a\u6b65\u9aa4\/epoch\u5c06\u6570\u636e\u9001\u5165’train_step’<\/strong><\/em><\/p>\n

              \u00a0 feed = { x: xs, y_: ys }<\/strong><\/em><\/p>\n

              \u00a0 sess.run(train_step, feed_dict=feed)<\/strong><\/em><\/p><\/blockquote>\n

              \u6211\u4eec\u7684\u7ebf\u6027\u6a21\u578b\u548c\u6210\u672c\u51fd\u6570[A]\u53ef\u4ee5\u8868\u793a\u6210\u4e0b\u9762\u7684 TensorFlow \u56fe\uff1a<\/p>\n

              <\/p>\n

              \u521b\u9020\u4e00\u4e2a\u5e26\u6709\u6a21\u578b\u548c\u6210\u672c\u51fd\u6570\u7684 TensorFlow \u56fe\uff0c\u5e76\u4f7f\u7528\u4e00\u4e9b\u503c\u521d\u59cb\u5316 W \u548c b<\/p>\n

              \u63a5\u4e0b\u6765\uff0c\u6211\u4eec\u9009\u62e9\u4e00\u4e2a\u6570\u636e\u70b9 (x, y_) [C]\uff0c\u7136\u540e\u5c06\u5176\u9001\u5165[D] TensorFlow \u56fe\uff0c\u4ece\u800c\u5f97\u5230\u9884\u6d4b y \u548c\u76f8\u5e94\u7684\u6210\u672c\u3002<\/p>\n

              <\/p>\n

              \u4f7f\u7528\u5355\u4e2a\u6570\u636e\u70b9\u8ba1\u7b97\u9884\u6d4b y \u548c\u6210\u672c<\/em><\/p>\n

              \u4e3a\u4e86\u5f97\u5230\u66f4\u597d\u7684 W \u548c b\uff0c\u6211\u4eec\u4f7f\u7528TensorFlow \u7684 tf.train.GradientDescentOptimizer [B]\u6267\u884c\u68af\u5ea6\u4e0b\u964d\u4ee5\u964d\u4f4e\u6210\u672c\u3002\u7528\u975e\u6280\u672f\u7684\u672f\u8bed\u6765\u8bf4\uff1a\u7ed9\u5b9a\u5f53\u524d\u6210\u672c\uff0c\u5e76\u57fa\u4e8e\u6210\u672c\u5c81\u5176\u5b83\u53d8\u91cf\uff08\u5373 W \u548c b\uff09\u7684\u53d8\u5316\u65b9\u5f0f\uff0c\u4f18\u5316\u5668\uff08optimizer\uff09\u5c06\u5bf9 W \u548c b \u6267\u884c\u4e00\u4e9b\u5c0f\u8c03\u6574\uff08\u9012\u589e\u6216\u9012\u51cf\uff09\u4ee5\u4f7f\u6211\u4eec\u7684\u9884\u6d4b\u66f4\u597d\u5730\u5951\u5408\u90a3\u4e2a\u5355\u4e2a\u6570\u636e\u70b9\u3002<\/p>\n

              <\/p>\n

              \u57fa\u4e8e\u5f53\u524d\u7684\u6210\u672c\uff0c\u51b3\u5b9a\u5982\u4f55\u8c03\u6574 W \u548c b \u4ee5\u63d0\u5347\u9884\u6d4b y \u548c\u964d\u4f4e\u6210\u672c<\/em><\/p>\n

              \u8bad\u7ec3\u5468\u671f\u4e2d\u7684\u6700\u540e\u6b65\u9aa4\u662f\u5728\u8c03\u6574 W \u548c b \u5bf9\u5b83\u4eec\u8fdb\u884c\u66f4\u65b0\u3002\u6ce8\u610f\u8fd9\u91cc\u7684\u300c\u5468\u671f\u300d\u7528\u673a\u5668\u5b66\u4e60\u7684\u672f\u8bed\u6765\u8bf4\u662f\u300cepoch\u300d\u3002<\/p>\n

              <\/p>\n

              \u5728\u4e0b\u4e00\u8bad\u7ec3 epoch \u7684\u8fed\u4ee3\u524d\uff0c\u901a\u8fc7\u8c03\u6574 W \u548c b \u5bf9\u5b83\u4eec\u8fdb\u884c\u66f4\u65b0<\/em><\/p>\n

