{"id":2951,"date":"2020-02-19T14:33:46","date_gmt":"2020-02-19T06:33:46","guid":{"rendered":"http:\/\/www.sniper97.cn\/?p=2951"},"modified":"2020-02-19T14:33:46","modified_gmt":"2020-02-19T06:33:46","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%e6%b7%b1%e5%ba%a6%e5%ad%a6%e4%b9%a0%e7%9a%84%e5%ae%9e%e7%94%a8%e5%b1%82%e9%9d%a2","status":"publish","type":"post","link":"http:\/\/www.sniper97.cn\/index.php\/note\/deep-learning\/2951\/","title":{"rendered":"\u3010\u5434\u6069\u8fbe\u6df1\u5ea6\u5b66\u4e60\u3011\u6df1\u5ea6\u5b66\u4e60\u7684\u5b9e\u7528\u5c42\u9762"},"content":{"rendered":"\n<p> \u5434\u6069\u8fbe\u6df1\u5ea6\u5b66\u4e60\u7b2c\u4e8c\u8bfe\u7b2c\u4e00\u5468 \u6df1\u5ea6\u5b66\u4e60\u7684\u5b9e\u7528\u5c42\u9762<\/p>\n\n\n<div class=\"wp-block-image\"><figure class=\"aligncenter\"><img loading=\"lazy\" decoding=\"async\" width=\"301\" height=\"374\" src=\"\/wp-content\/uploads\/2020\/02\/\u56fe\u7247-43.png\" alt=\"\" class=\"wp-image-2952\" srcset=\"http:\/\/www.sniper97.cn\/wp-content\/uploads\/2020\/02\/\u56fe\u7247-43.png 301w, http:\/\/www.sniper97.cn\/wp-content\/uploads\/2020\/02\/\u56fe\u7247-43-241x300.png 241w\" sizes=\"(max-width: 301px) 100vw, 301px\" \/><\/figure><\/div>\n\n\n<p> \u672c\u5468\u6211\u4eec\u5c06\u5b66\u4e60\u8d85\u53c2\u6570\u7684\u8c03\u4f18 \u5982\u4f55\u6784\u5efa\u6570\u636e&nbsp; \u5982\u4f55\u786e\u4fdd\u4f18\u5316\u7b97\u6cd5\u5feb\u901f\u8fd0\u884c\u3002<\/p>\n\n\n<h2 class=\"wp-block-heading\">1.\u6570\u636e\u7684\u5212\u5206<\/h2>\n\n\n<p>\n\u6570\u636e\u96c6\u7684\u79d1\u5b66\u5212\u5206\u5f80\u5f80\u53ef\u4ee5\u63d0\u9ad8\u7f51\u7edc\u7684\u51c6\u786e\u7387\u4ee5\u53ca\u8bc4\u4ef7\u7684\u4e2d\u80af\u6027\u3002\n<\/p>\n\n\n<p> \u673a\u5668\u5b66\u4e60\u65f6\u4ee3\uff0c\u5982\u679c\u6211\u4eec\u4e0d\u8bbe\u7f6e\u9a8c\u8bc1\u96c6\uff0c\u90a3\u4e48\u4e00\u822c\u8bad\u7ec3\u96c6\u536070%\uff0c\u6d4b\u8bd5\u96c6\u680830%\uff1b\u5982\u679c\u6211\u4eec\u8bbe\u7f6e\u9a8c\u8bc1\u96c6\uff0c\u90a3\u4ed6\u4eec\u7684\u6bd4\u4f8b\u4e00\u822c\u4e3a60%\uff0c20%\uff0c20%\u3002\u5f53\u7136\u8fd9\u5e76\u4e0d\u7edd\u5bf9\uff0c\u5982\u679c\u6211\u4eec\u6709100\u4e07\u6761\u6570\u636e\uff0c\u90a3\u4e48\u6211\u4eec98\u4e07\u6761\u8bad\u7ec3\u96c6\u3001\u9a8c\u8bc1\u96c6\u548c\u6d4b\u8bd5\u96c6\u54041\u4e07\u6761\u4e5f\u662f\u53ef\u4ee5\u7684\u3002 <\/p>\n\n\n<p>\n\u5bf9\u4e8e\u73b0\u4ee3\u6df1\u5ea6\u5b66\u4e60\uff0c\u5982\u679c\u6211\u4eec\u7684\u6570\u636e\u6765\u6e90\u5341\u5206\u590d\u6742\u7684\u60c5\u51b5\uff0c\u6bd4\u5982\u6709\u4e00\u4e9b\u7cbe\u7f8e\u5904\u7406\u8fc7\u7684\u6570\u636e\uff0c\u6709\u4e00\u4e9b\u722c\u53d6\u6765\u7684\u5dee\u4e00\u4e9b\u7684\u6570\u636e\uff0c\u8fd9\u65f6\u5019\u8bad\u7ec3\u96c6\u3001\u9a8c\u8bc1\u96c6\u4ee5\u53ca\u6d4b\u8bd5\u96c6\u4e2d\u5404\u79cd\u8d28\u91cf\u7684\u6570\u636e\u6bd4\u4f8b\u8981\u57fa\u672c\u4e00\u81f4\u3002\u5f53\u7136\u5bf9\u4e8e\u4e0d\u9700\u8981\u5bf9\u7ed3\u679c\u505a\u51fa\u65e0\u9519\u9a8c\u8bc1\u7684\u60c5\u51b5\u4e5f\u53ef\u4ee5\u4e0d\u4f7f\u7528\u6d4b\u8bd5\u96c6\uff08\u5f53\u7136\u8fd9\u65f6\u5019\u4f60\u53ef\u4ee5\u79f0\u9a8c\u8bc1\u96c6\u4e3a\u6d4b\u8bd5\u96c6\uff09\u3002\n<\/p>\n\n\n<h2 class=\"wp-block-heading\">2.\u504f\u5dee&amp;\u65b9\u5dee<\/h2>\n\n\n<div class=\"wp-block-image\"><figure class=\"aligncenter\"><img loading=\"lazy\" decoding=\"async\" width=\"902\" height=\"314\" src=\"\/wp-content\/uploads\/2020\/02\/\u56fe\u7247-44.png\" alt=\"\" class=\"wp-image-2953\" srcset=\"http:\/\/www.sniper97.cn\/wp-content\/uploads\/2020\/02\/\u56fe\u7247-44.png 902w, http:\/\/www.sniper97.cn\/wp-content\/uploads\/2020\/02\/\u56fe\u7247-44-300x104.png 300w, http:\/\/www.sniper97.cn\/wp-content\/uploads\/2020\/02\/\u56fe\u7247-44-768x267.png 768w\" sizes=\"(max-width: 902px) 100vw, 902px\" \/><\/figure><\/div>\n\n\n<p> \u5bf9\u4e8e\u9ad8\u504f\u5dee\u548c\u9ad8\u65b9\u5dee\u5728\u9519\u8bef\u7387\u4e0a\u7684\u4f53\u73b0\u5982\u4e0b\u8868\u6240\u793a\uff1a <\/p>\n\n\n<div class=\"wp-block-image\"><figure class=\"aligncenter\"><img loading=\"lazy\" decoding=\"async\" width=\"839\" height=\"217\" src=\"\/wp-content\/uploads\/2020\/02\/\u56fe\u7247-45.png\" alt=\"\" class=\"wp-image-2954\" srcset=\"http:\/\/www.sniper97.cn\/wp-content\/uploads\/2020\/02\/\u56fe\u7247-45.png 839w, http:\/\/www.sniper97.cn\/wp-content\/uploads\/2020\/02\/\u56fe\u7247-45-300x78.png 300w, http:\/\/www.sniper97.cn\/wp-content\/uploads\/2020\/02\/\u56fe\u7247-45-768x199.png 768w\" sizes=\"(max-width: 839px) 100vw, 839px\" \/><\/figure><\/div>\n\n\n<h2 class=\"wp-block-heading\">3.\u673a\u5668\u5b66\u4e60\u57fa\u7840<\/h2>\n\n\n<p>\u8be6\u89c1 <a rel=\"noreferrer noopener\" aria-label=\"\u673a\u5668\u5b66\u4e60\u7b97\u6cd5\u4e0e\u8c03\u4f18\uff08\u5728\u65b0\u7a97\u53e3\u6253\u5f00\uff09\" href=\"http:\/\/www.sniper97.cn\/index.php\/note\/machine-learning-in-action\/1390\/\" target=\"_blank\">\u673a\u5668\u5b66\u4e60\u7b97\u6cd5\u4e0e\u8c03\u4f18<\/a>\u3002<\/p>\n\n\n<h2 class=\"wp-block-heading\">4.\u6b63\u5219\u5316<\/h2>\n\n\n<p>\u5982\u679c\u6b20\u62df\u5408\u6211\u4eec\u5f80\u5f80\u9700\u8981\u589e\u5927\u6570\u636e\u91cf\uff0c\u4f46\u662f\u5982\u679c\u5f88\u96be\u518d\u589e\u52a0\u6570\u636e\u91cf\u6216\u8005\u83b7\u53d6\u6570\u636e\u7684\u4ee3\u4ef7\u5f88\u9ad8\uff0c\u6211\u4eec\u53ef\u4ee5\u4f7f\u7528\u6b63\u5219\u5316\u6765\u51cf\u5c11\u7f51\u7edc\u7684\u8bef\u5dee\u548c\u9632\u6b62\u8fc7\u62df\u5408\u3002 <\/p>\n\n\n<p> \u6b63\u5219\u5316\u5176\u5b9e\u5c31\u662f\u9650\u5236\u6743\u91cd\u7684\u53d8\u5316\u901f\u5ea6\u3002 <\/p>\n\n\n<p> <strong>L2\u6b63\u5219\u5316<\/strong>\uff1a\u5b9e\u9645\u4e0a\u5c31\u662f\u591a\u52a0\u4e86\u4e00\u4e2a\u6743\u91cd\u77e9\u9635\u7684\u4e8c\u8303\u6570\u7684\u5e73\u65b9\u3002 \uff08 \u76ee\u7684\u662f\u8ba9\u6743\u91cd\u66f4\u52a0\u63a5\u8fd1\u539f\u70b9\uff0c\u5176\u4ed6\u53eb\u6cd5\u4e5f\u6709\u5cad\u56de\u5f52\u6216\u8005Tikhonov\u6b63\u5219\uff09<\/p>\n\n\n<div class=\"wp-block-image\"><figure class=\"aligncenter\"><img loading=\"lazy\" decoding=\"async\" width=\"617\" height=\"162\" src=\"\/wp-content\/uploads\/2020\/02\/\u56fe\u7247-46.png\" alt=\"\" class=\"wp-image-2955\" srcset=\"http:\/\/www.sniper97.cn\/wp-content\/uploads\/2020\/02\/\u56fe\u7247-46.png 617w, http:\/\/www.sniper97.cn\/wp-content\/uploads\/2020\/02\/\u56fe\u7247-46-300x79.png 300w\" sizes=\"(max-width: 617px) 100vw, 617px\" \/><\/figure><\/div>\n\n\n<p><strong> L1\u6b63\u5219\u5316<\/strong>\uff1a\u76f8\u6bd4L2\u6b63\u5219\uff0cL1\u6b63\u5219\u4f1a\u4ea7\u751f\u66f4\u52a0\u7a00\u758f\u7684\u89e3\u3002\u4e00\u822c\u88ab\u7528\u6765\u7279\u5f81\u9009\u62e9\u673a\u5236\u3002 <\/p>\n\n\n<div class=\"wp-block-image\"><figure class=\"aligncenter\"><img loading=\"lazy\" decoding=\"async\" width=\"231\" height=\"66\" src=\"\/wp-content\/uploads\/2020\/02\/\u56fe\u7247-47.png\" alt=\"\" class=\"wp-image-2956\"\/><\/figure><\/div>\n\n\n<h2 class=\"wp-block-heading\">5.Dropout\u6b63\u5219\u5316<\/h2>\n\n\n<p> \u8fd8\u6709\u4e00\u4e2a\u975e\u5e38\u6709\u6548\u7684\u6b63\u5219\u5316\u65b9\u6cd5\uff1adropout\uff08\u968f\u673a\u5931\u6d3b\uff09\u3002 <\/p>\n\n\n<h3 class=\"wp-block-heading\">5.1\u53cd\u5411\u968f\u673a\u5931\u6d3b<\/h3>\n\n\n<p> \u6211\u4eec\u8bbe\u7f6e\u4e00\u4e2akeep_prob\u53c2\u6570\uff0c\u5047\u8bbe\u503c\u4e3a0.8\uff0c\u90a3\u4e48\u610f\u5473\u7740\u8fd9\u4e2a\u795e\u7ecf\u7f51\u7edc\u5c06\u4f1a\u670920%\u7684\u795e\u7ecf\u5143\u88ab\u6d88\u53bb\uff0c80%\u795e\u7ecf\u5143\u88ab\u4fdd\u7559\u3002 <\/p>\n\n\n<pre class=\"wp-block-code\"><code>3 = np.random.rand(a3.shape[0], a3.shape[1]) &lt; keep.prob<\/code><\/pre>\n\n\n<p>\u6240\u6709\u5c0f\u4e8e\u8be5\u503c\u7684\u795e\u7ecf\u5143\u90fd\u4f1a\u88ab\u7f6e\u4e3a0\uff0c\u5176\u4f59\u7f6e\u4e3a1\u3002\n<\/p>\n\n\n<pre class=\"wp-block-code\"><code>a3 = np.multiply(a3, d3)\na3 \/= keep_prob<\/code><\/pre>\n\n\n<p>\n\u9664\u4ee5keep_prob\u5b83\u4f1a\u5927\u81f4\u77eb\u6b63\u6216\u8005\u8865\u8db3\u4f60\u4e22\u5931\u768420%\uff0c\u4ee5\u786e\u4fdda3\u7684\u671f\u671b\u503c\u4ecd\u7136\u7ef4\u6301\u5728\u540c\u4e00\u6c34\u51c6\u3002\u8fd9\u4e5f\u4f1a\u4f7f\u4f60\u5728\u6d4b\u8bd5\u65f6\u8f7b\u677e\u4e00\u70b9\uff0c\u56e0\u4e3a\u4f60\u6ca1\u6709\u589e\u52a0\u989d\u5916\u7684\u7f29\u653e\u95ee\u9898(scaling problem)\u3002\n<\/p>\n\n\n<p>\n\u6d4b\u8bd5\u65f6\u6211\u4eec\u4e0d\u9700\u8981\u6267\u884cDropout\u64cd\u4f5c\uff0c\u4f60\u6ca1\u6709\u5fc5\u8981\u968f\u673a\u5316\u4f60\u7684\u8f93\u51fa\u3002\u5982\u679c\u4f60\u5728\u6d4b\u8bd5\u9636\u6bb5\u4f7f\u7528\u4e86Dropout\uff0c\u90a3\u53ea\u4f1a\u589e\u52a0\u4f60\u9884\u6d4b\u4e0a\u7684\u566a\u97f3\u3002\u7406\u8bba\u4e0a\uff0c\u4f60\u53ef\u4ee5\u591a\u6b21\u7528\u4e0d\u7528\u7684\u9690\u85cf\u5355\u5143\u8fd0\u884c\u968f\u673a\u5316\u7684Dropout\uff0c\u4f46\u4e5f\u53ea\u4f1a\u7ed9\u4f60\u4e0d\u7528Dropout\u4e00\u6837\u7684\u7ed3\u679c\u3002\u8fd9\u4f1a\u8017\u8d39\u4e86\u4f60\u7684\u8ba1\u7b97\u6548\u7387\uff0c\u4f46\u7ed9\u4e86\u4f60\u540c\u6837\u7684\u7ed3\u679c\u3002\n<\/p>\n\n\n<p> \u5982\u4e0b\u56fe\uff0c\u5b9e\u9645\u4e0adropout\u505a\u7684\u4e8b\u60c5\u662f\u4f7f\u7ed3\u679c\u4e0d\u4f1a\u8fc7\u5206\u4f9d\u8d56\u4e0e\u67d0\u4e00\u4e2a\u7ed3\u70b9\uff0c\u56e0\u4e3a\u4ed6\u4eec\u90fd\u4f1a\u968f\u673a\u5931\u6548\uff0c\u4e5f\u5c31\u9632\u6b62\u4e86\u8fc7\u62df\u5408\u7684\u53d1\u751f\uff0c\u4e00\u822c\u7528\u5728\u8ba1\u7b97\u673a\u89c6\u89c9\u4e0a\u504f\u591a\uff0c\u56e0\u4e3a\u7f51\u7edc\u5e38\u5e38\u6709\u8fc7\u62df\u5408\u7684\u95ee\u9898\u3002 <\/p>\n\n\n<div class=\"wp-block-image\"><figure class=\"aligncenter\"><img loading=\"lazy\" decoding=\"async\" width=\"219\" height=\"254\" src=\"\/wp-content\/uploads\/2020\/02\/\u56fe\u7247-48.png\" alt=\"\" class=\"wp-image-2957\"\/><\/figure><\/div>\n\n\n<h2 class=\"wp-block-heading\">6.