{"id":4406,"date":"2021-04-01T20:18:33","date_gmt":"2021-04-01T12:18:33","guid":{"rendered":"http:\/\/www.shumo.com\/home\/?p=4406"},"modified":"2021-04-01T20:18:33","modified_gmt":"2021-04-01T12:18:33","slug":"%e8%bd%ac%e8%bd%bd%ef%bc%9a%e4%b8%89%e9%a1%be%e7%a2%8e%e7%ba%b8%e5%a4%8d%e5%8e%9f%ef%bc%9a%e5%9f%ba%e4%ba%8ecnn%e7%9a%84%e7%a2%8e%e7%ba%b8%e5%a4%8d%e5%8e%9f","status":"publish","type":"post","link":"https:\/\/www.shumo.com\/home\/html\/4406.html","title":{"rendered":"\u8f6c\u8f7d\uff1a\u4e09\u987e\u788e\u7eb8\u590d\u539f\uff1a\u57fa\u4e8eCNN\u7684\u788e\u7eb8\u590d\u539f"},"content":{"rendered":"

<\/p>\n

\u539f\u6587\u5730\u5740\uff1ahttps:\/\/kexue.fm\/archives\/4100<\/a>
\n\u4f5c\u8005\uff1a\u82cf\u5251\u6797<\/p>\n

\u8d5b\u9898\u56de\u987e\u00a0#<\/a><\/h3>\n

\u4e0d\u5f97\u4e0d\u8bf4\uff0c2013\u5e74\u7684\u5168\u56fd\u6570\u5b66\u5efa\u6a21\u7ade\u8d5b\u4e2d\u7684B\u9898\u771f\u7684\u7b97\u662f\u6570\u5b66\u5efa\u6a21\u7ade\u8d5b\u4e2d\u767e\u5e74\u96be\u5f97\u4e00\u9047\u7684\u597d\u9898\uff1a\u9898\u76ee\u7b80\u6d01\u660e\u4e86\uff0c\u542b\u4e49\u4e30\u5bcc\uff0c\u505a\u6cd5\u591a\u6837\uff0c\u5ef6\u4f38\u6027\u5f3a\uff0c\u4ee5\u81f3\u4e8e\u6211\u4e00\u76f4\u5bf9\u5b83\u5ff5\u5ff5\u4e0d\u5fd8\u3002\u56e0\u4e3a\u8fd9\u4e2a\u9898\u76ee\uff0c\u6211\u5df2\u7ecf\u5728\u79d1\u5b66\u7a7a\u95f4\u5199\u4e86\u4e24\u7bc7\u6587\u7ae0\u4e86\uff0c\u5206\u522b\u662f\u300a\u4e00\u4e2a\u4eba\u7684\u6570\u5b66\u5efa\u6a21\uff1a\u788e\u7eb8\u590d\u539f\u300b<\/a>\u548c\u300a\u8fdf\u5230\u4e00\u5e74\u7684\u5efa\u6a21\uff1a\u518d\u63a2\u788e\u7eb8\u590d\u539f\u300b<\/a>\u3002\u4ee5\u524d\u505a\u8fd9\u9053\u9898\u7684\u65f6\u5019\uff0c\u8fd8\u53ea\u6709\u4e00\u70b9\u6570\u5b66\u5efa\u6a21\u7684\u77e5\u8bc6\uff0c\u800c\u81ea\u4ece\u5b66\u4e60\u4e86\u6570\u636e\u6316\u6398\u3001\u5c24\u5176\u662f\u6df1\u5ea6\u5b66\u4e60\u4e4b\u540e\uff0c\u6211\u4e00\u76f4\u60f3\u91cd\u505a\u8fd9\u9053\u9898\uff0c\u4f46\u4e00\u76f4\u5077\u61d2\u3002\u8fd9\u51e0\u5929\u7ec8\u4e8e\u628a\u5b83\u5b9e\u73b0\u4e86\u3002<\/p>\n

\u5982\u679c\u5bf9\u9898\u76ee\u8fd8\u4e0d\u6e05\u695a\u7684\u8bfb\u8005\uff0c\u53ef\u4ee5\u53c2\u8003\u524d\u9762\u4e24\u7bc7\u6587\u7ae0\u3002\u788e\u7eb8\u590d\u539f\u5171\u6709\u4e94\u4e2a\u9644\u4ef6\uff0c\u5206\u522b\u4ee3\u8868\u4e86\u4e94\u79cd\u201c\u788e\u7eb8\u7247\u201d\uff0c\u5373\u4e94\u79cd\u4e0d\u540c\u7c92\u5ea6\u7684\u788e\u7247\u3002\u5176\u4e2d\u9644\u4ef61\u548c2\u90fd\u4e0d\u56f0\u96be\uff0c\u96be\u5ea6\u4e3b\u8981\u96c6\u4e2d\u5728\u9644\u4ef63\u30014\u30015\uff0c\u800c3\u30014\u30015\u7684\u5b9e\u73b0\u96be\u5ea6\u57fa\u672c\u662f\u4e00\u6837\u7684\u3002\u505a\u8fd9\u9053\u9898\u6700\u5bb9\u6613\u60f3\u5230\u7684\u601d\u8def\u5c31\u662f\u8d2a\u5fc3\u7b97\u6cd5\uff0c\u5373\u968f\u4fbf\u9009\u4e00\u5f20\u56fe\u7247\uff0c\u7136\u540e\u627e\u5230\u4e0e\u5b83\u6700\u5339\u914d\u7684\u56fe\u7247\uff0c\u7136\u540e\u7ee7\u7eed\u5339\u914d\u4e0b\u4e00\u5f20\u3002\u8981\u60f3\u8d2a\u5fc3\u7b97\u6cd5\u6709\u6548\uff0c\u6700\u5173\u952e\u662f\u627e\u5230\u4e00\u4e2a\u826f\u597d\u7684\u8ddd\u79bb\u51fd\u6570\uff0c\u6765\u5224\u65ad\u4e24\u5f20\u788e\u7247\u662f\u5426\u76f8\u90bb\uff08\u6c34\u5e73\u76f8\u90bb\uff0c\u8fd9\u91cc\u4e0d\u8003\u8651\u5782\u76f4\u76f8\u90bb\uff09\u3002<\/p>\n

