{"id":3098,"date":"2020-02-29T18:21:05","date_gmt":"2020-02-29T10:21:05","guid":{"rendered":"http:\/\/www.sniper97.cn\/?p=3098"},"modified":"2020-02-29T18:21:05","modified_gmt":"2020-02-29T10:21:05","slug":"%e3%80%90%e5%90%b4%e6%81%a9%e8%be%be%e6%b7%b1%e5%ba%a6%e5%ad%a6%e4%b9%a0%e3%80%91%e8%b6%85%e5%8f%82%e6%95%b0%e8%b0%83%e8%af%95%e3%80%81batch%e6%ad%a3%e5%88%99%e5%8c%96%e5%92%8c%e7%a8%8b%e5%ba%8f","status":"publish","type":"post","link":"http:\/\/www.sniper97.cn\/index.php\/note\/deep-learning\/3098\/","title":{"rendered":"\u3010\u5434\u6069\u8fbe\u6df1\u5ea6\u5b66\u4e60\u3011\u8d85\u53c2\u6570\u8c03\u8bd5\u3001Batch\u6b63\u5219\u5316\u548c\u7a0b\u5e8f\u6846\u67b6"},"content":{"rendered":"\n

\u5434\u6069\u8fbe\u6df1\u5ea6\u5b66\u4e60\u7b2c\u4e8c\u8bfe\u7b2c\u4e09\u5468 \u8d85\u53c2\u6570\u8c03\u8bd5\u3001Batch\u6b63\u5219\u5316\u548c\u7a0b\u5e8f\u6846\u67b6 <\/p>\n\n\n

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

1.\u8c03\u8bd5\u5904\u7406<\/h2>\n\n\n

\u5173\u4e8e\u8bad\u7ec3\u6df1\u5ea6\u6700\u96be\u7684\u4e8b\u60c5\u4e4b\u4e00\u662f\u4f60\u8981\u5904\u7406\u7684\u53c2\u6570\u7684\u6570\u91cf\uff0c\u4ece\u5b66\u4e60\u901f\u7387alpha\u5230Momentum<\/strong>\uff08\u52a8\u91cf\u68af\u5ea6\u4e0b\u964d\u6cd5\uff09\u7684\u53c2\u6570theta\u3002 <\/p>\n\n\n

\u5982\u679c\u4f7f\u7528Momentum<\/strong>\u6216Adam<\/strong>\u4f18\u5316\u7b97\u6cd5\u7684\u53c2\u6570theta1\uff0ctheta\u548cEpsilon\uff0c\u4e5f\u8bb8\u4f60\u8fd8\u5f97\u9009\u62e9\u5c42\u6570\uff0c\u4e5f\u8bb8\u4f60\u8fd8\u5f97\u9009\u62e9\u4e0d\u540c\u5c42\u4e2d\u9690\u85cf\u5355\u5143\u7684\u6570\u91cf\uff0c\u4e5f\u8bb8\u4f60\u8fd8\u60f3\u4f7f\u7528\u5b66\u4e60\u7387\u8870\u51cf\u3002<\/p>\n\n\n

\u6240\u4ee5\uff0c\u4f60\u4f7f\u7528\u7684\u4e0d\u662f\u5355\u4e00\u7684\u5b66\u4e60\u7387alpha\uff0c\u5f53\u7136\u4f60\u53ef\u80fd\u8fd8\u9700\u8981\u9009\u62e9mini-batch<\/strong>\u7684\u5927\u5c0f\u3002 <\/p>\n\n\n

\u7ed3\u679c\u8bc1\u5b9e\u4e00\u4e9b\u8d85\u53c2\u6570\u6bd4\u5176\u5b83\u7684\u66f4\u4e3a\u91cd\u8981\uff0c\u5b66\u4e60\u901f\u7387\u5c31\u662f\u9700\u8981\u8c03\u8bd5\u7684\u6700\u91cd\u8981\u7684\u8d85\u53c2\u6570\u3002 <\/p>\n\n\n

\"\"
\u53f3\u4e0a\u89d2\u4e3a\u91cd\u8981\u7a0b\u5ea6\uff08\u4ece\u9ad8\u5230\u4f4e\uff09<\/figcaption><\/figure><\/div>\n\n\n

\u6211\u4eec\u5728\u8fdb\u884c\u8d85\u53c2\u8c03\u6574\u7684\u65f6\u5019\uff0c\u5728\u51e0\u4e2a\u51e0\u4f55\u8303\u56f4\u5185\u8fdb\u884c\u5c1d\u8bd5\uff0c\u5982\u679c\u627e\u5230\u4e00\u4e2a\u8f83\u597d\u7684\u8303\u56f4\uff0c\u5219\u5728\u8fd9\u4e2a\u8303\u56f4\u7ee7\u7eed\u7ec6\u5206\u5c1d\u8bd5 <\/p>\n\n\n

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

2. \u8d85\u53c2\u6570\u9009\u62e9\u5408\u9002\u7684\u8303\u56f4 <\/h2>\n\n\n

\u8fd9\u4e00\u8282\u8bb2\u600e\u6837\u9009\u62e9\u5408\u9002\u7684\u6807\u5c3a\u6765\u5bfb\u627e\u6700\u4f18\u8d85\u53c2 <\/p>\n\n\n

\u5047\u5982\u6211\u4eec\u8bbe\u5b9a\u5b66\u4e60\u7387\u5e94\u8be5\u57280.001~1\u8fd9\u4e2a\u533a\u95f4\u5185\u6bd4\u8f83\u597d\uff0c\u6211\u4eec\u600e\u4e48\u9009\u62e9\u6211\u4eec\u5c1d\u8bd5\u7684\u70b9\u5462\uff1f\u662f1-0.001\u518d\u966410\uff1f\u8fd9\u6837\u5bf9\u4e8e0.001~0.01\u8fd9\u4e2a\u533a\u95f4\u6211\u4eec\u4f7f\u7528\u4e86\u8fc7\u5c11\u7684\u8d44\u6e90\uff0c\u56e0\u6b64\u6211\u4eec\u91c7\u7528\u4e00\u79cd\u65b0\u7684\u65b9\u5f0f <\/p>\n\n\n

\u6211\u4eec\u4f7f\u75280.001\uff0c0.01\uff0c0.1\uff0c1\u8fd9\u51e0\u4e2a \u68af\u5ea6\u6765\u8fdb\u884c\u8ba1\u7b97\uff0c\u8fd9\u6837\u5728\u6bcf\u4e00\u4e2a\u68af\u5ea6\u4e4b\u95f4\u4f7f\u7528\u7684\u8d44\u6e90\u51e0\u4e4e\u76f8\u7b49\u3002 <\/p>\n\n\n

\u4e5f\u5c31\u662f\u6211\u4eec\u572810\u7684n\u6b21\u65b9\u4e0a\u53d6\u7b49\u4efd\u3002 <\/p>\n\n\n

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

3.\u5f52\u4e00\u5316\u7f51\u7edc\u7684\u6fc0\u6d3b\u51fd\u6570<\/h2>\n\n\n

batch\u5f52\u4e00\u5316\u4f1a\u4f7f\u4f60\u7684\u53c2\u6570\u641c\u7d22\u95ee\u9898\u53d8\u5f97\u66f4\u52a0\u5bb9\u6613\uff0c\u4f7f\u795e\u7ecf\u7f51\u7edc\u5bf9\u8d85\u53c2\u7684\u9009\u62e9\u66f4\u52a0\u7a33\u5b9a\uff0c\u8d85\u53c2\u7684\u8303\u56f4\u4f1a\u66f4\u52a0\u5e9e\u5927\uff0c\u5de5\u4f5c\u6548\u679c\u4f1a\u66f4\u597d\uff0c\u56e0\u6b64\u4f60\u8bad\u7ec3\u8d77\u6765\u4e5f\u66f4\u5bb9\u6613\u3002 <\/p>\n\n\n