              \u5728\u4e0b\u4e00\u8bad\u7ec3 epoch \u4e2d\uff0c\u91cd\u590d\u8fd9\u4e9b\u6b65\u9aa4\uff0c\u4f46\u4f7f\u7528\u4e00\u4e2a\u4e0d\u540c\u7684\u6570\u636e\u70b9\uff01<\/p>\n

              <\/p>\n

              \u4f7f\u7528\u4e0d\u540c\u7684\u6570\u636e\u70b9\u8fdb\u884c\u8bad\u7ec3<\/em><\/p>\n

              \u4f7f\u7528\u5404\u79cd\u6570\u636e\u70b9\u6cdb\u5316\uff08generalize\uff09\u6211\u4eec\u7684\u6a21\u578b\uff0c\u5373\u5b66\u4e60\u53ef\u88ab\u7528\u4e8e\u9884\u6d4b\u4efb\u4f55\u7279\u5f81\u503c\u7684 W \u548c b \u503c\u3002\u6ce8\u610f\uff1a<\/p>\n

                \n
              • \u5728\u5927\u90e8\u5206\u60c5\u51b5\u4e0b\uff0c\u6570\u636e\u70b9\u8d8a\u591a\uff0c\u6a21\u578b\u7684\u5b66\u4e60\u548c\u6cdb\u5316\u5c31\u8d8a\u597d<\/li>\n
              • \u5982\u679c\u4f60\u8bad\u7ec3\u7684 epoch \u6bd4\u6570\u636e\u70b9\u8fd8\u591a\uff0c\u4f60\u53ef\u4ee5\u91cd\u590d\u4f7f\u7528\u6570\u636e\u70b9\uff0c\u8fd9\u4e0d\u6210\u95ee\u9898\u3002\u68af\u5ea6\u4e0b\u964d\u4f18\u5316\u603b\u662f\u4f1a\u540c\u65f6\u4f7f\u7528\u6570\u636e\u70b9\u53ca\u5176\u6210\u672c\uff08\u6839\u636e\u8be5 epoch \u7684 W \u548c b \u503c\u4ece\u6570\u636e\u70b9\u4e2d\u8ba1\u7b97\u5f97\u5230\uff09\u6765\u5bf9 W \u548c b \u503c\u8fdb\u884c\u8c03\u6574\uff1b\u8be5\u4f18\u5316\u5668\u4e5f\u8bb8\u4e4b\u524d\u5df2\u7ecf\u89c1\u8fc7\u4e86\u8fd9\u4e2a\u6570\u636e\u70b9\uff0c\u4f46\u6210\u672c\u5e76\u4e0d\u4e00\u6837\uff0c\u56e0\u6b64\u5b83\u8fd8\u662f\u53ef\u4ee5\u5b66\u5230\u65b0\u7684\u4e1c\u897f\uff0c\u5e76\u4ee5\u4e0d\u540c\u7684\u65b9\u5f0f\u8c03\u6574 W \u548c b \u503c\u3002<\/li>\n<\/ul>\n

                \u4f60\u53ef\u4ee5\u7528\u56fa\u5b9a\u6570\u91cf\u7684 epoch \u8bad\u7ec3\u4e00\u4e2a\u6a21\u578b\uff0c\u76f4\u5230\u5176\u8fbe\u5230\u4ee4\u4eba\u6ee1\u610f\u7684\u6210\u672c\u9608\u503c\u3002<\/p>\n

                \u8bad\u7ec3\u53d8\u91cf<\/strong><\/p>\n

                1.\u968f\u673a\u3001mini-batch\u3001batch<\/strong><\/p>\n

                \u5728\u4e0a\u9762\u7684\u8bad\u7ec3\u4e2d\uff0c\u6211\u4eec\u5728\u6bcf\u4e2a epoch \u9001\u5165\u5355\u4e2a\u6570\u636e\u70b9\u3002\u8fd9\u88ab\u79f0\u4e3a\u968f\u673a\u68af\u5ea6\u4e0b\u964d\uff08stochastic gradient descent\uff09\u3002\u6211\u4eec\u4e5f\u53ef\u4ee5\u5728\u6bcf\u4e2a epoch \u9001\u5165\u4e00\u5806\u6570\u636e\u70b9\uff0c\u8fd9\u88ab\u79f0\u4e3a mini-batch \u68af\u5ea6\u4e0b\u964d\uff0c\u6216\u8005\u751a\u81f3\u5728\u4e00\u4e2a epoch \u4e00\u6b21\u6027\u9001\u5165\u6240\u6709\u7684\u6570\u636e\u70b9\uff0c\u8fd9\u88ab\u79f0\u4e3a batch \u68af\u5ea6\u4e0b\u964d\u3002\u8bf7\u770b\u4e0b\u56fe\u7684\u6bd4\u8f83\uff0c\u6ce8\u610f\u8fd9 3 \u5f20\u56fe\u7684 2 \u5904\u4e0d\u540c\uff1a<\/p>\n