\u524d\u4e24\u7bc7\u6587\u7ae0\u7528\u7684\u90fd\u662f\u8fb9\u7f18\u5411\u91cf\u7684\u6b27\u6c0f\u8ddd\u79bb\uff0c\u4e5f\u8fd8\u63d0\u5230\u8fc7\u8bf8\u5982\u76f8\u5173\u7cfb\u6570\u7684\u4e00\u4e9b\u6307\u6807\uff0c\u4f46\u5982\u679c\u7528\u5728\u9644\u4ef63\u30014\u30015\u4e2d\uff0c\u6548\u679c\u90fd\u4e0d\u597d\uff0c\u539f\u56e0\u5c31\u5728\u4e8e\u9644\u4ef63\u30014\u30015\u7684\u788e\u7247\u5e76\u4e0d\u5927\uff0c\u53ef\u7528\u8fb9\u7f18\u4fe1\u606f\u4e0d\u591a\u3002\u56e0\u6b64\uff0c\u53ea\u9760\u8fb9\u7f18\u662f\u4e0d\u884c\u7684\uff0c\u8fd8\u5fc5\u987b\u8003\u8651\u591a\u79cd\u56e0\u7d20\uff0c\u6bd4\u5982\u884c\u8ddd\u3001\u884c\u7684\u5e73\u5747\u4f4d\u7f6e\u7b49\u7b49\u3002\u7a76\u7adf\u600e\u4e48\u8003\u8651\uff1f\u4eba\u5de5\u5f88\u96be\u5199\u51fa\u4e00\u4e2a\u826f\u597d\u7684\u51fd\u6570\u3002\u65e2\u7136\u5982\u6b64\uff0c\u5e72\u561b\u4e0d\u4ea4\u7ed9\u6a21\u578b\u5462\uff1f\u76f4\u63a5\u4e0a\u5377\u79ef\u795e\u7ecf\u7f51\u7edc\uff08CNN\uff09\u5c31\u597d\u4e86\u5440\uff01<\/p>\n

\u6784\u9020\u6837\u672c\u00a0#<\/a><\/h3>\n

\u5177\u4f53\u6765\u8bb2\uff0c\u5c31\u662f\u6211\u4eec\u76ee\u6d4b\u4e00\u4e0b\u788e\u7247\u7684\u60c5\u51b5\uff0c\u5927\u6982\u5c31\u662f<\/p>\n

1\u300144\u53f7\u7684\u9ed1\u4f53\uff1b
\n2\u3001\u5982\u679c\u8003\u8651\u9644\u4ef63\uff0c\u5185\u5bb9\u5c31\u662f\u4e2d\u6587\uff0c\u5982\u679c\u9644\u4ef64\uff0c\u90a3\u4e48\u5c31\u662f\u82f1\u6587\uff1b
\n3\u3001\u788e\u7247\u56fa\u5b9a\u5927\u5c0f\u4e3a72×180<\/p><\/blockquote>\n

\u6709\u4e86\u8fd9\u4e2a\u7279\u5f81\uff0c\u6211\u4eec\u5c31\u53ef\u4ee5\u867d\u7136\u627e\u4e00\u5806\u6587\u672c\u6765\uff0c\u7136\u540e\u4eff\u7167\u8fd9\u4e2a\u89c4\u683c\uff0c\u81ea\u5df1\u6784\u9020\u4e00\u5927\u6279\u5177\u6709\u540c\u6837\u6027\u8d28\u7684\u3001\u76f8\u90bb\u548c\u4e0d\u76f8\u90bb\u7684\u6837\u672c\u6765\uff0c\u7136\u540e\u8bad\u7ec3\u4e00\u4e2a\u5377\u79ef\u795e\u7ecf\u7f51\u7edc\uff0c\u8fd9\u6837\u5c31\u53ef\u4ee5\u81ea\u52a8\u5730\u5f97\u5230\u4e00\u4e2a\u201c\u8ddd\u79bb\u51fd\u6570\u201d\u4e86\u3002\u8fd9\u79cd\u4e8b\u60c5\u5bf9\u4e8e\u719f\u6089\u6df1\u5ea6\u5b66\u4e60\u7684\u670b\u53cb\u518d\u7b80\u5355\u4e0d\u8fc7\u4e86\u3002\u6211\u662f\u76f4\u63a5\u627e\u4e86\u4e2d\u6587\u7684\u6587\u672c\uff0c\u7136\u540e\u751f\u6210\u4e86\u4e00\u6279\u4e2d\u6587\u7684\u6837\u672c\uff0c\u4ee3\u7801\u5927\u6982\u5982\u4e0b\uff1a<\/p>\n