\n\u4e4b\u524d\u6211\u4eec\u5b66\u8fc7\u8f93\u5165\u6570\u636e\u7684\u5f52\u4e00\u5316\uff0c\u8fd9\u6b21\u6211\u4eec\u5c06\u8f93\u5165\u7684\u5f52\u4e00\u5316\u5f15\u5165\u5230\u795e\u7ecf\u7f51\u7edc\u4e2d\uff0c\u6211\u4eec\u5bf9w\u548cb\u8fdb\u884c\u5f52\u4e00\u5316\uff0c\u4ee5\u8fbe\u5230\u4e00\u4e2a\u6bd4\u8f83\u597d\u7684\u6548\u679c\n<\/p>\n\n\n

4.\u5c06Batch Norm\u62df\u5408\u8fdb\u795e\u7ecf\u7f51\u7edc<\/h2>\n\n\n

Batch\u5f52\u4e00\u5316\u662f\u53d1\u751f\u5728\u8ba1\u7b97z\u548ca\u4e4b\u95f4\u3002 <\/p>\n\n\n

\u5b9e\u9645\u4e0a\u5c31\u662f\u5728\u6bcf\u4e00\u6b21\u8ba1\u7b97\u5b8cz\u4e4b\u540e\u8fdb\u884c\u4e00\u6b21\u6570\u636e\u6807\u51c6\u5316\uff08\u51cf\u53bb\u5747\u503c\u9664\u4ee5\u65b9\u5dee\uff09\uff0c\u7136\u540e\u5728\u8fdb\u884c\u4e0b\u4e00\u5c42\u8ba1\u7b97\uff0c\u81f3\u4e8e\u4f7f\u7528\u4ec0\u4e48\u6765\u6fc0\u6d3b\uff0c\u4f7f\u7528\u4ec0\u4e48\u65b9\u6cd5\u8fdb\u884c\u4e0b\u964d\uff0c\u90fd\u548c\u8fd9\u4e2a\u6ca1\u6709\u5173\u7cfb <\/p>\n\n\n

\u5728\u6846\u67b6\u4e2d\u5b9e\u9645\u4e0a\u53ea\u6709\u4e00\u884c\u4ee3\u7801\u5c31\u53ef\u4ee5\u5b8c\u6210\u8fd9\u4e2a\u64cd\u4f5c\uff0c\u6bd4\u5982\u5728TensorFlow\u4e2d\uff0c\u6211\u4eec\u53ea\u9700\u8981tf.nn.batch_normalization\u5373\u53ef\u5b8c\u6210batch Norm\u64cd\u4f5c<\/p>\n\n\n

\u6211\u4eec\u4e00\u822c\u5c06batch Norm\u4e0emini-batch\u7ec4\u5408\u4f7f\u7528\u3002 <\/p>\n\n\n

5.Batch Norm\u4e3a\u4ec0\u4e48\u594f\u6548<\/h2>\n\n\n

\u7b2c\u4e00\u4e2a\u539f\u56e0\uff0c\u7ecf\u8fc7\u6807\u51c6\u5316\u7684\u8f93\u5165\u4f3c\u7684\u5176\u5747\u503c\u4e3a0\uff0c\u65b9\u5dee\u4e3a1\uff0c\u56e0\u6b64\u6211\u4eec\u53ef\u4ee5\u5c06\u4e00\u4e2a\u5f88\u5927\u8303\u56f4\u7684\u8f93\u5165\u503c\u8f6c\u6362\u52300~1\u4e4b\u95f4\uff0c\u53ef\u4ee5\u52a0\u901f\u5b66\u4e60\u3002 <\/p>\n\n\n

\u53e6\u4e00\u4e2a\u539f\u56e0\u662f\u5b83\u53ef\u4ee5\u4f7f\u6743\u91cd\u6bd4\u4f60\u7684\u7f51\u7edc\u66f4\u6ede\u540e\u6216\u66f4\u6df1\u5c42\u3002 <\/p>\n\n\n

\u6bd4\u5982\u6211\u4eec\u4f7f\u7528\u4e00\u4e2a\u795e\u7ecf\u7f51\u7edc\u6765\u8bad\u7ec3\u732b\u7684\u8bc6\u522b\uff0c\u6211\u4eec\u6240\u6709\u7684\u4f8b\u5b50\u90fd\u662f\u9ed1\u732b <\/p>\n\n\n

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

\u8fd9\u65f6\u5019\u5982\u679c\u6211\u4eec\u6d4b\u8bd5\u5176\u4ed6\u989c\u8272\u7684\u732b <\/p>\n\n\n

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

\u6548\u679c\u5f88\u53ef\u80fd\u4e0d\u597d\uff0c\u4e3a\u4e86\u9632\u6b62\u8fd9\u79cd\u60c5\u51b5\uff0c\u6211\u4eec\u9700\u8981\u4f7f\u7f51\u7edc\u4e0d\u8fc7\u5206\u4f9d\u8d56\u8f93\u5165\u6570\u636e\uff0c\u8fd9\u65f6\u5019\u6211\u4eec\u5c31\u53ef\u4ee5\u5f15\u5165Batch Norm\uff0c <\/p>\n\n\n

\u6211\u4eec\u53ef\u4ee5\u628a\u6bcf\u4e00\u5c42\u795e\u7ecf\u7f51\u7edc\u62c6\u5f00\u6765\uff0c\u5f53\u524d\u5c42\u7684\u524d\u4e00\u5c42\u4e3a\u5f53\u524d\u5c42\u7684\u8f93\u5165\u5c42\uff0c\u90a3\u4e48\u600e\u4e48\u4f7f\u5f53\u524d\u5c42\u4e0d\u8fc7\u5206\u4f9d\u8d56\u524d\u4e00\u5c42\u7684\u503c\u5462\uff0c\u5c31\u662f\u6807\u51c6\u5316\uff0c \u5728\u524d\u4e00\u5c42\u503c\u7684\u57fa\u7840\u4e0a\u91cd\u65b0\u8ba1\u7b97\u4e00\u4e2a\u65b9\u5dee\u4e3a1\u5e73\u5747\u503c\u4e3a0\u7684\u4e00\u7ec4\u6570\uff0c\u8ba9\u6bcf\u4e00\u5c42\u7f51\u7edc\u90fd\u53ef\u4ee5\u201c\u72ec\u7acb\u201d\u5b66\u4e60\uff0c\u4ee5\u8fbe\u5230\u66f4\u597d\u7684\u6548\u679c\u3002 <\/p>\n\n\n

\u5f53\u7136\uff0cdropout\u4e5f\u6709\u4e00\u5b9a\u7684\u6548\u679c\uff0c\u6709\u65f6\u5019\u6211\u4eec\u5c06dropout\u548cbatch norm\u5408\u8d77\u6765\u4f7f\u7528\u3002\n<\/p>\n\n\n

6.\u6d4b\u8bd5\u65f6\u7684Batch Norm<\/h2>\n\n\n

\u6211\u4eec\u5728\u8bad\u7ec3\u65f6\u7531\u4e8e\u4f7f\u7528\u4e86Batch Norm\uff0c\u6bcf\u4e00\u5c42\u90fd\u8fdb\u884c\u6807\u51c6\u5316\uff0c\u4f46\u662f\u6211\u4eec\u5728\u6d4b\u8bd5\u7684\u65f6\u5019\uff0c\u4e00\u4e2a\u6570\u636e\u7684\u6807\u51c6\u5316\u662f\u6ca1\u6709\u610f\u4e49\u7684\uff0c\u56e0\u6b64\u6211\u4eec\u9700\u8981\u4f30\u8ba1\u8fd9\u4e24\u4e2a\u6807\u51c6\u5316\u7684\u53c2\u6570\u3002 <\/p>\n\n\n