                  \n
                • \u6bcf\u4e2a epoch \u9001\u5165 TensorFlow \u56fe\uff08TF.Graph\uff09\u7684\u6570\u636e\u70b9\u7684\u6570\u91cf\uff08\u56fe\u53f3\u4e0a\u65b9\uff09<\/li>\n
                • \u68af\u5ea6\u4e0b\u964d\u4f18\u5316\u5728\u8c03\u6574 W \u548c b \u503c\u65f6\u6240\u8003\u8651\u7684\u6570\u636e\u70b9\u7684\u6570\u91cf\uff08\u56fe\u53f3\u4e0b\u65b9\uff09<\/li>\n<\/ul>\n

                  <\/p>\n

                  \u968f\u673a\u68af\u5ea6\u4e0b\u964d<\/em><\/p>\n

                  <\/p>\n

                  mini-batch \u68af\u5ea6\u4e0b\u964d<\/em><\/p>\n

                  <\/p>\n

                  batch \u68af\u5ea6\u4e0b\u964d<\/em><\/p>\n

                  \u6bcf\u5f20\u56fe\u4e2d\u7684\u6570\u636e\u70b9\u7684\u6570\u91cf\u6709 2 \u4e2a\u542b\u4e49\u3002\u5f53\u6570\u636e\u70b9\u66f4\u591a\u65f6\uff1a<\/p>\n

                    \n
                  • \u8ba1\u7b97\u6210\u672c\u548c\u6267\u884c\u68af\u5ea6\u4e0b\u964d\u6240\u9700\u7684\u8ba1\u7b97\u8d44\u6e90\uff08\u51cf\u6cd5\u3001\u5e73\u65b9\u3001\u52a0\u6cd5\uff09\u4f1a\u589e\u52a0<\/li>\n
                  • \u6a21\u578b\u7684\u5b66\u4e60\u548c\u6cdb\u5316\u7684\u901f\u5ea6\u589e\u52a0<\/li>\n<\/ul>\n

                    \u9009\u62e9\u968f\u673a\u3001mini-batch\u3001batch \u68af\u5ea6\u4e0b\u964d\u7684\u4f18\u7f3a\u70b9\u603b\u7ed3\u5728\u4e0b\u56fe\u4e2d\uff1a<\/p>\n

                    <\/p>\n

                    \u9009\u62e9\u968f\u673a\u3001mini-batch\u3001batch \u68af\u5ea6\u4e0b\u964d\u7684\u4f18\u7f3a\u70b9<\/em><\/p>\n

                    \u8981\u5728\u968f\u673a\/mini-batch\/batch \u68af\u5ea6\u4e0b\u964d\u4e4b\u95f4\u5207\u6362\uff0c\u6211\u4eec\u53ea\u9700\u8981\u5728\u5c06\u6570\u636e\u70b9\u9001\u5165\u8bad\u7ec3\u6b65\u9aa4[D]\u4e4b\u524d\u5c06\u8fd9\u4e9b\u6570\u636e\u70b9\u5206\u6210\u4e0d\u540c\u7684 batch \u5927\u5c0f\uff0c\u5373\u4e3a [C] \u4f7f\u7528\u5982\u4e0b\u7684\u4ee3\u7801\u7247\u6bb5\uff1a<\/p>\n

                    # * all_xs: \u6240\u6709\u7684\u7279\u5f81\u503c<\/em><\/strong><\/p>\n

                    # * all_ys: \u6240\u6709\u7684\u8f93\u51fa\u503c<\/em><\/strong><\/p>\n

                    # datapoint_size: all_xs\/all_ys \u4e2d\u70b9\/\u9879\u7684\u6570\u91cf<\/em><\/strong><\/p>\n

                    # batch_size: \u914d\u7f6e\u5982\u4e0b:<\/em><\/strong><\/p>\n

                    # \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 1: \u968f\u673a\u6a21\u578b<\/em><\/strong><\/p>\n

                    # \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 integer < datapoint_size: mini-batch\u6a21\u5f0f<\/em><\/strong><\/p>\n