\u5176\u4ed6\u6b63\u5219\u5316<\/h2>\n\n\n<p><strong> early stop <\/strong><\/p>\n\n\n<p>\n\u7531\u4e8e\u521a\u5f00\u59cb\u521d\u59cb\u5316\u7684\u65f6\u5019w\u5f88\u5c0f\uff0c\u968f\u7740\u8fed\u4ee3\u7684\u6b21\u6570\u901a\u8fc7\u89c2\u5bdf\u8bad\u7ec3\u96c6\u4ee3\u4ef7\u548c\u9a8c\u8bc1\u96c6\u4ee3\u4ef7\uff0c\u5f53\u53d1\u73b0\u9a8c\u8bc1\u96c6\u4ee3\u4ef7\u5f00\u59cb\u5347\u9ad8\u7684\u65f6\u5019\uff0c\u8bf4\u660e\u63a5\u4e0b\u6765\u7684\u8fed\u4ee3\u5c31\u53ef\u80fd\u662f\u5f00\u59cb\u5bf9\u8bad\u7ec3\u96c6\u8fdb\u884c\u8fc7\u62df\u5408\u4e86\uff0c\u8fd9\u65f6\u5019\u65e9\u65e9\u7684\u7ed3\u675f\u7b97\u6cd5\u83b7\u5f97\u7684\u5f80\u5f80\u662f\u4e00\u4e2a\u6bd4\u8f83\u597d\u7684\u7ed3\u679c\u3002\n<\/p>\n\n\n<div class=\"wp-block-image\"><figure class=\"aligncenter\"><img loading=\"lazy\" decoding=\"async\" width=\"862\" height=\"396\" src=\"\/wp-content\/uploads\/2020\/02\/\u56fe\u7247-49.png\" alt=\"\" class=\"wp-image-2959\" srcset=\"http:\/\/www.sniper97.cn\/wp-content\/uploads\/2020\/02\/\u56fe\u7247-49.png 862w, http:\/\/www.sniper97.cn\/wp-content\/uploads\/2020\/02\/\u56fe\u7247-49-300x138.png 300w, http:\/\/www.sniper97.cn\/wp-content\/uploads\/2020\/02\/\u56fe\u7247-49-768x353.png 768w\" sizes=\"(max-width: 862px) 100vw, 862px\" \/><\/figure><\/div>\n\n\n<h2 class=\"wp-block-heading\">7.\u5f52\u4e00\u5316\u8f93\u5165<\/h2>\n\n\n<p><strong> \u6807\u51c6\u5316\u8bad\u7ec3\u96c6&nbsp; <\/strong><\/p>\n\n\n<p>\u5c06\u8bad\u7ec3\u96c6\u7684\u503c\u5747\u5300\u7684\u6563\u5217\u5728\u8fdc\u70b9\u9644\u8fd1\uff08\u7c7b\u4f3c\u4e8e\u505a\u6b63\u6001\u5206\u5e03\uff09\u3002<\/p>\n\n\n<div class=\"wp-block-image\"><figure class=\"aligncenter\"><img loading=\"lazy\" decoding=\"async\" width=\"918\" height=\"446\" src=\"\/wp-content\/uploads\/2020\/02\/\u56fe\u7247-50.png\" alt=\"\" class=\"wp-image-2960\" srcset=\"http:\/\/www.sniper97.cn\/wp-content\/uploads\/2020\/02\/\u56fe\u7247-50.png 918w, http:\/\/www.sniper97.cn\/wp-content\/uploads\/2020\/02\/\u56fe\u7247-50-300x146.png 300w, http:\/\/www.sniper97.cn\/wp-content\/uploads\/2020\/02\/\u56fe\u7247-50-768x373.png 768w\" sizes=\"(max-width: 918px) 100vw, 918px\" \/><\/figure><\/div>\n\n\n<h2 class=\"wp-block-heading\">8.\u68af\u5ea6\u6d88\u5931\u548c\u68af\u5ea6\u7206\u70b8<\/h2>\n\n\n<p> \u68af\u5ea6\u6d88\u5931\u662f\u5728\u4f7f\u7528\u4f8b\u5982sigmoid\u6fc0\u6d3b\u51fd\u6570\u8fd9\u7c7b\uff0c\u5f53\u7ed3\u679c\u8f83\u5927\u6216\u8005\u8f83\u5c0f\u65f6\uff0c\u5bfc\u6570\u8d8b\u8fd1\u4e8e0\uff0c\u8fd9\u6837\u5728\u4f20\u64ad\u7684\u65f6\u5019\uff0c\u968f\u7740\u7f51\u7edc\u7684\u52a0\u6df1\uff0cz\u7684\u7edd\u5bf9\u503c\u53d8\u5927\uff0c\u5bfc\u6570\u5f00\u59cb\u8d8b\u8fd1\u4e8e0\uff0c\u8fd9\u6837\u53cd\u5411\u4f20\u64ad\u7684\u65f6\u5019\u65e0\u6cd5\u88ab\u5feb\u901f\u66f4\u65b0\uff0c\u5bfc\u81f4\u795e\u7ecf\u7f51\u7edc\u65e0\u6cd5\u88ab\u5feb\u901f\u4f18\u5316\u751a\u81f3\u6c38\u8fdc\u4e0d\u4f1a\u6536\u655b\u3002 <\/p>\n\n\n<p> \u68af\u5ea6\u7206\u70b8\u5219\u76f8\u53cd\uff0c\u968f\u7740\u7f51\u7edc\u7684\u8d8a\u6765\u8d8a\u6df1\uff0cw\u7684\u503c\u8d8a\u8001\u8d8a\u5927\uff0c\u53ef\u80fd\u4f1a\u5bfc\u81f4\u4e00\u4e2a\u975e\u5e38\u5927\u7684\u6743\u91cd\u66f4\u65b0\u3002 <\/p>\n\n\n<p>\n\u5bf9\u4e8e\u68af\u5ea6\u6d88\u5931\u7684\u89e3\u51b3\u529e\u6cd5\u901a\u5e38\u66f4\u597d\u7684\u53c2\u6570\u521d\u59cb\u5316\u3001\u4f7f\u7528ReLU\u4ee3\u66ffSigmoid\u7b49\uff0c\u800c\u89e3\u51b3\u68af\u5ea6\u7206\u70b8\u7684\u65b9\u6cd5\u53ef\u4ee5\u6709\u68af\u5ea6\u9636\u6bb5\u3001\u66f4\u5feb\u7684\u4f18\u5316\u5668\u6216\u8005\u4f7f\u7528LSTM\u7b49\u3002\n<\/p>\n\n\n<h2 class=\"wp-block-heading\">9.\u68af\u5ea6\u7684\u6570\u503c\u903c\u8fd1\uff08\u53cc\u8fb9\u8bef\u5dee\u516c\u5f0f\uff09<\/h2>\n\n\n<p> \u53cc\u8fb9\u8bef\u5dee\u5c31\u662f\u66f4\u6539\u4e00\u4e0b\u6c42\u5bfc\u65b9\u6cd5\uff0c\u4ee5\u8fbe\u5230\u66f4\u4f18\u79c0\u7684\u7cbe\u5ea6\u3002 <\/p>\n\n\n<p> \u5177\u4f53\u63a8\u5bfc\u5982\u4e0b\uff1a <\/p>\n\n\n<p> \u6211\u4eec\u77e5\u9053\u5bfc\u6570\u7684\u5b9a\u4e49 <\/p>\n\n\n<div class=\"wp-block-image\"><figure class=\"aligncenter\"><img loading=\"lazy\" decoding=\"async\" width=\"238\" height=\"73\" src=\"\/wp-content\/uploads\/2020\/02\/\u56fe\u7247-51.png\" alt=\"\" class=\"wp-image-2961\"\/><\/figure><\/div>\n\n\n<p> \u4f7f\u7528\u6cf0\u52d2\u5c55\u5f00\u7684\u8bdd\u5982\u4e0b <\/p>\n\n\n<div class=\"wp-block-image\"><figure class=\"aligncenter\"><img loading=\"lazy\" decoding=\"async\" width=\"308\" height=\"50\" src=\"\/wp-content\/uploads\/2020\/02\/\u56fe\u7247-52.png\" alt=\"\" class=\"wp-image-2962\" srcset=\"http:\/\/www.sniper97.cn\/wp-content\/uploads\/2020\/02\/\u56fe\u7247-52.png 308w, http:\/\/www.sniper97.cn\/wp-content\/uploads\/2020\/02\/\u56fe\u7247-52-300x49.png 300w\" sizes=\"(max-width: 308px) 100vw, 308px\" \/><\/figure><\/div>\n\n\n<div class=\"wp-block-image\"><figure class=\"aligncenter\"><img loading=\"lazy\" decoding=\"async\" width=\"296\" height=\"50\" src=\"\/wp-content\/uploads\/2020\/02\/\u56fe\u7247-53.png\" alt=\"\" class=\"wp-image-2963\"\/><\/figure><\/div>\n\n\n<p> \u6211\u4eec\u53ef\u4ee5\u5f97\u51fa\u8bef\u5dee\u9879\u4e0eh\u540c\u9636\u3002 <\/p>\n\n\n<p>\n\u5982\u679c\u6211\u4eec\u4f7f\u7528\u4e24\u4e2a\u5bfc\u6570\n<\/p>\n\n\n<div class=\"wp-block-image\"><figure class=\"aligncenter\"><img loading=\"lazy\" decoding=\"async\" width=\"445\" height=\"160\" src=\"\/wp-content\/uploads\/2020\/02\/\u56fe\u7247-54.png\" alt=\"\" class=\"wp-image-2964\" srcset=\"http:\/\/www.sniper97.cn\/wp-content\/uploads\/2020\/02\/\u56fe\u7247-54.png 445w, http:\/\/www.sniper97.cn\/wp-content\/uploads\/2020\/02\/\u56fe\u7247-54-300x108.png 300w\" sizes=\"(max-width: 445px) 100vw, 445px\" \/><\/figure><\/div>\n\n\n<p> \u6211\u4eec\u53d1\u73b0\u73b0\u5728\u7684\u8bef\u5dee\u9879\u53d8\u4e3ah\u65b9\u540c\u9636\uff0c\u7cbe\u5ea6\u5f97\u5230\u4e86\u63d0\u5347\u3002 <\/p>\n\n\n<h2 class=\"wp-block-heading\">10.\u68af\u5ea6\u68c0\u9a8c<\/h2>\n\n\n<p> \u795e\u7ecf\u7f51\u7edc\u7b97\u6cd5\u4f7f\u7528\u53cd\u5411\u4f20\u64ad\u8ba1\u7b97\u76ee\u6807\u51fd\u6570\u5173\u4e8e\u6bcf\u4e2a\u53c2\u6570\u7684\u68af\u5ea6\uff0c\u53ef\u4ee5\u770b\u505a\u89e3\u6790\u68af\u5ea6\u3002\u7531\u4e8e\u8ba1\u7b97\u8fc7\u7a0b\u4e2d\u6d89\u53ca\u5230\u7684\u53c2\u6570\u5f88\u591a\uff0c\u53cd\u5411\u4f20\u64ad\u8ba1\u7b97\u7684\u68af\u5ea6\u5f88\u5bb9\u6613\u51fa\u73b0\u8bef\u5dee\uff0c\u5bfc\u81f4\u6700\u540e\u8fed\u4ee3\u5f97\u5230\u6548\u679c\u5f88\u5dee\u7684\u53c2\u6570\u503c\u3002 <\/p>\n\n\n<p> \u4e3a\u4e86\u786e\u8ba4\u4ee3\u7801\u4e2d\u53cd\u5411\u4f20\u64ad\u8ba1\u7b97\u7684\u68af\u5ea6\u662f\u5426\u6b63\u786e\uff0c\u53ef\u4ee5\u91c7\u7528\u68af\u5ea6\u68c0\u9a8c\uff08gradient check\uff09\u7684\u65b9\u6cd5\u3002\u901a\u8fc7\u8ba1\u7b97\u6570\u503c\u68af\u5ea6\uff0c\u5f97\u5230\u68af\u5ea6\u7684\u8fd1\u4f3c\u503c\uff0c\u7136\u540e\u548c\u53cd\u5411\u4f20\u64ad\u5f97\u5230\u7684\u68af\u5ea6\u8fdb\u884c\u6bd4\u8f83\uff0c\u82e5\u4e24\u8005\u76f8\u5dee\u5f88\u5c0f\u7684\u8bdd\u5219\u8bc1\u660e\u53cd\u5411\u4f20\u64ad\u7684\u4ee3\u7801\u662f\u6b63\u786e\u65e0\u8bef\u7684\u3002 <\/p>\n\n\n<div class=\"wp-block-image\"><figure class=\"aligncenter\"><img loading=\"lazy\" decoding=\"async\" width=\"983\" height=\"212\" src=\"\/wp-content\/uploads\/2020\/02\/\u56fe\u7247-55.png\" alt=\"\" class=\"wp-image-2965\" srcset=\"http:\/\/www.sniper97.cn\/wp-content\/uploads\/2020\/02\/\u56fe\u7247-55.png 983w, http:\/\/www.sniper97.cn\/wp-content\/uploads\/2020\/02\/\u56fe\u7247-55-300x65.png 300w, http:\/\/www.sniper97.cn\/wp-content\/uploads\/2020\/02\/\u56fe\u7247-55-768x166.png 768w\" sizes=\"(max-width: 983px) 100vw, 983px\" \/><\/figure><\/div>\n\n\n<p> \u4e00\u822c\u6211\u4eec\u8ba4\u4e3a\u5dee\u8ddd10^-7\u662f\u6bd4\u8f83\u597d\u7684\uff0c10^-5\u4e00\u822c\uff0c\u53ef\u80fd\u4ea7\u751fbug\uff0c\u800c10^-3\u53ef\u80fd\u5c31\u8981\u4ed4\u7ec6\u68c0\u67e5\u6709\u6ca1\u6709bug\u7684\u4ea7\u751f\u3002 <\/p>\n\n\n<p> \u9700\u8981\u6ce8\u610f\u7684\u662f\uff1a <\/p>\n\n\n<ul><li> \u4e0d\u8981\u5728\u8bad\u7ec3\u4e2d\u4f7f\u7528\u68af\u5ea6\u68c0\u9a8c\uff0c\u4ec5\u4ec5\u662f\u5728\u8c03\u8bd5\u9636\u6bb5\u4f7f\u7528\uff0c\u56e0\u4e3a\u5b83\u592a\u6162\u4e86 <\/li><li> \u5982\u679c\u68c0\u9a8c\u4e0d\u5408\u683c\uff0c\u8981\u68c0\u67e5\u6bcf\u4e00\u9879\u7684\u503c\u4ee5\u786e\u5b9abug\u7684\u4f4d\u7f6e <\/li><li> \u5728\u5b9e\u65bd\u68af\u5ea6\u68c0\u9a8c\u7684\u65f6\u5019\u4e0d\u8981\u5fd8\u8bb0\u4f7f\u7528\u6b63\u5219\u5316 <\/li><li> \u4e0d\u8981\u548cdropout\u540c\u65f6\u4f7f\u7528 <\/li><\/ul>\n\n\n<h2 class=\"wp-block-heading\">\u8bfe\u540e\u6d4b\u9a8c<\/h2>\n\n\n<p><strong>1. \u5982\u679c\u4f60\u670910,000,000\u4e2a\u4f8b\u5b50\uff0c\u4f60\u4f1a\u5982\u4f55\u5212\u5206\u8bad\u7ec3\/\u5f00\u53d1\/\u6d4b\u8bd5\u96c6\uff1f <\/strong><\/p>\n\n\n<p> \u8bad\u7ec3\u96c6\u536098% \uff0c \u5f00\u53d1\u96c6\u53601% \uff0c \u6d4b\u8bd5\u96c6\u53601% \u3002 <\/p>\n\n\n<p><strong>2. \u5f00\u53d1\u548c\u6d4b\u8bd5\u96c6\u5e94\u8be5\uff1a <\/strong><\/p>\n\n\n<p> \u6765\u81ea\u540c\u4e00\u5206\u5e03\u3002 <\/p>\n\n\n<p><strong>3. \u5982\u679c\u4f60\u7684\u795e\u7ecf\u7f51\u7edc\u6a21\u578b\u4f3c\u4e4e\u6709\u5f88\u9ad8\u7684\u65b9\u5dee\uff0c\u4e0b\u5217\u54ea\u4e2a\u5c1d\u8bd5\u662f\u53ef\u80fd\u89e3\u51b3\u95ee\u9898\u7684\uff1f <\/strong><\/p>\n\n\n<p> \u6dfb\u52a0\u6b63\u5219\u5316\uff0c \u83b7\u53d6\u66f4\u591a\u7684\u8bad\u7ec3\u6570\u636e \u3002<\/p>\n\n\n<p><strong>4. \u4f60\u5728\u4e00\u5bb6\u8d85\u5e02\u7684\u81ea\u52a8\u7ed3\u5e10\u4ead\u5de5\u4f5c\uff0c\u6b63\u5728\u4e3a\u82f9\u679c\uff0c\u9999\u8549\u548c\u6a58\u5b50\u5236\u4f5c\u5206\u7c7b\u5668\u3002 \u5047\u8bbe\u60a8\u7684\u5206\u7c7b\u5668\u5728\u8bad\u7ec3\u96c6\u4e0a\u67090.5\uff05\u7684\u9519\u8bef\uff0c\u4ee5\u53ca\u5f00\u53d1\u96c6\u4e0a\u67097\uff05\u7684\u9519\u8bef\u3002 \u4ee5\u4e0b\u54ea\u9879\u5c1d\u8bd5\u662f\u6709\u5e0c\u671b\u6539\u5584\u4f60\u7684\u5206\u7c7b\u5668\u7684\u5206\u7c7b\u6548\u679c\u7684\uff1f <\/strong><\/p>\n\n\n<p> \u589e\u52a0\u6b63\u5219\u5316\u53c2\u6570lambda\uff0c \u83b7\u53d6\u66f4\u591a\u7684\u8bad\u7ec3\u6570\u636e \u3002<\/p>\n\n\n<p><strong>5. \u4ec0\u4e48\u662f\u6743\u91cd\u8870\u51cf\uff1f <\/strong><\/p>\n\n\n<p> \u6b63\u5219\u5316\u6280\u672f\uff08\u4f8b\u5982L2\u6b63\u5219\u5316\uff09\u5bfc\u81f4\u68af\u5ea6\u4e0b\u964d\u5728\u6bcf\u6b21\u8fed\u4ee3\u65f6\u6743\u91cd\u6536\u7f29\u3002<\/p>\n\n\n<p><strong> 6.