\u6211\u4eec\u91c7\u7528\u6307\u6570\u52a0\u6743\u5e73\u5747\u7684\u65b9\u6cd5\u6765\u4f30\u8ba1\u8fd9\u4e24\u4e2a\u6570\uff0c\u6839\u636e\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u6bcf\u4e2amini-batch\u7684\u6570\u6c42\u51fa\u5e73\u5747\u503c\uff0c\u7528\u4e8e\u6d4b\u8bd5\u3002 <\/p>\n\n\n

\n\u5b9e\u9645\u4e0a\u91c7\u7528\u4f55\u79cd\u65b9\u6cd5\u6bd4\u5982\u76f4\u63a5\u53d6\u6700\u540e\u4e00\u6b21\u662f\u6ca1\u6709\u5f88\u5927\u7684\u5f71\u54cd\u7684\uff08\u4f46\u662f\u5b9e\u9645\u4e0a\u8fd0\u7528\u52a0\u6743\u5e73\u5747\u6bd4\u8f83\u591a\uff09\uff0c\u53ea\u8981\u5408\u7406\u7684\u53d6\u51fa\u8fd9\u4e24\u4e2a\u6570\u503c\u5728\u6d4b\u8bd5\u4e2d\u90fd\u4f1a\u6709\u6548\u3002\n<\/p>\n\n\n

7.Softmax\u56de\u5f52<\/h2>\n\n\n

\u5bf9\u4e8e\u591a\u5206\u7c7b\u7684\u95ee\u9898\uff0c\u6211\u4eec\u4e0d\u80fd\u50cf\u4e4b\u524d\u4e8c\u5206\u7c7b\u4e00\u6837\u8f93\u51fa\u662f\u4e0d\u662f\u7684\u6982\u7387\u3002<\/p>\n\n\n

\u56e0\u6b64\u5f15\u5165softmax\uff0c\u5b9e\u9645\u4e0a\u5c31\u662f\u8f93\u51fa\u5c42\u7684\u6fc0\u6d3b\u51fd\u6570\uff0c\u5148\u5bf9\u8f93\u51fa\u5c42\u53d6\u6307\u6570\uff0c\u7136\u540e\u6c42\u51fa\u4e00\u4e2a\u548c\u4e3a1\u7684\u6982\u7387\u503c\u3002 <\/p>\n\n\n

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

\u6d4b\u9a8c<\/h2>\n\n\n

1. \u5982\u679c\u5728\u5927\u91cf\u7684\u8d85\u53c2\u6570\u4e2d\u641c\u7d22\u6700\u4f73\u7684\u53c2\u6570\u503c\uff0c\u90a3\u4e48\u5e94\u8be5\u5c1d\u8bd5\u5728\u7f51\u683c\u4e2d\u641c\u7d22\u800c\u4e0d\u662f\u4f7f\u7528\u968f\u673a\u503c\uff0c\u4ee5\u4fbf\u66f4\u7cfb\u7edf\u7684\u641c\u7d22\uff0c\u800c\u4e0d\u662f\u4f9d\u9760\u8fd0\u6c14\uff0c\u8bf7\u95ee\u8fd9\u53e5\u8bdd\u662f\u6b63\u786e\u7684\u5417\uff1f <\/strong><\/p>\n\n\n

\u9519\u8bef\uff0c \u5e94\u5f53\u5c1d\u8bd5\u968f\u673a\u503c\uff0c\u4e0d\u8981\u4f7f\u7528\u7f51\u683c\u641c\u7d22\uff0c\u56e0\u4e3a\u4f60\u4e0d\u77e5\u9053\u54ea\u4e9b\u8d85\u53c2\u6570\u6bd4\u5176\u4ed6\u7684\u66f4\u91cd\u8981\u3002 <\/p>\n\n\n

2. \u6bcf\u4e2a\u8d85\u53c2\u6570\u5982\u679c\u8bbe\u7f6e\u5f97\u4e0d\u597d\uff0c\u90fd\u4f1a\u5bf9\u8bad\u7ec3\u4ea7\u751f\u5de8\u5927\u7684\u8d1f\u9762\u5f71\u54cd\uff0c\u56e0\u6b64\u6240\u6709\u7684\u8d85\u53c2\u6570\u90fd\u8981\u8c03\u6574\u597d\uff0c\u8bf7\u95ee\u8fd9\u662f\u6b63\u786e\u7684\u5417\uff1f <\/strong><\/p>\n\n\n

\u9519\u8bef\uff0c\u6bd4\u5982epsilon\uff0c\u5c31\u5c5e\u4e8e\u6bd4\u8f83\u65e0\u5173\u7d27\u8981\u7684\u53c2\u6570\u3002<\/p>\n\n\n

3. \u5728\u8d85\u53c2\u6570\u641c\u7d22\u8fc7\u7a0b\u4e2d\uff0c\u4f60\u5c1d\u8bd5\u53ea\u7167\u987e\u4e00\u4e2a\u6a21\u578b\uff08\u4f7f\u7528\u718a\u732b\u7b56\u7565\uff09\u8fd8\u662f\u4e00\u8d77\u8bad\u7ec3\u5927\u91cf\u7684\u6a21\u578b\uff08\u9c7c\u5b50\u9171\u7b56\u7565\uff09\u5728\u5f88\u5927\u7a0b\u5ea6\u4e0a\u53d6\u51b3\u4e8e <\/strong><\/p>\n\n\n

  1. \u662f\u5426\u4f7f\u7528\u6279\u91cf\uff08batch\uff09\u6216\u5c0f\u6279\u91cf\u4f18\u5316\uff08mini-batch optimization\uff09<\/li>
  2. \u795e\u7ecf\u7f51\u7edc\u4e2d\u5c40\u90e8\u6700\u5c0f\u503c\uff08\u978d\u70b9\uff09\u7684\u5b58\u5728\u6027 <\/li>
  3. \u5728\u4f60\u80fd\u529b\u8303\u56f4\u5185\uff0c\u4f60\u80fd\u591f\u62e5\u6709\u591a\u5927\u7684\u8ba1\u7b97\u80fd\u529b<\/li>
  4. \u9700\u8981\u8c03\u6574\u7684\u8d85\u53c2\u6570\u7684\u6570\u91cf <\/li><\/ol>\n\n\n

    3<\/p>\n\n\n

    4. \u5982\u679c\u60a8\u8ba4\u4e3a<\/strong>\u03b2<\/strong><\/em>\uff08\u52a8\u91cf\u8d85\u53c2\u6570\uff09\u4ecb\u4e8e0.9\u548c0.99\u4e4b\u95f4\uff0c\u90a3\u4e48\u63a8\u8350\u91c7\u7528\u4ee5\u4e0b\u54ea\u4e00\u79cd\u65b9\u6cd5\u6765\u5bf9<\/strong>\u03b2<\/strong><\/em>\u503c\u8fdb\u884c\u53d6\u6837\uff1f <\/strong><\/p>\n\n\n

