{"id":6815,"date":"2022-08-31T21:43:02","date_gmt":"2022-08-31T13:43:02","guid":{"rendered":"http:\/\/139.9.1.231\/?p=6815"},"modified":"2022-09-05T14:46:03","modified_gmt":"2022-09-05T06:46:03","slug":"nlp_mclass","status":"publish","type":"post","link":"http:\/\/139.9.1.231\/index.php\/2022\/08\/31\/nlp_mclass\/","title":{"rendered":"\u5173\u4e8eNLP\u591a\u6807\u7b7e\u6587\u672c\u5206\u7c7b\u7684\u4e00\u4e9b\u601d\u8def&#8211;\uff08\u5f85\u66f4\u65b0\uff09"},"content":{"rendered":"\n\n\n<p class=\"has-light-pink-background-color has-background\">    \u4f5c\u4e3a\u521a\u5165\u95e8\u7684\u5c0f\u767d\uff0c\u6709\u5fc5\u8981\u53bb\u8bb0\u5f55\u4e00\u4e9bNLP\u5206\u7c7b\u4efb\u52a1\u7684\u5c0ftrick\uff0c\u611f\u89c9\u5bf9\u4e8e\u6da8\u70b9\u63d0\u5206\u5341\u5206\u6709\u7528\u3002\u8fd9\u4e2a\u6587\u7ae0\u540e\u9762\u6709\u65b0\u7684\u60f3\u6cd5\u4f1a\u6301\u7eed\u66f4\u65b0<\/p>\n\n\n\n<p>   \u534e\u4e3a\u6709\u4e00\u4e2aNLP\u5173\u4e8e\u533b\u5b66\u7535\u5b50\u75c5\u5386\u7684\u75be\u75c5\u591a\u6807\u7b7e\u5206\u7c7b\u6bd4\u8d5b\uff0c\u56e0\u4e3a\u4e4b\u524d\u6bd4\u8f83\u5c11\u53bb\u505aNLP\u65b9\u5411\u7684\u4e1c\u897f\uff0c\u4ec5\u4ec5\u662f\u5b66\u4e60\u8fc7\u76f8\u5173rnn\u3001transformer\u3001bert\u8bba\u6587\u5462\uff0c\u6240\u4ee5\uff0c\u53c2\u8d5b\u7eaf\u7cb9\u662f\u4e3a\u4e86\u4e86\u89e3\u4e86\u89e3NLP\u65b9\u5411\uff0c\u597d\u5728nlp\u505a\u6587\u672c\u5206\u7c7b\u7b97\u662f\u6bd4\u8f83\u7b80\u5355\u7684\u4e0b\u6e38\u4efb\u52a1\uff0c\u4f46\u5728\u53c2\u8d5b\u8fc7\u7a0b\u4e2d\uff0c\u4f1a\u53d1\u73b0\uff0c\u5176\u5b9e\u5bf9\u4e8e\u6587\u672c\u5206\u7c7b\u6765\u8bf4\uff0c\u57fa\u672c\u7684bert-base\u7684\u6548\u679c\u4e0d\u662f\u5f88\u597d\uff0c\u4f46\u5176\u5b9e\u611f\u89c9\u4e0d\u662f\u51fa\u5728\u6a21\u578b\u67b6\u6784\u65b9\u9762\uff0c\u5bf9\u4e8e\u7b80\u5355\u7684\u5206\u7c7b\u4efb\u52a1\uff0c\u4e00\u4e2a12\u5c42\u7684bert\u5e94\u8be5\u9002\u8db3\u4ee5\u80dc\u4efb\u4e86\uff0c\u56e0\u6b64\u5c06\u6ce8\u610f\u529b\u4e0d\u8981\u8fc7\u591a\u7684\u653e\u5728\u6a21\u578b\u7ed3\u6784\u4e0a\u3002<\/p>\n\n\n\n<h3><strong>\u4efb\u52a1\u8bf4\u660e<\/strong><\/h3>\n\n\n\n<p>\u672c\u8d5b\u9898\u662f\u5229\u7528\u75c5\u4eba\u7535\u5b50\u75c5\u5386\u6587\u672c\u4fe1\u606f\u63a8\u65ad\u51fa\u5176\u53ef\u80fd\u60a3\u6709\u75be\u75c5\u7684\u75be\u75c5\u8bca\u65ad\u4efb\u52a1\u3002\u7535\u5b50\u75c5\u5386\u6587\u672c\u4fe1\u606f\u4e3b\u8981\u5305\u62ec\u75c5\u4eba\u7684\u6027\u522b\u3001\u5e74\u9f84\u3001\u4e3b\u8bc9\u3001\u73b0\u75c5\u53f2\u3001\u65e2\u5f80\u53f2\u3001\u4f53\u683c\u68c0\u67e5\u548c\u8f85\u52a9\u68c0\u67e5\u3002\u6807\u7b7e\u4fe1\u606f\u4e3a\u75c5\u4eba\u7684\u51fa\u9662\u8bca\u65ad\u75be\u75c5\u3002\u672c\u8d5b\u9898\u4efb\u52a1\u9700\u8981\u6839\u636e\u75c5\u4eba\u7684\u7535\u5b50\u75c5\u5386\u6587\u672c\u4fe1\u606f\u63a8\u65ad\u51fa\u75c5\u4eba\u6240\u60a3\u6709\u7684\u5168\u90e8\u75be\u75c5\u3002\uff08\u6ce8\uff1a\u75c5\u4eba\u7684\u51fa\u9662\u8bca\u65ad\u75be\u75c5\u5e76\u4e0d\u662f\u5355\u4e00\u7684\uff09&nbsp;<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" width=\"702\" height=\"450\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/08\/image-484.png\" alt=\"\" class=\"wp-image-6905\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/08\/image-484.png 702w, http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/08\/image-484-300x192.png 300w\" sizes=\"(max-width: 702px) 100vw, 702px\" \/><\/figure>\n\n\n\n<p>\u6a21\u578b\u8f93\u51fa\u683c\u5f0f\uff1a<\/p>\n\n\n\n<p>{ &#8220;ZY000001&#8221;: [&#8220;\u9ad8\u8840\u538b&#8221;, &#8220;\u80ba\u6c14\u80bf&#8221;, &#8220;\u5148\u5929\u6027\u5fc3\u810f\u75c5&#8221;]}<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" width=\"893\" height=\"667\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/08\/image-485.png\" alt=\"\" class=\"wp-image-6907\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/08\/image-485.png 893w, http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/08\/image-485-300x224.png 300w, http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/08\/image-485-768x574.png 768w\" sizes=\"(max-width: 893px) 100vw, 893px\" \/><\/figure>\n\n\n\n<h2><strong>\u8bc4\u5206\u6807\u51c6<\/strong><\/h2>\n\n\n\n<p>\u672c\u8d5b\u9898\u91c7\u7528macro F1\u4f5c\u4e3a\u8bc4\u4ef7\u6307\u6807\u3002\u8bc4\u4ef7\u6307\u6807\u8ba1\u7b97\u516c\u5f0f\u5982\u4e0b\uff1a<\/p>\n\n\n\n<p>\u5bf9\u4e8e\u6bcf\u4e00\u4e2a\u9884\u6d4b\u7684\u75be\u75c5\u6709\u771f\u9633\u6027\uff08True Positive\uff0cTP\uff09\uff0c\u5047\u9633\u6027\uff08False Positive\uff0cFP\uff09\uff0c\u5047\u9634\u6027\uff08False Negative\uff09\uff0c\u771f\u9634\u6027\uff08True Negative\uff09\uff0cn\u8868\u793an\u79cd\u75be\u75c5\u3002<\/p>\n\n\n\n<div class=\"wp-block-image\"><figure class=\"aligncenter size-full\"><img loading=\"lazy\" width=\"425\" height=\"323\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/08\/image-486.png\" alt=\"\" class=\"wp-image-6909\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/08\/image-486.png 425w, http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/08\/image-486-300x228.png 300w\" sizes=\"(max-width: 425px) 100vw, 425px\" \/><\/figure><\/div>\n\n\n\n<p class=\"has-light-pink-background-color has-background\">\u8fd9\u4e2a\u5f97\u5206\u6700\u9ad8\u5728 0.83\u5de6\u53f3\u3002\u6211\u4e5f\u8bd5\u4e86\u51e0\u6b21\uff0c\u4f46\u52300.57\u5c31\u6ca1\u518d\u52a8\u8fc7\u4e86&#8230;..\uff0c\u540e\u9762\u51c6\u5907\u53bb\u5c1d\u8bd5\u4e0b\u4e0b\u9762\u7684\u65b9\u6cd5\uff0c\u770b\u770b\u6709\u5565\u6548\u679c\u5417\u3002<\/p>\n\n\n\n<h2>\u601d\u8def\uff1a<\/h2>\n\n\n\n<p>\u5728github\u4e2d\u627e\u5230\u4e00\u4e2a\u5206\u7c7b\u4f1a\u8bae\u4efb\u52a1\u7684\u6bd4\u8d5bppt\u6a21\u578b\u8bb2\u89e3\uff1a<\/p>\n\n\n\n<p><a href=\"https:\/\/github.com\/TJBioMedNLP\/chip2019task3\">https:\/\/github.com\/TJBioMedNLP\/chip2019task3<\/a><\/p>\n\n\n\n<p>\u5e9f\u8bdd\u8bf4\u5b8c\uff0c\u6765\u70b9\u6216\u8bb8\u80fd\u63d0\u5206\u7684\u5e72\u8d27\uff1a<\/p>\n\n\n\n<p>\u6570\u636e\u65b9\u9762\uff1a<\/p>\n\n\n\n<h3><strong>1\u3001\u6570\u636e\u6e05\u6d17\uff08\u5f88\u591a\u810f\u6570\u636e\uff09\u3001\u6570\u636e\u589e\u5f3a<\/strong><\/h3>\n\n\n\n<p>     <strong>\u8bf4\u5b9e\u8bdd\uff0c\u611f\u89c9\u8fd9\u4e2a\u6bd4\u8d5b\u7684\u8981\u70b9\u5c31\u662f\u6570\u636e\u5904\u7406\uff0c\u60f3\u63d0\u5206\u5c31\u770b\u4f60\u7684\u6570\u636e\u7684\u597d\u574f\uff0c\u73b0\u5728\u662f\u771f\u7684\u610f\u8bc6\u5230\u6570\u636e\u5904\u7406\u5bf9\u4e8e\u4e00\u4e2a\u6a21\u578b\u7684\u5f71\u54cd\u4e4b\u5927\u4e86\uff0c\u540e\u9762\u8981\u7740\u91cd\u5173\u6ce8\u4e0b\u8fd9\u65b9\u9762\u4e86\u3002<\/strong><\/p>\n\n\n\n<p>\u672c\u6b21\u63d0\u4f9b\u7684\u8bad\u7ec3\u96c6\u4e2d\u51fa\u73b0\u4e86\u4e00\u4e9b\u4e0d\u9700\u8981\u8bca\u65ad\u7684\u75be\u75c5\uff1a\u777e\u4e38\u9798\u819c\u79ef\u6db2\u3001\u5bab\u9888\u708e\u6027\u75be\u75c5\u3001\u53e3\u8154\u7c98\u819c\u6e83\u75a1\u3001\u5934\u90e8\u5916\u4f24\u3001\u6025\u6027\u9634\u9053\u708e\u3001\u5973\u6027\u76c6\u8154\u708e\u3001\u6025\u6027\u6c14\u7ba1\u708e\uff0c\u9700\u8981\u81ea\u5df1\u53bb\u5c06\u8be5\u7c7b\u6570\u636e\u6e05\u6d17\uff0c\u53e6\u5916\uff0c\u901a\u8fc7\u6570\u636e\u7edf\u8ba1\u5206\u6790\uff0c\u53ef\u4ee5\u83b7\u5f97\u8bad\u7ec3\u96c6\u4e2d\u5404\u4e2alabel\u7684\u6570\u91cf\u4e25\u91cd\u4e0d\u5e73\u8861\uff0c\u5982\u4f55\u5904\u7406\u4e5f\u662f\u4e00\u4e2a\u95ee\u9898\uff0c\u662f\u5426\u53ef\u4ee5\u901a\u8fc7\u6570\u636e\u96c6\u589e\u5f3a\uff0c\u63d0\u9ad8\u67d0\u4e9b\u7c7b\u522b\u7684\u6d4b\u8bd5\u6570\u636e\u3002<\/p>\n\n\n\n<p>\u53e6\u5916 \u6027\u522b\u3001\u5e74\u9f84\u3001\u4e3b\u8bc9\u3001\u73b0\u75c5\u53f2\u3001\u65e2\u5f80\u53f2\u3001\u4f53\u683c\u68c0\u67e5\u548c\u8f85\u52a9\u68c0\u67e5 \u7b49\u957f\u5ea6\u4f1a\u8d85\u51fa\u6a21\u578b\u7684\u6700\u5927\u957f\u5ea6\uff0c\u5982\u4f55\u89e3\u51b3\u3001\u6700\u5927\u5316\u5229\u7528\u4e0a\u8ff0\u4fe1\u606f\u4e5f\u662f\u4e00\u4e2a\u95ee\u9898\u3002\u6211\u505a\u8fc7\u5bf9\u8fd9\u4e9b\u6570\u636e\u505a\u8fc7\u5206\u6790\uff0c\u5bf9\u4e8e \u6027\u522b\u3001\u5e74\u9f84\u3001\u4e3b\u8bc9\u3001\u73b0\u75c5\u53f2\u3001\u65e2\u5f80\u53f2\u3001\u4f53\u683c\u68c0\u67e5\u548c\u8f85\u52a9\u68c0\u67e5 \u7b49 \u7edf\u8ba1\u8fc7\u5e73\u5747\u957f\u5ea6\u3001\u6700\u5927\u6700\u5c0f\u957f\u5ea6\uff0c\u4ece\u51e0\u5341\u5230\u51e0\u767e\u4e0d\u7b49\u3002\u53e6\u5916\uff0c\u53bb\u770b\u4e0b\u6570\u636e\u96c6\u5c31\u53ef\u4ee5\u770b\u5230\uff0c\u6709\u5927\u91cf\u7684\u6807\u70b9\u7b26\u53f7\u548c\u77ed\u8bed\u3002\u53e6\u5916\uff0c\u636e\u8bf4emr_id\u8fd9\u4e2a\u4fe1\u606f\u4e5f\u662f\u4e00\u4e2a\u91cd\u8981\u7684\u4fe1\u606f\uff1f\uff1f\uff1f\u6211\u4e00\u8138\u9707\u60ca\u3002\u6b64\u5916\u5e74\u9f84\u548c\u6027\u522b\u4e5f\u4f1a\u5f71\u54cd\u3002<\/p>\n\n\n\n<h3>\u5176\u4ed6\uff1a<\/h3>\n\n\n\n<p class=\"has-light-pink-background-color has-background\">1\u3001\u5982\u679c\u8bad\u7ec3\u65f6\u4f7f\u7528\u6587\u672c\u957f\u5ea6\u4e3an\uff0c\u6d4b\u8bd5\u4f7f\u7528\u6bd4n\u957f\u4e00\u4e9b\u7684\u957f\u5ea6\uff0c\u53ef\u4ee5\u6da8\u70b9\u5206<\/p>\n\n\n\n<p class=\"has-light-pink-background-color has-background\">2\u3001\u6a21\u578b\u9884\u8bad\u7ec3\u4f1a\u63d0\u5206\uff08\u6216\u8005\u627e\u76f8\u5173\u9886\u57df\u9884\u8bad\u7ec3\u6a21\u578b\uff09<\/p>\n\n\n\n<p>\u8fd9\u91cc\u6211\u627e\u4e86\u4e24\u4e2a\u9884\u8bad\u7ec3\u6a21\u578b\uff1a<\/p>\n\n\n\n<p><a href=\"https:\/\/huggingface.co\/trueto\/medbert-base-chinese\">https:\/\/huggingface.co\/trueto\/medbert-base-chinese<\/a><\/p>\n\n\n\n<p><a href=\"https:\/\/huggingface.co\/nghuyong\/ernie-health-zh\">https:\/\/huggingface.co\/nghuyong\/ernie-health-zh<\/a><\/p>\n\n\n\n<p class=\"has-light-pink-background-color has-background\">3\u3001\u589e\u5927\u8bad\u7ec3\u65f6\u8f93\u5165\u7f16\u7801\u957f\u5ea6\uff08\u6587\u672c\u5e8f\u5217\u957f\u5ea6\uff09\uff0c\u5f53\u7136\uff0c\u9700\u8981\u663e\u5361\u7684\u6027\u80fd\u652f\u6301<\/p>\n\n\n\n<p>all_tokens = self.tokenizer.encode_plus(content, max_length=pad_size, padding=&#8221;max_length&#8221;, truncation=True) <\/p>\n\n\n\n<p> \u53ef\u4ee5\u63d0\u9ad8max_length\u7684\u5927\u5c0f\uff0c\u4f46\u662f\u6bd4\u8f83\u5403\u663e\u5361\u3002<\/p>\n\n\n\n<p class=\"has-light-pink-background-color has-background\">4\u3001\u4ea4\u53c9\u9a8c\u8bc1<\/p>\n\n\n\n<div class=\"wp-block-image\"><figure class=\"aligncenter size-full\"><img loading=\"lazy\" width=\"634\" height=\"368\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/08\/image-488.png\" alt=\"\" class=\"wp-image-6952\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/08\/image-488.png 634w, http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/08\/image-488-300x174.png 300w\" sizes=\"(max-width: 634px) 100vw, 634px\" \/><\/figure><\/div>\n\n\n\n<p>      \u4ea4\u53c9\u9a8c\u8bc1\u7ecf\u5e38\u7528\u4e8e\u7ed9\u5b9a\u7684\u6570\u636e\u96c6\u8bad\u7ec3\u3001\u8bc4\u4f30\u548c\u6700\u7ec8\u9009\u62e9\u673a\u5668\u5b66\u4e60\u6a21\u578b\uff0c\u56e0\u4e3a\u5b83\u6709\u52a9\u4e8e\u8bc4\u4f30\u6a21\u578b\u7684\u7ed3\u679c\u5728\u5b9e\u8df5\u4e2d\u5982\u4f55\u63a8\u5e7f\u5230\u72ec\u7acb\u7684\u6570\u636e\u96c6\uff0c\u6700\u91cd\u8981\u7684\u662f\uff0c\u4ea4\u53c9\u9a8c\u8bc1\u5df2\u7ecf\u88ab\u8bc1\u660e\u4ea7\u751f\u6bd4\u5176\u4ed6\u65b9\u6cd5\u66f4\u4f4e\u7684\u504f\u5dee\u7684\u6a21\u578b\u3002\u91cd\u590d\u7684k\u6298\u4ea4\u53c9\u9a8c\u8bc1\uff0c\u4e3b\u8981\u662f\u4f1a\u91cd\u590d\u8fdb\u884cn\u6b21\u7684k\u6298\u4ea4\u53c9\u9a8c\u8bc1\uff0c\u8fd9\u6837\u4f1a\u4ea7\u751fn\u6b21\u7ed3\u679c\uff0c\u4e00\u822c\u901a\u8fc7\u5e73\u5747\u65b9\u6cd5\u6216\u8005\uff08\u6295\u7968\u89c4\u5219\uff09\u5f97\u5230\u6700\u540e\u7684\u7ed3\u679c<\/p>\n\n\n\n<p>\u3000\u7b2c\u4e00\u79cd\u662f<strong>\u7b80\u5355\u4ea4\u53c9\u9a8c\u8bc1<\/strong>\uff0c\u6240\u8c13\u7684\u7b80\u5355\uff0c\u662f\u548c\u5176\u4ed6\u4ea4\u53c9\u9a8c\u8bc1\u65b9\u6cd5\u76f8\u5bf9\u800c\u8a00\u7684\u3002\u9996\u5148\uff0c\u6211\u4eec\u968f\u673a\u7684\u5c06\u6837\u672c\u6570\u636e\u5206\u4e3a\u4e24\u90e8\u5206\uff08\u6bd4\u5982\uff1a 70%\u7684\u8bad\u7ec3\u96c6\uff0c30%\u7684\u6d4b\u8bd5\u96c6\uff09\uff0c\u7136\u540e\u7528\u8bad\u7ec3\u96c6\u6765\u8bad\u7ec3\u6a21\u578b\uff0c\u5728\u6d4b\u8bd5\u96c6\u4e0a\u9a8c\u8bc1\u6a21\u578b\u53ca\u53c2\u6570\u3002\u63a5\u7740\uff0c\u6211\u4eec\u518d1\u628a\u6837\u672c\u6253\u4e71\uff0c\u91cd\u65b0\u9009\u62e9\u8bad\u7ec3\u96c6\u548c\u6d4b\u8bd5\u96c6\uff0c\u7ee7\u7eed\u8bad\u7ec3\u6570\u636e\u548c\u68c0\u9a8c\u6a21\u578b\u3002\u6700\u540e\u6211\u4eec\u9009\u62e9\u635f\u5931\u51fd\u6570\u8bc4\u4f30\u6700\u4f18\u7684\u6a21\u578b\u548c\u53c2\u6570\u3002<\/p>\n\n\n\n<p>\u9009\u62e9\u5206\u5c42k-\u6298\u4ea4\u53c9\u9a8c\u8bc1\uff1a<\/p>\n\n\n\n<p>\u5206\u5c42\u91c7\u6837\u5c31\u662f\u5728\u6bcf\u4e00\u4efd\u5b50\u96c6\u4e2d\u90fd\u4fdd\u6301\u539f\u59cb\u6570\u636e\u96c6\u7684\u7c7b\u522b\u6bd4\u4f8b\uff0c\u4fdd\u8bc1\u91c7\u6837\u6570\u636e\u8ddf\u539f\u59cb\u6570\u636e\u7684\u7c7b\u522b\u5206\u5e03\u4fdd\u6301\u4e00\u81f4\uff0c\u8be5\u65b9\u6cd5\u5728\u6709\u6548\u7684\u5e73\u8861\u65b9\u5dee\u548c\u504f\u5dee\u3002\u5f53\u9488\u5bf9\u4e0d\u5e73\u8861\u6570\u636e\u65f6\uff0c\u4f7f\u7528\u968f\u673a\u7684K-fold\u4ea4\u53c9\u9a8c\u8bc1\uff0c\u53ef\u80fd\u51fa\u73b0\u5728\u5b50\u96c6\u4e2d\u53eb\u5c11\u7684\u7c7b\u522b\u7684\u5206\u5e03\u4e0e\u539f\u59cb\u7c7b\u522b\u5206\u5e03\u4e0d\u4e00\u81f4\u3002\u56e0\u6b64\uff0c\u9488\u5bf9\u4e0d\u5e73\u8861\u6570\u636e\u5f80\u5f80\u4f7f\u7528stratified k-fold\u4ea4\u53c9\u9a8c\u8bc1\u3002<\/p>\n\n\n\n<p>\u5f53\u8bad\u7ec3\u6570\u636e\u96c6\u4e0d\u80fd\u4ee3\u8868\u6574\u4e2a\u6570\u636e\u96c6\u5206\u5e03\u662f\uff0c\u8fd9\u65f6\u5019\u4f7f\u7528stratified k\u6298\u4ea4\u53c9\u9a8c\u8bc1\u53ef\u80fd\u4e0d\u662f\u597d\u7684\u65b9\u6cd5\uff0c\u800c\u53ef\u80fd\u6bd4\u8f83\u9002\u5408\u4f7f\u7528\u7b80\u5355\u7684\u91cd\u590d\u968f\u673ak\u6298\u4ea4\u53c9\u9a8c\u8bc1\u3002<\/p>\n\n\n\n<ul><li>1.\u628a\u6574\u4e2a\u6570\u636e\u96c6\u968f\u673a\u5212\u5206\u6210k\u4efd<\/li><li>2.\u7528\u5176\u4e2dk-1\u4efd\u8bad\u7ec3\u6a21\u578b\uff0c\u7136\u540e\u7528\u7b2ck\u4efd\u9a8c\u8bc1\u6a21\u578b<\/li><li>3.\u8bb0\u5f55\u6bcf\u4e2a\u9884\u6d4b\u7ed3\u679c\u83b7\u5f97\u7684\u8bef\u5dee<\/li><li>4.\u91cd\u590d\u8fd9\u4e2a\u8fc7\u7a0b\uff0c\u77e5\u9053\u6bcf\u4efd\u6570\u636e\u90fd\u505a\u8fc7\u9a8c\u8bc1\u96c6<\/li><li>5.\u8bb0\u5f55\u4e0b\u7684k\u4e2a\u8bef\u5dee\u7684\u5e73\u5747\u503c\uff0c\u88ab\u79f0\u4e3a\u4ea4\u53c9\u9a8c\u8bc1\u8bef\u5dee\u3002\u53ef\u4ee5\u88ab\u7528\u505a\u8861\u91cf\u6a21\u578b\u6027\u80fd\u7684\u6807\u51c6<\/li><\/ul>\n\n\n\n<pre class=\"wp-block-code\"><code>&gt;&gt;&gt; from sklearn.model_selection import StratifiedKFold\n&gt;&gt;&gt; X = np.array(&#091;&#091;1, 2], &#091;3, 4], &#091;1, 2], &#091;3, 4]])\n&gt;&gt;&gt; y = np.array(&#091;0, 0, 1, 1])\n&gt;&gt;&gt; skf = StratifiedKFold(n_splits=2)\n&gt;&gt;&gt; skf.get_n_splits(X, y)\n2\n&gt;&gt;&gt; print(skf)  \nStratifiedKFold(n_splits=2, random_state=None, shuffle=False)\n&gt;&gt;&gt; for train_index, test_index in skf.split(X, y):\n...    print(\"TRAIN:\", train_index, \"TEST:\", test_index)\n...    