\u5f53\u4f60\u589e\u52a0\u6b63\u5219\u5316\u8d85\u53c2\u6570lambda\u65f6\u4f1a\u53d1\u751f\u4ec0\u4e48\uff1f  <\/strong><\/p>\n\n\n<p> \u6743\u91cd\u4f1a\u53d8\u5f97\u66f4\u5c0f\uff08\u63a5\u8fd10\uff09 <\/p>\n\n\n<p><strong>7. \u5728\u6d4b\u8bd5\u65f6\u5019\u4f7f\u7528dropout\u65f6<\/strong><\/p>\n\n\n<p> \u4e0d\u8981\u968f\u673a\u6d88\u9664\u8282\u70b9\uff0c\u4e5f\u4e0d\u8981\u5728\u8bad\u7ec3\u4e2d\u4f7f\u7528\u7684\u8ba1\u7b97\u4e2d\u4fdd\u75591 \/ keep_prob\u56e0\u5b50\u3002<\/p>\n\n\n<p><strong>8. \u5c06\u53c2\u6570keep_prob\u4ece\uff08\u6bd4\u5982\u8bf4\uff090.5\u589e\u52a0\u52300.6\u53ef\u80fd\u4f1a\u5bfc\u81f4\u4ee5\u4e0b\u60c5\u51b5 <\/strong><\/p>\n\n\n<p> \u6b63\u5219\u5316\u6548\u5e94\u88ab\u51cf\u5f31\u3001\u4f7f\u795e\u7ecf\u7f51\u7edc\u5728\u7ed3\u675f\u65f6\u4f1a\u5728\u8bad\u7ec3\u96c6\u4e0a\u8868\u73b0\u597d\u4e00\u4e9b\u3002 <\/p>\n\n\n<p><strong>9. \u4ee5\u4e0b\u54ea\u4e9b\u6280\u672f\u53ef\u7528\u4e8e\u51cf\u5c11\u65b9\u5dee\uff08\u51cf\u5c11\u8fc7\u62df\u5408\uff09\uff1a <\/strong><\/p>\n\n\n<p> Dropout\u3001 L2 \u6b63\u5219\u5316 \u3001 \u6269\u5145\u6570\u636e\u96c6 \u3002<\/p>\n\n\n<p><strong>10.\u4e3a\u4ec0\u4e48\u6211\u4eec\u8981\u5f52\u4e00\u5316\u8f93\u5165x\uff1f <\/strong><\/p>\n\n\n<p> \u5b83\u4f7f\u6210\u672c\u51fd\u6570\u66f4\u5feb\u5730\u8fdb\u884c\u4f18\u5316\u3002<\/p>\n\n\n<h2 class=\"wp-block-heading\">\u7f16\u7a0b\u4f5c\u4e1a<\/h2>\n\n\n<p>\u8fd9\u5468\u7684\u7f16\u7a0b\u4f5c\u4e1a\uff0c\u6211\u4eec\u9700\u8981\u505a\u521d\u59cb\u5316\u3001\u6b63\u5219\u5316\u4ee5\u53ca\u68af\u5ea6\u6821\u9a8c\u7684\u4ee3\u7801\u3002<\/p>\n\n\n<p>\u9996\u5148\u6211\u4eec\u5bfc\u5305<\/p>\n\n\n<pre class=\"wp-block-preformatted\">import numpy as np<br \/>import matplotlib.pyplot as plt<br \/>import sklearn<br \/>import sklearn.datasets<br \/>from course_2_week_1 import init_utils  # \u7b2c\u4e00\u90e8\u5206\uff0c\u521d\u59cb\u5316<br \/>from course_2_week_1 import reg_utils  # \u7b2c\u4e8c\u90e8\u5206\uff0c\u6b63\u5219\u5316<br \/>from course_2_week_1 import gc_utils  # \u7b2c\u4e09\u90e8\u5206\uff0c\u68af\u5ea6\u6821\u9a8c<\/pre>\n\n\n<p>\u6211\u4eec\u521d\u59cb\u5316\u4e00\u4e0b\u6570\u636e\uff0c\u7136\u540e\u6253\u5370\u51fa\u6765\u770b\u4e00\u4e0b\u6570\u636e\u6548\u679c\u3002<\/p>\n\n\n<pre class=\"wp-block-preformatted\">train_X, train_Y, test_X, test_Y = init_utils.load_dataset(is_plot=True)<br \/>plt.show()<\/pre>\n\n\n<div class=\"wp-block-image\"><figure class=\"aligncenter\"><img loading=\"lazy\" decoding=\"async\" width=\"492\" height=\"274\" src=\"\/wp-content\/uploads\/2020\/02\/\u56fe\u7247-57.png\" alt=\"\" class=\"wp-image-2978\" srcset=\"http:\/\/www.sniper97.cn\/wp-content\/uploads\/2020\/02\/\u56fe\u7247-57.png 492w, http:\/\/www.sniper97.cn\/wp-content\/uploads\/2020\/02\/\u56fe\u7247-57-300x167.png 300w\" sizes=\"(max-width: 492px) 100vw, 492px\" \/><\/figure><\/div>\n\n\n<p>\u6211\u4eec\u5728\u8fd9\u91cc\u5c1d\u8bd5\u4e09\u79cd\u53c2\u6570\u521d\u59cb\u5316\u65b9\u6cd5\uff0c\u5168\u90e8\u521d\u59cb\u5316\u4e3a0\u3001\u968f\u673a\u521d\u59cb\u5316\u4e0e\u6291\u68af\u5ea6\u5f02\u5e38\u521d\u59cb\u5316\u3002<\/p>\n\n\n<p>\u5728\u8fd9\u4e4b\u524d\uff0c\u6211\u4eec\u5148\u628a\u6a21\u578b\u642d\u5efa\u597d\uff08\u65b9\u6cd5\u4ee5\u540e\u518d\u8bf4\uff0c\u89e3\u91ca\u6027\u8bed\u8a00\u5c31\u662f\u597d\u554ax\uff09\uff0c\u7136\u540e\u4f7f\u7528\u4e0d\u540c\u7684\u53c2\u6570\u521d\u59cb\u5316\u65b9\u6cd5\u8fdb\u884c\u5bf9\u6bd4\uff08\u5176\u4ed6\u7684\u524d\u5411\u53cd\u5411\u4f20\u64ad\u7b49\u65b9\u6cd5\u5747\u5728\u5de5\u5177\u7c7b\u4e2d\u5b9e\u73b0\u597d\u4e86\uff09\u3002<\/p>\n\n\n<p>\uff09\uff0c\u7136\u540e\u4f7f\u7528\u4e0d\u540c\u7684\u53c2\u6570\u521d\u59cb\u5316\u65b9\u6cd5\u8fdb\u884c\u5bf9\u6bd4\uff08\u5176\u4ed6\u7684\u524d\u5411\u53cd\u5411\u4f20\u64ad\u7b49\u65b9\u6cd5\u5747\u5728\u5de5\u5177\u7c7b\u4e2d\u5b9e\u73b0\u597d\u4e86\uff09\u3002<\/p>\n\n\n<pre class=\"wp-block-preformatted\">def model(X, Y, learning_rate=0.01, num_iteration=15000, print_cost=True, initialization=\"he\", is_polt=True):<br \/>    <em>\"\"\"<br \/><\/em><em>    <\/em><em>\u795e\u7ecf\u7f51\u7edc\u6a21\u578b<br \/><\/em><em>    <\/em><strong><em>:param<\/em><\/strong><em> X:<br \/><\/em><em>    <\/em><strong><em>:param<\/em><\/strong><em> Y:<br \/><\/em><em>    <\/em><strong><em>:param<\/em><\/strong><em> learning_rate:<br \/><\/em><em>    <\/em><strong><em>:param<\/em><\/strong><em> num_iteration:<br \/><\/em><em>    <\/em><strong><em>:param<\/em><\/strong><em> print_cost:<br \/><\/em><em>    <\/em><strong><em>:param<\/em><\/strong><em> initialization:<br \/><\/em><em>    <\/em><strong><em>:param<\/em><\/strong><em> is_polt:<br \/><\/em><em>    <\/em><strong><em>:return<\/em><\/strong><em>:<br \/><\/em><em>    \"\"\"<br \/><\/em><em>    <\/em>grads = {}<br \/>    costs = []<br \/>    # \u8bad\u7ec3\u96c6\u4e2a\u6570<br \/>    m = X.shape[1]<br \/>    # \u4e09\u5c42\u795e\u7ecf\u7f51\u7edc\u7684\u7ed3\u6784<br \/>    layers_dims = [X.shape[0], 10, 5, 1]<br \/><br \/>    if initialization == 'zeros':<br \/>        parameters = initialize_parameters_zeros(layers_dims)<br \/>    elif initialization == 'random':<br \/>        parameters = initialize_parameters_random(layers_dims)<br \/>    elif initialization == \"he\":<br \/>        parameters = initialize_parameters_he(layers_dims)<br \/>    else:<br \/>        print(\"\u9519\u8bef\u7684\u521d\u59cb\u5316\u53c2\u6570\uff01\u7a0b\u5e8f\u9000\u51fa\")<br \/>        exit<br \/><br \/>    # \u5f00\u59cb\u5b66\u4e60<br \/>    for i in range(num_iteration):<br \/>        a3, cache = init_utils.forward_propagation(X, parameters)<br \/><br \/>        cost = init_utils.compute_loss(a3, Y)<br \/><br \/>        grads = init_utils.backward_propagation(X, Y, cache)<br \/><br \/>        parameters = init_utils.update_parameters(parameters, grads, learning_rate)<br \/><br \/>        if i % 1000 == 0:<br \/>            costs.append(cost)<br \/>            if print_cost:<br \/>                print(\"\u7b2c\" + str(i) + \"\u6b21\u8fed\u4ee3\uff0c\u6210\u672c\u503c\u4e3a\uff1a\" + str(cost))<br \/><br \/>    # \u5b66\u4e60\u5b8c\u6bd5\uff0c\u7ed8\u5236\u6210\u672c\u66f2\u7ebf<br \/>    if is_polt:<br \/>        plt.plot(costs)<br \/>        plt.ylabel('cost')<br \/>        plt.xlabel('iterations (per hundreds)')<br \/>        plt.title(\"Learning rate =\" + str(learning_rate))<br \/>        plt.show()<br \/><br \/>    # \u8fd4\u56de\u5b66\u4e60\u5b8c\u6bd5\u540e\u7684\u53c2\u6570<br \/>    return parameters<\/pre>\n\n\n<p>\u9996\u5148\u6211\u4eec\u5c1d\u8bd5\u4f7f\u7528\u521d\u59cb\u53160\uff1a<\/p>\n\n\n<pre class=\"wp-block-preformatted\">def initialize_parameters_zeros(layers_dims):<br \/>    <em>\"\"\"<br \/><\/em><em>    <\/em><em>\u6a21\u578b\u7684\u53c2\u6570\u5168\u90e8\u521d\u59cb\u5316\u4e3a<\/em><em>0<br \/><\/em><em>    <\/em><strong><em>:param<\/em><\/strong><em> layers_dims:<br \/><\/em><em>    <\/em><strong><em>:return<\/em><\/strong><em>:<br \/><\/em><em>    \"\"\"<br \/><\/em><em>    <\/em>parameters = {}<br \/><br \/>    # \u7f51\u7edc\u5c42\u6570<br \/>    L = len(layers_dims)<br \/><br \/>    for i in range(1, L):<br \/>        parameters['W' + str(i)] = np.zeros((layers_dims[i], layers_dims[i - 1]))<br \/>        parameters['b' + str(i)] = np.zeros((layers_dims[i], 1))<br \/><br \/>    return parameters<\/pre>\n\n\n<p>\u7136\u540e\u6211\u4eec\u8fd0\u884c\u6a21\u578b<\/p>\n\n\n<pre class=\"wp-block-preformatted\">parameters = model(train_X, train_Y, initialization=\"zero\", is_polt=True)\nprint(\"\u8bad\u7ec3\u96c6:\")\npredictions_train = init_utils.predict(train_X, train_Y, parameters)\nprint(\"\u6d4b\u8bd5\u96c6:\")\npredictions_test = init_utils.predict(test_X, test_Y, parameters)<\/pre>\n\n\n<p>\u8f93\u51fa\u5982\u4e0b\uff1a<\/p>\n\n\n<pre class=\"wp-block-code\"><code>\u7b2c0\u6b21\u8fed\u4ee3\uff0c\u6210\u672c\u503c\u4e3a\uff1a0.6931471805599453\n\u7b2c1000\u6b21\u8fed\u4ee3\uff0c\u6210\u672c\u503c\u4e3a\uff1a0.6931471805599453\n\u7b2c2000\u6b21\u8fed\u4ee3\uff0c\u6210\u672c\u503c\u4e3a\uff1a0.6931471805599453\n\u7b2c3000\u6b21\u8fed\u4ee3\uff0c\u6210\u672c\u503c\u4e3a\uff1a0.6931471805599453\n\u7b2c4000\u6b21\u8fed\u4ee3\uff0c\u6210\u672c\u503c\u4e3a\uff1a0.6931471805599453\n\u7b2c5000\u6b21\u8fed\u4ee3\uff0c\u6210\u672c\u503c\u4e3a\uff1a0.6931471805599453\n\u7b2c6000\u6b21\u8fed\u4ee3\uff0c\u6210\u672c\u503c\u4e3a\uff1a0.6931471805599453\n\u7b2c7000\u6b21\u8fed\u4ee3\uff0c\u6210\u672c\u503c\u4e3a\uff1a0.6931471805599453\n\u7b2c8000\u6b21\u8fed\u4ee3\uff0c\u6210\u672c\u503c\u4e3a\uff1a0.6931471805599453\n\u7b2c9000\u6b21\u8fed\u4ee3\uff0c\u6210\u672c\u503c\u4e3a\uff1a0.6931471805599453\n\u7b2c10000\u6b21\u8fed\u4ee3\uff0c\u6210\u672c\u503c\u4e3a\uff1a0.6931471805599455\n\u7b2c11000\u6b21\u8fed\u4ee3\uff0c\u6210\u672c\u503c\u4e3a\uff1a0.6931471805599453\n\u7b2c12000\u6b21\u8fed\u4ee3\uff0c\u6210\u672c\u503c\u4e3a\uff1a0.6931471805599453\n\u8bad\u7ec3\u96c6:\nAccuracy: 0.5\n\u6d4b\u8bd5\u96c6:\nAccuracy: 0.5<\/code><\/pre>\n\n\n<p>\u53ef\u4ee5\u770b\u5230\u968f\u7740\u8fed\u4ee3\u6b21\u6570\u7684\u589e\u52a0\uff0c\u4ee3\u4ef7\u6ca1\u6709\u4efb\u4f55\u53d8\u5316\u3002<\/p>\n\n\n<div class=\"wp-block-image\"><figure class=\"aligncenter\"><img loading=\"lazy\" decoding=\"async\" width=\"489\" height=\"277\" src=\"\/wp-content\/uploads\/2020\/02\/\u56fe\u7247-58.png\" alt=\"\" class=\"wp-image-2979\" srcset=\"http:\/\/www.sniper97.cn\/wp-content\/uploads\/2020\/02\/\u56fe\u7247-58.png 489w, http:\/\/www.sniper97.cn\/wp-content\/uploads\/2020\/02\/\u56fe\u7247-58-300x170.png 300w\" sizes=\"(max-width: 489px) 100vw, 489px\" \/><\/figure><\/div>\n\n\n<p>\u5c31\u50cf\u4e4b\u524d\u8bf4\u7684\uff0c\u5f53\u53c2\u6570\u521d\u59cb\u5316\u4e3a0\u7684\u65f6\u5019\uff0c\u795e\u7ecf\u7f51\u7edc\u5c31\u9000\u5316\u6210\u903b\u8f91\u56de\u5f52\u4e86\u3002<\/p>\n\n\n<p>\u7136\u540e\u662f\u4f7f\u7528\u968f\u673a\u521d\u59cb\u5316\uff1a<\/p>\n\n\n<pre class=\"wp-block-preformatted\">def initialize_parameters_random(layers_dims):<br \/>    <em>\"\"\"<br \/><\/em><em>    <\/em><em>\u6a21\u578b\u968f\u673a\u521d\u59cb\u5316\u53c2\u6570<br \/><\/em><em>    <\/em><strong><em>:param<\/em><\/strong><em> layers_dims:<br \/><\/em><em>    <\/em><strong><em>:return<\/em><\/strong><em>:<br \/><\/em><em>    \"\"\"<br \/><\/em><em>    <\/em>np.random.seed(3)<br \/><br \/>    parameters = {}<br \/><br \/>    L = len(layers_dims)<br \/><br \/>    for i in range(1, L):<br \/>        parameters['W' + str(i)] = np.random.randn(layers_dims[i], layers_dims[i - 1]) * 10<br \/>        parameters['b' + str(i)] = np.zeros((layers_dims[i], 