    r = np.random.rand()\nbeta = 1 - 10 ** ( - r - 1 )<\/code><\/pre>\n\n\n
    5. \u627e\u5230\u597d\u7684\u8d85\u53c2\u6570\u7684\u503c\u662f\u975e\u5e38\u8017\u65f6\u7684\uff0c\u6240\u4ee5\u901a\u5e38\u60c5\u51b5\u4e0b\u4f60\u5e94\u8be5\u5728\u9879\u76ee\u5f00\u59cb\u65f6\u505a\u4e00\u6b21\uff0c\u5e76\u5c1d\u8bd5\u627e\u5230\u975e\u5e38\u597d\u7684\u8d85\u53c2\u6570\uff0c\u8fd9\u6837\u4f60\u5c31\u4e0d\u5fc5\u518d\u6b21\u91cd\u65b0\u8c03\u6574\u5b83\u4eec\u3002\u8bf7\u95ee\u8fd9\u6b63\u786e\u5417\uff1f <\/strong><\/p>\n\n\n
    \u9519\u8bef\uff0c \u6a21\u578b\u4e2d\u7684\u7ec6\u5fae\u53d8\u5316\u53ef\u80fd\u5bfc\u81f4\u60a8\u9700\u8981\u4ece\u5934\u5f00\u59cb\u91cd\u65b0\u627e\u5230\u597d\u7684\u8d85\u53c2\u6570\u3002 <\/p>\n\n\n
    6. \u5728\u89c6\u9891\u4e2d\u4ecb\u7ecd\u7684\u6279\u91cf\u6807\u51c6\u5316\u4e2d\uff0c\u5982\u679c\u5c06\u5176\u5e94\u7528\u4e8e\u795e\u7ecf\u7f51\u7edc\u7684\u7b2c<\/strong>l<\/strong><\/em>\u5c42\uff0c\u90a3\u4e48\u60a8\u4f7f\u7528\u4ec0\u4e48\u8fdb\u884c\u6807\u51c6\u5316\uff1f <\/strong><\/p>\n\n\n
    z[l]<\/p>\n\n\n
     7.\u5728\u6807\u51c6\u5316\u516c\u5f0f\u4e2d\uff0c\u4e3a\u4ec0\u4e48\u8981\u4f7f\u7528epsilon\uff08<\/strong>\u03f5<\/strong><\/em>\uff09\uff1f <\/strong><\/p>\n\n\n
    \u4e3a\u4e86\u907f\u514d\u96640\u64cd\u4f5c\u3002<\/p>\n\n\n
    8. \u6279\u5904\u7406\u89c4\u8303\u4e2d\u5173\u4e8e <\/strong>\u03b3<\/strong><\/em> \u548c <\/strong>\u03b2<\/strong><\/em> \u7684\u4ee5\u4e0b\u54ea\u4e9b\u9648\u8ff0\u662f\u6b63\u786e\u7684\uff1f<\/strong><\/p>\n\n\n
    1.  \u5b83\u4eec\u53ef\u4ee5\u5728Adam\u3001\u5177\u6709\u52a8\u91cf\u7684\u68af\u5ea6\u4e0b\u964d\u6216RMSprop\u4f7f\u4e2d\u7528\uff0c\u800c\u4e0d\u4ec5\u4ec5\u662f\u7528\u68af\u5ea6\u4e0b\u964d\u6765\u5b66\u4e60\u3002 <\/li>
    2.  \u5b83\u4eec\u8bbe\u5b9a\u7ed9\u5b9a\u5c42\u7684\u7ebf\u6027\u53d8\u91cf z<\/em>[l<\/em>] \u7684\u5747\u503c\u548c\u65b9\u5dee\u3002 <\/li><\/ol>\n\n\n
      1\uff0c2<\/p>\n\n\n
      9. \u5728\u8bad\u7ec3\u5177\u6709\u6279\u5904\u7406\u89c4\u8303\u7684\u795e\u7ecf\u7f51\u7edc\u4e4b\u540e\uff0c\u5728\u6d4b\u8bd5\u65f6\u95f4\uff0c\u5728\u65b0\u6837\u672c\u4e0a\u8bc4\u4f30\u795e\u7ecf\u7f51\u7edc\uff0c\u60a8\u5e94\u8be5\uff1a <\/strong><\/p>\n\n\n
       \u6267\u884c\u6240\u9700\u7684\u6807\u51c6\u5316\uff0c\u5728\u8bad\u7ec3\u671f\u95f4\u4f7f\u7528\u4f7f\u7528\u4e86\u03bc<\/em>\u548c\u03c3<\/em>2\u7684\u6307\u6570\u52a0\u6743\u5e73\u5747\u503c\u6765\u4f30\u8ba1mini-batches\u7684\u60c5\u51b5\u3002 <\/p>\n\n\n
      10. \u5173\u4e8e\u6df1\u5ea6\u5b66\u4e60\u7f16\u7a0b\u6846\u67b6\u7684\u8fd9\u4e9b\u9648\u8ff0\u4e2d\uff0c\u54ea\u4e00\u4e2a\u662f\u6b63\u786e\u7684\uff1f <\/strong><\/p>\n\n\n
      1.  \u901a\u8fc7\u7f16\u7a0b\u6846\u67b6\uff0c\u60a8\u53ef\u4ee5\u4f7f\u7528\u6bd4\u4f4e\u7ea7\u8bed\u8a00\uff08\u5982Python\uff09\u66f4\u5c11\u7684\u4ee3\u7801\u6765\u7f16\u5199\u6df1\u5ea6\u5b66\u4e60\u7b97\u6cd5\u3002 <\/li>
      2.  \u5373\u4f7f\u4e00\u4e2a\u9879\u76ee\u76ee\u524d\u662f\u5f00\u6e90\u7684\uff0c\u9879\u76ee\u7684\u826f\u597d\u7ba1\u7406\u6709\u52a9\u4e8e\u786e\u4fdd\u5b83\u5373\u4f7f\u5728\u957f\u671f\u5185\u4ecd\u7136\u4fdd\u6301\u5f00\u653e\uff0c\u800c\u4e0d\u662f\u4ec5\u4ec5\u4e3a\u4e86\u4e00\u4e2a\u516c\u53f8\u800c\u5173\u95ed\u6216\u4fee\u6539\u3002 <\/li>
      3.  \u6df1\u5ea6\u5b66\u4e60\u7f16\u7a0b\u6846\u67b6\u7684\u8fd0\u884c\u9700\u8981\u57fa\u4e8e\u4e91\u7684\u673a\u5668\u3002 <\/li><\/ol>\n\n\n
        1\uff0c2<\/p>\n\n\n
        \u7f16\u7a0b\u4f5c\u4e1a<\/h2>\n\n\n
        \u9996\u5148\u8fdb\u884cTensorFlow\u5165\u95e8\uff1a<\/p>\n\n\n
        \u5bfc\u5305\uff1a<\/p>\n\n\n
        import tensorflow as tf<\/pre>\n\n\n
        \u5b9a\u4e49\u635f\u5931\uff1a<\/p>\n\n\n
        # \u8ba1\u7b97\u635f\u5931
        # \u5b9a\u4e49y_hat\u4e3a36
        y_hat = tf.constant(36, name='y_hat')
        # \u5b9a\u4e49y\u4e3a39
        y = tf.constant(39, name='y')

        # \u5b9a\u4e49\u635f\u5931\u51fd\u6570
        loss = tf.Variable((y - y_hat) ** 2, name='loss')

        # \u521d\u59cb\u5316\u53c2\u6570
        init = tf.global_variables_initializer()

        with tf.Session() as session:
            session.run(init)
            print(session.run(loss))
        <\/pre>\n\n\n
        \u7a0b\u5e8f\u8f93\u51fa\u4e3a9\u3002<\/p>\n\n\n
        placeholder\u5360\u4f4d\u7b26<\/p>\n\n\n
        # \u4f7f\u7528placeholder\u5360\u4f4d\u7b26
        x = tf.placeholder(tf.int64, name=\"x\")
        with tf.Session() as session:
            print(session.run(2 * x, feed_dict={x: 3}))<\/pre>\n\n\n
        \u7a0b\u5e8f\u8f93\u51fa\u4e3a6\u3002<\/p>\n\n\n
        \u4e0b\u9762\u662f\u8ba1\u7b97\u7ebf\u6027\u51fd\u6570\uff1a<\/p>\n\n\n
        # \u8ba1\u7b97\u7ebf\u6027\u51fd\u6570
        def linear_function():
            \"\"\"
        <\/em>    <\/em>\u5b9e\u73b0\u4e00\u4e2a\u8ba1\u7b97\u7ebf\u6027\u51fd\u6570\u7684\u529f\u80fd
        <\/em>    <\/em>:return<\/em><\/strong>:
        <\/em>    \"\"\"
        <\/em>    <\/em>X = np.random.randn(3, 1)
            W = np.random.randn(4, 3)
            b = np.random.randn(4, 1)