                    # \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 datapoint_size: batch\u6a21\u5f0f<\/em><\/strong><\/p>\n

                    # i: \u5f53\u524depoch\u6570\u91cf<\/em><\/strong><\/p>\n

                    if datapoint_size == batch_size:<\/em><\/strong><\/p>\n

                    \u00a0 # Batch \u6a21\u5f0f\uff0c\u6240\u4ee5\u9009\u62e9\u6240\u6709\u6570\u636e\u70b9\u4ece index 0 \u5f00\u59cb<\/em><\/strong><\/p>\n

                    \u00a0 batch_start_idx = 0<\/em><\/strong><\/p>\n

                    elif datapoint_size < batch_size:<\/em><\/strong><\/p>\n

                    \u00a0 # \u4e0d\u53ef\u80fd<\/em><\/strong><\/p>\n

                    \u00a0 raise ValueError(\u201cdatapoint_size: %d, must be greater than \u00a0 \u00a0 \u00a0 \u00a0\u00a0<\/em><\/strong><\/p>\n

                    \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 batch_size: %d\u201d % (datapoint_size, batch_size))<\/em><\/strong><\/p>\n

                    else:<\/em><\/strong><\/p>\n

                    \u00a0 # \u968f\u673a\/mini-batch\u6a21\u5f0f: \u4ece\u6240\u6709\u53ef\u80fd\u7684\u6570\u636e\u70b9\u4e2d\u5206\u6279\u9009\u62e9\u6570\u636e\u70b9<\/em><\/strong><\/p>\n

                    \u00a0 batch_start_idx = (i * batch_size) % (datapoint_size \u2014 batch_size)<\/em><\/strong><\/p>\n

                    \u00a0 batch_end_idx = batch_start_idx + batch_size<\/em><\/strong><\/p>\n

                    \u00a0 batch_xs = all_xs[batch_start_idx:batch_end_idx]<\/em><\/strong><\/p>\n

                    \u00a0 batch_ys = all_ys[batch_start_idx:batch_end_idx]<\/em><\/strong><\/p>\n

                    # \u5c06\u5206\u6279\u7684\u6570\u636e\u70b9\u5b9a\u4e49\u4e3axs, ys, \u5b83\u4eec\u4f1a\u88ab\u9001\u5165 ‘train_step’\u8bad\u7ec3\u6b65\u9aa4<\/em><\/strong><\/p>\n

                    xs = np.array(batch_xs)<\/em><\/strong><\/p>\n

                    ys = np.array(batch_ys)<\/em><\/strong><\/p><\/blockquote>\n

                    2.\u5b66\u4e60\u7387\u53d8\u5316<\/strong><\/p>\n

                    \u5b66\u4e60\u7387\uff08learn rate\uff09\u662f\u6307\u68af\u5ea6\u4e0b\u964d\u8c03\u6574 W \u548c b \u9012\u589e\u6216\u9012\u51cf\u7684\u901f\u5ea6\u3002\u5b66\u4e60\u7387\u8f83\u5c0f\u65f6\uff0c\u5904\u7406\u8fc7\u7a0b\u4f1a\u66f4\u6162\uff0c\u4f46\u80af\u5b9a\u80fd\u5f97\u5230\u66f4\u5c0f\u6210\u672c\uff1b\u800c\u5f53\u5b66\u4e60\u7387\u66f4\u5927\u65f6\uff0c\u6211\u4eec\u53ef\u4ee5\u66f4\u5feb\u5730\u5f97\u5230\u6700\u5c0f\u6210\u672c\uff0c\u4f46\u6709\u300c\u51b2\u8fc7\u5934\u300d\u7684\u98ce\u9669\uff0c\u5bfc\u81f4\u6211\u4eec\u6ca1\u6cd5\u627e\u5230\u6700\u5c0f\u6210\u672c\u3002<\/p>\n

