{"id":556,"date":"2021-10-23T21:42:05","date_gmt":"2021-10-23T13:42:05","guid":{"rendered":"http:\/\/hk.khscience.cn\/docs\/?p=556"},"modified":"2024-04-07T21:37:08","modified_gmt":"2024-04-07T13:37:08","slug":"rnns_class","status":"publish","type":"post","link":"https:\/\/www.khsci.com\/docs\/index.php\/2021\/10\/23\/rnns_class\/","title":{"rendered":"\u4f7f\u7528LSTM\/BiLSTM\/GRU\u505a\u6a21\u5f0f\u8bc6\u522b\uff08\u5206\u7c7b\uff09"},"content":{"rendered":"

\u4e00\u3001\u7a0b\u5e8f\u4ecb\u7ecd<\/strong><\/p>\n

-\u7a0b\u5e8f\u6db5\u76d6\u4e86LSTM\/BiLSTM\/GRU<\/span>\u4e09\u79cd\u6df1\u5ea6\u5b66\u4e60\u65b9\u6cd5<\/span><\/p>\n

-\u53ef\u4ee5\u81ea\u7531\u8bbe\u7f6e\u795e\u7ecf\u7f51\u7edcLSTM\/BiLSTM\/GRU Layer\u7684\u5c42\u6570<\/span>\u548c\u6bcf\u5c42\u7684\u9690\u85cf\u5355\u5143\u6570<\/span><\/span><\/p>\n

-\u53ef\u4ee5\u8bbe\u7f6e\u4e22\u5f03\u7387<\/span><\/span><\/p>\n

-\u53ef\u4ee5\u5feb\u901f\u5212\u5206\u8bad\u7ec3\u96c6\u3001\u9a8c\u8bc1\u96c6\u3001\u6d4b\u8bd5\u96c6<\/span>\uff0c\u7a0b\u5e8f\u8fd0\u884c\u5b8c\u540e\u7ed8\u5236\u51fa\u6d4b\u8bd5\u96c6\u6df7\u6dc6\u77e9\u9635<\/span><\/p>\n

-\u53ef\u4ee5\u8bbe\u7f6e\u81ea\u52a8\u7ea0\u6b63<\/span>\u8f93\u5165\u6570\u636e\u7684\u884c\u5217\u65b9\u5411<\/span><\/p>\n

-\u652f\u6301\u5411\u91cf<\/span>\u548c\u7d22\u5f15<\/span>\u4e24\u79cd\u5f62\u5f0f\u7684\u6807\u7b7e\u8f93\u5165<\/span><\/p>\n

-\u591a\u79cd\u8bad\u7ec3\u53c2\u6570\u53ef\u8bbe\u7f6e\uff0c\u5305\u62ec\u6c42\u89e3\u5668\u3001\u8fed\u4ee3\u6b21\u6570\u3001\u521d\u59cb\u5b66\u4e60\u7387\u7b49\u7b49<\/span><\/span><\/p>\n

-\u53ef\u8bbe\u7f6e\u968f\u673a\u79cd\u5b50<\/span>\uff0c\u4fdd\u8bc1\u6bcf\u6b21\u8fd0\u884c\u7ed3\u679c\u4fdd\u6301\u4e00\u81f4<\/span><\/p>\n

–\u9700\u8981\u4f60\u505a\u7684\u57fa\u672c\u53ea\u6709\u5bfc\u5165\u6570\u636e\u548c\u8c03\u53c2\u3002<\/strong><\/span><\/p>\n

–\u7edd\u5927\u591a\u6570\u6d41\u7a0b\u90fd\u88ab\u5c01\u88c5\u56fa\u5316\u5230\u51fd\u6570\u4e2d\uff0c\u4eff\u7167\u6848\u4f8b\u5bfc\u5165\u4f60\u7684\u6570\u636e\u5373\u523b\u5f97\u5230\u7ed3\u679c~<\/strong><\/span><\/p>\n