            Y = tf.add(tf.matmul(W, X), b)
            # \u8fd9\u4fe9\u662f\u4e00\u6837\u7684\uff0c\u56e0\u4e3a\u5982\u679ctf\u4f1a\u91cd\u8f7d\u8fd0\u7b97\u7b26
            Y = tf.matmul(W, X) + b

            with tf.Session() as session:
                result = session.run(Y)
                session.close()
            return result


        print(\"result = \" + str(linear_function()))<\/pre>\n\n\n
        \u8f93\u51fa\uff1a<\/p>\n\n\n
        result = [[-2.15657382]\n [ 2.95891446]\n [-1.08926781]\n [-0.84538042]]<\/pre>\n\n\n
        sigmoid\uff1a<\/p>\n\n\n
        # sigmoid
        def sigmoid(z):
            \"\"\"
        <\/em>    sigmoid
        <\/em>    <\/em>:param<\/em><\/strong> z:
        <\/em>    <\/em>:return<\/em><\/strong>:
        <\/em>    \"\"\"
        <\/em>    <\/em>x = tf.placeholder(tf.float32, name='x')
            sigmoid = tf.sigmoid(x)
            with tf.Session() as session:
                result = session.run(sigmoid, feed_dict={x: z})
            return result


        print(\"sigmoid(0) = \" + str(sigmoid(0)))
        print(\"sigmoid(12) = \" + str(sigmoid(12)))<\/pre>\n\n\n
        \u8f93\u51fa\uff1a<\/p>\n\n\n
        # sigmoid(0) = 0.5\n# sigmoid(12) = 0.9999938<\/pre>\n\n\n
        one-hot\u77e9\u9635\uff0c\u5c31\u662f\u5c06\u7ed3\u679c\u5411\u91cf\u8f6c\u5316\u4e3a\u77e9\u9635\u7684\u5f62\u5f0f\uff1a<\/p>\n\n\n
        # one-hot\u77e9\u9635
        def one_hot_matrix(labels, C):
            \"\"\"
        <\/em>    <\/em>\u751f\u6210<\/em>onehot<\/em>\u77e9\u9635
        <\/em>    <\/em>:param<\/em><\/strong> labels:<\/em>\u6807\u7b7e\u5411\u91cf
        <\/em>    <\/em>:param<\/em><\/strong> C:<\/em>\u5206\u7c7b\u6570
        <\/em>    <\/em>:return<\/em><\/strong>:
        <\/em>    \"\"\"
        <\/em>    <\/em>C = tf.constant(C, name='C')
            one_hot_matrix = tf.one_hot(indices=labels, depth=C, axis=0)
            with tf.Session() as session:
                result = session.run(one_hot_matrix)
                session.close()
            return result


        labels = np.array([1, 2, 3, 0, 2, 1])
        one_hot = one_hot_matrix(labels, C=4)
        print(str(one_hot))<\/pre>\n\n\n
        \u8f93\u51fa\uff1a<\/p>\n\n\n
        [[0. 0. 0. 1. 0. 0.]
         [1. 0. 0. 0. 0. 1.]
         [0. 1. 0. 0. 1. 0.]
         [0. 0. 1. 0. 0. 0.]]<\/pre>\n\n\n
        \u4e0b\u9762\u5c31\u662f\u4f7f\u7528TensorFlow\u6765\u5199\u4e00\u4e2a\u8bc6\u522b\u624b\u52bf\u7684\u795e\u7ecf\u7f51\u7edc\uff0c\u9996\u5148\u5bfc\u5305\uff1a<\/p>\n\n\n
        import numpy as np\nimport matplotlib.pyplot as plt\nimport tensorflow as tf\nfrom course_2_week_3 import tf_utils\nimport time<\/pre>\n\n\n
        \u9996\u5148\u52a0\u8f7d\u6570\u636e\uff0c\u5e76\u968f\u673a\u6253\u5370\u4e00\u4e2a\u770b\u770b\uff1a<\/p>\n\n\n

        # \u52a0\u8f7d\u6570\u636e
        X_train_orig, Y_train_orig, X_test_orig, Y_test_orig, classes = tf_utils.load_dataset()

        plt.imshow(X_train_orig[11])
        print(\"Y = \" + str(np.squeeze(Y_train_orig[:, 11])))
        plt.show()<\/pre>\n\n\n
        \u53ef\u4ee5\u770b\u5230\u8fd9\u4e2a\u662f1\uff1a<\/p>\n\n\n
        \"\"<\/figure><\/div>\n\n\n
        \u7136\u540e\u6211\u4eec\u5c06\u8bad\u7ec3\u6570\u636e\u8f6c\u7f6e\u3001\u6807\u51c6\u5316\u5e76\u5c06\u7ed3\u679c\u96c6\u53d8\u6210one-hot\u77e9\u9635\u3002<\/p>\n\n\n
        X_train_flatten = X_train_orig.reshape(X_train_orig.shape[0], -1).T
        X_test_flatten = X_test_orig.reshape(X_test_orig.shape[0], -1).T
        # \u5f52\u4e00\u5316\u6570\u636e
        X_train = X_train_flatten \/ 255
        X_test = X_test_flatten \/ 255

        # \u8f6c\u6362onehot\u77e9\u9635
        Y_train = tf_utils.convert_to_one_hot(Y_train_orig, 6)
        Y_test = tf_utils.convert_to_one_hot(Y_test_orig, 6)<\/pre>\n\n\n
        \u7136\u540e\u6211\u4eec\u5b9a\u4e49placeholder\u751f\u6210\u51fd\u6570\uff1a<\/p>\n\n\n
        def create_placeholders(n_x, n_y):
            \"\"\"
        <\/em>    <\/em>\u4e3a<\/em>TensorFlow<\/em>\u4f1a\u8bdd\u521b\u5efa\u5360\u4f4d\u7b26
        <\/em>    <\/em>:param<\/em><\/strong> n_x:
        <\/em>    <\/em>:param<\/em><\/strong> n_y:
        <\/em>    <\/em>:return<\/em><\/strong>:
        <\/em>    \"\"\"
        <\/em>    <\/em>X = tf.placeholder(tf.float32, [n_x, None], name='X')
            Y = tf.placeholder(tf.float32, [n_y, None], name='Y')
            return X, Y<\/pre>\n\n\n
        \u53c2\u6570\u7684\u521d\u59cb\u5316\uff1a<\/p>\n\n\n
        def initialize_parameters():
            \"\"\"
        <\/em>    <\/em>\u521d\u59cb\u5316\u53c2\u6570
        <\/em>    <\/em>:return<\/em><\/strong>:
        <\/em>    \"\"\"
        <\/em>
        <\/em>    <\/em>W1 = tf.get_variable('W1', [25, 12288], initializer=tf.contrib.layers.xavier_initializer(seed=1))
            b1 = tf.get_variable('b1', [25, 1], initializer=tf.zeros_initializer())

            W2 = tf.get_variable('W2', [12, 25], initializer=tf.contrib.layers.xavier_initializer(seed=1))
            b2 = tf.get_variable('b2', [12, 1], initializer=tf.zeros_initializer())