1))<br \/><br \/>    return parameters<br \/><\/pre>\n\n\n<p>\u7136\u540e\u8c03\u7528\u8f93\u51fa\u5e76\u7ed8\u5236\u51b3\u7b56\u8fb9\u754c\u3002<\/p>\n\n\n<pre class=\"wp-block-preformatted\">parameters = model(train_X, train_Y, initialization=\"random\", is_polt=True)\nprint(\"\u8bad\u7ec3\u96c6:\")\npredictions_train = init_utils.predict(train_X, train_Y, parameters)\nprint(\"\u6d4b\u8bd5\u96c6:\")\npredictions_test = init_utils.predict(test_X, test_Y, parameters)\nplt.title(\"Model with large random initialization\")\naxes = plt.gca()\naxes.set_xlim([-1.5, 1.5])\naxes.set_ylim([-1.5, 1.5])\ninit_utils.plot_decision_boundary(lambda x: init_utils.predict_dec(parameters, x.T), train_X, train_Y)<\/pre>\n\n\n<pre class=\"wp-block-code\"><code>\u7b2c0\u6b21\u8fed\u4ee3\uff0c\u6210\u672c\u503c\u4e3a\uff1ainf\n\u7b2c1000\u6b21\u8fed\u4ee3\uff0c\u6210\u672c\u503c\u4e3a\uff1a0.6250982793959966\n\u7b2c2000\u6b21\u8fed\u4ee3\uff0c\u6210\u672c\u503c\u4e3a\uff1a0.5981216596703697\n\u7b2c3000\u6b21\u8fed\u4ee3\uff0c\u6210\u672c\u503c\u4e3a\uff1a0.5638417572298645\n\u7b2c4000\u6b21\u8fed\u4ee3\uff0c\u6210\u672c\u503c\u4e3a\uff1a0.5501703049199763\n\u7b2c5000\u6b21\u8fed\u4ee3\uff0c\u6210\u672c\u503c\u4e3a\uff1a0.5444632909664456\n\u7b2c6000\u6b21\u8fed\u4ee3\uff0c\u6210\u672c\u503c\u4e3a\uff1a0.5374513807000807\n\u7b2c7000\u6b21\u8fed\u4ee3\uff0c\u6210\u672c\u503c\u4e3a\uff1a0.4764042074074983\n\u7b2c8000\u6b21\u8fed\u4ee3\uff0c\u6210\u672c\u503c\u4e3a\uff1a0.39781492295092263\n\u7b2c9000\u6b21\u8fed\u4ee3\uff0c\u6210\u672c\u503c\u4e3a\uff1a0.3934764028765484\n\u7b2c10000\u6b21\u8fed\u4ee3\uff0c\u6210\u672c\u503c\u4e3a\uff1a0.3920295461882659\n\u7b2c11000\u6b21\u8fed\u4ee3\uff0c\u6210\u672c\u503c\u4e3a\uff1a0.38924598135108\n\u7b2c12000\u6b21\u8fed\u4ee3\uff0c\u6210\u672c\u503c\u4e3a\uff1a0.3861547485712325\n\u7b2c13000\u6b21\u8fed\u4ee3\uff0c\u6210\u672c\u503c\u4e3a\uff1a0.384984728909703\n\u7b2c14000\u6b21\u8fed\u4ee3\uff0c\u6210\u672c\u503c\u4e3a\uff1a0.3827828308349524\n\u8bad\u7ec3\u96c6:\nAccuracy: 0.83\n\u6d4b\u8bd5\u96c6:\nAccuracy: 0.86<\/code><\/pre>\n\n\n<p>\u4ee3\u4ef7\u66f2\u7ebf\uff1a<\/p>\n\n\n<div class=\"wp-block-image\"><figure class=\"aligncenter\"><img loading=\"lazy\" decoding=\"async\" width=\"487\" height=\"273\" src=\"\/wp-content\/uploads\/2020\/02\/\u56fe\u7247-59.png\" alt=\"\" class=\"wp-image-2981\" srcset=\"http:\/\/www.sniper97.cn\/wp-content\/uploads\/2020\/02\/\u56fe\u7247-59.png 487w, http:\/\/www.sniper97.cn\/wp-content\/uploads\/2020\/02\/\u56fe\u7247-59-300x168.png 300w\" sizes=\"(max-width: 487px) 100vw, 487px\" \/><\/figure><\/div>\n\n\n<p>\u51b3\u7b56\u8fb9\u754c\uff1a<\/p>\n\n\n<div class=\"wp-block-image\"><figure class=\"aligncenter\"><img loading=\"lazy\" decoding=\"async\" width=\"611\" height=\"340\" src=\"\/wp-content\/uploads\/2020\/02\/\u56fe\u7247-60.png\" alt=\"\" class=\"wp-image-2983\" srcset=\"http:\/\/www.sniper97.cn\/wp-content\/uploads\/2020\/02\/\u56fe\u7247-60.png 611w, http:\/\/www.sniper97.cn\/wp-content\/uploads\/2020\/02\/\u56fe\u7247-60-300x167.png 300w\" sizes=\"(max-width: 611px) 100vw, 611px\" \/><\/figure><\/div>\n\n\n<p>\u521d\u59cb\u5316\u7684\u65f6\u5019\u4e58\u4e86\u4e2a10\uff0c\u5982\u679c\u6211\u4eec\u5c06\u8fd9\u4e2a10\u53bb\u6389\u53d8\u62101\uff0c\u90a3\u4e48\u51c6\u786e\u7387\u5c31\u4f1a\u8fbe\u523096%\u3002\u8bf4\u660e\u8fd9\u4e2a\u8d85\u53c2\u4e5f\u80fd\u5f71\u54cd\u51c6\u786e\u5ea6\uff0c\u8fd9\u4e0e\u8fd9\u4e2a\u6570\u53d8\u6210\u591a\u5c11\u6bd4\u8f83\u597d\uff0c\u4e0b\u9762\u4f1a\u8bf4\u3002<\/p>\n\n\n<div class=\"wp-block-image\"><figure class=\"aligncenter\"><img loading=\"lazy\" decoding=\"async\" width=\"604\" height=\"338\" src=\"\/wp-content\/uploads\/2020\/02\/\u56fe\u7247-61.png\" alt=\"\" class=\"wp-image-2984\" srcset=\"http:\/\/www.sniper97.cn\/wp-content\/uploads\/2020\/02\/\u56fe\u7247-61.png 604w, http:\/\/www.sniper97.cn\/wp-content\/uploads\/2020\/02\/\u56fe\u7247-61-300x168.png 300w\" sizes=\"(max-width: 604px) 100vw, 604px\" \/><\/figure><\/div>\n\n\n<p>\u7136\u540e\u662f\u6291\u68af\u5ea6\u5f02\u5e38\u521d\u59cb\u5316\uff1a<\/p>\n\n\n<pre class=\"wp-block-preformatted\">def initialize_parameters_he(layers_dims):<br \/>    <em>\"\"\"<br \/><\/em><em>    <\/em><em>\u6291\u68af\u5ea6\u5f02\u5e38\u521d\u59cb\u5316<br \/><\/em><em>    <\/em><strong><em>:param<\/em><\/strong><em> layers_dims:<br \/><\/em><em>    <\/em><strong><em>:return<\/em><\/strong><em>:<br \/><\/em><em>    \"\"\"<br \/><\/em><em>    <\/em>np.random.seed(3)<br \/>    parameters = {}<br \/>    L = len(layers_dims)<br \/>    for i in range(1, L):<br \/>        print(np.sqrt(2 \/ layers_dims[i - 1]))<br \/>        parameters['W' + str(i)] = np.random.randn(layers_dims[i], layers_dims[i - 1]) * np.sqrt(2 \/ layers_dims[i - 1])<br \/>        parameters['b' + str(i)] = np.zeros((layers_dims[i], 1))<br \/>    return parameters<\/pre>\n\n\n<p>\u7136\u540e\u8f93\u51fa\u7ed3\u679c\u548c\u7ed8\u5236\u51b3\u7b56\u8fb9\u754c\uff1a<\/p>\n\n\n<pre class=\"wp-block-preformatted\"><br \/>parameters = model(train_X, train_Y, initialization=\"he\", is_polt=True)<br \/>print(\"\u8bad\u7ec3\u96c6:\")<br \/>predictions_train = init_utils.predict(train_X, train_Y, parameters)<br \/>print(\"\u6d4b\u8bd5\u96c6:\")<br \/>predictions_test = init_utils.predict(test_X, test_Y, parameters)<br \/><br \/>plt.title(\"Model with large random initialization\")<br \/>axes = plt.gca()<br \/>axes.set_xlim([-1.5, 1.5])<br \/>axes.set_ylim([-1.5, 1.5])<br \/>init_utils.plot_decision_boundary(lambda x: init_utils.predict_dec(parameters, x.T), train_X, train_Y)<\/pre>\n\n\n<pre class=\"wp-block-code\"><code>\u7b2c0\u6b21\u8fed\u4ee3\uff0c\u6210\u672c\u503c\u4e3a\uff1a0.8830537463419761\n\u7b2c1000\u6b21\u8fed\u4ee3\uff0c\u6210\u672c\u503c\u4e3a\uff1a0.6879825919728063\n\u7b2c2000\u6b21\u8fed\u4ee3\uff0c\u6210\u672c\u503c\u4e3a\uff1a0.6751286264523371\n\u7b2c3000\u6b21\u8fed\u4ee3\uff0c\u6210\u672c\u503c\u4e3a\uff1a0.6526117768893807\n\u7b2c4000\u6b21\u8fed\u4ee3\uff0c\u6210\u672c\u503c\u4e3a\uff1a0.6082958970572937\n\u7b2c5000\u6b21\u8fed\u4ee3\uff0c\u6210\u672c\u503c\u4e3a\uff1a0.5304944491717495\n\u7b2c6000\u6b21\u8fed\u4ee3\uff0c\u6210\u672c\u503c\u4e3a\uff1a0.4138645817071793\n\u7b2c7000\u6b21\u8fed\u4ee3\uff0c\u6210\u672c\u503c\u4e3a\uff1a0.3117803464844441\n\u7b2c8000\u6b21\u8fed\u4ee3\uff0c\u6210\u672c\u503c\u4e3a\uff1a0.23696215330322556\n\u7b2c9000\u6b21\u8fed\u4ee3\uff0c\u6210\u672c\u503c\u4e3a\uff1a0.18597287209206828\n\u7b2c10000\u6b21\u8fed\u4ee3\uff0c\u6210\u672c\u503c\u4e3a\uff1a0.15015556280371808\n\u7b2c11000\u6b21\u8fed\u4ee3\uff0c\u6210\u672c\u503c\u4e3a\uff1a0.12325079292273548\n\u7b2c12000\u6b21\u8fed\u4ee3\uff0c\u6210\u672c\u503c\u4e3a\uff1a0.09917746546525937\n\u7b2c13000\u6b21\u8fed\u4ee3\uff0c\u6210\u672c\u503c\u4e3a\uff1a0.08457055954024274\n\u7b2c14000\u6b21\u8fed\u4ee3\uff0c\u6210\u672c\u503c\u4e3a\uff1a0.07357895962677366\n\u8bad\u7ec3\u96c6:\nAccuracy: 0.9933333333333333\n\u6d4b\u8bd5\u96c6:\nAccuracy: 0.96<\/code><\/pre>\n\n\n<p>\u4ee3\u4ef7\u53d8\u5316\u66f2\u7ebf\uff1a<\/p>\n\n\n<div class=\"wp-block-image\"><figure class=\"aligncenter\"><img loading=\"lazy\" decoding=\"async\" width=\"589\" height=\"334\" src=\"\/wp-content\/uploads\/2020\/02\/\u56fe\u7247-62.png\" alt=\"\" class=\"wp-image-2985\" srcset=\"http:\/\/www.sniper97.cn\/wp-content\/uploads\/2020\/02\/\u56fe\u7247-62.png 589w, http:\/\/www.sniper97.cn\/wp-content\/uploads\/2020\/02\/\u56fe\u7247-62-300x170.png 300w\" sizes=\"(max-width: 589px) 100vw, 589px\" \/><\/figure><\/div>\n\n\n<p>\u51b3\u7b56\u8fb9\u754c<\/p>\n\n\n<div class=\"wp-block-image\"><figure class=\"aligncenter\"><img loading=\"lazy\" decoding=\"async\" width=\"588\" height=\"337\" src=\"\/wp-content\/uploads\/2020\/02\/\u56fe\u7247-63.png\" alt=\"\" class=\"wp-image-2986\" srcset=\"http:\/\/www.sniper97.cn\/wp-content\/uploads\/2020\/02\/\u56fe\u7247-63.png 588w, http:\/\/www.sniper97.cn\/wp-content\/uploads\/2020\/02\/\u56fe\u7247-63-300x172.png 300w\" sizes=\"(max-width: 588px) 100vw, 588px\" \/><\/figure><\/div>\n\n\n<p>\u6211\u4eec\u53d1\u73b0\u51c6\u786e\u5ea6\u5f88\u9ad8\uff0c\u5e76\u4e14\u51b3\u7b56\u8fb9\u754c\u548c\u4e0a\u9762\u8d85\u53c2\u4e3a1\u7684\u65f6\u5019\u5f88\u50cf\uff0c\u6211\u4eec\u8f93\u51fa\u4e00\u4e0b\u8fd9\u4e2a\u8d85\u53c2\u770b\u770b\u3002<\/p>\n\n\n<pre class=\"wp-block-code\"><code>1.0\n0.4472135954999579\n0.6324555320336759<\/code><\/pre>\n\n\n<p>\u53d1\u73b0\u57fa\u672c\u4e5f\u57281\u9644\u8fd1\u3002<\/p>\n\n\n<p>\u5904\u7406\u5b8c\u53c2\u6570\u521d\u59cb\u5316\uff0c\u6211\u4eec\u63a5\u4e0b\u6765\u5b66\u4e60\u6b63\u5219\u5316\u7684\u76f8\u5173\u77e5\u8bc6\u3002<\/p>\n\n\n<div class=\"wp-block-image\"><figure class=\"aligncenter\"><img loading=\"lazy\" decoding=\"async\" width=\"535\" height=\"397\" src=\"\/wp-content\/uploads\/2020\/02\/\u56fe\u7247-64.png\" alt=\"\" class=\"wp-image-2991\" srcset=\"http:\/\/www.sniper97.cn\/wp-content\/uploads\/2020\/02\/\u56fe\u7247-64.png 535w, http:\/\/www.sniper97.cn\/wp-content\/uploads\/2020\/02\/\u56fe\u7247-64-300x223.png 300w\" sizes=\"(max-width: 535px) 100vw, 535px\" \/><\/figure><\/div>\n\n\n<pre class=\"wp-block-code\"><code>\u7b2c0\u6b21\u8fed\u4ee3\uff0c\u6210\u672c\u503c\u4e3a\uff1a0.6557412523481002\n\u7b2c5000\u6b21\u8fed\u4ee3\uff0c\u6210\u672c\u503c\u4e3a\uff1a0.17620471758400447\n\u7b2c10000\u6b21\u8fed\u4ee3\uff0c\u6210\u672c\u503c\u4e3a\uff1a0.1632998752572419\n\u7b2c15000\u6b21\u8fed\u4ee3\uff0c\u6210\u672c\u503c\u4e3a\uff1a0.14796400922574113\n\u7b2c20000\u6b21\u8fed\u4ee3\uff0c\u6210\u672c\u503c\u4e3a\uff1a0.13851642423239133\n\u7b2c25000\u6b21\u8fed\u4ee3\uff0c\u6210\u672c\u503c\u4e3a\uff1a0.13285370211487862\n\u8bad\u7ec3\u96c6:\nAccuracy: 0.9478672985781991\n\u6d4b\u8bd5\u96c6:\nAccuracy: 0.915<\/code><\/pre>\n\n\n<div class=\"wp-block-image\"><figure class=\"aligncenter\"><img loading=\"lazy\" decoding=\"async\" width=\"534\" height=\"397\" src=\"\/wp-content\/uploads\/2020\/02\/\u56fe\u7247-65.png\" alt=\"\" class=\"wp-image-2992\" srcset=\"http:\/\/www.sniper97.cn\/wp-content\/uploads\/2020\/02\/\u56fe\u7247-65.png 534w, http:\/\/www.sniper97.cn\/wp-content\/uploads\/2020\/02\/\u56fe\u7247-65-300x223.png 300w\" sizes=\"(max-width: 534px) 100vw, 