\u4e8c\u3001\u4ee3\u7801\u8fd0\u884c\u73af\u5883<\/strong><\/p>\n

MATLAB2020a\u53ca\u66f4\u65b0\u7248\u672c\u3002<\/p>\n

\u4e09\u3001\u7a0b\u5e8f\u5177\u4f53\u5185\u5bb9<\/strong><\/p>\n

<\/p>\n

\u6ce8\uff1a\u56fe\u6807\u4ee3\u8868\u8be5m\u6587\u4ef6\u4e3a\u811a\u672c\u6587\u4ef6\uff0c\u53ef\u4ee5\u76f4\u63a5\u8fd0\u884c\uff1b\u56fe\u6807\u4ee3\u8868\u51fd\u6570\u6587\u4ef6\uff0c\u5728\u6ca1\u6709\u8f93\u5165\u53d8\u91cf\u7684\u60c5\u51b5\u4e0b\u65e0\u6cd5\u76f4\u63a5\u8fd0\u884c\u3002\u66f4\u8be6\u7ec6\u7684\u89e3\u91ca\u53ef\u4ee5\u770b\u8fd9\u91cc<\/a>\u3002<\/span><\/p>\n

1.scriptClassRNNs.m<\/strong><\/span><\/p>\n

\u4f7f\u7528MNIST\u6570\u636e\u96c6\u8fdb\u884c\u624b\u5199\u5b57\u4f53\u6570\u5b57\u7684\u8bc6\u522b\u7684demo\u811a\u672c\u6587\u4ef6\uff0c\u4e3b\u8981\u8c03\u7528\u4e86FunClassRNNs\u51fd\u6570\u3002<\/p>\n

2.FunClassRNNs.m<\/span><\/strong><\/p>\n

\u4f7f\u7528RNN\u884d\u751f\u65b9\u6cd5\uff08\u5305\u62ecGRU\u3001LSTM\u3001BiLSTM\uff09\u8fdb\u884c\u6a21\u5f0f\u8bc6\u522b\uff08\u5206\u7c7b\uff09\u7684\u5feb\u901f\u5b9e\u73b0\u51fd\u6570\u3002<\/p>\n

\u51fd\u6570\u53ef\u8c03\u53c2\u6570\u3001\u76f8\u5173\u8bbe\u7f6e\u4ee5\u53ca\u8f93\u51fa\u53c2\u91cf\u5982\u4e0b\uff1a<\/p>\n

[accuracy,recall,precision,net] = FunClassRNNs(dataX,dataY,divideR,rnnName,hiddenSizes,dropProb,auto,options,figflag)<\/span><\/strong>
\n% \u4f7f\u7528RNN\u884d\u751f\u65b9\u6cd5\uff08\u5305\u62ecGRU\u3001LSTM\u3001BiLSTM\uff09\u8fdb\u884c\u6a21\u5f0f\u8bc6\u522b\uff08\u5206\u7c7b\uff09\u7684\u5feb\u901f\u5b9e\u73b0\u51fd\u6570<\/span><\/span><\/p>\n