            W3 = tf.get_variable('W3', [6, 12], initializer=tf.contrib.layers.xavier_initializer(seed=1))
            b3 = tf.get_variable('b3', [6, 1], initializer=tf.zeros_initializer())

            return {
                'W1': W1,
                'b1': b1,
                'W2': W2,
                'b2': b2,
                'W3': W3,
                'b3': b3
            }<\/pre>\n\n\n
        \u524d\u5411\u4f20\u64ad\uff1a<\/p>\n\n\n
        def forward_propagation(X, parameters):
            \"\"\"
        <\/em>    <\/em>\u524d\u5411\u4f20\u64ad
        <\/em>    <\/em>:param<\/em><\/strong> X:
        <\/em>    <\/em>:param<\/em><\/strong> parameters:
        <\/em>    <\/em>:return<\/em><\/strong>:
        <\/em>    \"\"\"
        <\/em>    <\/em>W1 = parameters['W1']
            b1 = parameters['b1']
            W2 = parameters['W2']
            b2 = parameters['b2']
            W3 = parameters['W3']
            b3 = parameters['b3']

            Z1 = tf.matmul(W1, X) + b1
            A1 = tf.nn.relu(Z1)

            Z2 = tf.matmul(W2, A1) + b2
            A2 = tf.nn.relu(Z2)

            Z3 = tf.matmul(W3, A2) + b3

            return Z3<\/pre>\n\n\n
        \u8ba1\u7b97\u4ee3\u4ef7\uff1a<\/p>\n\n\n
        def compute_cost(Z3, Y):
            \"\"\"
        <\/em>    <\/em>\u8ba1\u7b97\u4ee3\u4ef7
        <\/em>    <\/em>:param<\/em><\/strong> Z3:
        <\/em>    <\/em>:param<\/em><\/strong> Y:
        <\/em>    <\/em>:return<\/em><\/strong>:
        <\/em>    \"\"\"
        <\/em>    <\/em>logits = tf.transpose(Z3)
            labels = tf.transpose(Y)

            cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=logits, labels=labels))
            return cost<\/pre>\n\n\n
        \u7f16\u5199\u6a21\u578b\u5e76\u5bf9\u6a21\u578b\u8fdb\u884c\u4fdd\u5b58\uff1a<\/p>\n\n\n
        def model(X_train, Y_train, X_test, Y_test, learning_rate=0.0001, num_epochs=1300, minibatch_size=32,
                  print_cost=True,
                  is_plot=True):
            \"\"\"
        <\/em>    <\/em>\u795e\u7ecf\u7f51\u7edc\u6a21\u578b<\/em>
        <\/em>    LINEAR->RELU->LINEAR->RELU->LINEAR->SOFTMAX
        <\/em>    <\/em>:param<\/em><\/strong> X_train:
        <\/em>    <\/em>:param<\/em><\/strong> Y_train:
        <\/em>    <\/em>:param<\/em><\/strong> X_test:
        <\/em>    <\/em>:param<\/em><\/strong> Y_test:
        <\/em>    <\/em>:param<\/em><\/strong> learning_rate:
        <\/em>    <\/em>:param<\/em><\/strong> num_epochs:
        <\/em>    <\/em>:param<\/em><\/strong> minbatch_size:
        <\/em>    <\/em>:param<\/em><\/strong> print_cost:
        <\/em>    <\/em>:param<\/em><\/strong> is_plot:
        <\/em>    <\/em>:return<\/em><\/strong>:
        <\/em>    \"\"\"
        <\/em>    <\/em>n_x, m = X_train.shape
            n_y = Y_train.shape[0]
            costs = []

            # \u521b\u5efaplaceholder
            X, Y = create_placeholders(n_x, n_y)

            # \u521d\u59cb\u5316\u53c2\u6570
            parameters = initialize_parameters()

            # \u524d\u884c\u4f20\u64ad
            Z3 = forward_propagation(X, parameters)

            # \u8ba1\u7b97\u6210\u672c
            cost = compute_cost(Z3, Y)

            # \u53cd\u5411\u4f20\u64ad,Adam\u4f18\u5316
            optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate).minimize(cost)

            saver = tf.train.Saver()

            # \u521d\u59cb\u5316\u6240\u6709\u7684\u53d8\u91cf
            init = tf.global_variables_initializer()

            # \u5f00\u59cb\u4f1a\u8bdd\u5e76\u8ba1\u7b97
            with tf.Session() as session:
                session.run(init)

                # \u5faa\u73af
                for epoch in range(num_epochs):
                    epoch_cost = 0
                    num_minbatches = int(m \/ minibatch_size)
                    minibatches = tf_utils.random_mini_batches(X_train, Y_train, minibatch_size)

                    for minibatch in minibatches:
                        minibatch_X, minibatch_Y = minibatch
                        # \u5f00\u59cb\u8fd0\u884c
                        _, minibatch_cost = session.run([optimizer, cost], feed_dict={X: minibatch_X, Y: minibatch_Y})
                        # \u8ba1\u7b97\u4ee3\u4ef7
                        epoch_cost = epoch_cost + minibatch_cost \/ num_minbatches

                    # \u8bb0\u5f55\u4ee3\u4ef7
                    if epoch % 5 == 0:
                        costs.append(epoch_cost)
                        if print_cost and epoch % 100 == 0:
                            print(\"epoch = \" + str(epoch) + \"    epoch_cost = \" + str(epoch_cost))

                    saver.save(session, '.\/model\/my_model', global_step=100)

                # \u662f\u5426\u7ed8\u5236\u56fe\u8c31
                if is_plot:
                    plt.plot(np.squeeze(costs))
                    plt.ylabel('cost')
                    plt.xlabel('iterations (per tens)')
                    plt.title(\"Learning rate =\" + str(learning_rate))
                    plt.show()

                # \u4fdd\u5b58\u53c2\u6570
                parameters = session.run(parameters)

                # \u8ba1\u7b97\u9884\u6d4b\u7ed3\u679c
                current_prediction = tf.equal(tf.argmax(Z3), tf.argmax(Y))

                # \u8ba1\u7b97\u51c6\u786e\u7387
                accuracy = tf.reduce_mean(tf.cast(current_prediction, 'float'))

                print(\"\u8bad\u7ec3\u96c6\u7684\u51c6\u786e\u7387\uff1a\", accuracy.eval({X: X_train, Y: Y_train}))
                print(\"\u6d4b\u8bd5\u96c6\u7684\u51c6\u786e\u7387:\", accuracy.eval({X: X_test, Y: Y_test}))

                return parameters<\/pre>\n\n\n
        \u8c03\u7528\u7f51\u7edc\uff1a<\/p>\n\n\n
        # \u5f00\u59cb\u65f6\u95f4
        start_time = time.clock()
        # \u5f00\u59cb\u8bad\u7ec3
        parameters = model(X_train, Y_train, X_test, Y_test)
        # \u7ed3\u675f\u65f6\u95f4
        end_time = time.clock()
        # \u8ba1\u7b97\u65f6\u5dee
        print(\"CPU\u7684\u6267\u884c\u65f6\u95f4 = \" + str(end_time - start_time) + \" \u79d2\")<\/pre>\n\n\n
        \u6211\u4eec\u53ef\u4ee5\u770b\u5230\u8bad\u7ec3\u7ed3\u679c\uff08\u8fc7\u7a0b\u8f83\u6162\uff09\uff1a<\/p>\n\n\n
        epoch = 0    epoch_cost = 1.8670177423592766
        epoch = 100    epoch_cost = 1.0305518833073701
        epoch = 200    epoch_cost = 0.7247187675851763
        epoch = 300    epoch_cost = 0.49531238006822986
        epoch = 400    epoch_cost = 0.33661927914980694
        epoch = 500    epoch_cost = 0.2284338934855028
        epoch = 600    epoch_cost = 0.17487916463252268
        epoch = 700    epoch_cost = 0.11813201111826033
        epoch = 800    epoch_cost = 0.08827614908417065
        epoch = 900    epoch_cost = 0.06095291397562531
        epoch = 1000    epoch_cost = 0.04344996804315032
        epoch = 1100    epoch_cost = 0.11286204381648339
        epoch = 1200    epoch_cost = 0.04022079737236102
        epoch = 1300    epoch_cost = 0.03377958937463435
        epoch = 1400    epoch_cost = 0.07520042944022201
        \u8bad\u7ec3\u96c6\u7684\u51c6\u786e\u7387\uff1a 0.9962963
        \u6d4b\u8bd5\u96c6\u7684\u51c6\u786e\u7387: 0.825
        CPU\u7684\u6267\u884c\u65f6\u95f4 = 433.66559440000003 \u79d2<\/pre>\n\n\n
        \u56fe\u7247\uff0c\u53ef\u4ee5\u770b\u5230\u7531\u4e8emini-batch\u9020\u6210\u7684\u8d77\u4f0f\u3002<\/p>\n\n\n
        \"\"<\/figure><\/div>\n\n\n
        \u5b8c\u6574\u4ee3\u7801\uff1a<\/p>\n\n\n
        # -*- coding:utf-8 -*-