from<\/span> PIL import<\/span> Image,<\/span> ImageFont,<\/span> ImageDraw\r\nimport<\/span> numpy as<\/span> np\r\nfrom<\/span> scipy import<\/span> misc\r\nimport<\/span> pymongo\r\nfrom<\/span> tqdm import<\/span> tqdm\r\n\r\ntexts =<\/span> list(<\/span>pymongo.<\/span>MongoClient(<\/span>)<\/span>.<\/span>weixin.<\/span>text_articles.<\/span>find(<\/span>)<\/span>.<\/span>limit(<\/span>1000<\/span>)<\/span>)<\/span>\r\ntext =<\/span> texts[<\/span>0<\/span>]<\/span>[<\/span>'text'<\/span>]<\/span>\r\nline_words =<\/span> 30<\/span>\r\nfont_size =<\/span> 44<\/span>\r\nnb_columns =<\/span> line_words*<\/span>font_size\/<\/span>72<\/span>+<\/span>1<\/span>\r\n\r\ndef<\/span> gen_img<\/span>(<\/span>text)<\/span>:<\/span>\r\n    n =<\/span> len(<\/span>text)<\/span> \/<\/span> line_words +<\/span> 1<\/span>\r\n    size =<\/span> (<\/span>nb_columns*<\/span>72<\/span>,<\/span> (<\/span>n*<\/span>font_size\/<\/span>180<\/span>+<\/span>1<\/span>)<\/span>*<\/span>180<\/span>)<\/span>\r\n    im =<\/span> Image.<\/span>new(<\/span>'L'<\/span>,<\/span> size,<\/span> 255<\/span>)<\/span>\r\n    dr =<\/span> ImageDraw.<\/span>Draw(<\/span>im)<\/span>\r\n    font =<\/span> ImageFont.<\/span>truetype(<\/span>'simhei.ttf'<\/span>,<\/span> font_size)<\/span>\r\n    for<\/span> i in<\/span> range(<\/span>n)<\/span>:<\/span>\r\n        dr.<\/span>text(<\/span>(<\/span>0<\/span>,<\/span> 70<\/span>*<\/span>i)<\/span>,<\/span> text[<\/span>line_words*<\/span>i:<\/span> line_words*<\/span>(<\/span>i+<\/span>1<\/span>)<\/span>]<\/span>,<\/span> font=<\/span>font)<\/span>\r\n    im =<\/span> np.<\/span>array(<\/span>im.<\/span>getdata(<\/span>)<\/span>)<\/span>.<\/span>reshape(<\/span>(<\/span>size[<\/span>1<\/span>]<\/span>,<\/span> size[<\/span>0<\/span>]<\/span>)<\/span>)<\/span>\r\n    r =<\/span> [<\/span>]<\/span>\r\n    for<\/span> j in<\/span> range(<\/span>size[<\/span>1<\/span>]<\/span>\/<\/span>180<\/span>)<\/span>:<\/span>\r\n        for<\/span> i in<\/span> range(<\/span>size[<\/span>0<\/span>]<\/span>\/<\/span>72<\/span>)<\/span>:<\/span>\r\n            r.<\/span>append(<\/span>1<\/span>-<\/span>im[<\/span>j*<\/span>180<\/span>:<\/span>(<\/span>j+<\/span>1<\/span>)<\/span>*<\/span>180<\/span>,<\/span> i*<\/span>72<\/span>:<\/span>(<\/span>i+<\/span>1<\/span>)<\/span>*<\/span>72<\/span>]<\/span>.<\/span>T\/<\/span>255.0<\/span>)<\/span>\r\n    return<\/span> r\r\n\r\nsample =<\/span> [<\/span>]<\/span>\r\nfor<\/span> i in<\/span> tqdm(<\/span>iter(<\/span>texts)<\/span>)<\/span>:<\/span>\r\n    sample.<\/span>extend(<\/span>gen_img(<\/span>i[<\/span>'text'<\/span>]<\/span>)<\/span>)<\/span>\r\n\r\nnp.<\/span>save(<\/span>'sample.npy'<\/span>,<\/span> sample)<\/span>\r\nnb_samples =<\/span> len(<\/span>sample)<\/span> -<\/span> len(<\/span>sample)<\/span>\/<\/span>nb_columns\r\n\r\ndef<\/span> data<\/span>(<\/span>sample,<\/span> batch_size)<\/span>:<\/span>\r\n    sample_shuffle_totally =<\/span> sample[<\/span>:<\/span>]<\/span>\r\n    sample_shuffle_in_line =<\/span> sample[<\/span>:<\/span>]<\/span>\r\n    while<\/span> True<\/span>:<\/span>\r\n        np.<\/span>random.<\/span>shuffle(<\/span>sample_shuffle_totally)<\/span>\r\n        for<\/span> i in<\/span> range(<\/span>0<\/span>,<\/span> len(<\/span>sample_shuffle_in_line)<\/span>,<\/span> nb_columns)<\/span>:<\/span>\r\n            subsample =<\/span> sample_shuffle_in_line[<\/span>i:<\/span> i+<\/span>nb_columns]<\/span>\r\n            np.<\/span>random.