                    \u4e3a\u4e86\u514b\u670d\u8fd9\u4e00\u95ee\u9898\uff0c\u8bb8\u591a\u673a\u5668\u5b66\u4e60\u5b9e\u8df5\u8005\u9009\u62e9\u5f00\u59cb\u65f6\u4f7f\u7528\u8f83\u5927\u7684\u5b66\u4e60\u7387\uff08\u5047\u8bbe\u5f00\u59cb\u65f6\u7684\u6210\u672c\u79bb\u6700\u5c0f\u6210\u672c\u8fd8\u5f88\u8fdc\uff09\uff0c\u7136\u540e\u968f\u6bcf\u4e2a epoch \u800c\u9010\u6e10\u964d\u4f4e\u5b66\u4e60\u7387\u3002<\/p>\n

                    TensorFlow \u63d0\u4f9b\u4e86 2 \u79cd\u65b9\u6cd5\u53ef\u4ee5\u505a\u5230\u8fd9\u4e00\u70b9\uff0c\u8be6\u7ec6\u89e3\u91ca\u53ef\u53c2\u8003\uff1ahttp:\/\/stackoverflow.com\/questions\/33919948\/how-to-set-adaptive-learning-rate-for-gradientdescentoptimizer\uff1b\u4f46\u8fd9\u91cc\u8fdb\u884c\u4e86\u603b\u7ed3\u3002<\/p>\n

                      \n
                    • \u4f7f\u7528\u68af\u5ea6\u4e0b\u964d\u4f18\u5316\u7684\u53d8\u4f53<\/li>\n<\/ul>\n

                      TensorFlow \u5e26\u6709\u591a\u79cd\u652f\u6301\u5b66\u4e60\u7387\u53d8\u5316\u7684\u68af\u5ea6\u4e0b\u964d\u4f18\u5316\u5668\uff0c\u4f8b\u5982 \u00a0tf.train.AdagradientOptimizer \u548c tf.train.AdamOptimizer.<\/p>\n

                        \n
                      • \u4f7f\u7528 tf.placeholder \u8c03\u6574\u5b66\u4e60\u7387<\/li>\n<\/ul>\n

                        \u5982\u540c\u524d\u9762\u6240\u770b\u5230\u7684\uff0c\u5982\u679c\u6211\u4eec\u5728\u8fd9\u4e2a\u4f8b\u5b50\u4e2d\u58f0\u660e\u4e86 tf.placeholder \u6765\u8bbe\u7f6e\u5b66\u4e60\u7387\uff0c\u7136\u540e\u5728 tf.train.GradientDescentOptimizer \u4e2d\u4f7f\u7528\u5b83\uff0c\u6211\u4eec\u53ef\u4ee5\u5728\u6bcf\u4e2a\u8bad\u7ec3 epoch \u5411\u5176\u9001\u5165\u4e00\u4e2a\u4e0d\u540c\u7684\u503c\uff0c\u8fd9\u5f88\u50cf\u6211\u4eec\u7ed9 x \u548c y_ \u9001\u5165\u4e0d\u540c\u7684\u6570\u636e\u70b9\uff0c\u8fd9\u4e5f\u662f\u6bcf\u4e2a epoch \u7684 tf.placeholders.<\/p>\n

                        \u6211\u4eec\u9700\u8981 2 \u4e2a\u5c0f\u4fee\u6539\uff1a<\/p>\n

                        # \u4fee\u6539 [B] \uff0c\u5c06 ‘learn_rate’ \u8bbe\u7f6e\u4e3a’tf.placeholder’<\/strong><\/p>\n

                        # \u5e76\u5c06\u5176\u63d0\u4f9b\u7ed9’learning_rate’\u53c2\u6570\u540dtf.train.GradientDescentOptimizer<\/strong><\/p>\n

                        learn_rate = tf.placeholder(tf.float32, shape=[])<\/strong><\/p>\n

                        train_step = tf.train.GradientDescentOptimizer(<\/strong><\/p>\n

                        \u00a0 \u00a0 learning_rate=learn_rate).minimize(cost)<\/strong><\/p>\n

                        # \u4fee\u6539[D]\uff0c\u5305\u542b\u9001\u5165\u4e00\u4e2a’learn_rate’\u503c,<\/strong><\/p>\n

                        # \u5373 ‘initial_learn_rate’\uff08\u521d\u59cb\u5b66\u4e60\u7387\uff09\u9664\u4ee5’i’ (\u5f53\u524depoch\u6570)<\/strong><\/p>\n