        \"\"\"
        <\/em>      \u250f\u251b \u253b\u2501\u2501\u2501\u2501\u2501\u251b \u253b\u2513
        <\/em>      \u2503<\/em>\u3000\u3000\u3000\u3000\u3000\u3000<\/em> \u2503
        <\/em>      \u2503<\/em>\u3000\u3000\u3000<\/em>\u2501<\/em>\u3000\u3000\u3000<\/em>\u2503
        <\/em>      \u2503<\/em>\u3000<\/em>\u2533\u251b<\/em>\u3000<\/em>  \u2517\u2533<\/em>\u3000<\/em>\u2503
        <\/em>      \u2503<\/em>\u3000\u3000\u3000\u3000\u3000\u3000<\/em> \u2503
        <\/em>      \u2503<\/em>\u3000\u3000\u3000<\/em>\u253b<\/em>\u3000\u3000\u3000<\/em>\u2503
        <\/em>      \u2503<\/em>\u3000\u3000\u3000\u3000\u3000\u3000<\/em> \u2503
        <\/em>      \u2517\u2501\u2513<\/em>\u3000\u3000\u3000<\/em>\u250f\u2501\u2501\u2501\u251b
        <\/em>        \u2503<\/em>\u3000\u3000\u3000<\/em>\u2503   <\/em>\u795e\u517d\u4fdd\u4f51<\/em>
        <\/em>        \u2503<\/em>\u3000\u3000\u3000<\/em>\u2503   <\/em>\u4ee3\u7801\u65e0<\/em>BUG<\/em>\uff01<\/em>
        <\/em>        \u2503<\/em>\u3000\u3000\u3000<\/em>\u2517\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2513
        <\/em>        \u2503<\/em>\u3000\u3000\u3000\u3000\u3000\u3000\u3000<\/em>    \u2523\u2513
        <\/em>        \u2503<\/em>\u3000\u3000\u3000\u3000<\/em>         \u250f\u251b
        <\/em>        \u2517\u2501\u2513 \u2513 \u250f\u2501\u2501\u2501\u2533 \u2513 \u250f\u2501\u251b
        <\/em>          \u2503 \u252b \u252b   \u2503 \u252b \u252b
        <\/em>          \u2517\u2501\u253b\u2501\u251b   \u2517\u2501\u253b\u2501\u251b
        <\/em>\"\"\"
        <\/em>
        <\/em>import numpy as np
        import matplotlib.pyplot as plt
        import tensorflow as tf
        from course_2_week_3 import tf_utils
        import time

        import os
        os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'

        # \u52a0\u8f7d\u6570\u636e
        X_train_orig, Y_train_orig, X_test_orig, Y_test_orig, classes = tf_utils.load_dataset()

        # plt.imshow(X_train_orig[11])
        # print(\"Y = \" + str(np.squeeze(Y_train_orig[:, 11])))
        # plt.show()

        X_train_flatten = X_train_orig.reshape(X_train_orig.shape[0], -1).T
        X_test_flatten = X_test_orig.reshape(X_test_orig.shape[0], -1).T
        # \u5f52\u4e00\u5316\u6570\u636e
        X_train = X_train_flatten \/ 255
        X_test = X_test_flatten \/ 255

        # \u8f6c\u6362onehot\u77e9\u9635
        Y_train = tf_utils.convert_to_one_hot(Y_train_orig, 6)
        Y_test = tf_utils.convert_to_one_hot(Y_test_orig, 6)


        def create_placeholders(n_x, n_y):
            \"\"\"
        <\/em>    <\/em>\u4e3a<\/em>TensorFlow<\/em>\u4f1a\u8bdd\u521b\u5efa\u5360\u4f4d\u7b26
        <\/em>    <\/em>:param<\/em><\/strong> n_x:
        <\/em>    <\/em>:param<\/em><\/strong> n_y:
        <\/em>    <\/em>:return<\/em><\/strong>:
        <\/em>    \"\"\"
        <\/em>    <\/em>X = tf.placeholder(tf.float32, [n_x, None], name='X')
            Y = tf.placeholder(tf.float32, [n_y, None], name='Y')
            return X, Y


        # \u521d\u59cb\u5316\u53c2\u6570
        def initialize_parameters():
            \"\"\"
        <\/em>    <\/em>\u521d\u59cb\u5316\u53c2\u6570
        <\/em>    <\/em>:return<\/em><\/strong>:
        <\/em>    \"\"\"
        <\/em>
        <\/em>    <\/em>W1 = tf.get_variable('W1', [25, 12288], initializer=tf.contrib.layers.xavier_initializer(seed=1))
            b1 = tf.get_variable('b1', [25, 1], initializer=tf.zeros_initializer())

            W2 = tf.get_variable('W2', [12, 25], initializer=tf.contrib.layers.xavier_initializer(seed=1))
            b2 = tf.get_variable('b2', [12, 1], initializer=tf.zeros_initializer())

            W3 = tf.get_variable('W3', [6, 12], initializer=tf.contrib.layers.xavier_initializer(seed=1))
            b3 = tf.get_variable('b3', [6, 1], initializer=tf.zeros_initializer())

            return {
                'W1': W1,
                'b1': b1,
                'W2': W2,
                'b2': b2,
                'W3': W3,
                'b3': b3
            }


        def forward_propagation(X, parameters):
            \"\"\"
        <\/em>    <\/em>\u524d\u5411\u4f20\u64ad
        <\/em>    <\/em>:param<\/em><\/strong> X:
        <\/em>    <\/em>:param<\/em><\/strong> parameters:
        <\/em>    <\/em>:return<\/em><\/strong>:
        <\/em>    \"\"\"
        <\/em>    <\/em>W1 = parameters['W1']
            b1 = parameters['b1']
            W2 = parameters['W2']
            b2 = parameters['b2']
            W3 = parameters['W3']
            b3 = parameters['b3']

            Z1 = tf.matmul(W1, X) + b1
            A1 = tf.nn.relu(Z1)

            Z2 = tf.matmul(W2, A1) + b2
            A2 = tf.nn.relu(Z2)