<\/span>shuffle(<\/span>subsample)<\/span>\r\n            sample_shuffle_in_line[<\/span>i:<\/span> i+<\/span>nb_columns]<\/span> =<\/span> subsample\r\n        x =<\/span> [<\/span>]<\/span>\r\n        y =<\/span> [<\/span>]<\/span>\r\n        for<\/span> i in<\/span> range(<\/span>0<\/span>,<\/span> len(<\/span>sample)<\/span>,<\/span> nb_columns)<\/span>:<\/span>\r\n            subsample_1 =<\/span> sample[<\/span>i:<\/span> i+<\/span>nb_columns]<\/span>\r\n            for<\/span> j in<\/span> range(<\/span>0<\/span>,<\/span> nb_columns-1<\/span>)<\/span>:<\/span>\r\n                x.<\/span>append(<\/span>np.<\/span>vstack(<\/span>(<\/span>subsample_1[<\/span>j]<\/span>,<\/span> subsample_1[<\/span>j+<\/span>1<\/span>]<\/span>)<\/span>)<\/span>)<\/span>\r\n                y.<\/span>append(<\/span>[<\/span>1<\/span>]<\/span>)<\/span>\r\n            subsample_2 =<\/span> sample_shuffle_totally[<\/span>i:<\/span> i+<\/span>nb_columns]<\/span>\r\n            for<\/span> j in<\/span> range(<\/span>0<\/span>,<\/span> nb_columns-1<\/span>)<\/span>:<\/span>\r\n                x.<\/span>append(<\/span>np.<\/span>vstack(<\/span>(<\/span>subsample_2[<\/span>j]<\/span>,<\/span> subsample_2[<\/span>j+<\/span>1<\/span>]<\/span>)<\/span>)<\/span>)<\/span>\r\n                y.<\/span>append(<\/span>[<\/span>0<\/span>]<\/span>)<\/span>\r\n            subsample_3 =<\/span> sample_shuffle_in_line[<\/span>i:<\/span> i+<\/span>nb_columns]<\/span>\r\n            for<\/span> j in<\/span> range(<\/span>0<\/span>,<\/span> nb_columns-1<\/span>)<\/span>:<\/span>\r\n                x.<\/span>append(<\/span>np.<\/span>vstack(<\/span>(<\/span>subsample_3[<\/span>j]<\/span>,<\/span> subsample_3[<\/span>j+<\/span>1<\/span>]<\/span>)<\/span>)<\/span>)<\/span>\r\n                y.<\/span>append(<\/span>[<\/span>0<\/span>]<\/span>)<\/span>\r\n            if<\/span> len(<\/span>y)<\/span> >=<\/span> batch_size:<\/span>\r\n                yield<\/span> np.<\/span>array(<\/span>x)<\/span>,<\/span> np.<\/span>array(<\/span>y)<\/span>\r\n                x =<\/span> [<\/span>]<\/span>\r\n                y =<\/span> [<\/span>]<\/span>\r\n        if<\/span> y:<\/span>\r\n            yield<\/span> np.<\/span>array(<\/span>x)<\/span>,<\/span> np.<\/span>array(<\/span>y)<\/span>\r\n            x =<\/span> [<\/span>]<\/span>\r\n            y =<\/span> [<\/span>]<\/span>\r\n<\/code><\/pre>\n
\u8fc7\u7a0b\u662f\u8fd9\u6837\u7684\uff1a\u6211\u627e\u4e861000\u7bc7\u6587\u7ae0\uff0c\u6bcf\u7bc7\u82e5\u5e72\u5343\u5b57\uff0c\u5c06\u6bcf\u7bc7\u6587\u7ae0\u5747\u5300\u6253\u5370\u5230\u4e00\u5f20\u56fe\u7247\u4e0a\u53bb\uff0c\u7136\u540e\u88c1\u526a\u5f00\uff0c\u5c31\u5f97\u5230\u4e00\u6279\u6837\u672c\u4e86\uff08sample\uff09\uff0c\u540e\u9762\u7684data\u662f\u4e00\u4e2a\u8fed\u4ee3\u5668\uff0c\u7528\u6765\u751f\u6210\u6b63\u8d1f\u6837\u672c\uff0c\u56e0\u4e3a\u4e00\u6b21\u6027\u8f7d\u5165\u5185\u5b58\u5403\u4e0d\u6d88\uff0c\u6240\u4ee5\u53ea\u80fd\u901a\u8fc7\u8fed\u4ee3\u5668\u6765\u505a\u4e86\u3002\u4e0d\u8fc7\u6211\u8fd8\u