% \u8f93\u5165\uff1a<\/span>
\n% dataX\uff1a\u8f93\u5165\u6570\u636e\uff0cR*Q\u7684\u77e9\u9635\uff0cR\u4e3a\u7279\u5f81\u7ef4\u5ea6\uff0cQ\u4e3a\u6279\u6b21\u6570\uff0c\u8f93\u5165\u8be5\u53d8\u91cf\u65f6\u4e00\u5b9a\u8981\u6ce8\u610f\u884c\u5217\u65b9\u5411\u662f\u5426\u6b63\u786e<\/span><\/span>
\n% dataY\uff1a\u6807\u7b7e\u503c\uff0c\u53ef\u4ee5\u4e3a\u4e24\u79cd\u65b9\u5f0f\uff1a<\/span><\/span>
\n% \u5411\u91cf\u578b\uff1aU*Q\u7684\u77e9\u9635\uff0cU\u4e3a\u6807\u7b7e\u79cd\u7c7b\u6570\uff0cQ\u4e3a\u6279\u6b21\u6570<\/span><\/span>
\n% \u7d22\u5f15\u578b\uff1a1*Q\u7684\u77e9\u9635\uff0cQ\u4e3a\u6279\u6b21\u6570<\/span><\/span>
\n% divideR\uff1a\u6570\u636e\u96c6\uff08\u8bad\u7ec3\u96c6\u3001\u9a8c\u8bc1\u96c6\u3001\u6d4b\u8bd5\u96c6\uff09\u5212\u5206\u6bd4\u4f8b\uff0c\u5982\uff1adivideR =[0.6,0.2,0.2]\uff0c<\/span><\/span>
\n% \u5219\u4ee3\u886860%\u6570\u636e\u7528\u4e8e\u8bad\u7ec3\u96c6\uff0c20%\u6570\u636e\u7528\u4e8e\u9a8c\u8bc1\u96c6\uff0c20%\u6570\u636e\u7528\u4e8e\u6d4b\u8bd5\u96c6<\/span><\/span>
\n% rnnName\uff1a\u4f7f\u7528\u7684\u7f51\u7edc\u540d\u79f0\uff0c\u53ef\u9009\u9879\u5305\u62ec\uff1a’GRU’,’LSTM’,’BiLSTM’\uff0c\u6ce8\u610f\u5927\u5c0f\u5199<\/span><\/span>
\n% hiddenSizes\uff1a\u6bcf\u4e2a\u5c42\uff08\u5982LSTM\u5c42\uff09\u7684\u5355\u5143\u6570\uff0c\u5982\u679c\u8981\u8bbe\u7f6e\u591a\u4e2a\u5c42\uff0c\u5219\u4f7f\u7528\u6570\u7ec4\u8868\u793a\uff0c\u6570\u503c\u4e00\u822c\u5728\u51e0\u5341\u5230\u51e0\u767e\u4e4b\u95f4<\/span><\/span>
\n% \u5982\u7b2c\u4e00\u5c42\u9690\u85cf\u5355\u5143\u6570\u4e3a100\uff0c\u7b2c\u4e8c\u5c42\u9690\u85cf\u5355\u5143\u6570\u4e3a60\uff0c\u5219\u8bbe\u7f6e\uff1ahiddenSizes=[100,20]<\/span><\/span>
\n% dropProb\uff1a\u4e22\u5f03\u5c42\u7684\u4e22\u5f03\u6982\u7387\uff0c\u5982\u679c\u4e0d\u4e22\u5f03\uff0c\u5219\u8bbe\u7f6e\u4e3a0\uff0c\u6700\u5927\u4e3a1<\/span><\/span>
\n% auto\uff1a\u662f\u5426\u8fdb\u884c\u81ea\u52a8\u7ea0\u9519\uff0c’on’\u4e3a\u662f\uff0c\u5426\u5219\u4e3a\u5426\u3002\u5f00\u542f\u81ea\u52a8\u7ea0\u9519\u540e\u4f1a\u667a\u80fd\u8c03\u6574\u8bad\u7ec3\u3001\u6d4b\u8bd5\u96c6\u7684\u884c\u5217\u65b9\u5411\u3002<\/span><\/span>
\n% options\uff1a\u4e00\u4e9b\u4e0e\u7f51\u7edc\u8bad\u7ec3\u7b49\u76f8\u5173\u7684\u8bbe\u7f6e\uff0c\u4f7f\u7528\u7ed3\u6784\u4f53\u65b9\u5f0f\u8d4b\u503c\uff0c\u6bd4\u5982 options.MaxEpochs = 1000\uff0c\u5177\u4f53\u5305\u62ec\uff1a<\/span><\/span>