X_train, X_test = X&#091;train_index], X&#091;test_index]\n...    y_train, y_test = y&#091;train_index], y&#091;test_index]\nTRAIN: &#091;1 3] TEST: &#091;0 2]\nTRAIN: &#091;0 2] TEST: &#091;1 3]\n<\/code><\/pre>\n\n\n\n<p class=\"has-light-pink-background-color has-background\">\u5177\u4f53\u6765\u8bf4\uff1a<\/p>\n\n\n\n<p><strong>\u4ee5k-fold CV\u4e3a\u4f8b\uff1a<\/strong>\u4ecd\u7136\u662f\u628a\u539f\u59cb\u6570\u636e\u96c6\u5206\u6210\u8bad\u7ec3\u96c6\u548c\u6d4b\u8bd5\u96c6\uff0c\u4f46\u662f\u8bad\u7ec3\u6a21\u578b\u7684\u65f6\u5019\u4e0d\u4f7f\u7528\u6d4b\u8bd5\u96c6\u3002\u6700\u5e38\u89c1\u7684\u4e00\u4e2a\u53eb\u505a<em>k_fold CV<\/em>\u3002<\/p>\n\n\n\n<ul><li>\u5177\u4f53\u6765\u8bf4\u5c31\u662f\u628a\u8bad\u7ec3\u96c6\u5e73\u5206\u4e3ak\u4e2afold\uff0c\u5176\u4e2d\u6bcf\u4e2afold\u4f9d\u6b21\u4f5c\u4e3a\u6d4b\u8bd5\u96c6\u3001\u4f59\u4e0b\u7684\u4f5c\u4e3a\u8bad\u7ec3\u96c6\uff0c\u8fdb\u884ck\u6b21\u8bad\u7ec3\uff0c\u5f97\u5230\u5171\u8ba1k\u7ec4\u53c2\u6570\u3002<strong>\u53d6k\u7ec4\u53c2\u6570\u7684\u5747\u503c\u4f5c\u4e3a\u6a21\u578b\u7684\u6700\u7ec8\u53c2\u6570<\/strong>\u3002<\/li><\/ul>\n\n\n\n<ol><li><strong>\u4f18\u70b9\uff1a<\/strong>\u5145\u5206\u538b\u69a8\u4e86\u6570\u636e\u96c6\u7684\u4ef7\u503c\u3002\u5728\u6837\u672c\u96c6\u4e0d\u591f\u5927\u7684\u60c5\u51b5\u4e0b\u5c24\u5176\u73cd\u8d35\u3002<\/li><li><strong>\u7f3a\u70b9<\/strong>\uff1a\u8fd0\u7b97\u8d77\u6765\u82b1\u65f6\u95f4\u3002<\/li><\/ol>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" width=\"973\" height=\"482\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/08\/image-487.png\" alt=\"\" class=\"wp-image-6948\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/08\/image-487.png 973w, http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/08\/image-487-300x149.png 300w, http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/08\/image-487-768x380.png 768w\" sizes=\"(max-width: 973px) 100vw, 973px\" \/><\/figure>\n\n\n\n<p>K\u6298\u4ea4\u53c9\u9a8c\u8bc1\u8bad\u7ec3\u5355\u4e2a\u6a21\u578b\uff1a<\/p>\n\n\n\n<p>\u901a\u8fc7\u5bf9 k \u4e2a\u4e0d\u540c\u5206\u7ec4\u8bad\u7ec3\u7684\u7ed3\u679c\u8fdb\u884c\u5e73\u5747\u6765\u51cf\u5c11\u65b9\u5dee\uff0c\u56e0\u6b64\u6a21\u578b\u7684\u6027\u80fd\u5bf9\u6570\u636e\u7684\u5212\u5206\u5c31\u4e0d\u90a3\u4e48\u654f\u611f\uff0c\u7ecf\u8fc7\u591a\u6b21\u5212\u5206<a href=\"https:\/\/so.csdn.net\/so\/search?q=%E6%95%B0%E6%8D%AE%E9%9B%86&amp;spm=1001.2101.3001.7020\" target=\"_blank\" rel=\"noreferrer noopener\">\u6570\u636e\u96c6<\/a>\uff0c\u5927\u5927\u964d\u4f4e\u4e86\u7ed3\u679c\u7684\u5076\u7136\u6027\uff0c\u4ece\u800c\u63d0\u9ad8\u4e86\u6a21\u578b\u7684\u51c6\u786e\u6027\u3002\u5177\u4f53\u505a\u6cd5\u5982\u4e0b\uff1a<\/p>\n\n\n\n<ul><li>step1:\u4e0d\u91cd\u590d\u62bd\u6837\u5c06\u539f\u59cb\u6570\u636e\u968f\u673a\u5206\u4e3a k \u4efd\u3002<\/li><li>step2:\u6bcf\u4e00\u6b21\u6311\u9009\u5176\u4e2d 1 \u4efd\u4f5c\u4e3a\u9a8c\u8bc1\u96c6\uff0c\u5269\u4f59 k-1 \u4efd\u4f5c\u4e3a\u8bad\u7ec3\u96c6\u7528\u4e8e\u6a21\u578b\u8bad\u7ec3\u3002\u4e00\u5171\u8bad\u7ec3k\u4e2a\u6a21\u578b\u3002<\/li><li>step3\uff1a\u5728\u6bcf\u4e2a\u8bad\u7ec3\u96c6\u4e0a\u8bad\u7ec3\u540e\u5f97\u5230\u4e00\u4e2a\u6a21\u578b\uff0c\u7528\u8fd9\u4e2a\u6a21\u578b\u5728\u6d4b\u8bd5\u96c6\u4e0a\u6d4b\u8bd5\uff0c\u8ba1\u7b97\u5e76\u4fdd\u5b58\u6a21\u578b\u7684\u8bc4\u4f30\u6307\u6807\uff0c<\/li><li>step4\uff1a\u8ba1\u7b97 k \u7ec4\u6d4b\u8bd5\u7ed3\u679c\u7684\u5e73\u5747\u503c\u4f5c\u4e3a\u6a21\u578b\u6700\u7ec8\u5728\u6d4b\u8bd5\u96c6\u4e0a\u7684\u9884\u6d4b\u503c\uff0c\u6c42k \u4e2a\u6a21\u578b\u8bc4\u4f30\u6307\u6807\u7684\u5e73\u5747\u503c\uff0c\u5e76\u4f5c\u4e3a\u5f53\u524d k \u6298\u4ea4\u53c9\u9a8c\u8bc1\u4e0b\u6a21\u578b\u7684\u6027\u80fd\u6307\u6807\u3002<\/li><\/ul>\n\n\n\n<p>6\u3001\u6a21\u578b\u878d\u5408<\/p>\n\n\n\n<p><strong>\u6a21\u578b\u878d\u5408\uff1a<\/strong>\u901a\u8fc7\u878d\u5408\u591a\u4e2a\u4e0d\u540c\u7684\u6a21\u578b\uff0c\u53ef\u80fd\u63d0\u5347\u673a\u5668\u5b66\u4e60\u7684\u6027\u80fd\u3002\u8fd9\u4e00\u65b9\u6cd5\u5728\u5404\u79cd\u673a\u5668\u5b66\u4e60\u6bd4\u8d5b\u4e2d\u5e7f\u6cdb\u5e94\u7528\uff0c \u4e5f\u662f\u5728\u6bd4\u8d5b\u7684\u653b\u575a\u65f6\u523b\u51b2\u523aTop\u7684\u5173\u952e\u3002\u800c\u878d\u5408\u6a21\u578b\u5f80\u5f80\u53c8\u53ef\u4ee5\u4ece\u6a21\u578b\u7ed3\u679c\uff0c\u6a21\u578b\u81ea\u8eab\uff0c\u6837\u672c\u96c6\u7b49\u4e0d\u540c\u7684\u89d2\u5ea6\u8fdb\u884c\u878d\u5408\u3002\u5373\u591a\u4e2a\u6a21\u578b\u7684\u7ec4\u5408\u53ef\u4ee5\u6539\u5584\u6574\u4f53\u7684\u8868\u73b0\u3002\u96c6\u6210\u6a21\u578b\u662f\u4e00\u79cd\u80fd\u5728\u5404\u79cd\u7684\u673a\u5668\u5b66\u4e60\u4efb\u52a1\u4e0a\u63d0\u9ad8\u51c6\u786e\u7387\u7684\u5f3a\u6709\u529b\u6280\u672f\u3002<\/p>\n\n\n\n<p>\u6a21\u578b\u878d\u5408\u662f\u6bd4\u8d5b\u540e\u671f\u4e00\u4e2a\u91cd\u8981\u7684\u73af\u8282\uff0c\u5927\u4f53\u6765\u8bf4\u6709\u5982\u4e0b\u7684\u7c7b\u578b\u65b9\u5f0f\uff1a<\/p>\n\n\n\n<p>1. \u7b80\u5355\u52a0\u6743\u878d\u5408\uff1a<\/p>\n\n\n\n<ul><li>\u56de\u5f52\uff08\u5206\u7c7b\u6982\u7387\uff09\uff1a\u7b97\u672f\u5e73\u5747\u878d\u5408\uff08Arithmetic mean\uff09\uff0c\u51e0\u4f55\u5e73\u5747\u878d\u5408\uff08Geometric mean\uff09\uff1b<\/li><li>\u5206\u7c7b\uff1a\u6295\u7968\uff08Voting\uff09\uff1b<\/li><li>\u7efc\u5408\uff1a\u6392\u5e8f\u878d\u5408(Rank averaging)\uff0clog\u878d\u5408\u3002<\/li><\/ul>\n\n\n\n<p>2. stacking\/blending:<\/p>\n\n\n\n<ul><li>\u6784\u5efa\u591a\u5c42\u6a21\u578b\uff0c\u5e76\u5229\u7528\u9884\u6d4b\u7ed3\u679c\u518d\u62df\u5408\u9884\u6d4b\u3002<\/li><\/ul>\n\n\n\n<p>3. boosting\/bagging:<\/p>\n\n\n\n<ul><li>\u591a\u6811\u7684\u63d0\u5347\u65b9\u6cd5\uff0c\u5728xgboost\uff0cAdaboost,GBDT\u4e2d\u5df2\u7ecf\u7528\u5230\u3002<\/li><\/ul>\n\n\n\n<p id=\"%E5%B9%B3%E5%9D%87%E6%B3%95%EF%BC%88Averaging%EF%BC%89\"><strong>\u5e73\u5747\u6cd5\uff08Averaging\uff09<\/strong><\/p>\n\n\n\n<p><strong>\u57fa\u672c\u601d\u60f3\uff1a<\/strong>\u5bf9\u4e8e\u56de\u5f52\u95ee\u9898\uff0c\u4e00\u4e2a\u7b80\u5355\u76f4\u63a5\u7684\u601d\u8def\u662f\u53d6\u5e73\u5747\u3002\u7a0d\u7a0d\u6539\u8fdb\u7684\u65b9\u6cd5\u662f\u8fdb\u884c\u52a0\u6743\u5e73\u5747\u3002\u6743\u503c\u53ef\u4ee5\u7528\u6392\u5e8f\u7684\u65b9\u6cd5\u786e\u5b9a\uff0c\u4e3e\u4e2a\u4f8b\u5b50\uff0c\u6bd4\u5982A\u3001B\u3001C\u4e09\u79cd\u57fa\u672c\u6a21\u578b\uff0c\u6a21\u578b\u6548\u679c\u8fdb\u884c\u6392\u540d\uff0c\u5047\u8bbe\u6392\u540d\u5206\u522b\u662f1\uff0c2\uff0c3\uff0c\u90a3\u4e48\u7ed9\u8fd9\u4e09\u4e2a\u6a21\u578b\u8d4b\u4e88\u7684\u6743\u503c\u5206\u522b\u662f3\/6\u30012\/6\u30011\/6\u3002<\/p>\n\n\n\n<p>\u5e73\u5747\u6cd5\u6216\u52a0\u6743\u5e73\u5747\u6cd5\u770b\u4f3c\u7b80\u5355\uff0c\u5176\u5b9e\u540e\u9762\u7684\u9ad8\u7ea7\u7b97\u6cd5\u4e5f\u53ef\u4ee5\u8bf4\u662f\u57fa\u4e8e\u6b64\u800c\u4ea7\u751f\u7684\uff0cBagging\u6216\u8005Boosting\u90fd\u662f\u4e00\u79cd\u628a\u8bb8\u591a\u5f31\u5206\u7c7b\u5668\u8fd9\u6837\u878d\u5408\u6210\u5f3a\u5206\u7c7b\u5668\u7684\u601d\u60f3\u3002<\/p>\n\n\n\n<p><strong>\u7b80\u5355\u7b97\u672f\u5e73\u5747\u6cd5\uff1a<\/strong>Averaging\u65b9\u6cd5\u5c31\u591a\u4e2a\u6a21\u578b\u9884\u6d4b\u7684\u7ed3\u679c\u8fdb\u884c\u5e73\u5747\u3002\u8fd9\u79cd\u65b9\u6cd5\u65e2\u53ef\u4ee5\u7528\u4e8e\u56de\u5f52\u95ee\u9898\uff0c\u4e5f\u53ef\u4ee5\u7528\u4e8e\u5bf9\u5206\u7c7b\u95ee\u9898\u7684\u6982\u7387\u8fdb\u884c\u5e73\u5747\u3002<\/p>\n\n\n\n<p><strong>\u52a0\u6743\u7b97\u672f\u5e73\u5747\u6cd5\uff1a<\/strong>\u8fd9\u79cd\u65b9\u6cd5\u662f\u5e73\u5747\u6cd5\u7684\u6269\u5c55\u3002\u8003\u8651\u4e0d\u540c\u6a21\u578b\u7684\u80fd\u529b\u4e0d\u540c\uff0c\u5bf9\u6700\u7ec8\u7ed3\u679c\u7684\u8d21\u732e\u4e5f\u6709\u5dee\u5f02\uff0c\u9700\u8981\u7528\u6743\u91cd\u6765\u8868\u5f81\u4e0d\u540c\u6a21\u578b\u7684\u91cd\u8981\u6027importance\u3002<\/p>\n\n\n\n<p