662f\u5f88\u5077\u61d2\u7684\uff0c\u56e0\u4e3a\u5373\u4fbf\u8fd9\u6837\uff0c\u5728\u6211\u7684\u670d\u52a1\u5668\u4e0a\u4e5f\u8017\u4e8618G\u5185\u5b58\u3002sample_shuffle_totally\u662f\u5c06sample\u5168\u90e8\u6253\u4e71\u540e\uff0c\u5f97\u5230\u7684\u4efb\u610f\u8d1f\u6837\u672c\uff1bsample_shuffle_in_line\u5219\u53ea\u662f\u5728\u540c\u4e00\u884c\u5185\u6253\u4e71\uff0c\u8fd9\u6837\u5f97\u5230\u7684\u662f\u884c\u8ddd\u3001\u884c\u4f4d\u7f6e\u90fd\u76f8\u540c\uff0c\u4ec5\u4ec5\u662f\u5185\u5bb9\u4e0d\u540c\u800c\u5bfc\u81f4\u7684\u8d1f\u6837\u672c\u3002\u6bcf\u4e00\u8f6e\u8fed\u4ee3\u6211\u90fd\u6d17\u4e71\u4e00\u6b21\uff0c\u8fd9\u6837\u80fd\u591f\u63d0\u5347\u6570\u636e\u7684\u6027\u80fd\uff08data argument\uff09\uff0c\u4f7f\u5f97\u5b9e\u9645\u53c2\u4e0e\u8bad\u7ec3\u7684\u8d1f\u6837\u672c\u6570\u5927\u5927\u589e\u52a0\u3002<\/p>\n
\u8981\u6ce8\u610f\u7684\u662f\uff0c\u6211\u4eec\u8003\u8651\u6c34\u5e73\u76f8\u90bb\uff0c\u4f46python\u8bfb\u77e9\u9635\u7684\u65f6\u5019\uff0c\u662f\u4ece\u4e0a\u5f80\u4e0b\u8bfb\u7684\uff0c\u4e0d\u662f\u4ece\u5de6\u5f80\u53f3\uff0c\u56e0\u6b64\u6211\u4eec\u8981\u5bf9\u56fe\u7247\u77e9\u9635\u8fdb\u884c\u8f6c\u7f6e\uff0c\u6b64\u5916\uff0c\u8fd8\u5bf9\u56fe\u7247\u8fdb\u884c\u4e86\u5f52\u4e00\u5316\uff0c\u539f\u6765\u662f0\uff5e255\u8303\u56f4\u7684\u7070\u5ea6\u56fe\uff0c\u5f52\u4e00\u5316\u52300\uff5e1\u4e4b\u95f4\uff0c\u8fd9\u6837\u80fd\u591f\u52a0\u5feb\u6536\u655b\uff0c\u6700\u540e\u5c31\u662f\u75281\u51cf\u53bb\u4e86\u56fe\u7247\u77e9\u9635\uff0c\u5b9e\u9645\u4e0a\u5bf9\u56fe\u7247\u505a\u4e86\u4e2a\u989c\u8272\u53cd\u8f6c\uff0c\u201c\u767d\u5e95\u9ed1\u5b57\u201d\u53d8\u6210\u4e86\u201c\u9ed1\u5e95\u767d\u5b57\u201d\u3002\u56e0\u4e3a\u5728\u8bad\u7ec3\u7f51\u7edc\u65f6\uff0c\u6211\u4eec\u5e0c\u671b\u8f93\u5165\u6709\u5927\u91cf\u76840\uff0c\u8fd9\u6837\u53ef\u4ee5\u52a0\u5feb\u6536\u655b\uff0c\u4f46\u662f\u989c\u8272\u4e2d\uff0c\u767d\u8272\u7684255\uff0c\u9ed1\u8272\u624d\u662f0\uff0c\u56e0\u6b64\uff0c\u201c\u9ed1\u5e95\u767d\u5b57\u201d\u6bd4\u201c\u767d\u5e95\u9ed1\u5b57\u201d\u66f4\u5feb\u6536\u655b\u3002<\/p>\n
\u8bad\u7ec3\u6a21\u578b\u00a0#<\/a><\/h3>\n
\u7136\u540e\u5c31\u7528\u5b83\u4eec\u8bad\u7ec3\u4e00\u4e2aCNN\u4e86\uff0c\u8fd9\u4e2a\u8fc7\u7a0b\u592a\u5e38\u89c4\u4e86\uff0c\u7528\u7684\u5c31\u662f\u4e09\u4e2a\u5377\u79ef\u5c42\u548c\u6c60\u5316\u5c42\u7684\u5806\u53e0\uff0c\u7136\u540e\u505a\u4e00\u4e2asoftmax\u5206\u7c7b\u5668\uff0c\u5c31\u8fd9\u4e48\u7b80\u5355\uff5e<\/p>\n
\u6a21\u578b\u7684\u7ed3\u6784\uff1a<\/p>\n
______________________________________________________________
\nLayer (type) Output Shape Param # Connected to
\n==============================================================
\ninput_2 (InputLayer) (None, 144, 180) 0
\n______________________________________________________________
\nconvolution1d_4 (Convolution1D) (None, 143, 32) 11552 input_2[0][0]
\n______________________________________________________________
\nmaxpooling1d_4 (MaxPooling1D) (None, 71, 32) 0 convolution1d_4[0][0]