\n% solverName\uff1a\u6c42\u89e3\u5668\uff0c’sgdm’\uff08\u9ed8\u8ba4\uff09 | ‘rmsprop’ | ‘adam’<\/span><\/span>
\n% MaxEpochs\uff1a\u6700\u5927\u8fed\u4ee3\u6b21\u6570\uff0c\u9ed8\u8ba430<\/span><\/span>
\n% MiniBatchSize\uff1a\u6279\u5c3a\u5bf8\uff0c\u9ed8\u8ba4128<\/span><\/span>
\n% GradientThreshold\uff1a\u68af\u5ea6\u6781\u9650\uff0c\u9ed8\u8ba4\u4e3aInf<\/span><\/span>
\n% InitialLearnRate\uff1a\u521d\u59cb\u5316\u5b66\u4e60\u901f\u7387\uff08\u9ed8\u8ba40.005\uff09<\/span><\/span>
\n% Plots\uff1a\u662f\u5426\u663e\u793a\u8bad\u7ec3\u8fc7\u7a0b\uff0c’none’ \u4e3a\u4e0d\u663e\u793a\uff08\u9ed8\u8ba4\uff09 | ‘training-progress’\u4e3a\u663e\u793a<\/span><\/span>
\n% ValidationFrequency\uff1a\u9a8c\u8bc1\u9891\u7387\uff0c\u5373\u6bcf\u95f4\u9694\u591a\u5c11\u6b21\u8fed\u4ee3\u8fdb\u884c\u4e00\u6b21\u9a8c\u8bc1\uff0c\u9ed8\u8ba450<\/span><\/span>
\n% LearnRateSchedule\uff1a\u5373LearnRateSchedule\u662f\u5426\u5728\u4e00\u5b9a\u8fed\u4ee3\u6b21\u6570\u540e\u5b66\u4e60\u901f\u7387\u4e0b\u964d\uff0c LearnRateSchedule =’piecewise’\u4e3a\u4f7f\u7528\uff0c’none’\u4e3a\u4e0d\u4f7f\u7528\uff08\u9ed8\u8ba4\uff09<\/span><\/span>
\n% LearnRateDropPeriod\uff1a\u5373LearnRateDropPeriod\u5b66\u4e60\u901f\u7387\u4e0b\u964d\u65f6\u7684\u8fed\u4ee3\u6570\uff0c\u9ed8\u8ba4\u4e3a10<\/span><\/span>
\n% LearnRateDropFactor\uff1a\u5373LearnRateDropFactor\u5b66\u4e60\u901f\u7387\u4e0b\u964d\u56e0\u5b50\uff0c\u4e0b\u964d\u540e\u53d8\u4e3aLearnRateDropFactor*InitialLearnRate\uff0cLearnRateSchedule\u4e3a0\u65f6\u53ef\u4ee5\u8d4b0\uff0c\u9ed8\u8ba4\u4e3a0.1<\/span><\/span>
\n% \uff08\u672a\u542f\u7528\uff09NorFlag\uff1a\u5373Normalization Flag\uff0c\u8bbe\u7f6e\u4e3a1\u65f6\u5219\u5728\u7a0b\u5e8f\u4e2d\u8fdb\u884c\u6570\u636e\u5f52\u4e00\u5316\u548c\u53cd\u5f52\u4e00\u5316\u64cd\u4f5c\uff0c\u5426\u5219\u4e0d\u8fdb\u884c\uff0c\u5efa\u8bae\u8bbe\u7f6e\u4e3a1<\/span><\/span>
\n% SeedFlag\uff1a\u968f\u673a\u79cd\u5b50\u6807\u5fd7\uff0c\u8bbe\u7f6e\u4e3a1\u65f6\u542f\u7528\u968f\u673a\u79cd\u5b50,(\u9ed8\u8ba4\u4e3a1)<\/span><\/span>
\n% figflag\uff1a\u662f\u5426\u753b\u56fe\uff0c’on’\u4e3a\u753b\u56fe\uff0c’off’\u4e3a\u4e0d\u753b<\/span><\/span><\/p>\n