id=\"%E6%8A%95%E7%A5%A8%E6%B3%95%EF%BC%88voting%EF%BC%89\"><strong>\u6295\u7968\u6cd5\uff08voting\uff09<\/strong><\/p>\n\n\n\n<p><strong>\u57fa\u672c\u601d\u60f3\uff1a<\/strong>\u5047\u8bbe\u5bf9\u4e8e\u4e00\u4e2a\u4e8c\u5206\u7c7b\u95ee\u9898\uff0c\u67093\u4e2a\u57fa\u7840\u6a21\u578b\uff0c\u73b0\u5728\u6211\u4eec\u53ef\u4ee5\u5728\u8fd9\u4e9b\u57fa\u5b66\u4e60\u5668\u7684\u57fa\u7840\u4e0a\u5f97\u5230\u4e00\u4e2a\u6295\u7968\u7684\u5206\u7c7b\u5668\uff0c\u628a\u7968\u6570\u6700\u591a\u7684\u7c7b\u4f5c\u4e3a\u6211\u4eec\u8981\u9884\u6d4b\u7684\u7c7b\u522b\u3002<\/p>\n\n\n\n<p><strong>\u7edd\u5bf9\u591a\u6570\u6295\u7968\u6cd5\uff1a<\/strong>\u6700\u7ec8\u7ed3\u679c\u5fc5\u987b\u5728\u6295\u7968\u4e2d\u5360\u4e00\u534a\u4ee5\u4e0a\u3002<\/p>\n\n\n\n<p><strong>\u76f8\u5bf9\u591a\u6570\u6295\u7968\u6cd5\uff1a<\/strong>\u6700\u7ec8\u7ed3\u679c\u5728\u6295\u7968\u4e2d\u7968\u6570\u6700\u591a\u3002<\/p>\n\n\n\n<p><strong>\u52a0\u6743\u6295\u7968\u6cd5\uff1a<\/strong>\u6bcf\u4e2a\u5f31\u5b66\u4e60\u5668\u7684\u5206\u7c7b\u7968\u6570\u4e58\u4ee5\u6743\u91cd\uff0c\u5e76\u5c06\u5404\u4e2a\u7c7b\u522b\u7684\u52a0\u6743\u7968\u6570\u6c42\u548c\uff0c\u6700\u5927\u503c\u5bf9\u5e94\u7684\u7c7b\u522b\u5373\u6700\u7ec8\u7c7b\u522b\u3002<\/p>\n\n\n\n<p><strong>\u786c\u6295\u7968\uff1a<\/strong>\u5bf9\u591a\u4e2a\u6a21\u578b\u76f4\u63a5\u8fdb\u884c\u6295\u7968\uff0c\u4e0d\u533a\u5206\u6a21\u578b\u7ed3\u679c\u7684\u76f8\u5bf9\u91cd\u8981\u5ea6\uff0c\u6700\u7ec8\u6295\u7968\u6570\u6700\u591a\u7684\u7c7b\u4e3a\u6700\u7ec8\u88ab\u9884\u6d4b\u7684\u7c7b\u3002<\/p>\n\n\n\n<p><strong>\u8f6f\u6295\u7968\uff1a<\/strong>\u589e\u52a0\u4e86\u8bbe\u7f6e\u6743\u91cd\u7684\u529f\u80fd\uff0c\u53ef\u4ee5\u4e3a\u4e0d\u540c\u6a21\u578b\u8bbe\u7f6e\u4e0d\u540c\u6743\u91cd\uff0c\u8fdb\u800c\u533a\u522b\u6a21\u578b\u4e0d\u540c\u7684\u91cd\u8981\u5ea6\u3002<\/p>\n\n\n\n<p id=\"%E5%A0%86%E5%8F%A0%E6%B3%95%EF%BC%88Stacking%EF%BC%89\"><strong>\u5806\u53e0\u6cd5\uff08Stacking\uff09<\/strong><\/p>\n\n\n\n<p><strong>\u57fa\u672c\u601d\u60f3<\/strong><\/p>\n\n\n\n<p>stacking \u5c31\u662f\u5f53\u7528\u521d\u59cb\u8bad\u7ec3\u6570\u636e\u5b66\u4e60\u51fa\u82e5\u5e72\u4e2a\u57fa\u5b66\u4e60\u5668\u540e\uff0c\u5c06\u8fd9\u51e0\u4e2a\u5b66\u4e60\u5668\u7684\u9884\u6d4b\u7ed3\u679c\u4f5c\u4e3a\u65b0\u7684\u8bad\u7ec3\u96c6\uff0c\u6765\u5b66\u4e60\u4e00\u4e2a\u65b0\u7684\u5b66\u4e60\u5668\u3002\u5bf9\u4e0d\u540c\u6a21\u578b\u9884\u6d4b\u7684\u7ed3\u679c\u518d\u8fdb\u884c\u5efa\u6a21\u3002<\/p>\n\n\n\n<p>\u5c06\u4e2a\u4f53\u5b66\u4e60\u5668\u7ed3\u5408\u5728\u4e00\u8d77\u7684\u65f6\u5019\u4f7f\u7528\u7684\u65b9\u6cd5\u53eb\u505a\u7ed3\u5408\u7b56\u7565\u3002\u5bf9\u4e8e\u5206\u7c7b\u95ee\u9898\uff0c\u6211\u4eec\u53ef\u4ee5\u4f7f\u7528\u6295\u7968\u6cd5\u6765\u9009\u62e9\u8f93\u51fa\u6700\u591a\u7684\u7c7b\u3002\u5bf9\u4e8e\u56de\u5f52\u95ee\u9898\uff0c\u6211\u4eec\u53ef\u4ee5\u5c06\u5206\u7c7b\u5668\u8f93\u51fa\u7684\u7ed3\u679c\u6c42\u5e73\u5747\u503c\u3002<\/p>\n\n\n\n<p>\u4e0a\u9762\u8bf4\u7684\u6295\u7968\u6cd5\u548c\u5e73\u5747\u6cd5\u90fd\u662f\u5f88\u6709\u6548\u7684\u7ed3\u5408\u7b56\u7565\uff0c\u8fd8\u6709\u4e00\u79cd\u7ed3\u5408\u7b56\u7565\u662f\u4f7f\u7528\u53e6\u5916\u4e00\u4e2a\u673a\u5668\u5b66\u4e60\u7b97\u6cd5\u6765\u5c06\u4e2a\u4f53\u673a\u5668\u5b66\u4e60\u5668\u7684\u7ed3\u679c\u7ed3\u5408\u5728\u4e00\u8d77\uff0c\u8fd9\u4e2a\u65b9\u6cd5\u5c31\u662fStacking\u3002\u5728stacking\u65b9\u6cd5\u4e2d\uff0c\u6211\u4eec\u628a\u4e2a\u4f53\u5b66\u4e60\u5668\u53eb\u505a\u521d\u7ea7\u5b66\u4e60\u5668\uff0c\u7528\u4e8e\u7ed3\u5408\u7684\u5b66\u4e60\u5668\u53eb\u505a\u6b21\u7ea7\u5b66\u4e60\u5668\u6216\u5143\u5b66\u4e60\u5668\uff08metalearner\uff09\uff0c\u6b21\u7ea7\u5b66\u4e60\u5668\u7528\u4e8e\u8bad\u7ec3\u7684\u6570\u636e\u53eb\u505a\u6b21\u7ea7\u8bad\u7ec3\u96c6\u3002\u6b21\u7ea7\u8bad\u7ec3\u96c6\u662f\u5728\u8bad\u7ec3\u96c6\u4e0a\u7528\u521d\u7ea7\u5b66\u4e60\u5668\u5f97\u5230\u7684\u3002<\/p>\n\n\n\n<ul><li>step1\uff1a\u8bad\u7ec3T\u4e2a\u521d\u7ea7\u5b66\u4e60\u5668\uff0c\u8981\u4f7f\u7528\u4ea4\u53c9\u9a8c\u8bc1\u7684\u65b9\u6cd5\u5728Train Set\u4e0a\u9762\u8bad\u7ec3\uff08\u56e0\u4e3a\u7b2c\u4e8c\u9636\u6bb5\u5efa\u7acb\u5143\u5b66\u4e60\u5668\u7684\u6570\u636e\u662f\u521d\u7ea7\u5b66\u4e60\u5668\u8f93\u51fa\u7684\uff0c\u5982\u679c\u521d\u7ea7\u5b66\u4e60\u5668\u7684\u6cdb\u5316\u80fd\u529b\u4f4e\u4e0b\uff0c\u5143\u5b66\u4e60\u5668\u4e5f\u4f1a\u8fc7\u62df\u5408\uff09<\/li><li>step2\uff1aT\u4e2a\u521d\u7ea7\u5b66\u4e60\u5668\u5728Train Set\u4e0a\u8f93\u51fa\u7684\u9884\u6d4b\u503c\uff0c\u4f5c\u4e3a\u5143\u5b66\u4e60\u5668\u7684\u8bad\u7ec3\u6570\u636eD\uff0c\u6709T\u4e2a\u521d\u7ea7\u5b66\u4e60\u5668\uff0cD\u4e2d\u5c31\u6709T\u4e2a\u7279\u5f81\u3002D\u7684label\u548c\u8bad\u7ec3\u521d\u7ea7\u5b66\u4e60\u5668\u65f6\u7684label\u4e00\u81f4\u3002<\/li><li>step3\uff1aT\u4e2a\u521d\u7ea7\u5b66\u4e60\u5668\u5728Test Set\u4e0a\u8f93\u51fa\u7684\u9884\u6d4b\u503c\uff0c\u4f5c\u4e3a\u8bad\u7ec3\u5143\u5b66\u4e60\u5668\u65f6\u7684\u6d4b\u8bd5\u96c6\uff0c\u540c\u6837\u4e5f\u662f\u6709T\u4e2a\u6a21\u578b\u5c31\u6709T\u4e2a\u7279\u5f81\u3002<\/li><li>step4\uff1a\u8bad\u7ec3\u5143\u5b66\u4e60\u5668\uff0c\u5143\u5b66\u4e60\u5668\u8bad\u7ec3\u96c6D\u7684label\u548c\u8bad\u7ec3\u521d\u7ea7\u5b66\u4e60\u5668\u65f6\u7684label\u4e00\u81f4\u3002<\/li><\/ul>\n\n\n\n<p id=\"%E6%B7%B7%E5%90%88%E6%B3%95%EF%BC%88Blending%EF%BC%89\"><strong>\u6df7\u5408\u6cd5\uff08Blending\uff09<\/strong><\/p>\n\n\n\n<p><strong>\u57fa\u672c\u601d\u60f3\uff1a<\/strong>Blending\u91c7\u7528\u4e86\u548cstacking\u540c\u6837\u7684\u65b9\u6cd5\uff0c\u4e0d\u8fc7\u53ea\u4ece\u8bad\u7ec3\u96c6\u4e2d\u9009\u62e9\u4e00\u4e2afold\u7684\u7ed3\u679c\uff0c\u518d\u548c\u539f\u59cb\u7279\u5f81\u8fdb\u884cconcat\u4f5c\u4e3a\u5143\u5b66\u4e60\u5668meta learner\u7684\u7279\u5f81\uff0c\u6d4b\u8bd5\u96c6\u4e0a\u8fdb\u884c\u540c\u6837\u7684\u64cd\u4f5c\u3002<\/p>\n\n\n\n<p>\u628a\u539f\u59cb\u7684\u8bad\u7ec3\u96c6\u5148\u5206\u6210\u4e24\u90e8\u5206\uff0c\u6bd4\u598270%\u7684\u6570\u636e\u4f5c\u4e3a\u65b0\u7684\u8bad\u7ec3\u96c6\uff0c\u5269\u4e0b30%\u7684\u6570\u636e\u4f5c\u4e3a\u6d4b\u8bd5\u96c6\u3002<\/p>\n\n\n\n<ul><li>\u7b2c\u4e00\u5c42\uff0c\u6211\u4eec\u5728\u8fd970%\u7684\u6570\u636e\u4e0a\u8bad\u7ec3\u591a\u4e2a\u6a21\u578b\uff0c\u7136\u540e\u53bb\u9884\u6d4b\u90a330%\u6570\u636e\u7684label\uff0c\u540c\u65f6\u4e5f\u9884\u6d4btest\u96c6\u7684label\u3002<\/li><li>\u5728\u7b2c\u4e8c\u5c42\uff0c\u6211\u4eec\u5c31\u76f4\u63a5\u7528\u8fd930%\u6570\u636e\u5728\u7b2c\u4e00\u5c42\u9884\u6d4b\u7684\u7ed3\u679c\u505a\u4e3a\u65b0\u7279\u5f81\u7ee7\u7eed\u8bad\u7ec3\uff0c\u7136\u540e\u7528test\u96c6\u7b2c\u4e00\u5c42\u9884\u6d4b\u7684label\u505a\u7279\u5f81\uff0c\u7528\u7b2c\u4e8c\u5c42\u8bad\u7ec3\u7684\u6a21\u578b\u505a\u8fdb\u4e00\u6b65\u9884\u6d4b\u3002<\/li><\/ul>\n\n\n\n<p><strong>Blending\u8bad\u7ec3\u8fc7\u7a0b\uff1a<\/strong><\/p>\n\n\n\n<ol><li>\u6574\u4e2a\u8bad\u7ec3\u96c6\u5212\u5206\u6210\u8bad\u7ec3\u96c6training sets\u548c\u9a8c\u8bc1\u96c6validation sets\u4e24\u4e2a\u90e8\u5206\uff1b<\/li><li>\u5728training sets\u4e0a\u8bad\u7ec3\u6a21\u578b\uff1b<\/li><li>\u5728validation sets\u548ctest sets\u4e0a\u5f97\u5230\u9884\u6d4b\u7ed3\u679c\uff1b<\/li><li>\u5c06validation sets\u7684\u539f\u59cb\u7279\u5f81\u548c\u4e0d\u540c\u57fa\u6a21\u578bbase model\u9884\u6d4b\u5f97\u5230\u7684\u7ed3\u679c\u4f5c\u4e3a\u65b0\u7684\u5143\u5b66\u4e60\u5668meta