\n______________________________________________________________
\nconvolution1d_5 (Convolution1D) (None, 70, 32) 2080 maxpooling1d_4[0][0]
\n______________________________________________________________
\nmaxpooling1d_5 (MaxPooling1D) (None, 35, 32) 0 convolution1d_5[0][0]
\n______________________________________________________________
\nconvolution1d_6 (Convolution1D) (None, 34, 32) 2080 maxpooling1d_5[0][0]
\n______________________________________________________________
\nmaxpooling1d_6 (MaxPooling1D) (None, 17, 32) 0 convolution1d_6[0][0]
\n______________________________________________________________
\nflatten_2 (Flatten) (None, 544) 0 maxpooling1d_6[0][0]
\n______________________________________________________________
\ndense_3 (Dense) (None, 32) 17440 flatten_2[0][0]
\n______________________________________________________________
\ndense_4 (Dense) (None, 1) 33 dense_3[0][0]
\n==============================================================
\nTotal params: 33185
\n______________________________________________________________<\/p><\/blockquote>\n
\u4ee3\u7801\uff1a<\/p>\n
from<\/span> keras.<\/span>layers import<\/span> Input,<\/span> Convolution1D,<\/span> MaxPooling1D,<\/span> Flatten,<\/span> Dense\r\nfrom<\/span> keras.<\/span>models import<\/span> Model\r\n\r\ninput =<\/span> Input(<\/span>(<\/span>144<\/span>,<\/span> 180<\/span>)<\/span>)<\/span>\r\ncnn =<\/span> Convolution1D(<\/span>32<\/span>,<\/span> 2<\/span>)<\/span>(<\/span>input)<\/span>\r\ncnn =<\/span> MaxPooling1D(<\/span>2<\/span>)<\/span>(<\/span>cnn)<\/span>\r\ncnn =<\/span> Convolution1D(<\/span>32<\/span>,<\/span> 2<\/span>)<\/span>(<\/span>cnn)<\/span>\r\ncnn =<\/span> MaxPooling1D(<\/span>2<\/span>)<\/span>(<\/span>cnn)<\/span>\r\ncnn =<\/span> Convolution1D(<\/span>32<\/span>,<\/span> 2<\/span>)<\/span>(<\/span>cnn)<\/span>\r\ncnn =<\/span> MaxPooling1D(<\/span>2<\/span>)<\/span>(<\/span>cnn)<\/span>\r\ncnn =<\/span> Flatten(<\/span>)<\/span>(<\/span>cnn)<\/span>\r\ndense =<\/span> Dense(<\/span>32<\/span>,<\/span> activation=<\/span>'relu'<\/span>)<\/span>(<\/span>cnn)<\/span>\r\ndense =<\/span> Dense(<\/span>1<\/span>,<\/span> activation=<\/span>'sigmoid'<\/span>)<\/span>(<\/span>dense)<\/span>\r\n\r\nmodel =<\/span> Model(<\/span>input=<\/span>input,<\/span> output=<\/span>dense)<\/span>\r\nmodel.<\/span>compile(<\/span>loss=<\/span>'binary_crossentropy'<\/span>,<\/span> optimizer=<\/span>'adam'<\/span>,<\/span> metrics=<\/span>[<\/span>'accuracy'<\/span>]<\/span>)<\/span>\r\nmodel.<\/span>summary(<\/span>)<\/span>\r\n\r\nmodel.<\/span>fit_generator(<\/span>data(<\/span>sample,<\/span> batch_size=<\/span>1024<\/span>)<\/span>,<\/span> \r\n                    nb_epoch=<\/span>100<\/span>,<\/span> \r\n                    samples_per_epoch=<\/span>nb_samples*<\/span>3<\/span>\r\n                   )<\/span>\r\nmodel.<\/span>save_weights(<\/span>'2013_suizhifuyuan_cnn.model'<\/span>)<\/span>\r\n<\/code><\/pre>\n