534px\" \/><\/figure><\/div>\n\n\n<div class=\"wp-block-image\"><figure class=\"aligncenter\"><img loading=\"lazy\" decoding=\"async\" width=\"534\" height=\"399\" src=\"\/wp-content\/uploads\/2020\/02\/\u56fe\u7247-66.png\" alt=\"\" class=\"wp-image-2993\" srcset=\"http:\/\/www.sniper97.cn\/wp-content\/uploads\/2020\/02\/\u56fe\u7247-66.png 534w, http:\/\/www.sniper97.cn\/wp-content\/uploads\/2020\/02\/\u56fe\u7247-66-300x224.png 300w\" sizes=\"(max-width: 534px) 100vw, 534px\" \/><\/figure><\/div>\n\n\n<p>\u7136\u540e\u662f\u6b63\u5219\u5316\uff1a<\/p>\n\n\n<p>\u9996\u5148\u6211\u4eec\u5bfc\u5165\u6570\u636e<\/p>\n\n\n<pre class=\"wp-block-preformatted\">train_X, train_Y, test_X, test_Y = reg_utils.load_2D_dataset()<br \/>plt.show()<\/pre>\n\n\n<p>\u8fd8\u662f\u548c\u4e4b\u524d\u4e00\u6837\uff0c\u6211\u4eec\u5148\u5199\u6a21\u578b\uff0c\u5bf9\u4e8e\u65b9\u6cd5\u4ee5\u540e\u5b9e\u73b0\uff1a<\/p>\n\n\n<pre class=\"wp-block-preformatted\">def model(X, Y, learning_rate=0.3, num_iterations=30000, print_cost=True, is_plot=True, lambd=0.0, keep_prob=1.0):<br \/>    <em>\"\"\"<br \/><\/em><em>    <\/em><em>\u6a21\u578b\u4e3b\u4f53<br \/><\/em><em>    <\/em><strong><em>:param<\/em><\/strong><em> X:<br \/><\/em><em>    <\/em><strong><em>:param<\/em><\/strong><em> Y:<br \/><\/em><em>    <\/em><strong><em>:param<\/em><\/strong><em> learning_rate:<br \/><\/em><em>    <\/em><strong><em>:param<\/em><\/strong><em> num_iterations:<br \/><\/em><em>    <\/em><strong><em>:param<\/em><\/strong><em> print_cost:<br \/><\/em><em>    <\/em><strong><em>:param<\/em><\/strong><em> is_plot:<br \/><\/em><em>    <\/em><strong><em>:param<\/em><\/strong><em> lambd:<br \/><\/em><em>    <\/em><strong><em>:param<\/em><\/strong><em> keep_prob:<br \/><\/em><em>    <\/em><strong><em>:return<\/em><\/strong><em>:<br \/><\/em><em>    \"\"\"<br \/><\/em><em>    <\/em>grads = {}<br \/>    costs = []<br \/>    m = X.shape[1]<br \/>    layers_dims = [X.shape[0], 20, 3, 1]<br \/><br \/>    # \u8fd9\u4e2a\u521d\u59cb\u5316\u540e\u53f0\u4e5f\u662f\u6291\u68af\u5ea6\u5f02\u5e38\u521d\u59cb\u5316<br \/>    parameters = reg_utils.initialize_parameters(layers_dims)<br \/><br \/>    for i in range(num_iterations):<br \/>        # dropout<br \/>        if keep_prob == 1:<br \/>            a3, cache = reg_utils.forward_propagation(X, parameters)<br \/>        elif keep_prob &lt; 1:<br \/>            a3, cache = forward_propagation_with_dropout(X, parameters, keep_prob)<br \/><br \/>        # \u6b63\u5219\u5316<br \/>        if lambd == 0:<br \/>            # \u4e0d\u4f7f\u7528L2\u6b63\u5219\u5316<br \/>            cost = reg_utils.compute_cost(a3, Y)<br \/>        else:<br \/>            # \u4f7f\u7528L2\u6b63\u5219\u5316<br \/>            cost = compute_cost_with_regularization(a3, Y, parameters, lambd)<br \/><br \/>            # \u53cd\u5411\u4f20\u64ad<br \/>            # \u53ef\u4ee5\u540c\u65f6\u4f7f\u7528L2\u6b63\u5219\u5316\u548c\u968f\u673a\u5220\u9664\u8282\u70b9\uff0c\u4f46\u662f\u672c\u6b21\u5b9e\u9a8c\u4e0d\u540c\u65f6\u4f7f\u7528\u3002<br \/>        assert (lambd == 0 or keep_prob == 1)<br \/><br \/>        # \u4e24\u4e2a\u53c2\u6570\u7684\u4f7f\u7528\u60c5\u51b5<br \/>        if (lambd == 0 and keep_prob == 1):<br \/>            # \u4e0d\u4f7f\u7528L2\u6b63\u5219\u5316\u548c\u4e0d\u4f7f\u7528\u968f\u673a\u5220\u9664\u8282\u70b9<br \/>            grads = reg_utils.backward_propagation(X, Y, cache)<br \/>        elif lambd != 0:<br \/>            # \u4f7f\u7528L2\u6b63\u5219\u5316\uff0c\u4e0d\u4f7f\u7528\u968f\u673a\u5220\u9664\u8282\u70b9<br \/>            grads = backward_propagation_with_regularization(X, Y, cache, lambd)<br \/>        elif keep_prob &lt; 1:<br \/>            # \u4f7f\u7528\u968f\u673a\u5220\u9664\u8282\u70b9\uff0c\u4e0d\u4f7f\u7528L2\u6b63\u5219\u5316<br \/>            grads = backward_propagation_with_dropout(X, Y, cache, keep_prob)<br \/><br \/>        parameters = reg_utils.update_parameters(parameters, grads, learning_rate)<br \/><br \/>        if i % 1000 == 0:<br \/>            costs.append(cost)<br \/>            if (print_cost and i % 5000 == 0):<br \/>                print(\"\u7b2c\" + str(i) + \"\u6b21\u8fed\u4ee3\uff0c\u6210\u672c\u503c\u4e3a\uff1a\" + str(cost))<br \/><br \/>    if is_plot:<br \/>        plt.plot(costs)<br \/>        plt.ylabel('cost')<br \/>        plt.xlabel('iterations (x1,000)')<br \/>        plt.title(\"Learning rate =\" + str(learning_rate))<br \/>        plt.show()<br \/><br \/>    return parameters<\/pre>\n\n\n<p>\u9996\u5148\u662f\u6b63\u5219\u5316\u7684\u524d\u5411\u4f20\u64ad\uff1a<\/p>\n\n\n<pre class=\"wp-block-preformatted\">def compute_cost_with_regularization(A3, Y, parameters, lambd):<br \/>    <em>\"\"\"<br \/><\/em><em>    <\/em><em>\u6b63\u5219\u5316\u524d\u5411\u4f20\u64ad<br \/><\/em><em>    <\/em><strong><em>:return<\/em><\/strong><em>:<br \/><\/em><em>    \"\"\"<br \/><\/em><em>    <\/em>m = Y.shape[1]<br \/>    W1 = parameters['W1']<br \/>    W2 = parameters['W2']<br \/>    W3 = parameters['W3']<br \/><br \/>    cost = reg_utils.compute_cost(A3, Y)<br \/>    L2_regularization_cost = lambd * (np.sum(np.square(W1)) + np.sum(np.square(W2)) + np.sum(np.square(W3))) \/ (2 * m)<br \/>    cost = cost + L2_regularization_cost<br \/><br \/>    return cost<\/pre>\n\n\n<p>\u7136\u540e\u662f\u6b63\u5219\u5316\u7684\u53cd\u5411\u4f20\u64ad\uff1a<\/p>\n\n\n<pre class=\"wp-block-preformatted\">def backward_propagation_with_regularization(X, Y, cache, lambd):<br \/>    <em>\"\"\"<br \/><\/em><em>    <\/em><em>\u56e0\u4e3a\u4ee3\u4ef7\u51fd\u6570\u7684\u6539\u53d8\uff0c\u53cd\u5411\u4f20\u64ad\u7684\u8fc7\u7a0b\u4e5f\u8981\u6539\u53d8<br \/><\/em><em>    <\/em><strong><em>:param<\/em><\/strong><em> X:<br \/><\/em><em>    <\/em><strong><em>:param<\/em><\/strong><em> Y:<br \/><\/em><em>    <\/em><strong><em>:param<\/em><\/strong><em> cache:<br \/><\/em><em>    <\/em><strong><em>:param<\/em><\/strong><em> lambd:<br \/><\/em><em>    <\/em><strong><em>:return<\/em><\/strong><em>:<br \/><\/em><em>    \"\"\"<br \/><\/em><em>    <\/em>m = X.shape[1]<br \/><br \/>    (Z1, A1, W1, b1, Z2, A2, W2, b2, Z3, A3, W3, b3) = cache<br \/><br \/>    dZ3 = A3 - Y<br \/><br \/>    dW3 = (1 \/ m) * np.dot(dZ3, A2.T) + ((lambd * W3) \/ m)<br \/>    db3 = (1 \/ m) * np.sum(dZ3, axis=1, keepdims=True)<br \/><br \/>    dA2 = np.dot(W3.T, dZ3)<br \/>    dZ2 = np.multiply(dA2, np.int64(A2 &gt; 0))<br \/>    dW2 = (1 \/ m) * np.dot(dZ2, A1.T) + ((lambd * W2) \/ m)<br \/>    db2 = (1 \/ m) * np.sum(dZ2, axis=1, keepdims=True)<br \/><br \/>    dA1 = np.dot(W2.T, dZ2)<br \/>    dZ1 = np.multiply(dA1, np.int64(A1 &gt; 0))<br \/>    dW1 = (1 \/ m) * np.dot(dZ1, X.T) + ((lambd * W1) \/ m)<br \/>    db1 = (1 \/ m) * np.sum(dZ1, axis=1, keepdims=True)<br \/><br \/>    gradients = {\"dZ3\": dZ3, \"dW3\": dW3, \"db3\": db3, \"dA2\": dA2,<br \/>                 \"dZ2\": dZ2, \"dW2\": dW2, \"db2\": db2, \"dA1\": dA1,<br \/>                 \"dZ1\": dZ1, \"dW1\": dW1, \"db1\": db1}<br \/><br \/>    return gradients<\/pre>\n\n\n<p>\u7136\u540e\u6211\u4eec\u8c03\u7528\u8bad\u7ec3<\/p>\n\n\n<pre class=\"wp-block-preformatted\">parameters = model(train_X, train_Y, is_plot=True, lambd=0.7)\nprint(\"\u8bad\u7ec3\u96c6:\")\npredictions_train = reg_utils.predict(train_X, train_Y, parameters)\nprint(\"\u6d4b\u8bd5\u96c6:\")\npredictions_test = reg_utils.predict(test_X, test_Y, parameters)\nplt.title(\"Model without regularization\")\naxes = plt.gca()\naxes.set_xlim([-0.75, 0.40])\naxes.set_ylim([-0.75, 0.65])\nreg_utils.plot_decision_boundary(lambda x: reg_utils.predict_dec(parameters, x.T), train_X, train_Y)\n<\/pre>\n\n\n<p>\u6211\u4eec\u53ef\u4ee5\u770b\u5230\u8f93\u51fa\u5982\u4e0b\uff1a<\/p>\n\n\n<pre class=\"wp-block-code\"><code>\u7b2c0\u6b21\u8fed\u4ee3\uff0c\u6210\u672c\u503c\u4e3a\uff1a0.6974484493131264\n\u7b2c5000\u6b21\u8fed\u4ee3\uff0c\u6210\u672c\u503c\u4e3a\uff1a0.2690430474349705\n\u7b2c10000\u6b21\u8fed\u4ee3\uff0c\u6210\u672c\u503c\u4e3a\uff1a0.2684918873282239\n\u7b2c15000\u6b21\u8fed\u4ee3\uff0c\u6210\u672c\u503c\u4e3a\uff1a0.2682199033729048\n\u7b2c20000\u6b21\u8fed\u4ee3\uff0c\u6210\u672c\u503c\u4e3a\uff1a0.2680916337127301\n\u7b2c25000\u6b21\u8fed\u4ee3\uff0c\u6210\u672c\u503c\u4e3a\uff1a0.26794285663887024\n\u8bad\u7ec3\u96c6:\nAccuracy: 0.9383886255924171\n\u6d4b\u8bd5\u96c6:\nAccuracy: 0.93<\/code><\/pre>\n\n\n<p>\u4ee3\u4ef7\u53d8\u5316\u66f2\u7ebf\uff1a<\/p>\n\n\n<div class=\"wp-block-image\"><figure class=\"aligncenter\"><img loading=\"lazy\" decoding=\"async\" width=\"530\" height=\"397\" src=\"\/wp-content\/uploads\/2020\/02\/\u56fe\u7247-67.png\" alt=\"\" class=\"wp-image-2994\" srcset=\"http:\/\/www.sniper97.cn\/wp-content\/uploads\/2020\/02\/\u56fe\u7247-67.png 530w, http:\/\/www.sniper97.cn\/wp-content\/uploads\/2020\/02\/\u56fe\u7247-67-300x225.png 300w\" sizes=\"(max-width: 530px) 100vw, 530px\" \/><\/figure><\/div>\n\n\n<p>\u51b3\u7b56\u8fb9\u754c\uff1a<\/p>\n\n\n<div class=\"wp-block-image\"><figure class=\"aligncenter\"><img loading=\"lazy\" decoding=\"async\" width=\"534\" height=\"389\" src=\"\/wp-content\/uploads\/2020\/02\/\u56fe\u7247-68.png\" alt=\"\" class=\"wp-image-2995\" srcset=\"http:\/\/www.sniper97.cn\/wp-content\/uploads\/2020\/02\/\u56fe\u7247-68.png 534w, http:\/\/www.sniper97.cn\/wp-content\/uploads\/2020\/02\/\u56fe\u7247-68-300x219.png 300w\" sizes=\"(max-width: 534px) 100vw, 534px\" \/><\/figure><\/div>\n\n\n<p>\u7136\u540e\u5c31\u662fdropout\u7684\u524d\u5411\u548c\u53cd\u5411\u4f20\u64ad\u8fc7\u7a0b\uff1a<\/p>\n\n\n<pre class=\"wp-block-preformatted\">def forward_propagation_with_dropout(X, parameters, keep_prob=0.5):<br \/>    <em>\"\"\"<br \/><\/em><em>    <\/em><em>\u968f\u673a\u820d\u5f03\u7ed3\u70b9\u7684\u524d\u5411\u4f20\u64ad<br \/><\/em><em>    <\/em><strong><em>:param<\/em><\/strong><em> X:<br \/><\/em><em>    <\/em><strong><em>:param<\/em><\/strong><em> parameters:<br \/><\/em><em>    <\/em><strong><em>:param<\/em><\/strong><em> keep_prob:<br \/><\/em><em>    <\/em><strong><em>:return<\/em><\/strong><em>:<br \/><\/em><em>    \"\"\"<br \/><\/em><em>    <\/em>np.random.seed(1)<br \/><br \/>    W1 = parameters[\"W1\"]<br \/>    b1 = parameters[\"b1\"]<br \/>    W2 = parameters[\"W2\"]<br \/>    