learner\u7684\u8f93\u5165\uff0c\u8fdb\u884c\u8bad\u7ec3 \uff1b<\/li><li>\u4f7f\u7528\u8bad\u7ec3\u597d\u7684\u6a21\u578bmeta learner\u5728test sets\u4ee5\u53ca\u5728base model\u4e0a\u7684\u9884\u6d4b\u7ed3\u679c\u4e0a\u8fdb\u884c\u9884\u6d4b\uff0c\u5f97\u5230\u6700\u7ec8\u7ed3\u679c\u3002<\/li><\/ol>\n\n\n\n<p><strong>Stacking\u4e0eBlending\u7684\u5bf9\u6bd4\uff1a<\/strong><\/p>\n\n\n\n<p>\u4f18\u70b9\u5728\u4e8e\uff1a<\/p>\n\n\n\n<ul><li>blending\u6bd4stacking\u7b80\u5355\uff0c\u56e0\u4e3a\u4e0d\u7528\u8fdb\u884ck\u6b21\u7684\u4ea4\u53c9\u9a8c\u8bc1\u6765\u83b7\u5f97stacker feature<\/li><li>blending\u907f\u5f00\u4e86\u4e00\u4e2a\u4fe1\u606f\u6cc4\u9732\u95ee\u9898\uff1agenerlizers\u548cstacker\u4f7f\u7528\u4e86\u4e0d\u4e00\u6837\u7684\u6570\u636e\u96c6<\/li><\/ul>\n\n\n\n<p>\u7f3a\u70b9\u5728\u4e8e\uff1a<\/p>\n\n\n\n<ul><li>blending\u4f7f\u7528\u4e86\u5f88\u5c11\u7684\u6570\u636e\uff08\u7b2c\u4e8c\u9636\u6bb5\u7684blender\u53ea\u4f7f\u7528training set10%\u7684\u91cf\uff09<\/li><li>blender\u53ef\u80fd\u4f1a\u8fc7\u62df\u5408<\/li><li>stacking\u4f7f\u7528\u591a\u6b21\u7684\u4ea4\u53c9\u9a8c\u8bc1\u4f1a\u6bd4\u8f83\u7a33\u5065<\/li><\/ul>\n\n\n\n<p id=\"Bagging\"><strong>Bagging<\/strong><\/p>\n\n\n\n<p><strong>\u57fa\u672c\u601d\u60f3\uff1a<\/strong>Bagging\u57fa\u4e8ebootstrap\uff08\u81ea\u91c7\u6837\uff09\uff0c\u4e5f\u5c31\u662f\u6709\u653e\u56de\u7684\u91c7\u6837\u3002\u8bad\u7ec3\u5b50\u96c6\u7684\u5927\u5c0f\u548c\u539f\u59cb\u6570\u636e\u96c6\u7684\u5927\u5c0f\u76f8\u540c\u3002Bagging\u7684\u6280\u672f\u4f7f\u7528\u5b50\u96c6\u6765\u4e86\u89e3\u6574\u4e2a\u6837\u672c\u96c6\u7684\u5206\u5e03\uff0c\u901a\u8fc7bagging\u91c7\u6837\u7684\u5b50\u96c6\u7684\u5927\u5c0f\u8981\u5c0f\u4e8e\u539f\u59cb\u96c6\u5408\u3002<\/p>\n\n\n\n<ul><li>\u91c7\u7528bootstrap\u7684\u65b9\u6cd5\u57fa\u4e8e\u539f\u59cb\u6570\u636e\u96c6\u4ea7\u751f\u5927\u91cf\u7684\u5b50\u96c6<\/li><li>\u57fa\u4e8e\u8fd9\u4e9b\u5b50\u96c6\u8bad\u7ec3\u5f31\u6a21\u578bbase model<\/li><li>\u6a21\u578b\u662f\u5e76\u884c\u8bad\u7ec3\u5e76\u4e14\u76f8\u4e92\u72ec\u7acb\u7684<\/li><li>\u6700\u7ec8\u7684\u9884\u6d4b\u7ed3\u679c\u53d6\u51b3\u4e8e\u591a\u4e2a\u6a21\u578b\u7684\u9884\u6d4b\u7ed3\u679c<\/li><\/ul>\n\n\n\n<p>Bagging\u662f\u4e00\u79cd\u5e76\u884c\u5f0f\u7684\u96c6\u6210\u5b66\u4e60\u65b9\u6cd5\uff0c\u5373\u57fa\u5b66\u4e60\u5668\u7684\u8bad\u7ec3\u4e4b\u95f4\u6ca1\u6709\u524d\u540e\u987a\u5e8f\u53ef\u4ee5\u540c\u65f6\u8fdb\u884c\uff0cBagging\u4f7f\u7528\u201c\u6709\u653e\u56de\u201d\u91c7\u6837\u7684\u65b9\u5f0f\u9009\u53d6\u8bad\u7ec3\u96c6\uff0c\u5bf9\u4e8e\u5305\u542bm\u4e2a\u6837\u672c\u7684\u8bad\u7ec3\u96c6\uff0c\u8fdb\u884cm\u6b21\u6709\u653e\u56de\u7684\u968f\u673a\u91c7\u6837\u64cd\u4f5c\uff0c\u4ece\u800c\u5f97\u5230m\u4e2a\u6837\u672c\u7684\u91c7\u6837\u96c6\uff0c\u8fd9\u6837\u8bad\u7ec3\u96c6\u4e2d\u6709\u63a5\u8fd136.8%\u7684\u6837\u672c\u6ca1\u6709\u88ab\u91c7\u5230\u3002\u6309\u7167\u76f8\u540c\u7684\u65b9\u5f0f\u91cd\u590d\u8fdb\u884c\uff0c\u6211\u4eec\u5c31\u53ef\u4ee5\u91c7\u96c6\u5230T\u4e2a\u5305\u542bm\u4e2a\u6837\u672c\u7684\u6570\u636e\u96c6\uff0c\u4ece\u800c\u8bad\u7ec3\u51faT\u4e2a\u57fa\u5b66\u4e60\u5668\uff0c\u6700\u7ec8\u5bf9\u8fd9T\u4e2a\u57fa\u5b66\u4e60\u5668\u7684\u8f93\u51fa\u8fdb\u884c\u7ed3\u5408\u3002<\/p>\n\n\n\n<p id=\"Boosting\"><strong>Boosting<\/strong><\/p>\n\n\n\n<p><strong>\u57fa\u7840\u601d\u60f3\uff1a<\/strong>Boosting\u662f\u4e00\u79cd\u4e32\u884c\u7684\u5de5\u4f5c\u673a\u5236\uff0c\u5373\u4e2a\u4f53\u5b66\u4e60\u5668\u7684\u8bad\u7ec3\u5b58\u5728\u4f9d\u8d56\u5173\u7cfb\uff0c\u5fc5\u987b\u4e00\u6b65\u4e00\u6b65\u5e8f\u5217\u5316\u8fdb\u884c\u3002Boosting\u662f\u4e00\u4e2a\u5e8f\u5217\u5316\u7684\u8fc7\u7a0b\uff0c\u540e\u7eed\u6a21\u578b\u4f1a\u77eb\u6b63\u4e4b\u524d\u6a21\u578b\u7684\u9884\u6d4b\u7ed3\u679c\u3002\u4e5f\u5c31\u662f\u8bf4\uff0c\u4e4b\u540e\u7684\u6a21\u578b\u4f9d\u8d56\u4e8e\u4e4b\u524d\u7684\u6a21\u578b\u3002<\/p>\n\n\n\n<p>\u5176\u57fa\u672c\u601d\u60f3\u662f\uff1a\u589e\u52a0\u524d\u4e00\u4e2a\u57fa\u5b66\u4e60\u5668\u5728\u8bad\u7ec3\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u9884\u6d4b\u9519\u8bef\u6837\u672c\u7684\u6743\u91cd\uff0c\u4f7f\u5f97\u540e\u7eed\u57fa\u5b66\u4e60\u5668\u66f4\u52a0\u5173\u6ce8\u8fd9\u4e9b\u6253\u6807\u9519\u8bef\u7684\u8bad\u7ec3\u6837\u672c\uff0c\u5c3d\u53ef\u80fd\u7ea0\u6b63\u8fd9\u4e9b\u9519\u8bef\uff0c\u4e00\u76f4\u5411\u4e0b\u4e32\u884c\u76f4\u81f3\u4ea7\u751f\u9700\u8981\u7684T\u4e2a\u57fa\u5b66\u4e60\u5668\uff0cBoosting\u6700\u7ec8\u5bf9\u8fd9T\u4e2a\u5b66\u4e60\u5668\u8fdb\u884c\u52a0\u6743\u7ed3\u5408\uff0c\u4ea7\u751f\u5b66\u4e60\u5668\u59d4\u5458\u4f1a\u3002<\/p>\n\n\n\n<p><strong>Boosting\u8bad\u7ec3\u8fc7\u7a0b\uff1a<\/strong><\/p>\n\n\n\n<ul><li>\u57fa\u4e8e\u539f\u59cb\u6570\u636e\u96c6\u6784\u9020\u5b50\u96c6<\/li><li>\u521d\u59cb\u7684\u65f6\u5019\uff0c\u6240\u6709\u7684\u6570\u636e\u70b9\u90fd\u7ed9\u76f8\u540c\u7684\u6743\u91cd<\/li><li>\u57fa\u4e8e\u8fd9\u4e2a\u5b50\u96c6\u521b\u5efa\u4e00\u4e2a\u57fa\u6a21\u578b<\/li><li>\u4f7f\u7528\u8fd9\u4e2a\u6a21\u578b\u5728\u6574\u4e2a\u6570\u636e\u96c6\u4e0a\u8fdb\u884c\u9884\u6d4b<\/li><li>\u57fa\u4e8e\u771f\u5b9e\u503c\u548c\u9884\u6d4b\u503c\u8ba1\u7b97\u8bef\u5dee<\/li><li>\u88ab\u9884\u6d4b\u9519\u7684\u89c2\u6d4b\u503c\u4f1a\u8d4b\u4e88\u66f4\u5927\u7684\u6743\u91cd<\/li><li>\u518d\u6784\u9020\u4e00\u4e2a\u6a21\u578b\u57fa\u4e8e\u4e4b\u524d\u9884\u6d4b\u7684\u8bef\u5dee\u8fdb\u884c\u9884\u6d4b\uff0c\u8fd9\u4e2a\u6a21\u578b\u4f1a\u5c1d\u8bd5\u77eb\u6b63\u4e4b\u524d\u7684\u6a21\u578b<\/li><li>\u7c7b\u4f3c\u5730\uff0c\u6784\u9020\u591a\u4e2a\u6a21\u578b\uff0c\u6bcf\u4e00\u4e2a\u90fd\u4f1a\u77eb\u6b63\u4e4b\u524d\u7684\u8bef\u5dee<\/li><li>\u6700\u7ec8\u7684\u6a21\u578b\uff08strong learner\uff09\u662f\u6240\u6709\u5f31\u5b66\u4e60\u5668\u7684\u52a0\u6743\u878d\u5408<\/li><\/ul>\n\n\n\n<p class=\"has-light-pink-background-color has-background\">7\u3001\u635f\u5931\u51fd\u6570,\u6ce8\u610f\u4e0d\u540c\u7c7b\u522b\u7684\u6743\u91cd\uff08\u4f7f\u7528F1_loss\u3001Hamming Loss\u3001\u6570\u636e\u7c7b\u522b\u5206\u5e03\u4e0d\u5747\uff0c\u5982\u4f55\u89e3\u51b3\u957f\u5c3e\u5206\u5e03\uff08\u52a0\u6743\u635f\u5931\u3001\u5148\u9a8c\u6743\u91cd\uff09\uff09<\/p>\n\n\n\n<p class=\"has-light-gray-background-color has-background\">\u76f8\u5173\u8bba\u6587: <a href=\"https:\/\/arxiv.org\/abs\/2108.10566\" target=\"_blank\" rel=\"noreferrer noopener\">sigmoidF1: A Smooth F1 Score Surrogate Loss for Multilabel Classification<\/a><\/p>\n\n\n\n<p>      \u5e73\u65f6\u6211\u4eec\u5728\u505a\u591a\u6807\u7b7e\u5206\u7c7b\uff0c\u6216\u8005\u662f\u591a\u5206\u7c7b\u7684\u65f6\u5019\uff0c\u7ecf\u5e38\u4f7f\u7528\u7684loss\u51fd\u6570\u4e00\u822c\u662fbinary_crossentropy\uff08\u4e5f\u5c31\u662flog_loss)\u6216\u8005\u662fcategorical_crossentropy,\u4e0d\u8fc7\u4ea4\u53c9\u71b5\u5176\u5b9e\u8fd8\u662f\u6709\u70b9\u95ee\u9898\u7684,\u5728\u591a\u6807\u7b7e\u5206\u7c7b\u7684\u95ee\u9898\u91cc\uff0c\u4ea4\u53c9\u71b5\u5e76\u975e\u662f\u6700\u5408\u7406\u7684\u635f\u5931\u51fd\u6570\u3002\u5728\u591a\u6807\u7b7e\u5206\u7c7b\u7684\u95ee\u9898\u4e2d\uff0c\u6211\u4eec\u6700\u7ec8\u8bc4\u4ef7\u5f80\u5f80\u4f1a\u9009\u62e9F1\u5206\u6570\u4f5c\u4e3a\u8bc4\u4ef7\u6307\u6807\uff0c\u90a3\u4e48\u662f\u5426\u80fd\u76f4\u63a5\u5c06F1-score\u5236\u4f5c\u6210\u4e3a\u4e00\u4e2aloss\u51fd\u6570\u5462\uff1f\u5f53\u7136\u662f\u53ef\u4ee5\u7684\u3002<\/p>\n\n\n\n<p>\u5728\u591a\u5206\u7c7b\/\u591a\u6807\u7b7e\u5206\u7c7b\u4e2d\uff0cF1-score\u6709\u4e24\u79cd\u884d\u751f\u683c\u5f0f\uff0c\u5206\u522b\u662fmicro-F1\u548cmacro-F1\u3002\u662f\u4e24\u79cd\u4e0d\u540c\u7684\u8ba1\u7b97\u65b9\u5f0f\u3002<\/p>\n\n\n\n<p>micro-F1\u662f\u5148\u8ba1\u7b97\u5148\u62ff\u603b\u4f53\u6837\u672c\u6765\u8ba1\u7b97\u51faTP\u3001TN\u3001FP\u3001FN\u7684\u503c\uff0c\u518d\u4f7f\u7528\u8fd9\u4e9b\u503c\u8ba1\u7b97\u51fapercision\u548crecall\uff0c\u518d\u6765\u8ba1\u7b97\u51faF1\u503c\u3002<\/p>\n\n\n\n<p>macro-F1\u5219\u662f\u5148\u5bf9\u6bcf\u4e00\u79cd\u5206\u7c7b\uff0c\u89c6\u4f5c\u4e8c\u5206\u7c7b\uff0c\u8ba1\u7b97\u5176F1\u503c\uff0c\u6700\u540e\u518d\u5bf9\u6bcf\u4e00\u4e2a\u5206\u7c7b\u8fdb\u884c\u7b80\u5355\u5e73\u5747\u3002<\/p>\n\n\n\n<p>\u7b80\u5355\u7684\u8bb0\u7684\u8bdd\u5176\u5b9e\u662f\u8fd9\u6837\u7684\uff0cmicro\uff08\u5fae\u89c2\uff09\u4e0emacro(\u5b8f\u89c2)\u7684\u542b\u4e49\u5176\u5b9e\u662f\uff0cmicro-F1\u662f\u5728\u6837\u672c\u7684\u7b49\u7ea7\u4e0a\u505a\u5e73\u5747\uff0c\u662f\u6700\u5c0f\u9897\u7c92\u5ea6\u4e0a\u7684\u5e73\u5747\u4e86\uff0c\u6240\u4ee5\u662f\u5fae\u89c2\u3002macro-F1\u662f\u5728\u6bcf\u4e00\u4e2a\u5206\u7c7b\u7684\u5c42\u9762\u4e0a\u505a\u5e73\u5747\uff0c\u6bcf\u4e00\u4e2a\u5206\u7c7b\u90fd\u5305\u542b\u5f88\u591a\u6837\u672c\uff0c\u6240\u4ee5\u76f8\u5bf9\u662f\u5b8f\u89c2\u3002<\/p>\n\n\n\n<p id=\"\u4f5c\u4e3aloss\u51fd\u6570\u7684F1\"><strong>\u4f5c\u4e3aloss\u51fd\u6570\u7684F1<\/strong><\/p>\n\n\n\n<p>F1-score\u6539\u9020\u6210loss\u51fd\u6570\u76f8\u5bf9\u8f83\u4e3a\u7b80\u5355\uff0cF1\u662f\u8303\u56f4\u57280~1\u4e4b\u95f4\u7684\u6307\u6807\uff0c\u8d8a\u5927\u4ee3\u8868\u6027\u80fd\u8d8a\u597d\uff0c\u5728\u4f5c\u4e3aloss\u65f6\u53ea\u9700\u8981\u53d6\uff081-F1\uff09\u5373\u53ef\u3002<\/p>\n\n\n\n<p>\u4e00\u4e0b\u662fkeras\u4e2d\u7684\u5b9e\u73b0\uff1a<\/p>\n\n\n\n<p>\uff08\u8fd9\u91cc\u7684K\u5c31\u662fkeras\u7684\u540e\u7aef\uff0c\u4e00\u822c\u6765\u8bf4\u5c31\u662ftensorflow\uff09<\/p>\n\n\n\n<figure class=\"wp-block-table\"><table><tbody><tr><td>1<br>2<br>3<br>4<br>5<br>6<br>7<br>8<br>9<br>10<br>11<br>12<br>13<br>14<br>15<br>16<\/td><td>K = keras.backend<br>def f1_loss(y_true, y_pred):<br>#\u8ba1\u7b97tp\u3001tn\u3001fp\u3001fn<br>tp = K.sum(K.cast(y_true*y_pred, &#8216;float&#8217;), axis=0)<br>tn = K.sum(K.cast((1-y_true)*(1-y_pred), &#8216;float&#8217;), axis=0)<br>fp = K.sum(K.cast((1-y_true)*y_pred, &#8216;float&#8217;), axis=0)<br>fn = K.sum(K.cast(y_true*(1-y_pred), &#8216;float&#8217;), axis=0)<br><br>#percision\u4e0erecall\uff0c\u8fd9\u91cc\u7684K.epsilon\u4ee3\u8868\u4e00\u4e2a\u5c0f\u6b63\u6570\uff0c\u7528\u6765\u907f\u514d\u5206\u6bcd\u4e3a\u96f6<br>p = tp \/ (tp + fp + K.epsilon())<br>r = tp \/ (tp + fn + K.epsilon())<br><br>#\u8ba1\u7b97f1<br>f1 = 2*p*r \/ (p+r+K.epsilon())<br>f1 = tf.where(tf.is_nan(f1), tf.zeros_like(f1), f1)#\u5176\u5b9e\u5c31\u662f\u628anan\u6362\u62100<br>return 1 &#8211; K.mean(f1)<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<p>\u8fd9\u4e2a\u51fd\u6570\u53ef\u4ee5\u76f4\u63a5\u5728keras\u6a21\u578b\u7f16\u8bd1\u65f6\u4f7f\u7528\uff0c\u5982\u4e0b\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-table\"><table><tbody><tr><td>1<br>2<br>3<br>4<\/td><td># \u7c7b\u4f3c\u8fd9\u6837<br>model.compile(optimizer=tf.train.AdamOptimizer(0.003),<br>loss=f1_loss,<br>metrics=[&#8216;acc&#8217;,&#8217;mae&#8217;])<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<pre class=\"wp-block-code\"><code> \ndef f1_loss(predict, target):\n    predict = torch.sigmoid(predict)\n    predict = torch.clamp(predict * (1-target), min=0.01) + predict * target\n    tp = predict * target\n    tp = tp.sum(dim=0)\n    precision = tp \/ (predict.sum(dim=0) + 1e-8)\n    recall = tp \/ (target.sum(dim=0) + 1e-8)\n    f1 = 2 * (precision * recall \/ (precision + recall + 1e-8))\n    return 1 - f1.mean()<\/code><\/pre>\n\n\n\n<p class=\"has-light-pink-background-color has-background\">8\u3001\u8003\u8651\u5c06\u591a\u5206\u7c7b \u53d8\u6210\u591a\u4e2a\u4e8c\u5206\u7c7b\u4efb\u52a1<\/p>\n\n\n\n<p class=\"has-light-pink-background-color has-background\">9\u3001\u9664\u4e86bert\u6a21\u578b\uff0c\u8fd8\u53ef\u4ee5\u5c1d\u8bd5Performer\u3001ernie-health<\/p>\n\n\n\n<p>Performer \u662fICLR 2021\u7684\u65b0paper\uff0c\u5728\u5904\u7406\u957f\u5e8f\u5217\u9884\u6d4b\u65b9\u9762\u6709\u975e\u5e38\u4e0d\u9519\u7684\u7ed3\u679c\uff0c\u901f\u5ea6\u5feb\uff0c\u5185\u5b58\u5c0f\uff0c\u5728LRA\uff08long range arena \u4e00\u4e2a\u7edf\u4e00\u7684benchmark\uff09\u4e0a\u7efc\u5408\u5f97\u5206\u4e0d\u9519\u3002<\/p>\n\n\n\n<p>\u8bba\u6587\uff1a<a href=\"https:\/\/arxiv.org\/pdf\/2009.14794.pdf\">https:\/\/arxiv.org\/pdf\/2009.14794.pdf<\/a><\/p>\n\n\n\n<p> ernie-health \uff1aBuilding Chinese Biomedical Language Models via Multi-Level Text Discrimination<br>\u4e2d\u6587\u9898\u76ee\uff1a\u57fa\u4e8e\u591a\u5c42\u6b21\u6587\u672c\u8fa8\u6790\u6784\u5efa\u4e2d\u6587\u751f\u7269\u533b\u5b66\u8bed\u8a00\u6a21\u578b<br>\u8bba\u6587\u5730\u5740\uff1a<a rel=\"noreferrer noopener\" href=\"https:\/\/links.jianshu.com\/go?to=https%3A%2F%2Farxiv.org%2Fpdf%2F2110.07244.pdf\" target=\"_blank\">https:\/\/arxiv.org\/pdf\/2110.07244.pdf<\/a><br>\u9886\u57df\uff1a\u81ea\u7136\u8bed\u8a00\u5904\u7406\uff0c\u751f\u7269\u533b\u5b66<br>\u53d1\u8868\u65f6\u95f4\uff1a2021<br>\u4f5c\u8005\uff1aQuan Wang\u7b49\uff0c\u767e\u5ea6<br>\u6a21\u578b\u4e0b\u8f7d\uff1a<a rel=\"noreferrer noopener\" href=\"https:\/\/links.jianshu.com\/go?to=https%3A%2F%2Fhuggingface.co%2Fnghuyong%2Fernie-health-zh\" target=\"_blank\">https:\/\/huggingface.co\/nghuyong\/ernie-health-zh<\/a><br>\u6a21\u578b\u4ecb\u7ecd\uff1a<a rel=\"noreferrer noopener\" href=\"https:\/\/links.jianshu.com\/go?to=https%3A%2F%2Fgithub.com%2FPaddlePaddle%2FResearch%2Ftree%2Fmaster%2FKG%2FeHealth\" target=\"_blank\">https:\/\/github.com\/PaddlePaddle\/Research\/tree\/master\/KG\/eHealth<\/a><br>\u6a21\u578b\u4ee3\u7801\uff1a<a rel=\"noreferrer noopener\" href=\"https:\/\/links.jianshu.com\/go?to=https%3A%2F%2Fgithub.com%2FPaddlePaddle%2FPaddleNLP%2Ftree%2Fdevelop%2Fmodel_zoo%2Fernie-health\" target=\"_blank\">https:\/\/github.com\/PaddlePaddle\/PaddleNLP\/tree\/develop\/model_zoo\/ernie-health<\/a><\/p>\n\n\n\n<p class=\"has-light-pink-background-color has-background\">10\u3001<strong>NLP\u4e2d\u7684\u5bf9\u6297\u8bad\u7ec3&nbsp;<\/strong>\uff08\u6dfb\u52a0\u7684\u6270\u52a8\u662f\u5fae\u5c0f\uff09<\/p>\n\n\n\n<ol><li>\u63d0\u9ad8\u6a21\u578b\u5e94\u5bf9\u6076\u610f\u5bf9\u6297\u6837\u672c\u65f6\u7684\u9c81\u68d2\u6027\uff1b<\/li><li>\u4f5c\u4e3a\u4e00\u79cdregularization\uff0c\u51cf\u5c11overfitting\uff0c\u63d0\u9ad8\u6cdb\u5316\u80fd\u529b\u3002<\/li><\/ol>\n\n\n\n<p><strong>\u5bf9\u6297\u8bad\u7ec3<\/strong>\u5176\u5b9e\u662f\u201c\u5bf9\u6297\u201d\u5bb6\u65cf\u4e2d<strong>\u9632\u5fa1<\/strong>\u7684\u4e00\u79cd\u65b9\u5f0f\uff0c\u5176\u57fa\u672c\u7684\u539f\u7406\u5462\uff0c\u5c31\u662f\u901a\u8fc7\u6dfb\u52a0<strong>\u6270\u52a8<\/strong>\u6784\u9020\u4e00\u4e9b\u5bf9\u6297\u6837\u672c\uff0c\u653e\u7ed9\u6a21\u578b\u53bb\u8bad\u7ec3\uff0c<strong>\u4ee5\u653b\u4e3a\u5b88<\/strong>\uff0c\u63d0\u9ad8\u6a21\u578b\u5728\u9047\u5230\u5bf9\u6297\u6837\u672c\u65f6\u7684\u9c81\u68d2\u6027\uff0c\u540c\u65f6\u4e00\u5b9a\u7a0b\u5ea6\u4e5f\u80fd\u63d0\u9ad8\u6a21\u578b\u7684\u8868\u73b0\u548c\u6cdb\u5316\u80fd\u529b\u3002<\/p>\n\n\n\n<p>\u90a3\u4e48\uff0c\u4ec0\u4e48\u6837\u7684\u6837\u672c\u624d\u662f\u597d\u7684\u5bf9\u6297\u6837\u672c\u5462\uff1f\u5bf9\u6297\u6837\u672c\u4e00\u822c\u9700\u8981\u5177\u6709\u4e24\u4e2a\u7279\u70b9\uff1a<\/p>\n\n\n\n<ol><li>\u76f8\u5bf9\u4e8e\u539f\u59cb\u8f93\u5165\uff0c\u6240\u6dfb\u52a0\u7684\u6270\u52a8\u662f\u5fae\u5c0f\u7684\uff1b<\/li><li>\u80fd\u4f7f\u6a21\u578b\u72af\u9519\u3002<\/li><\/ol>\n\n\n\n<p>NLP\u4e2d\u7684\u4e24\u79cd\u5bf9\u6297\u8bad\u7ec3 + PyTorch\u5b9e\u73b0<\/p>\n\n\n\n<p><strong>a. Fast Gradient Method\uff08FGM\uff09<\/strong><\/p>\n\n\n\n<p>\u4e0a\u9762\u6211\u4eec\u63d0\u5230\uff0cGoodfellow\u572815\u5e74\u7684ICLR [7] \u4e2d\u63d0\u51fa\u4e86Fast Gradient Sign Method\uff08FGSM\uff09\uff0c\u968f\u540e\uff0c\u572817\u5e74\u7684ICLR [9]\u4e2d\uff0cGoodfellow\u5bf9FGSM\u4e2d\u8ba1\u7b97\u6270\u52a8\u7684\u90e8\u5206\u505a\u4e86\u4e00\u70b9\u7b80\u5355\u7684\u4fee\u6539\u3002\u5047\u8bbe\u8f93\u5165\u7684\u6587\u672c\u5e8f\u5217\u7684embedding vectors&nbsp;[v1,v2,&#8230;,vT]&nbsp;\u4e3a&nbsp;x&nbsp;\uff0cembedding\u7684\u6270\u52a8\u4e3a\uff1a<\/p>\n\n\n\n<div class=\"wp-block-image\"><figure class=\"aligncenter size-full\"><img loading=\"lazy\" width=\"347\" height=\"124\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/08\/image-489.png\" alt=\"\" class=\"wp-image-6994\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/08\/image-489.png 347w, http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/08\/image-489-300x107.png 300w\" sizes=\"(max-width: 347px) 100vw, 347px\" \/><\/figure><\/div>\n\n\n\n<p>\u5b9e\u9645\u4e0a\u5c31\u662f\u53d6\u6d88\u4e86\u7b26\u53f7\u51fd\u6570\uff0c\u7528\u4e8c\u8303\u5f0f\u505a\u4e86\u4e00\u4e2ascale\uff0c\u9700\u8981\u6ce8\u610f\u7684\u662f\uff1a\u8fd9\u91cc\u7684norm\u8ba1\u7b97\u7684\u662f\uff0c\u6bcf\u4e2a\u6837\u672c\u7684\u8f93\u5165\u5e8f\u5217\u4e2d\u51fa\u73b0\u8fc7\u7684\u8bcd\u7ec4\u6210\u7684\u77e9\u9635\u7684\u68af\u5ea6norm\u3002\u539f\u4f5c\u8005\u63d0\u4f9b\u4e86\u4e00\u4e2aTensorFlow\u7684\u5b9e\u73b0 [10]\uff0c\u5728\u4ed6\u7684\u5b9e\u73b0\u4e2d\uff0c\u516c\u5f0f\u91cc\u7684&nbsp;x&nbsp;\u662fembedding\u540e\u7684\u4e2d\u95f4\u7ed3\u679c\uff08batch_size, timesteps, hidden_dim\uff09\uff0c\u5bf9\u5176\u68af\u5ea6&nbsp;g&nbsp;\u7684\u540e\u9762\u4e24\u7ef4\u8ba1\u7b97norm\uff0c\u5f97\u5230\u7684\u662f\u4e00\u4e2a(batch_size, 1, 1)\u7684\u5411\u91cf&nbsp;||g||2&nbsp;\u3002\u4e3a\u4e86\u5b9e\u73b0\u63d2\u4ef6\u5f0f\u7684\u8c03\u7528\uff0c\u7b14\u8005\u5c06\u4e00\u4e2abatch\u62bd\u8c61\u6210\u4e00\u4e2a\u6837\u672c\uff0c\u4e00\u4e2abatch\u7edf\u4e00\u7528\u4e00\u4e2anorm\uff0c\u7531\u4e8e\u672c\u6765norm\u4e5f\u53ea\u662f\u4e00\u4e2ascale\u7684\u4f5c\u7528\uff0c\u5f71\u54cd\u4e0d\u5927\u3002\u5b9e\u73b0\u5982\u4e0b\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>import torch\nclass FGM():\n    def __init__(self, model):\n        self.model = model\n        self.backup = {}\n\n    def attack(self, epsilon=1., emb_name='emb.'):\n        # emb_name\u8fd9\u4e2a\u53c2\u6570\u8981\u6362\u6210\u4f60\u6a21\u578b\u4e2dembedding\u7684\u53c2\u6570\u540d\n        for name, param in self.model.named_parameters():\n            if param.requires_grad and emb_name in name:\n                self.backup&#091;name] = param.data.clone()\n                norm = torch.norm(param.grad)\n                if norm != 0 and not torch.isnan(norm):\n                    r_at = epsilon * param.grad \/ norm\n                    param.data.add_(r_at)\n\n    def restore(self, emb_name='emb.'):\n        # emb_name\u8fd9\u4e2a\u53c2\u6570\u8981\u6362\u6210\u4f60\u6a21\u578b\u4e2dembedding\u7684\u53c2\u6570\u540d\n        for name, param in self.model.named_parameters():\n            if param.requires_grad and emb_name in name: \n                assert name in self.backup\n                param.data = self.backup&#091;name]\n        self.backup = {}<\/code><\/pre>\n\n\n\n<p>\u9700\u8981\u4f7f\u7528\u5bf9\u6297\u8bad\u7ec3\u7684\u65f6\u5019\uff0c\u53ea\u9700\u8981\u6dfb\u52a0\u4e94\u884c\u4ee3\u7801\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code><em># \u521d\u59cb\u5316<\/em>\nfgm = FGM(model)\nfor batch_input, batch_label in data:\n    <em># \u6b63\u5e38\u8bad\u7ec3<\/em>\n    loss = model(batch_input, batch_label)\n    loss.backward() <em># \u53cd\u5411\u4f20\u64ad\uff0c\u5f97\u5230\u6b63\u5e38\u7684grad<\/em>\n    <em># \u5bf9\u6297\u8bad\u7ec3<\/em>\n    fgm.attack() <em># \u5728embedding\u4e0a\u6dfb\u52a0\u5bf9\u6297\u6270\u52a8<\/em>\n    loss_adv = model(batch_input, batch_label)\n    loss_adv.backward() <em># \u53cd\u5411\u4f20\u64ad\uff0c\u5e76\u5728\u6b63\u5e38\u7684grad\u57fa\u7840\u4e0a\uff0c\u7d2f\u52a0\u5bf9\u6297\u8bad\u7ec3\u7684\u68af\u5ea6<\/em>\n    fgm.restore() <em># \u6062\u590dembedding\u53c2\u6570<\/em>\n    <em># \u68af\u5ea6\u4e0b\u964d\uff0c\u66f4\u65b0\u53c2\u6570<\/em>\n    optimizer.step()\n    model.zero_grad()<\/code><\/pre>\n\n\n\n<p>PyTorch\u4e3a\u4e86\u8282\u7ea6\u5185\u5b58\uff0c\u5728backward\u7684\u65f6\u5019\u5e76\u4e0d\u4fdd\u5b58\u4e2d\u95f4\u53d8\u91cf\u7684\u68af\u5ea6\u3002\u56e0\u6b64\uff0c\u5982\u679c\u9700\u8981\u5b8c\u5168\u7167\u642c\u539f\u4f5c\u7684\u5b9e\u73b0\uff0c\u9700\u8981\u7528<code>register_hook<\/code>\u63a5\u53e3[11]\u5c06embedding\u540e\u7684\u4e2d\u95f4\u53d8\u91cf\u7684\u68af\u5ea6\u4fdd\u5b58\u6210\u5168\u5c40\u53d8\u91cf\uff0cnorm\u540e\u9762\u4e24\u7ef4\uff0c\u8ba1\u7b97\u51fa\u6270\u52a8\u540e\uff0c\u5728\u5bf9\u6297\u8bad\u7ec3forward\u65f6\u4f20\u5165\u6270\u52a8\uff0c\u7d2f\u52a0\u5230embedding\u540e\u7684\u4e2d\u95f4\u53d8\u91cf\u4e0a\uff0c\u5f97\u5230\u65b0\u7684loss\uff0c\u518d\u8fdb\u884c\u68af\u5ea6\u4e0b\u964d\u3002<\/p>\n\n\n\n<p><strong>b. Projected Gradient Descent\uff08PGD\uff09<\/strong><\/p>\n\n\n\n<p>\u5185\u90e8max\u7684\u8fc7\u7a0b\uff0c\u672c\u8d28\u4e0a\u662f\u4e00\u4e2a\u975e\u51f9\u7684\u7ea6\u675f\u4f18\u5316\u95ee\u9898\uff0cFGM\u89e3\u51b3\u7684\u601d\u8def\u5176\u5b9e\u5c31\u662f\u68af\u5ea6\u4e0a\u5347\uff0c<strong>\u90a3\u4e48FGM\u7b80\u5355\u7c97\u66b4\u7684\u201c\u4e00\u6b65\u5230\u4f4d\u201d\uff0c\u662f\u4e0d\u662f\u6709\u53ef\u80fd\u5e76\u4e0d\u80fd\u8d70\u5230\u7ea6\u675f\u5185\u7684\u6700\u4f18\u70b9\u5462\uff1f<\/strong>\u5f53\u7136\u662f\u6709\u53ef\u80fd\u7684\u3002\u4e8e\u662f\uff0c\u4e00\u4e2a\u5f88intuitive\u7684\u6539\u8fdb\u8bde\u751f\u4e86\uff1aMadry\u572818\u5e74\u7684ICLR\u4e2d[8]\uff0c\u63d0\u51fa\u4e86\u7528Projected Gradient Descent\uff08PGD\uff09\u7684\u65b9\u6cd5\uff0c\u7b80\u5355\u7684\u8bf4\uff0c\u5c31\u662f<strong>\u201c\u5c0f\u6b65\u8d70\uff0c\u591a\u8d70\u51e0\u6b65\u201d<\/strong>\uff0c\u5982\u679c\u8d70\u51fa\u4e86\u6270\u52a8\u534a\u5f84\u4e3a&nbsp;\u03f5&nbsp;\u7684\u7a7a\u95f4\uff0c\u5c31\u6620\u5c04\u56de\u201c\u7403\u9762\u201d\u4e0a\uff0c\u4ee5\u4fdd\u8bc1\u6270\u52a8\u4e0d\u8981\u8fc7\u5927\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" width=\"947\" height=\"190\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/08\/image-490.png\" alt=\"\" class=\"wp-image-6995\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/08\/image-490.png 947w, http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/08\/image-490-300x60.png 300w, http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/08\/image-490-768x154.png 768w\" sizes=\"(max-width: 947px) 100vw, 947px\" \/><\/figure>\n\n\n\n<pre class=\"wp-block-code\"><code>import torch\nclass PGD():\n    def __init__(self, model):\n        self.model = model\n        self.emb_backup = {}\n        self.grad_backup = {}\n\n    def attack(self, epsilon=1., alpha=0.3, emb_name='emb.', is_first_attack=False):\n        # emb_name\u8fd9\u4e2a\u53c2\u6570\u8981\u6362\u6210\u4f60\u6a21\u578b\u4e2dembedding\u7684\u53c2\u6570\u540d\n        for name, param in self.model.named_parameters():\n            if param.requires_grad and emb_name in name:\n                if is_first_attack:\n                    self.emb_backup&#091;name] = param.data.clone()\n                