\u5176\u5b9e\u8fed\u4ee33\u6b21\u5c31\u53ef\u4ee5\u5f97\u523095%\u7684\u51c6\u786e\u7387\u4e86\uff0c\u8fd9\u4e2anb_epoch=100\u662f\u968f\u610f\u5199\u7684\uff5e\u5982\u679c\u8ba1\u7b97\u8d44\u6e90\u5145\u8db3\u800c\u53c8\u4e0d\u7740\u6025\u7684\u8bdd\uff0c\u591a\u8fed\u4ee3\u51e0\u6b21\u65e0\u59a8\u3002\u6211\u6700\u7ec8\u5f97\u5230\u4e8697.7%\u7684\u51c6\u786e\u7387\u3002<\/p>\n
\u62fc\u63a5\u6548\u679c\u00a0#<\/a><\/h3>\n
\u73b0\u5728\u5c31\u53ef\u4ee5\u6d4b\u8bd5\u4e00\u4e0b\u6211\u4eec\u8bad\u7ec3\u51fa\u6765\u7684\u201c\u8ddd\u79bb\u51fd\u6570\u201d\u7684\u6027\u80fd\u4e86\u3002<\/p>\n
import<\/span> glob\r\n\r\nimg_names =<\/span> glob.<\/span>glob(<\/span>u'\u9644\u4ef63\/*'<\/span>)<\/span>\r\nimages =<\/span> {<\/span>}<\/span>\r\nfor<\/span> i in<\/span> img_names:<\/span>\r\n    images[<\/span>i]<\/span> =<\/span> 1<\/span> -<\/span> misc.<\/span>imread(<\/span>i,<\/span> flatten=<\/span>True<\/span>)<\/span>.<\/span>T\/<\/span>255.0<\/span>\r\n\r\ndef<\/span> find_most_similar<\/span>(<\/span>img,<\/span> images)<\/span>:<\/span>\r\n    imgs_ =<\/span> np.<\/span>array(<\/span>[<\/span>np.<\/span>vstack(<\/span>(<\/span>images[<\/span>img]<\/span>,<\/span> images[<\/span>i]<\/span>)<\/span>)<\/span> for<\/span> i in<\/span> images if<\/span> i !=<\/span> img]<\/span>)<\/span>\r\n    img_names_ =<\/span> [<\/span>i for<\/span> i in<\/span> images if<\/span> i !=<\/span> img]<\/span>\r\n    sims =<\/span> model.<\/span>predict(<\/span>imgs_)<\/span>.<\/span>reshape(<\/span>-<\/span>1<\/span>)<\/span>\r\n    return<\/span> img_names_[<\/span>sims.<\/span>argmax(<\/span>)<\/span>]<\/span>\r\n\r\nimg =<\/span> img_names[<\/span>14<\/span>]<\/span>\r\nresult =<\/span> [<\/span>img]<\/span>\r\nimages_ =<\/span> images.<\/span>copy(<\/span>)<\/span>\r\nwhile<\/span> len(<\/span>images_)<\/span> ><\/span> 1<\/span>:<\/span>\r\n    print<\/span> len(<\/span>images_)<\/span>\r\n    img_ =<\/span> find_most_similar(<\/span>img,<\/span> images_)<\/span>\r\n    result.<\/span>append(<\/span>img_)<\/span>\r\n    del<\/span> images_[<\/span>img]<\/span>\r\n    img =<\/span> img_\r\n\r\nimages_ =<\/span> [<\/span>images[<\/span>i]<\/span>.<\/span>T for<\/span> i in<\/span> result]<\/span>\r\ncompose =<\/span> (<\/span>1<\/span> -<\/span> np.<\/span>hstack(<\/span>images_)<\/span>)<\/span>*<\/span>255<\/span>\r\nmisc.<\/span>imsave(<\/span>'result.png'<\/span>,<\/span> compose)<\/span>\r\n<\/code><\/pre>\n
\u5bf9\u4e8e\u9644\u4ef63\uff0c\u4e00\u6b21\u6027\u62fc\u63a5\u7684\u7ed3\u679c\u662f\uff08\u653e\u5927\u770b\u539f\u56fe\uff09\uff1a<\/p>\n
\n
\n
\"\u9644\u4ef63\u590d\u539f\u7a0b\u5ea6\uff08CNN+\u8d2a\u5fc3\u7b97\u6cd5\uff09\"<\/a><\/div>\n
\u9644\u4ef63\u590d\u539f\u7a0b\u5ea6\uff08CNN+\u8d2a\u5fc3\u7b97\u6cd5\uff09<\/p>\n<\/div>\n<\/div>\n
 <\/p>\n
\u4e3a\u4e86\u5bf9\u6bd4\uff0c\u4ee5\u524d\u4f7f\u7528\u6b27\u6c0f\u8ddd\u79bb\uff0c\u4e00\u6b21\u6027\u62fc\u63a5\u7684\u7ed3\u679c\u662f\uff1a<\/p>\n
\n
\n
\"\u9644\u4ef63\u590d\u539f\u7a0b\u5ea6\uff08\u6b27\u6c0f\u8ddd\u79bb+\u8d2a\u5fc3\u7b97\u6cd5\uff09\"<\/a><\/div>\n