b2 = parameters[\"b2\"]<br \/>    W3 = parameters[\"W3\"]<br \/>    b3 = parameters[\"b3\"]<br \/><br \/>    Z1 = np.dot(W1, X) + b1<br \/>    A1 = reg_utils.relu(Z1)<br \/><br \/>    # \u968f\u673a\u521d\u59cb\u5316\u4e00\u4e2a0~1\u7684\u77e9\u9635\uff0c\u7136\u540e\u6839\u636e\u521d\u59cb\u5316\u7684\u503c\u91cd\u65b0\u8ba1\u7b97A1\uff0c\u6700\u540e\u7f29\u5c0f\u4e00\u70b9A<br \/>    D1 = np.random.rand(A1.shape[0], A1.shape[1])<br \/>    D1 = D1 &lt; keep_prob<br \/>    A1 = A1 * D1<br \/>    A1 = A1 \/ keep_prob<br \/><br \/>    Z2 = np.dot(W2, A1) + b2<br \/>    A2 = reg_utils.relu(Z2)<br \/><br \/>    D2 = np.random.rand(A2.shape[0], A2.shape[1])<br \/>    D2 = D2 &lt; keep_prob<br \/>    A2 = A2 * D2<br \/>    A2 = A2 \/ keep_prob<br \/><br \/>    Z3 = np.dot(W3, A2) + b3<br \/>    A3 = reg_utils.sigmoid(Z3)<br \/><br \/>    cache = (Z1, D1, A1, W1, b1, Z2, D2, A2, W2, b2, Z3, A3, W3, b3)<br \/><br \/>    return A3, cache<br \/><br \/><br \/>def backward_propagation_with_dropout(X, Y, cache, keep_prob):<br \/>    <em>\"\"\"<br \/><\/em><em>    <\/em><em>\u968f\u673a\u820d\u5f03\u7ed3\u70b9\u7684\u53cd\u5411\u4f20\u64ad<br \/><\/em><em>    <\/em><strong><em>:param<\/em><\/strong><em> X:<br \/><\/em><em>    <\/em><strong><em>:param<\/em><\/strong><em> Y:<br \/><\/em><em>    <\/em><strong><em>:param<\/em><\/strong><em> cache:<br \/><\/em><em>    <\/em><strong><em>:param<\/em><\/strong><em> keep_prob:<br \/><\/em><em>    <\/em><strong><em>:return<\/em><\/strong><em>:<br \/><\/em><em>    \"\"\"<br \/><\/em><em>    <\/em>m = X.shape[1]<br \/>    (Z1, D1, A1, W1, b1, Z2, D2, A2, W2, b2, Z3, A3, W3, b3) = cache<br \/><br \/>    dZ3 = A3 - Y<br \/>    dW3 = (1 \/ m) * np.dot(dZ3, A2.T)<br \/>    db3 = 1. \/ m * np.sum(dZ3, axis=1, keepdims=True)<br \/>    dA2 = np.dot(W3.T, dZ3)<br \/><br \/>    # \u6839\u636e\u524d\u5411\u968f\u673a\u820d\u5f03\u7684\u7ed3\u70b9\u91cd\u65b0\u8ba1\u7b97A\uff0c\u5e76\u7f29\u653eA<br \/>    dA2 = dA2 * D2<br \/>    dA2 = dA2 \/ keep_prob<br \/><br \/>    dZ2 = np.multiply(dA2, np.int64(A2 &gt; 0))<br \/>    dW2 = 1. \/ m * np.dot(dZ2, A1.T)<br \/>    db2 = 1. \/ m * np.sum(dZ2, axis=1, keepdims=True)<br \/><br \/>    dA1 = np.dot(W2.T, dZ2)<br \/><br \/>    dA1 = dA1 * D1<br \/>    dA1 = dA1 \/ keep_prob<br \/><br \/>    dZ1 = np.multiply(dA1, np.int64(A1 &gt; 0))<br \/>    dW1 = 1. \/ m * np.dot(dZ1, X.T)<br \/>    db1 = 1. \/ m * np.sum(dZ1, axis=1, keepdims=True)<br \/><br \/>    gradients = {\"dZ3\": dZ3, \"dW3\": dW3, \"db3\": db3, \"dA2\": dA2,<br \/>                 \"dZ2\": dZ2, \"dW2\": dW2, \"db2\": db2, \"dA1\": dA1,<br \/>                 \"dZ1\": dZ1, \"dW1\": dW1, \"db1\": db1}<br \/><br \/>    return gradients<\/pre>\n\n\n<p>\u8f93\u51fa\u5982\u4e0b\uff1a<\/p>\n\n\n<pre class=\"wp-block-code\"><code>\u7b2c0\u6b21\u8fed\u4ee3\uff0c\u6210\u672c\u503c\u4e3a\uff1a0.6543912405149825\n\u7b2c5000\u6b21\u8fed\u4ee3\uff0c\u6210\u672c\u503c\u4e3a\uff1a0.06466905008519824\n\u7b2c10000\u6b21\u8fed\u4ee3\uff0c\u6210\u672c\u503c\u4e3a\uff1a0.061016986574905605\n\u7b2c15000\u6b21\u8fed\u4ee3\uff0c\u6210\u672c\u503c\u4e3a\uff1a0.060664572161287754\n\u7b2c20000\u6b21\u8fed\u4ee3\uff0c\u6210\u672c\u503c\u4e3a\uff1a0.060582435798513114\n\u7b2c25000\u6b21\u8fed\u4ee3\uff0c\u6210\u672c\u503c\u4e3a\uff1a0.06050179002362491\n\u8bad\u7ec3\u96c6:\nAccuracy: 0.9289099526066351\n\u6d4b\u8bd5\u96c6:\nAccuracy: 0.95<\/code><\/pre>\n\n\n<p>\u4ee3\u4ef7\u53d8\u5316\u66f2\u7ebf<\/p>\n\n\n<div class=\"wp-block-image\"><figure class=\"aligncenter\"><img loading=\"lazy\" decoding=\"async\" width=\"523\" height=\"393\" src=\"\/wp-content\/uploads\/2020\/02\/\u56fe\u7247-69.png\" alt=\"\" class=\"wp-image-2996\" srcset=\"http:\/\/www.sniper97.cn\/wp-content\/uploads\/2020\/02\/\u56fe\u7247-69.png 523w, http:\/\/www.sniper97.cn\/wp-content\/uploads\/2020\/02\/\u56fe\u7247-69-300x225.png 300w\" sizes=\"(max-width: 523px) 100vw, 523px\" \/><\/figure><\/div>\n\n\n<p>\u51b3\u7b56\u8fb9\u754c\uff1a<\/p>\n\n\n<div class=\"wp-block-image\"><figure class=\"aligncenter\"><img loading=\"lazy\" decoding=\"async\" width=\"525\" height=\"408\" src=\"\/wp-content\/uploads\/2020\/02\/\u56fe\u7247-70.png\" alt=\"\" class=\"wp-image-2997\" srcset=\"http:\/\/www.sniper97.cn\/wp-content\/uploads\/2020\/02\/\u56fe\u7247-70.png 525w, http:\/\/www.sniper97.cn\/wp-content\/uploads\/2020\/02\/\u56fe\u7247-70-300x233.png 300w\" sizes=\"(max-width: 525px) 100vw, 525px\" \/><\/figure><\/div>\n\n\n<p>\u6211\u4eec\u53ef\u4ee5\u770b\u5230\uff0c\u867d\u7136\u6574\u4f53\u7684\u8bad\u7ec3\u96c6\u51c6\u786e\u7387\u964d\u4f4e\u4e86\uff0c\u4f46\u662f\u6d4b\u8bd5\u96c6\u51c6\u786e\u7387\u5374\u4e0a\u5347\u4e86\uff0c\u8bf4\u660e\u6b63\u5219\u5316\u4e00\u5b9a\u7a0b\u5ea6\u4e0a\u6d88\u9664\u4e86\u8fc7\u62df\u5408\u3002<\/p>\n\n\n<p>\u5b8c\u6574\u4ee3\u7801\uff1a<\/p>\n\n\n<pre class=\"wp-block-preformatted\"># -*- coding:utf-8 -*-<br \/><br \/><em>\"\"\"<br \/><\/em><em>      \u250f\u251b \u253b\u2501\u2501\u2501\u2501\u2501\u251b \u253b\u2513<br \/><\/em><em>      \u2503<\/em><em>\u3000\u3000\u3000\u3000\u3000\u3000<\/em><em> \u2503<br \/><\/em><em>      \u2503<\/em><em>\u3000\u3000\u3000<\/em><em>\u2501<\/em><em>\u3000\u3000\u3000<\/em><em>\u2503<br \/><\/em><em>      \u2503<\/em><em>\u3000<\/em><em>\u2533\u251b<\/em><em>\u3000<\/em><em>  \u2517\u2533<\/em><em>\u3000<\/em><em>\u2503<br \/><\/em><em>      \u2503<\/em><em>\u3000\u3000\u3000\u3000\u3000\u3000<\/em><em> \u2503<br \/><\/em><em>      \u2503<\/em><em>\u3000\u3000\u3000<\/em><em>\u253b<\/em><em>\u3000\u3000\u3000<\/em><em>\u2503<br \/><\/em><em>      \u2503<\/em><em>\u3000\u3000\u3000\u3000\u3000\u3000<\/em><em> \u2503<br \/><\/em><em>      \u2517\u2501\u2513<\/em><em>\u3000\u3000\u3000<\/em><em>\u250f\u2501\u2501\u2501\u251b<br \/><\/em><em>        \u2503<\/em><em>\u3000\u3000\u3000<\/em><em>\u2503   <\/em><em>\u795e\u517d\u4fdd\u4f51<\/em><em><br \/><\/em><em>        \u2503<\/em><em>\u3000\u3000\u3000<\/em><em>\u2503   <\/em><em>\u4ee3\u7801\u65e0<\/em><em>BUG<\/em><em>\uff01<\/em><em><br \/><\/em><em>        \u2503<\/em><em>\u3000\u3000\u3000<\/em><em>\u2517\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2513<br \/><\/em><em>        \u2503<\/em><em>\u3000\u3000\u3000\u3000\u3000\u3000\u3000<\/em><em>    \u2523\u2513<br \/><\/em><em>        \u2503<\/em><em>\u3000\u3000\u3000\u3000<\/em><em>         \u250f\u251b<br \/><\/em><em>        \u2517\u2501\u2513 \u2513 \u250f\u2501\u2501\u2501\u2533 \u2513 \u250f\u2501\u251b<br \/><\/em><em>          \u2503 \u252b \u252b   \u2503 \u252b \u252b<br \/><\/em><em>          \u2517\u2501\u253b\u2501\u251b   \u2517\u2501\u253b\u2501\u251b<br \/><\/em><em>\"\"\"<br \/><\/em><em><br \/><\/em>import numpy as np<br \/>import matplotlib.pyplot as plt<br \/>import sklearn<br \/>import sklearn.datasets<br \/>from course_2_week_1 import init_utils  # \u7b2c\u4e00\u90e8\u5206\uff0c\u521d\u59cb\u5316<br \/>from course_2_week_1 import reg_utils  # \u7b2c\u4e8c\u90e8\u5206\uff0c\u6b63\u5219\u5316<br \/>from course_2_week_1 import gc_utils  # \u7b2c\u4e09\u90e8\u5206\uff0c\u68af\u5ea6\u6821\u9a8c<br \/><br \/><br \/># plt.rcParams['figure.figsize'] = (7.0, 4.0)  # set default size of plots<br \/># plt.rcParams['image.interpolation'] = 'nearest'<br \/># plt.rcParams['image.cmap'] = 'gray'<br \/><br \/># \u521d\u59cb\u5316\u6570\u636e<br \/># train_X, train_Y, test_X, test_Y = init_utils.load_dataset(is_plot=True)<br \/># plt.show()<br \/><br \/><br \/>def initialize_parameters_zeros(layers_dims):<br \/>    <em>\"\"\"<br \/><\/em><em>    <\/em><em>\u6a21\u578b\u7684\u53c2\u6570\u5168\u90e8\u521d\u59cb\u5316\u4e3a<\/em><em>0<br \/><\/em><em>    <\/em><strong><em>:param<\/em><\/strong><em> layers_dims:<br \/><\/em><em>    <\/em><strong><em>:return<\/em><\/strong><em>:<br \/><\/em><em>    \"\"\"<br \/><\/em><em>    <\/em>parameters = {}<br \/><br \/>    # \u7f51\u7edc\u5c42\u6570<br \/>    L = len(layers_dims)<br \/><br \/>    for i in range(1, L):<br \/>        parameters['W' + str(i)] = np.zeros((layers_dims[i], layers_dims[i - 1]))<br \/>        parameters['b' + str(i)] = np.zeros((layers_dims[i], 1))<br \/><br \/>    return parameters<br \/><br \/><br \/>def initialize_parameters_random(layers_dims):<br \/>    <em>\"\"\"<br \/><\/em><em>    <\/em><em>\u6a21\u578b\u968f\u673a\u521d\u59cb\u5316\u53c2\u6570<br \/><\/em><em>    <\/em><strong><em>:param<\/em><\/strong><em> layers_dims:<br \/><\/em><em>    <\/em><strong><em>:return<\/em><\/strong><em>:<br \/><\/em><em>    \"\"\"<br \/><\/em><em>    <\/em>np.random.seed(3)<br \/><br \/>    parameters = {}<br \/><br \/>    L = len(layers_dims)<br \/><br \/>    for i in range(1, L):<br \/>        parameters['W' + str(i)] = np.random.randn(layers_dims[i], layers_dims[i - 1]) * 10<br \/>        parameters['b' + str(i)] = np.zeros((layers_dims[i], 1))<br \/><br \/>    return parameters<br \/><br \/><br \/>def initialize_parameters_he(layers_dims):<br \/>    <em>\"\"\"<br \/><\/em><em>    <\/em><em>\u6291\u68af\u5ea6\u5f02\u5e38\u521d\u59cb\u5316<br \/><\/em><em>    <\/em><strong><em>:param<\/em><\/strong><em> layers_dims:<br \/><\/em><em>    <\/em><strong><em>:return<\/em><\/strong><em>:<br \/><\/em><em>    \"\"\"<br \/><\/em><em>    <\/em>np.random.seed(3)<br \/>    parameters = {}<br \/>    L = len(layers_dims)<br \/>    for i in range(1, L):<br \/>        print(np.sqrt(2 \/ layers_dims[i - 1]))<br \/>        parameters['W' + str(i)] = np.random.randn(layers_dims[i], layers_dims[i - 1]) * np.sqrt(2 \/ layers_dims[i - 1])<br \/>        parameters['b' + str(i)] = np.zeros((layers_dims[i], 1))<br \/>    return parameters<br \/><br \/><br \/>def model(X, Y, learning_rate=0.01, num_iteration=15000, print_cost=True, initialization=\"he\", is_polt=True):<br \/>    <em>\"\"\"<br \/><\/em><em>    <\/em><em>\u795e\u7ecf\u7f51\u7edc\u6a21\u578b<br \/><\/em><em>    <\/em><strong><em>:param<\/em><\/strong><em> X:<br \/><\/em><em>    <\/em><strong><em>:param<\/em><\/strong><em> Y:<br \/><\/em><em>    <\/em><strong><em>:param<\/em><\/strong><em> learning_rate:<br \/><\/em><em>    <\/em><strong><em>:param<\/em><\/strong><em> num_iteration:<br \/><\/em><em>    <\/em><strong><em>:param<\/em><\/strong><em> print_cost:<br \/><\/em><em>    <\/em><strong><em>:param<\/em><\/strong><em> initialization:<br \/><\/em><em>    <\/em><strong><em>:param<\/em><\/strong><em> is_polt:<br \/><\/em><em>    <\/em><strong><em>:return<\/em><\/strong><em>:<br \/><\/em><em>    \"\"\"<br \/><\/em><em>    <\/em>grads = {}<br \/>    costs = []<br \/>    # \u8bad\u7ec3\u96c6\u4e2a\u6570<br \/>    m = X.shape[1]<br \/>    # \u4e09\u5c42\u795e\u7ecf\u7f51\u7edc\u7684\u7ed3\u6784<br \/>    layers_dims = [X.shape[0], 10, 5, 1]<br \/><br \/>    if initialization == 'zeros':<br \/>        parameters = initialize_parameters_zeros(layers_dims)<br \/>    elif initialization == 'random':<br \/>        parameters = initialize_parameters_random(layers_dims)<br \/>    elif initialization == \"he\":<br \/>        parameters = initialize_parameters_he(layers_dims)<br \/>    else:<br \/>        print(\"\u9519\u8bef\u7684\u521d\u59cb\u5316\u53c2\u6570\uff01\u7a0b\u5e8f\u9000\u51fa\")<br \/>        exit<br \/><br \/>    # \u5f00\u59cb\u5b66\u4e60<br \/>    for i in range(num_iteration):<br \/>        a3, cache = init_utils.forward_propagation(X, parameters)<br \/><br \/>        cost = init_utils.compute_loss(a3, Y)<br \/><br \/>        grads = init_utils.backward_propagation(X, Y, cache)<br \/><br \/>        parameters = init_utils.update_parameters(parameters, grads, learning_rate)<br \/><br \/>        if i % 1000 == 0:<br \/>            costs.append(cost)<br \/>            if print_cost:<br \/>                print(\"\u7b2c\" + str(i) + \"\u6b21\u8fed\u4ee3\uff0c\u6210\u672c\u503c\u4e3a\uff1a\" + str(cost))<br \/><br \/>    # \u5b66\u4e60\u5b8c\u6bd5\uff0c\u7ed8\u5236\u6210\u672c\u66f2\u7ebf<br \/>    if is_polt:<br \/>        plt.plot(costs)<br \/>        plt.ylabel('cost')<br \/>        plt.xlabel('iterations (per hundreds)')<br \/>        plt.title(\"Learning rate =\" + str(learning_rate))<br \/>        plt.show()<br \/><br \/>    # \u8fd4\u56de\u5b66\u4e60\u5b8c\u6bd5\u540e\u7684\u53c2\u6570<br \/>    return parameters<br \/><br \/><br \/># parameters = model(train_X, train_Y, initialization=\"he\", is_polt=True)<br \/># print(\"\u8bad\u7ec3\u96c6:\")<br \/># predictions_train = init_utils.predict(train_X, train_Y, parameters)<br \/># print(\"\u6d4b\u8bd5\u96c6:\")<br \/># predictions_test = init_utils.predict(test_X, test_Y, parameters)<br \/>#<br \/># plt.title(\"Model with large random initialization\")<br \/># axes = plt.gca()<br \/># axes.set_xlim([-1.5, 1.5])<br \/># axes.set_ylim([-1.5, 1.5])<br \/># init_utils.plot_decision_boundary(lambda x: init_utils.predict_dec(parameters, x.T), train_X, train_Y)<br \/><br \/># \u6b63\u5219\u5316<br \/># \u9996\u5148\u662f\u8bfb\u53d6\u6570\u636e<br \/># train_X, train_Y, test_X, test_Y = reg_utils.load_2D_dataset()<br \/># plt.show()<br \/><br \/><br \/>def compute_cost_with_regularization(A3, Y, parameters, lambd):<br \/>    <em>\"\"\"<br \/><\/em><em>    <\/em><em>\u6b63\u5219\u5316\u524d\u5411\u4f20\u64ad<br \/><\/em><em>    <\/em><strong><em>:return<\/em><\/strong><em>:<br \/><\/em><em>    \"\"\"<br \/><\/em><em>    <\/em>m = Y.shape[1]<br \/>    W1 = parameters['W1']<br \/>    W2 = parameters['W2']<br \/>    W3 = parameters['W3']<br \/><br \/>    cost = reg_utils.compute_cost(A3, Y)<br \/>    L2_regularization_cost = lambd * (np.sum(np.square(W1)) + np.sum(np.square(W2)) + np.sum(np.square(W3))) \/ (2 * m)<br \/>    cost = cost + L2_regularization_cost<br \/><br \/>    return cost<br \/><br \/><br \/>def backward_propagation_with_regularization(X, Y, cache, lambd):<br \/>    <em>\"\"\"<br \/><\/em><em>    <\/em><em>\u56e0\u4e3a\u4ee3\u4ef7\u51fd\u6570\u7684\u6539\u53d8\uff0c\u53cd\u5411\u4f20\u64ad\u7684\u8fc7\u7a0b\u4e5f\u8981\u6539\u53d8<br \/><\/em><em>    <\/em><strong><em>:param<\/em><\/strong><em> X:<br \/><\/em><em>    <\/em><strong><em>:param<\/em><\/strong><em> Y:<br \/><\/em><em>    <\/em><strong><em>:param<\/em><\/strong><em> cache:<br \/><\/em><em>    <\/em><strong><em>:param<\/em><\/strong><em> lambd:<br \/><\/em><em>    <\/em><strong><em>:return<\/em><\/strong><em>:<br \/><\/em><em>    \"\"\"<br \/><\/em><em>    <\/em>m = X.shape[1]<br \/><br \/>    (Z1, A1, W1, b1, Z2, A2, W2, b2, Z3, A3, W3, b3) = cache<br \/><br \/>    dZ3 = A3 - Y<br \/><br \/>    dW3 = (1 \/ m) * np.dot(dZ3, A2.T) + ((lambd * W3) \/ m)<br \/>    db3 = (1 \/ m) * np.sum(dZ3, axis=1, keepdims=True)<br \/><br \/>    dA2 = np.dot(W3.T, dZ3)<br \/>    dZ2 = np.multiply(dA2, np.int64(A2 &gt; 0))<br \/>    dW2 = (1 \/ m) * np.dot(dZ2, A1.T) + ((lambd * W2) \/ m)<br \/>    db2 = (1 \/ m) * np.sum(dZ2, axis=1, keepdims=True)<br \/><br \/>    dA1 = np.dot(W2.T, dZ2)<br \/>    dZ1 = np.multiply(dA1, np.int\n64(A1 &gt; 0))<br \/>    dW1 = (1 \/ m) * np.dot(dZ1, X.T) + ((lambd * W1) \/ m)<br \/>    db1 = (1 \/ m) * np.sum(dZ1, axis=1, keepdims=True)<br \/><br \/>    gradients = {\"dZ3\": dZ3, \"dW3\": dW3, \"db3\": db3, \"dA2\": dA2,<br \/>                 \"dZ2\": dZ2, \"dW2\": dW2, \"db2\": db2, \"dA1\": dA1,<br \/>                 \"dZ1\": dZ1, \"dW1\": dW1, \"db1\": db1}<br \/><br \/>    return gradients<br \/><br \/><br \/>def forward_propagation_with_dropout(X, parameters, keep_prob=0.5):<br \/>    <em>\"\"\"<br \/><\/em><em>    <\/em><em>\u968f\u673a\u820d\u5f03\u7ed3\u70b9\u7684\u524d\u5411\u4f20\u64ad<\/em><em><br \/><\/em><em>    :param X:<br \/><\/em><em>    :param parameters:<br \/><\/em><em>    :param keep_prob:<br \/><\/em><em>    :return:<br \/><\/em><em>    \"\"\"<br \/><\/em><em>    <\/em>np.random.seed(1)<br \/><br \/>    W1 = parameters[\"W1\"]<br \/>    b1 = parameters[\"b1\"]<br \/>    W2 = parameters[\"W2\"]<br \/>    b2 = parameters[\"b2\"]<br \/>    W3 = parameters[\"W3\"]<br \/>    b3 = parameters[\"b3\"]<br \/><br \/>    Z1 = np.dot(W1, X) + b1<br \/>    A1 = reg_utils.relu(Z1)<br \/><br \/>    # \u968f\u673a\u521d\u59cb\u5316\u4e00\u4e2a0~1\u7684\u77e9\u9635\uff0c\u7136\u540e\u6839\u636e\u521d\u59cb\u5316\u7684\u503c\u91cd\u65b0\u8ba1\u7b97A1\uff0c\u6700\u540e\u7f29\u5c0f\u4e00\u70b9A<br \/>    D1 = np.random.rand(A1.shape[0], A1.shape[1])<br \/>    D1 = D1 &lt; keep_prob<br \/>    A1 = A1 * D1<br \/>    A1 = A1 \/ keep_prob<br \/><br \/>    Z2 = np.dot(W2, A1) + b2<br \/>    A2 = reg_utils.relu(Z2)<br \/><br \/>    D2 = np.random.rand(A2.shape[0], A2.shape[1])<br \/>    D2 = D2 &lt; keep_prob<br \/>    A2 = A2 * D2<br \/>    A2 = A2 \/ keep_prob<br \/><br \/>    Z3 = np.dot(W3, A2) + b3<br \/>    A3 = reg_utils.sigmoid(Z3)<br \/><br \/>    cache = (Z1, D1, A1, W1, b1, Z2, D2, A2, W2, b2, Z3, A3, W3, b3)<br \/><br \/>    return A3, cache<br \/><br \/><br \/>def backward_propagation_with_dropout(X, Y, cache, keep_prob):<br \/>    <em>\"\"\"<br \/><\/em><em>    <\/em><em>\u968f\u673a\u820d\u5f03\u7ed3\u70b9\u7684\u53cd\u5411\u4f20\u64ad<\/em><em><br \/><\/em><em>    :param X:<br \/><\/em><em>    :param Y:<br \/><\/em><em>    :param cache:<br \/><\/em><em>    :param keep_prob:<br \/><\/em><em>    :return:<br \/><\/em><em>    \"\"\"<br \/><\/em><em>    <\/em>m = X.shape[1]<br \/>    (Z1, D1, A1, W1, b1, Z2, D2, A2, W2, b2, Z3, A3, W3, b3) = cache<br \/><br \/>    dZ3 = A3 - Y<br \/>    dW3 = (1 \/ m) * np.dot(dZ3, A2.T)<br \/>    db3 = 1. \/ m * np.sum(dZ3, axis=1, keepdims=True)<br \/>    dA2 = np.dot(W3.T, dZ3)<br \/><br \/>    # \u6839\u636e\u524d\u5411\u968f\u673a\u820d\u5f03\u7684\u7ed3\u70b9\u91cd\u65b0\u8ba1\u7b97A\uff0c\u5e76\u7f29\u653eA<br \/>    dA2 = dA2 * D2<br \/>    dA2 = dA2 \/ keep_prob<br \/><br \/>    dZ2 = np.multiply(dA2, np.int64(A2 &gt; 0))<br \/>    dW2 = 1. \/ m * np.dot(dZ2, A1.T)<br \/>    db2 = 1. \/ m * np.sum(dZ2, axis=1, keepdims=True)<br \/><br \/>    dA1 = np.dot(W2.T, dZ2)<br \/><br \/>    dA1 = dA1 * D1<br \/>    dA1 = dA1 \/ keep_prob<br \/><br \/>    dZ1 = np.multiply(dA1, np.int64(A1 &gt; 0))<br \/>    dW1 = 1. \/ m * np.dot(dZ1, X.T)<br \/>    db1 = 1. \/ m * np.sum(dZ1, axis=1, keepdims=True)<br \/><br \/>    gradients = {\"dZ3\": dZ3, \"dW3\": dW3, \"db3\": db3, \"dA2\": dA2,<br \/>                 \"dZ2\": dZ2, \"dW2\": dW2, \"db2\": db2, \"dA1\": dA1,<br \/>                 \"dZ1\": dZ1, \"dW1\": dW1, \"db1\": db1}<br \/><br \/>    return gradients<br \/><br \/><br \/>def model(X, Y, learning_rate=0.3, num_iterations=30000, print_cost=True, is_plot=True, lambd=0.0, keep_prob=1.0):<br \/>    <em>\"\"\"<br \/><\/em><em>    <\/em><em>\u6a21\u578b\u4e3b\u4f53<\/em><em><br \/><\/em><em>    :param X:<br \/><\/em><em>    :param Y:<br \/><\/em><em>    :param learning_rate:<br \/><\/em><em>    :param num_iterations:<br \/><\/em><em>    :param print_cost:<br \/><\/em><em>    :param is_plot:<br \/><\/em><em>    :param lambd:<br \/><\/em><em>    :param keep_prob:<br \/><\/em><em>    :return:<br \/><\/em><em>    \"\"\"<br \/><\/em><em>    <\/em>grads = {}<br \/>    costs = []<br \/>    m = X.shape[1]<br \/>    layers_dims = [X.shape[0], 20, 3, 1]<br \/><br \/>    # \u8fd9\u4e2a\u521d\u59cb\u5316\u540e\u53f0\u4e5f\u662f\u6291\u68af\u5ea6\u5f02\u5e38\u521d\u59cb\u5316<br \/>    parameters = reg_utils.initialize_parameters(layers_dims)<br \/><br \/>    for i in range(num_iterations):<br \/>        # dropout<br \/>        if keep_prob == 1:<br \/>            a3, cache = reg_utils.forward_propagation(X, parameters)<br \/>        elif keep_prob &lt; 1:<br \/>            a3, cache = forward_propagation_with_dropout(X, parameters, keep_prob)<br \/><br \/>        # \u6b63\u5219\u5316<br \/>        if lambd == 0:<br \/>            # \u4e0d\u4f7f\u7528L2\u6b63\u5219\u5316<br \/>            cost = reg_utils.compute_cost(a3, Y)<br \/>        else:<br \/>            # \u4f7f\u7528L2\u6b63\u5219\u5316<br \/>            cost = compute_cost_with_regularization(a3, Y, parameters, lambd)<br \/><br \/>            # \u53cd\u5411\u4f20\u64ad<br \/>            # \u53ef\u4ee5\u540c\u65f6\u4f7f\u7528L2\u6b63\u5219\u5316\u548c\u968f\u673a\u5220\u9664\u8282\u70b9\uff0c\u4f46\u662f\u672c\u6b21\u5b9e\u9a8c\u4e0d\u540c\u65f6\u4f7f\u7528\u3002<br \/>        assert (lambd == 0 or keep_prob == 1)<br \/><br \/>        # \u4e24\u4e2a\u53c2\u6570\u7684\u4f7f\u7528\u60c5\u51b5<br \/>        if (lambd == 0 and keep_prob == 1):<br \/>            # \u4e0d\u4f7f\u7528L2\u6b63\u5219\u5316\u548c\u4e0d\u4f7f\u7528\u968f\u673a\u5220\u9664\u8282\u70b9<br \/>            