                        # \u6ce8: \u8fd9\u662f\u8fc7\u4e8e\u7b80\u5316\u7684\uff0c\u4ec5\u7528\u4f5c\u793a\u4f8b<\/strong><\/p>\n

                        feed = { x: xs, y_: ys, learn_rate: initial_learn_rate\/i }<\/strong><\/p>\n

                        sess.run(train_step, feed_dict=feed)<\/strong><\/p><\/blockquote>\n

                        \u5c0f\u7ed3<\/strong><\/p>\n

                        \u6211\u4eec\u89e3\u91ca\u4e86\u673a\u5668\u5b66\u4e60\u4e2d\u300c\u8bad\u7ec3\uff08training\uff09\u300d\u7684\u542b\u4e49\uff0c\u4ee5\u53ca\u5728 TensorFlow \u4e2d\u901a\u8fc7\u6a21\u578b\u548c\u6210\u672c\u5b9a\u4e49\u3001\u7136\u540e\u5faa\u73af\u901a\u8fc7\u8bad\u7ec3\u6b65\u9aa4\uff08\u5c06\u6570\u636e\u70b9\u9001\u5165\u68af\u5ea6\u4e0b\u964d\u4f18\u5316\u5668\uff09\u6765\u8fdb\u884c\u8bad\u7ec3\u7684\u65b9\u5f0f\u3002\u6211\u4eec\u8fd8\u8ba8\u8bba\u4e86\u8bad\u7ec3\u4e2d\u7684\u5e38\u89c1\u53d8\u91cf\uff0c\u5373\u6539\u53d8\u6a21\u578b\u5b66\u4e60\u65f6\u6bcf\u4e2a epoch \u6240\u7528\u7684\u6570\u636e\u70b9\u7684\u5927\u5c0f\u548c\u6539\u53d8\u68af\u5ea6\u4e0b\u964d\u4f18\u5316\u5668\u7684\u5b66\u4e60\u7387\u3002<\/p>\n

                        \u540e\u7eed\u5185\u5bb9<\/strong><\/p>\n

                          \n
                        • \u521b\u5efa Tensor Board \u6765\u53ef\u89c6\u5316 Tensorflow \u7684\u6267\u884c\uff0c\u4ece\u800c\u68c0\u6d4b\u6211\u4eec\u7684\u6a21\u578b\u3001\u6210\u672c\u51fd\u6570\u6216\u68af\u5ea6\u4e0b\u964d\u4e2d\u7684\u95ee\u9898<\/li>\n
                        • \u4f7f\u7528\u591a\u4e2a\u7279\u5f81\u8868\u8fbe\u7ebf\u6027\u56de\u5f52<\/li>\n<\/ul>\n

                          \u00a9\u672c\u6587\u7531\u673a\u5668\u4e4b\u5fc3\u7f16\u8bd1\uff0c\u8f6c\u8f7d\u8bf7\u8054\u7cfb\u672c\u516c\u4f17\u53f7\u83b7\u5f97\u6388\u6743\u3002<\/em><\/strong><\/p>\n<\/div>\n","protected":false},"excerpt":{"rendered":"

                          \u6df1\u5ea6 | \u673a\u5668\u5b66\u4e60\u6572\u95e8\u7816\uff1a\u4efb\u4f55\u4eba\u90fd\u80fd\u770b\u61c2<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"colormag_page_container_layout":"default_layout","colormag_page_sidebar_layout":"default_layout","footnotes":""},"categories":[24],"tags":[121],"class_list":["post-1128","post","type-post","status-publish","format-standard","hentry","category-value_docs","tag-05-"],"_links":{"self":[{"href":"https:\/\/i007.cc\/wordpress\/wp-json\/wp\/v2\/posts\/1128","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/i007.cc\/wordpress\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/i007.cc\/wordpress\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/i007.cc\/wordpress\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/i007.cc\/wordpress\/wp-json\/wp\/v2\/comments?post=1128"}],"version-history":[{"count":0,"href":"https:\/\/i007.cc\/wordpress\/wp-json\/wp\/v2\/posts\/1128\/revisions"}],"wp:attachment":[{"href":"https:\/\/i007.cc\/wordpress\/wp-json\/wp\/v2\/media?parent=1128"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/i007.cc\/wordpress\/wp-json\/wp\/v2\/categories?post=1128"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/i007.cc\/wordpress\/wp-json\/wp\/v2\/tags?post=1128"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}