% \u8f93\u51fa\uff1a<\/span><\/span>
\n% accuracy\uff1a\u6d4b\u8bd5\u96c6\u5206\u7c7b\u6b63\u786e\u7387<\/span><\/span>
\n% recall\uff1a \u53ec\u56de\u7387<\/span><\/span>
\n% precision\uff1a \u7cbe\u786e\u7387<\/span><\/span>
\n% net\uff1a\u8bad\u7ec3\u597d\u7684\u7f51\u7edc<\/span><\/span><\/p>\n

\u5f53figflag<\/span>\u8bbe\u7f6e\u4e3a<\/span><\/span>‘on’\u65f6\uff0c\u5c06\u753b\u51fa\u5982\u4e0b\u6df7\u6dc6\u77e9\u9635\u56fe\uff1a<\/span><\/span><\/strong><\/p>\n

<\/p>\n

\u5f53Plots<\/span>\u8bbe\u7f6e\u4e3a<\/span><\/span>\u00a0‘training-progress’\u65f6\uff0c\u5c06\u753b\u51fa\u5982\u4e0b\u8bad\u7ec3\u8fc7\u7a0b\u56fe\uff1a<\/span><\/span><\/strong><\/p>\n

<\/p>\n

3.EvaluteClassResults.m<\/strong><\/span><\/p>\n

\u8bc4\u4f30\u5206\u7c7b\u7ed3\u679c\u7684\u51fd\u6570\u6587\u4ef6\uff0c\u5176\u4e2d\u5305\u62ec\u7ed8\u5236\u6df7\u6dc6\u77e9\u9635\uff0c\u8ba1\u7b97\u53ec\u56de\u7387\u548c\u7cbe\u786e\u7387\u3002\u51fd\u6570\u53c2\u6570\u5982\u4e0b\uff1a<\/p>\n

[recall, precision] = EvaluteClassResults(outputs,targets,figflag)<\/span>
\n% \u5bf9\u5206\u7c7b\u7ed3\u679c\u8fdb\u884c\u8bc4\u4ef7\uff0c\u5305\u62ec\uff1a<\/span>
\n% \u753b\u6df7\u6dc6\u77e9\u9635\u56fe<\/span>
\n% \u8ba1\u7b97 weigthed-Precision, weigthed-Recall\u548cF1-score \u53c2\u8003 https:\/\/zhuanlan.zhihu.com\/p\/147663370?from_voters_page=true<\/span>
\n% \u8f93\u5165\uff1a<\/span>
\n% outputs\uff1a\u7ecf\u8fc7\u5206\u7c7b\u7b97\u6cd5\u5f97\u5230\u7684\u5206\u7c7b\u7ed3\u679c\uff0c\u4e00\u7ef4\u6b63\u6574\u6570<\/span>
\n% targets\uff1a\u76ee\u6807\u5206\u7c7b\u7ed3\u679c\uff0c\u4e00\u7ef4\u6b63\u6574\u6570<\/span>
\n% figflag\uff1a\u662f\u5426\u753b\u56fe\uff0c’on’\u4e3a\u753b\u56fe\uff0c’off’\u4e3a\u4e0d\u753b<\/span>
\n% \u8f93\u51fa\uff1a<\/span>
\n% recall\uff1a \u53ec\u56de\u7387<\/span>
\n% precision\uff1a \u7cbe\u786e\u7387<\/span><\/p>\n