norm = torch.norm(param.grad)\n                if norm != 0 and not torch.isnan(norm):\n                    r_at = alpha * param.grad \/ norm\n                    param.data.add_(r_at)\n                    param.data = self.project(name, param.data, epsilon)\n\n    def restore(self, emb_name='emb.'):\n        # emb_name\u8fd9\u4e2a\u53c2\u6570\u8981\u6362\u6210\u4f60\u6a21\u578b\u4e2dembedding\u7684\u53c2\u6570\u540d\n        for name, param in self.model.named_parameters():\n            if param.requires_grad and emb_name in name: \n                assert name in self.emb_backup\n                param.data = self.emb_backup&#091;name]\n        self.emb_backup = {}\n\n    def project(self, param_name, param_data, epsilon):\n        r = param_data - self.emb_backup&#091;param_name]\n        if torch.norm(r) &gt; epsilon:\n            r = epsilon * r \/ torch.norm(r)\n        return self.emb_backup&#091;param_name] + r\n\n    def backup_grad(self):\n        for name, param in self.model.named_parameters():\n            if param.requires_grad:\n                self.grad_backup&#091;name] = param.grad.clone()\n\n    def restore_grad(self):\n        for name, param in self.model.named_parameters():\n            if param.requires_grad:\n                param.grad = self.grad_backup&#091;name]<\/code><\/pre>\n\n\n\n<p>\u4f7f\u7528\u7684\u65f6\u5019\uff0c\u8981\u9ebb\u70e6\u4e00\u70b9\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>pgd = PGD(model)\nK = 3\nfor batch_input, batch_label in data:\n    <em># \u6b63\u5e38\u8bad\u7ec3<\/em>\n    loss = model(batch_input, batch_label)\n    loss.backward() <em># \u53cd\u5411\u4f20\u64ad\uff0c\u5f97\u5230\u6b63\u5e38\u7684grad<\/em>\n    pgd.backup_grad()\n    <em># \u5bf9\u6297\u8bad\u7ec3<\/em>\n    for t in range(K):\n        pgd.attack(is_first_attack=(t==0)) <em># \u5728embedding\u4e0a\u6dfb\u52a0\u5bf9\u6297\u6270\u52a8, first attack\u65f6\u5907\u4efdparam.data<\/em>\n        if t != K-1:\n            model.zero_grad()\n        else:\n            pgd.restore_grad()\n        loss_adv = model(batch_input, batch_label)\n        loss_adv.backward() <em># \u53cd\u5411\u4f20\u64ad\uff0c\u5e76\u5728\u6b63\u5e38\u7684grad\u57fa\u7840\u4e0a\uff0c\u7d2f\u52a0\u5bf9\u6297\u8bad\u7ec3\u7684\u68af\u5ea6<\/em>\n    pgd.restore() <em># \u6062\u590dembedding\u53c2\u6570<\/em>\n    <em># \u68af\u5ea6\u4e0b\u964d\uff0c\u66f4\u65b0\u53c2\u6570<\/em>\n    optimizer.step()\n    model.zero_grad()<\/code><\/pre>\n\n\n\n<p>\u5b9e\u9a8c\u5bf9\u7167<\/p>\n\n\n\n<p>\u4e3a\u4e86\u8bf4\u660e\u5bf9\u6297\u8bad\u7ec3\u7684\u4f5c\u7528\uff0c\u7b14\u8005\u9009\u4e86\u56db\u4e2aGLUE\u4e2d\u7684\u4efb\u52a1\u8fdb\u884c\u4e86\u5bf9\u7167\u8bd5\u9a8c\u3002\u5b9e\u9a8c\u4ee3\u7801\u662f\u7528\u7684Huggingface\u7684<code>transfomers\/examples\/run_glue.py<\/code>&nbsp;[12]\uff0c\u8d85\u53c2\u90fd\u662f\u9ed8\u8ba4\u7684\uff0c\u5bf9\u6297\u8bad\u7ec3\u7528\u7684\u4e5f\u662f\u76f8\u540c\u7684\u8d85\u53c2\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" width=\"1024\" height=\"310\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/08\/image-491-1024x310.png\" alt=\"\" class=\"wp-image-6997\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/08\/image-491-1024x310.png 1024w, http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/08\/image-491-300x91.png 300w, http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/08\/image-491-768x232.png 768w, http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/08\/image-491.png 1186w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p>\u9664\u4e86\u76d1\u7763\u8bad\u7ec3\uff0c\u5bf9\u6297\u8bad\u7ec3\u8fd8\u53ef\u4ee5\u7528\u5728\u534a\u76d1\u7763\u4efb\u52a1\u4e2d\uff0c\u5c24\u5176\u5bf9\u4e8eNLP\u4efb\u52a1\u6765\u8bf4\uff0c\u5f88\u591a\u65f6\u5019\u8f93\u5165\u7684\u65e0\u76d1\u7763\u6587\u672c\u591a\u7684\u5f88\uff0c\u4f46\u662f\u5f88\u96be\u5927\u89c4\u6a21\u5730\u8fdb\u884c\u6807\u6ce8\uff0cDistributional Smoothing with Virtual Adversarial Training.&nbsp;<a rel=\"noreferrer noopener\" href=\"https:\/\/arxiv.org\/abs\/1507.00677\" target=\"_blank\">https:\/\/arxiv.org\/abs\/1507.00677<\/a>   \u63d0\u5230\u7684 Virtual Adversarial Training\u8fdb\u884c\u534a\u76d1\u7763\u8bad\u7ec3\u3002<\/p>\n\n\n\n<p class=\"has-light-pink-background-color has-background\">1<strong>1\u3001Pseudo Labeling\uff08\u4f2a\u6807\u7b7e\uff09\u63d0\u9ad8\u6a21\u578b\u7684\u5206\u7c7b\u6548\u679c<\/strong><\/p>\n\n\n\n<p>\u7b80\u800c\u8a00\u4e4b\uff0cPseudo Labeling\u5c06\u6d4b\u8bd5\u96c6\u4e2d\u5224\u65ad\u7ed3\u679c\u6b63\u786e\u7684\u7f6e\u4fe1\u5ea6\u9ad8\u7684\u6837\u672c\u52a0\u5165\u5230\u8bad\u7ec3\u96c6\u4e2d\uff0c\u4ece\u800c\u6a21\u62df\u4e00\u90e8\u5206\u4eba\u7c7b\u5bf9\u65b0\u5bf9\u8c61\u8fdb\u884c\u5224\u65ad\u63a8\u6f14\u7684\u8fc7\u7a0b\u3002\u6548\u679c\u6bd4\u4e0d\u4e0a\u4eba\u8111\u90a3\u4e48\u597d\uff0c\u4f46\u662f\u5728\u76d1\u7763\u5b66\u4e60\u95ee\u9898\u4e2d\uff0cPseudo Labeling\u51e0\u4e4e\u662f\u4e07\u91d1\u6cb9\uff0c\u51e0\u4e4e\u80fd\u591f\u8ba9\u4f60\u6a21\u578b\u5404\u4e2a\u65b9\u9762\u7684\u8868\u73b0\u90fd\u5f97\u5230\u63d0\u5347\u3002<\/p>\n\n\n\n<ol><li>\u4f7f\u7528\u539f\u59cb\u8bad\u7ec3\u96c6\u8bad\u7ec3\u5e76\u5efa\u7acb\u6a21\u578b<\/li><li>\u4f7f\u7528\u8bad\u7ec3\u597d\u7684\u6a21\u578b\u5bf9\u6d4b\u8bd5\u96c6\u8fdb\u884c\u5206\u7c7b<\/li><li>\u5c06\u9884\u6d4b\u6b63\u786e\u7f6e\u4fe1\u5ea6\u9ad8\u7684\u6837\u672c\u52a0\u5165\u5230\u8bad\u7ec3\u96c6\u4e2d<\/li><li>\u4f7f\u7528\u7ed3\u5408\u4e86\u90e8\u5206\u6d4b\u8bd5\u96c6\u6837\u672c\u7684\u65b0\u8bad\u7ec3\u96c6\u518d\u6b21\u8bad\u7ec3\u6a21\u578b<\/li><li>\u4f7f\u7528\u65b0\u6a21\u578b\u518d\u6b21\u8fdb\u884c\u9884\u6d4b<\/li><\/ol>\n\n\n\n<p>\u603b\u4e4b\uff1a\u63d0\u5206\u70b9\u5f88\u591a\uff0c\u4f46\u80fd\u5426\u6709\u6548\u4ee5\u53ca\u80fd\u5426\u5b9e\u73b0\u53c8\u662f\u53e6\u4e00\u4e2a\u4e8b\u60c5\u4e86\uff0c\u6bd5\u7adf\u6709\u65f6\u5019\u662f\u5426\u6709\u6548\u4e5f\u53d6\u51b3\u4e8e\u6570\u636e\u96c6\uff0c\u6bd5\u7adf\u7f18\u5206\uff0c\u5999\u4e0d\u53ef\u8a00~\uff0c\u540e\u7eed\u6211\u4f1a\u62bd\u65f6\u95f4\u5c06\u4e0a\u9762\u7684tricks\u90fd\u5c1d\u8bd5\u5c1d\u8bd5\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u4f5c\u4e3a\u521a\u5165\u95e8\u7684\u5c0f\u767d\uff0c\u6709\u5fc5\u8981\u53bb\u8bb0\u5f55\u4e00\u4e9bNLP\u5206\u7c7b\u4efb\u52a1\u7684\u5c0ftrick\uff0c\u611f\u89c9\u5bf9\u4e8e\u6da8\u70b9\u63d0\u5206\u5341\u5206\u6709\u7528\u3002\u8fd9\u4e2a\u6587\u7ae0\u540e\u9762\u6709\u65b0\u7684\u60f3 &hellip; <a href=\"http:\/\/139.9.1.231\/index.php\/2022\/08\/31\/nlp_mclass\/\" class=\"more-link\">\u7ee7\u7eed\u9605\u8bfb<span class=\"screen-reader-text\">\u5173\u4e8eNLP\u591a\u6807\u7b7e\u6587\u672c\u5206\u7c7b\u7684\u4e00\u4e9b\u601d\u8def&#8211;\uff08\u5f85\u66f4\u65b0\uff09<\/span><\/a><\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":[],"categories":[21],"tags":[],"_links":{"self":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/6815"}],"collection":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/comments?post=6815"}],"version-history":[{"count":120,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/6815\/revisions"}],"predecessor-version":[{"id":7327,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/6815\/revisions\/7327"}],"wp:attachment":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/media?parent=6815"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/categories?post=6815"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/tags?post=6815"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}