            Z3 = tf.matmul(W3, A2) + b3

            return Z3


        def compute_cost(Z3, Y):
            \"\"\"
        <\/em>    <\/em>\u8ba1\u7b97\u4ee3\u4ef7
        <\/em>    <\/em>:param<\/em><\/strong> Z3:
        <\/em>    <\/em>:param<\/em><\/strong> Y:
        <\/em>    <\/em>:return<\/em><\/strong>:
        <\/em>    \"\"\"
        <\/em>    <\/em>logits = tf.transpose(Z3)
            labels = tf.transpose(Y)

            cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=logits, labels=labels))
            return cost


        def model(X_train, Y_train, X_test, Y_test, learning_rate=0.0001, num_epochs=1300, minibatch_size=32,
                  print_cost=True,
                  is_plot=True):
            \"\"\"
        <\/em>    <\/em>\u795e\u7ecf\u7f51\u7edc\u6a21\u578b<\/em>
        <\/em>    LINEAR->RELU->LINEAR->RELU->LINEAR->SOFTMAX
        <\/em>    <\/em>:param<\/em><\/strong> X_train:
        <\/em>    <\/em>:param<\/em><\/strong> Y_train:
        <\/em>    <\/em>:param<\/em><\/strong> X_test:
        <\/em>    <\/em>:param<\/em><\/strong> Y_test:
        <\/em>    <\/em>:param<\/em><\/strong> learning_rate:
        <\/em>    <\/em>:param<\/em><\/strong> num_epochs:
        <\/em>    <\/em>:param<\/em><\/strong> minbatch_size:
        <\/em>    <\/em>:param<\/em><\/strong> print_cost:
        <\/em>    <\/em>:param<\/em><\/strong> is_plot:
        <\/em>    <\/em>:return<\/em><\/strong>:
        <\/em>    \"\"\"
        <\/em>    <\/em>n_x, m = X_train.shape
            n_y = Y_train.shape[0]
            costs = []

            # \u521b\u5efaplaceholder
            X, Y = create_placeholders(n_x, n_y)

            # \u521d\u59cb\u5316\u53c2\u6570
            parameters = initialize_parameters()

            # \u524d\u884c\u4f20\u64ad
            Z3 = forward_propagation(X, parameters)

            # \u8ba1\u7b97\u6210\u672c
            cost = compute_cost(Z3, Y)

            # \u53cd\u5411\u4f20\u64ad,Adam\u4f18\u5316
            optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate).minimize(cost)

            saver = tf.train.Saver()

            # \u521d\u59cb\u5316\u6240\u6709\u7684\u53d8\u91cf
            init = tf.global_variables_initializer()

            # \u5f00\u59cb\u4f1a\u8bdd\u5e76\u8ba1\u7b97
            with tf.Session() as session:
                session.run(init)

                # \u5faa\u73af
                for epoch in range(num_epochs):
                    epoch_cost = 0
                    num_minbatches = int(m \/ minibatch_size)
                    minibatches = tf_utils.random_mini_batches(X_train, Y_train, minibatch_size)

                    for minibatch in minibatches:
                        minibatch_X, minibatch_Y = minibatch
                        # \u5f00\u59cb\u8fd0\u884c
                        _, minibatch_cost = session.run([optimizer, cost], feed_dict={X: minibatch_X, Y: minibatch_Y})
                        # \u8ba1\u7b97\u4ee3\u4ef7
                        epoch_cost = epoch_cost + minibatch_cost \/ num_minbatches

                    # \u8bb0\u5f55\u4ee3\u4ef7
                    if epoch % 5 == 0:
                        costs.append(epoch_cost)
                        if print_cost and epoch % 100 == 0:
                            print(\"epoch = \" + str(epoch) + \"    epoch_cost = \" + str(epoch_cost))

                    saver.save(session, '.\/model\/my_model', global_step=100)

                # \u662f\u5426\u7ed8\u5236\u56fe\u8c31
                if is_plot:
                    plt.plot(np.squeeze(costs))
                    plt.ylabel('cost')
                    plt.xlabel('iterations (per tens)')
                    plt.title(\"Learning rate =\" + str(learning_rate))
                    plt.show()

                # \u4fdd\u5b58\u53c2\u6570
                parameters = session.run(parameters)

                # \u8ba1\u7b97\u9884\u6d4b\u7ed3\u679c
                current_prediction = tf.equal(tf.argmax(Z3), tf.argmax(Y))

                # \u8ba1\u7b97\u51c6\u786e\u7387
                accuracy = tf.reduce_mean(tf.cast(current_prediction, 'float'))

                print(\"\u8bad\u7ec3\u96c6\u7684\u51c6\u786e\u7387\uff1a\", accuracy.eval({X: X_train, Y: Y_train}))
                print(\"\u6d4b\u8bd5\u96c6\u7684\u51c6\u786e\u7387:\", accuracy.eval({X: X_test, Y: Y_test}))

                return parameters


        # \u5f00\u59cb\u65f6\u95f4
        start_time = time.cl\nock()
        # \u5f00\u59cb\u8bad\u7ec3
        parameters = model(X_train, Y_train, X_test, Y_test)
        # \u7ed3\u675f\u65f6\u95f4
        end_time = time.clock()
        # \u8ba1\u7b97\u65f6\u5dee
        print(\"CPU\u7684\u6267\u884c\u65f6\u95f4 = \" + str(end_time - start_time) + \" \u79d2\")

        \"\"\"
        epoch = 0    epoch_cost = 1.8670177423592766
        epoch = 100    epoch_cost = 1.0305518833073701
        epoch = 200    epoch_cost = 0.7247187675851763
        epoch = 300    epoch_cost = 0.49531238006822986
        epoch = 400    epoch_cost = 0.33661927914980694
        epoch = 500    epoch_cost = 0.2284338934855028
        epoch = 600    epoch_cost = 0.17487916463252268
        epoch = 700    epoch_cost = 0.11813201111826033
        epoch = 800    epoch_cost = 0.08827614908417065
        epoch = 900    epoch_cost = 0.06095291397562531
        epoch = 1000    epoch_cost = 0.04344996804315032
        epoch = 1100    epoch_cost = 0.11286204381648339
        epoch = 1200    epoch_cost = 0.04022079737236102
        epoch = 1300    epoch_cost = 0.03377958937463435
        epoch = 1400    epoch_cost = 0.07520042944022201
        \u8bad\u7ec3\u96c6\u7684\u51c6\u786e\u7387\uff1a 0.9962963
        \u6d4b\u8bd5\u96c6\u7684\u51c6\u786e\u7387: 0.825
        CPU\u7684\u6267\u884c\u65f6\u95f4 = 433.66559440000003 \u79d2
        \"\"\"
        <\/pre>\n","protected":false},"excerpt":{"rendered":"
        \u5434\u6069\u8fbe\u6df1\u5ea6\u5b66\u4e60\u7b2c\u4e8c\u8bfe\u7b2c\u4e09\u5468 \u8d85\u53c2\u6570\u8c03\u8bd5\u3001Batch\u6b63\u5219\u5316\u548c\u7a0b\u5e8f\u6846\u67b6 1.\u8c03\u8bd5\u5904\u7406 \u5173\u4e8e\u8bad\u7ec3\u6df1\u5ea6\u6700\u96be\u7684 […]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"om_disable_all_campaigns":false,"_mi_skip_tracking":false,"_monsterinsights_sitenote_active":false,"_monsterinsights_sitenote_note":"","_monsterinsights_sitenote_category":0,"footnotes":""},"categories":[7],"tags":[],"views":3699,"_links":{"self":[{"href":"http:\/\/www.sniper97.cn\/index.php\/wp-json\/wp\/v2\/posts\/3098"}],"collection":[{"href":"http:\/\/www.sniper97.cn\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"http:\/\/www.sniper97.cn\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"http:\/\/www.sniper97.cn\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"http:\/\/www.sniper97.cn\/index.php\/wp-json\/wp\/v2\/comments?post=3098"}],"version-history":[{"count":0,"href":"http:\/\/www.sniper97.cn\/index.php\/wp-json\/wp\/v2\/posts\/3098\/revisions"}],"wp:attachment":[{"href":"http:\/\/www.sniper97.cn\/index.php\/wp-json\/wp\/v2\/media?parent=3098"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/www.sniper97.cn\/index.php\/wp-json\/wp\/v2\/categories?post=3098"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/www.sniper97.cn\/index.php\/wp-json\/wp\/v2\/tags?post=3098"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}