\u9644\u4ef63\u590d\u539f\u7a0b\u5ea6\uff08\u6b27\u6c0f\u8ddd\u79bb+\u8d2a\u5fc3\u7b97\u6cd5\uff09<\/p>\n<\/div>\n<\/div>\n
 <\/p>\n
\u53ef\u89c1\u6539\u8fdb\u662f\u5f88\u660e\u663e\u7684\u3002\u867d\u7136\u6211\u4eec\u8fd9\u4e2a\u6a21\u578b\u662f\u7528\u4e2d\u6587\u8bed\u6599\u8bad\u7ec3\u7684\uff0c\u4f46\u662f\u76f4\u63a5\u7528\u4e8e\u9644\u4ef64\u7684\u5224\u65ad\uff0c\u5f97\u5230\u7684\u7ed3\u679c\u4e5f\u4e0d\u9519\uff1a<\/p>\n
\n
\n
\"\u9644\u4ef64\u590d\u539f\u7a0b\u5ea6\uff08CNN+\u8d2a\u5fc3\u7b97\u6cd5\uff09\"<\/a><\/div>\n
\u9644\u4ef64\u590d\u539f\u7a0b\u5ea6\uff08CNN+\u8d2a\u5fc3\u7b97\u6cd5\uff09<\/p>\n<\/div>\n<\/div>\n
 <\/p>\n
\u540c\u6837\u7684\uff0c\u4f5c\u4e3a\u5bf9\u6bd4\uff0c\u9644\u4ef64\u4f7f\u7528\u6b27\u6c0f\u8ddd\u79bb\u7684\u62fc\u63a5\u7ed3\u679c\u662f\uff1a<\/p>\n
\n
\n
\"\u9644\u4ef64\u590d\u539f\u7a0b\u5ea6\uff08\u6b27\u6c0f\u8ddd\u79bb+\u8d2a\u5fc3\u7b97\u6cd5\uff09\"<\/a><\/div>\n
\u9644\u4ef64\u590d\u539f\u7a0b\u5ea6\uff08\u6b27\u6c0f\u8ddd\u79bb+\u8d2a\u5fc3\u7b97\u6cd5\uff09<\/p>\n<\/div>\n<\/div>\n
 <\/p>\n
\u56e0\u6b64\u53ef\u4ee5\u53d1\u73b0\uff0c\u6211\u4eec\u5f97\u5230\u7684\u6a21\u578b\u6548\u679c\u7684\u786e\u4e0d\u9519\uff0c\u800c\u4e14\u5177\u6709\u5f88\u5f3a\u7684\u6cdb\u5316\u80fd\u529b\u3002\u5f53\u7136\uff0c\u76f4\u63a5\u62fc\u63a5\u82f1\u6587\u7684\u6548\u679c\u4e0d\u5927\u597d\uff0c\u6700\u597d\u540c\u65f6\u52a0\u5165\u82f1\u6587\u8bed\u6599\u4e00\u8d77\u8bad\u7ec3\uff0c\u624d\u80fd\u5f97\u5230\u66f4\u597d\u6548\u679c\u3002<\/p>\n
\u540e\u6587\u00a0#<\/a><\/h3>\n
\u4e00\u9053\u597d\u9898\u5fc5\u7136\u662f\u5185\u6db5\u4e30\u5bcc\uff0c\u7ecf\u4e45\u4e0d\u8870\u7684\uff0c\u8fd9\u5df2\u7ecf\u662f\u6211\u5199\u7684\u5173\u4e8e\u788e\u7eb8\u590d\u539f\u8fd9\u9053\u9898\u7684\u7b2c\u4e09\u7bc7\u6587\u7ae0\u4e86\uff0c\u4e5f\u8bb8\u8fd8\u4f1a\u6709\u7b2c4\u7bc7\u3001\u7b2c5\u7bc7\uff0c\u6bcf\u4e00\u6b21\u7814\u7a76\u90fd\u662f\u4e00\u6b21\u6df1\u5165…<\/p>\n","protected":false},"excerpt":{"rendered":"","protected":false},"author":2,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":[],"categories":[165,54],"tags":[],"_links":{"self":[{"href":"https:\/\/www.shumo.com\/home\/wp-json\/wp\/v2\/posts\/4406"}],"collection":[{"href":"https:\/\/www.shumo.com\/home\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.shumo.com\/home\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.shumo.com\/home\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.shumo.com\/home\/wp-json\/wp\/v2\/comments?post=4406"}],"version-history":[{"count":1,"href":"https:\/\/www.shumo.com\/home\/wp-json\/wp\/v2\/posts\/4406\/revisions"}],"predecessor-version":[{"id":4407,"href":"https:\/\/www.shumo.com\/home\/wp-json\/wp\/v2\/posts\/4406\/revisions\/4407"}],"wp:attachment":[{"href":"https:\/\/www.shumo.com\/home\/wp-json\/wp\/v2\/media?parent=4406"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.shumo.com\/home\/wp-json\/wp\/v2\/categories?post=4406"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.shumo.com\/home\/wp-json\/wp\/v2\/tags?post=4406"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}