grads = reg_utils.backward_propagation(X, Y, cache)<br \/>        elif lambd != 0:<br \/>            # \u4f7f\u7528L2\u6b63\u5219\u5316\uff0c\u4e0d\u4f7f\u7528\u968f\u673a\u5220\u9664\u8282\u70b9<br \/>            grads = backward_propagation_with_regularization(X, Y, cache, lambd)<br \/>        elif keep_prob &lt; 1:<br \/>            # \u4f7f\u7528\u968f\u673a\u5220\u9664\u8282\u70b9\uff0c\u4e0d\u4f7f\u7528L2\u6b63\u5219\u5316<br \/>            grads = backward_propagation_with_dropout(X, Y, cache, keep_prob)<br \/><br \/>        parameters = reg_utils.update_parameters(parameters, grads, learning_rate)<br \/><br \/>        if i % 1000 == 0:<br \/>            costs.append(cost)<br \/>            if (print_cost and i % 5000 == 0):<br \/>                print(\"\u7b2c\" + str(i) + \"\u6b21\u8fed\u4ee3\uff0c\u6210\u672c\u503c\u4e3a\uff1a\" + str(cost))<br \/><br \/>    if is_plot:<br \/>        plt.plot(costs)<br \/>        plt.ylabel('cost')<br \/>        plt.xlabel('iterations (x1,000)')<br \/>        plt.title(\"Learning rate =\" + str(learning_rate))<br \/>        plt.show()<br \/><br \/>    return parameters<br \/><br \/><br \/># parameters = model(train_X, train_Y, is_plot=True)<br \/># parameters = model(train_X, train_Y, is_plot=True, lambd=0.7)<br \/># # parameters = model(train_X, train_Y, is_plot=True, keep_prob=0.86)<br \/># print(\"\u8bad\u7ec3\u96c6:\")<br \/># predictions_train = reg_utils.predict(train_X, train_Y, parameters)<br \/># print(\"\u6d4b\u8bd5\u96c6:\")<br \/># predictions_test = reg_utils.predict(test_X, test_Y, parameters)<br \/>#<br \/># plt.title(\"Model without regularization\")<br \/># axes = plt.gca()<br \/># axes.set_xlim([-0.75, 0.40])<br \/># axes.set_ylim([-0.75, 0.65])<br \/># reg_utils.plot_decision_boundary(lambda x: reg_utils.predict_dec(parameters, x.T), train_X, train_Y)<br \/><br \/><br \/># \u68af\u5ea6\u6821\u9a8c<br \/># \u73b0\u5728\u5047\u8bbe\u5b9a\u4e49\u4e00\u4e2a\u4e00\u7ef4\u7684\u7ebf\u6027\u51fd\u6570<br \/>def forward_propagation(x, theta):<br \/>    <em>\"\"\"<br \/><\/em><em>    <\/em><em>\u4e00\u7ef4\u7ebf\u6027\u51fd\u6570\u7684\u524d\u5411\u4f20\u64ad<\/em><em><br \/><\/em><em>    :param x:<br \/><\/em><em>    :param theta:<br \/><\/em><em>    :return:<br \/><\/em><em>    \"\"\"<br \/><\/em><em>    <\/em>J = np.dot(theta, x)<br \/>    return J<br \/><br \/><br \/>def backward_propagation(x, theta):<br \/>    <em>\"\"\"<br \/><\/em><em>    <\/em><em>\u4e00\u9636\u7ebf\u6027\u7684\u53cd\u5411\u4f20\u64ad<\/em><em><br \/><\/em><em>    :param x:<br \/><\/em><em>    :param theta:<br \/><\/em><em>    :return:<br \/><\/em><em>    \"\"\"<br \/><\/em><em>    <\/em>dtheta = x<br \/>    return dtheta<br \/><br \/><br \/>def gradient_check(x, theta, epsilon=1e-7):<br \/>    <em>\"\"\"<br \/><\/em><em>    <\/em><em>\u9a8c\u8bc1\u68af\u5ea6\u6821\u9a8c<\/em><em><br \/><\/em><em>    :param x:<br \/><\/em><em>    :param theta:<br \/><\/em><em>    :param epsilon:<br \/><\/em><em>    :return:<br \/><\/em><em>    \"\"\"<br \/><\/em><em><br \/><\/em><em>    <\/em># \u4f7f\u7528\u516c\u5f0f\uff083\uff09\u7684\u5de6\u4fa7\u8ba1\u7b97gradapprox\u3002<br \/>    thetaplus = theta + epsilon  # Step 1<br \/>    thetaminus = theta - epsilon  # Step 2<br \/>    J_plus = forward_propagation(x, thetaplus)  # Step 3<br \/>    J_minus = forward_propagation(x, thetaminus)  # Step 4<br \/>    gradapprox = (J_plus - J_minus) \/ (2 * epsilon)  # Step 5<br \/><br \/>    # \u68c0\u67e5gradapprox\u662f\u5426\u8db3\u591f\u63a5\u8fd1backward_propagation\uff08\uff09\u7684\u8f93\u51fa<br \/>    grad = backward_propagation(x, theta)<br \/><br \/>    # \u6c42\u4e8c\u8303\u6570<br \/>    numerator = np.linalg.norm(grad - gradapprox)  # Step 1'<br \/>    denominator = np.linalg.norm(grad) + np.linalg.norm(gradapprox)  # Step 2'<br \/>    difference = numerator \/ denominator  # Step 3'<br \/><br \/>    if difference &lt; 1e-7:<br \/>        print(\"\u68af\u5ea6\u68c0\u67e5\uff1a\u68af\u5ea6\u6b63\u5e38!\")<br \/>    else:<br \/>        print(\"\u68af\u5ea6\u68c0\u67e5\uff1a\u68af\u5ea6\u8d85\u51fa\u9608\u503c!\")<br \/><br \/>    return difference<br \/><br \/><br \/># print(\"-----------------\u6d4b\u8bd5gradient_check-----------------\")<br \/># x, theta = 2, 4<br \/># difference = gradient_check(x, theta)<br \/># print(\"difference = \" + str(difference))<br \/><br \/><br \/>def forward_propagation_n(X, Y, parameters):<br \/>    <em>\"\"\"<br \/><\/em><em>\n    <\/em><em>\u524d\u5411\u4f20\u64ad\u5e76\u8ba1\u7b97\u4ee3\u4ef7<\/em><em><br \/><\/em><em>    :param X:<br \/><\/em><em>    :param Y:<br \/><\/em><em>    :param parameters:<br \/><\/em><em>    :return:<br \/><\/em><em>    \"\"\"<br \/><\/em><em>    <\/em>m = X.shape[1]<br \/>    W1 = parameters[\"W1\"]<br \/>    b1 = parameters[\"b1\"]<br \/>    W2 = parameters[\"W2\"]<br \/>    b2 = parameters[\"b2\"]<br \/>    W3 = parameters[\"W3\"]<br \/>    b3 = parameters[\"b3\"]<br \/><br \/>    Z1 = np.dot(W1, X) + b1<br \/>    A1 = gc_utils.relu(Z1)<br \/><br \/>    Z2 = np.dot(W2, A1) + b2<br \/>    A2 = gc_utils.relu(Z2)<br \/><br \/>    Z3 = np.dot(W3, A2) + b3<br \/>    A3 = gc_utils.sigmoid(Z3)<br \/><br \/>    # \u8ba1\u7b97\u6210\u672c<br \/>    logprobs = np.multiply(-np.log(A3), Y) + np.multiply(-np.log(1 - A3), 1 - Y)<br \/>    cost = (1 \/ m) * np.sum(logprobs)<br \/><br \/>    cache = (Z1, A1, W1, b1, Z2, A2, W2, b2, Z3, A3, W3, b3)<br \/><br \/>    return cost, cache<br \/><br \/><br \/>def backward_propagation_n(X, Y, cache):<br \/>    <em>\"\"\"<br \/><\/em><em>    <\/em><em>\u53cd\u5411\u4f20\u64ad<\/em><em><br \/><\/em><em>    :param X:<br \/><\/em><em>    :param Y:<br \/><\/em><em>    :param cache:<br \/><\/em><em>    :return:<br \/><\/em><em>    \"\"\"<br \/><\/em><em>    <\/em>m = X.shape[1]<br \/>    (Z1, A1, W1, b1, Z2, A2, W2, b2, Z3, A3, W3, b3) = cache<br \/><br \/>    dZ3 = A3 - Y<br \/>    dW3 = 1. \/ m * np.dot(dZ3, A2.T)<br \/>    db3 = 1. \/ m * np.sum(dZ3, axis=1, keepdims=True)<br \/><br \/>    dA2 = np.dot(W3.T, dZ3)<br \/>    dZ2 = np.multiply(dA2, np.int64(A2 &gt; 0))<br \/>    dW2 = 1. \/ m * np.dot(dZ2, A1.T)<br \/>    db2 = 1. \/ m * np.sum(dZ2, axis=1, keepdims=True)<br \/><br \/>    dA1 = np.dot(W2.T, dZ2)<br \/>    dZ1 = np.multiply(dA1, np.int64(A1 &gt; 0))<br \/>    dW1 = 1. \/ m * np.dot(dZ1, X.T)<br \/>    db1 = 1. \/ m * np.sum(dZ1, axis=1, keepdims=True)<br \/><br \/>    gradients = {\"dZ3\": dZ3, \"dW3\": dW3, \"db3\": db3,<br \/>                 \"dA2\": dA2, \"dZ2\": dZ2, \"dW2\": dW2, \"db2\": db2,<br \/>                 \"dA1\": dA1, \"dZ1\": dZ1, \"dW1\": dW1, \"db1\": db1}<br \/><br \/>    return gradients<br \/><br \/><br \/>def gradient_check_n(parameters, gradients, X, Y, epsilon=1e-7):<br \/>    <em>\"\"\"<br \/><\/em><em>    <\/em><em>\u68af\u5ea6\u6821\u9a8c<\/em><em><br \/><\/em><em>    :param parameters:<br \/><\/em><em>    :param gradients:<br \/><\/em><em>    :param X:<br \/><\/em><em>    :param Y:<br \/><\/em><em>    :param epsilon:<br \/><\/em><em>    :return:<br \/><\/em><em>    \"\"\"<br \/><\/em><em>    <\/em># \u521d\u59cb\u5316\u53c2\u6570<br \/>    parameters_values, keys = gc_utils.dictionary_to_vector(parameters)  # keys\u7528\u4e0d\u5230<br \/>    grad = gc_utils.gradients_to_vector(gradients)<br \/>    num_parameters = parameters_values.shape[0]<br \/>    J_plus = np.zeros((num_parameters, 1))<br \/>    J_minus = np.zeros((num_parameters, 1))<br \/>    gradapprox = np.zeros((num_parameters, 1))<br \/><br \/>    # \u8ba1\u7b97gradapprox<br \/>    for i in range(num_parameters):<br \/>        # \u8ba1\u7b97J_plus [i]\u3002\u8f93\u5165\uff1a\u201cparameters_values\uff0cepsilon\u201d\u3002\u8f93\u51fa=\u201cJ_plus [i]\u201d<br \/>        thetaplus = np.copy(parameters_values)  # Step 1<br \/>        thetaplus[i][0] = thetaplus[i][0] + epsilon  # Step 2<br \/>        J_plus[i], cache = forward_propagation_n(X, Y, gc_utils.vector_to_dictionary(thetaplus))  # Step 3 \uff0ccache\u7528\u4e0d\u5230<br \/><br \/>        # \u8ba1\u7b97J_minus [i]\u3002\u8f93\u5165\uff1a\u201cparameters_values\uff0cepsilon\u201d\u3002\u8f93\u51fa=\u201cJ_minus [i]\u201d\u3002<br \/>        thetaminus = np.copy(parameters_values)  # Step 1<br \/>        thetaminus[i][0] = thetaminus[i][0] - epsilon  # Step 2<br \/>        J_minus[i], cache = forward_propagation_n(X, Y, gc_utils.vector_to_dictionary(thetaminus))  # Step 3 \uff0ccache\u7528\u4e0d\u5230<br \/><br \/>        # \u8ba1\u7b97gradapprox[i]<br \/>        gradapprox[i] = (J_plus[i] - J_minus[i]) \/ (2 * epsilon)<br \/><br \/>    # \u901a\u8fc7\u8ba1\u7b97\u5dee\u5f02\u6bd4\u8f83gradapprox\u548c\u540e\u5411\u4f20\u64ad\u68af\u5ea6\u3002<br \/>    numerator = np.linalg.norm(grad - gradapprox)  # Step 1'<br \/>    denominator = np.linalg.norm(grad) + np.linalg.norm(gradapprox)  # Step 2'<br \/>    difference = numerator \/ denominator  # Step 3'<br \/><br \/>    if difference &lt; 1e-7:<br \/>        print(\"\u68af\u5ea6\u68c0\u67e5\uff1a\u68af\u5ea6\u6b63\u5e38!\")<br \/>    else:<br \/>        print(\"\u68af\u5ea6\u68c0\u67e5\uff1a\u68af\u5ea6\u8d85\u51fa\u9608\u503c!\")<br \/><br \/>    return difference<br \/><\/pre>\n\n\n<p><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u5434\u6069\u8fbe\u6df1\u5ea6\u5b66\u4e60\u7b2c\u4e8c\u8bfe\u7b2c\u4e00\u5468 \u6df1\u5ea6\u5b66\u4e60\u7684\u5b9e\u7528\u5c42\u9762 \u672c\u5468\u6211\u4eec\u5c06\u5b66\u4e60\u8d85\u53c2\u6570\u7684\u8c03\u4f18 \u5982\u4f55\u6784\u5efa\u6570\u636e&nbsp; [&hellip;]<\/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":5492,"_links":{"self":[{"href":"http:\/\/www.sniper97.cn\/index.php\/wp-json\/wp\/v2\/posts\/2951"}],"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=2951"}],"version-history":[{"count":0,"href":"http:\/\/www.sniper97.cn\/index.php\/wp-json\/wp\/v2\/posts\/2951\/revisions"}],"wp:attachment":[{"href":"http:\/\/www.sniper97.cn\/index.php\/wp-json\/wp\/v2\/media?parent=2951"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/www.sniper97.cn\/index.php\/wp-json\/wp\/v2\/categories?post=2951"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/www.sniper97.cn\/index.php\/wp-json\/wp\/v2\/tags?post=2951"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}