4.clcEvaluteCalss<\/strong><\/span><\/p>\n

\u8ba1\u7b97 \u53ec\u56de\u7387\u548c\u7cbe\u786e\u7387\u7684\u5e95\u5c42\u51fd\u6570\u3002\u51fd\u6570\u53c2\u6570\u5982\u4e0b\uff1a<\/p>\n

function [recall, precision] = clcEvaluteCalss(outputs,targets)<\/span>
\n% \u8ba1\u7b97 weigthed-Precision, weigthed-Recall \u53c2\u8003 https:\/\/zhuanlan.zhihu.com\/p\/147663370?from_voters_page=true<\/span>
\n% \u6ce8\u610f\uff0coutputs\u548ctargets\u4e2d\u7684\u5206\u7c7b\u4e3a\u6574\u578b\uff0c\u4ece1\u6392\u5e8f\u5230n\uff0c\u4e0d\u53ef\u8df3\u6570<\/span><\/p>\n

\u56db\u3001\u5feb\u901f\u5f00\u59cb<\/h4>\n

1.\u5bfc\u5165\u4f60\u7684\u6570\u636e\u5feb\u901f\u83b7\u53d6\u9884\u6d4b\u7ed3\u679c<\/strong><\/p>\n

\u6839\u636e\u4f60\u7684\u6587\u4ef6\u7c7b\u578b\u7684\u4e0d\u540c\uff08excel,txt,csv\u7b49\uff09\uff0c\u5c06\u6570\u636e\u5bfc\u5165MATLAB\u7684\u65b9\u6cd5\u6709\u6240\u4e0d\u540c\u3002\u540c\u5b66\u4eec\u53ef\u4ee5\u770b\u535a\u4e3b\u9488\u5bf9\u5e38\u7528\u6587\u4ef6\u7684\u5bfc\u5165\u65b9\u6cd5\u7684\u8fd9\u4e2a\u6559\u7a0b<\/a>\u3002<\/p>\n

\u9700\u8981\u6ce8\u610f\u4f60\u5bfc\u5165\u7684\u6570\u636e\u9700\u8981\u6ee1\u8db3FunClassRNNs\u51fd\u6570\u5bf9\u8f93\u5165\u6570\u636e\u7684\u8981\u6c42\uff0c\u5373\uff1a<\/span><\/p>\n

\u8f93\u5165\u6570\u636e\uff1aR*Q\u7684\u77e9\u9635\uff0cR\u4e3a\u7279\u5f81\u7ef4\u5ea6\uff0cQ\u4e3a\u6279\u6b21\u6570\uff0c\u8f93\u5165\u8be5\u53d8\u91cf\u65f6\u4e00\u5b9a\u8981\u6ce8\u610f\u884c\u5217\u65b9\u5411\u662f\u5426\u6b63\u786e<\/span><\/span>
\n\u6807\u7b7e\u503c\uff1a\u53ef\u4ee5\u4e3a\u4e24\u79cd\u65b9\u5f0f\uff1a<\/span><\/span>
\n\u00a0 \u00a0 \u00a0 \u00a0 – \u5411\u91cf\u578b\uff1aU*Q\u7684\u77e9\u9635\uff0cU\u4e3a\u6807\u7b7e\u79cd\u7c7b\u6570\uff0cQ\u4e3a\u6279\u6b21\u6570<\/span><\/span>
\n\u00a0 \u00a0 \u00a0 \u00a0 – \u7d22\u5f15\u578b\uff1a1*Q\u7684\u77e9\u9635\uff0cQ\u4e3a\u6279\u6b21\u6570<\/span><\/span><\/p>\n

2.\u5b9e\u73b0\u5206\u7c7b<\/strong><\/p>\n

\u53c2\u7167scriptClassRNNs\u6587\u4ef6\u4e2d\u7684\u51fd\u6570\u8c03\u7528\u65b9\u6cd5\u7f16\u5199\u81ea\u5df1\u7684\u811a\u672c\u6587\u4ef6\u3002<\/p>\n

\u4e94\u3001\u83b7\u53d6\u7a0b\u5e8f\uff08\u9700\u4f7f\u7528\u7535\u8111\u6d4f\u89c8\u5668\u6253\u5f00\uff09<\/h4>\n