“`” \u4e00\u3001\u53c2\u6570\u521d\u59cb\u5316 \u4e0b\u9762\u51e0\u79cd\u65b9\u5f0f,\u968f\u4fbf\u9009\u4e00\u4e2a,\u7ed3\u679c\u57fa\u672c\u90fd\u5dee\u4e0d\u591a\u3002\u4f46\u662f\u4e00\u5b9a\u8981\u505a\u3002\u5426\u5219\u53ef\u80fd\u4f1a\u51cf\u6162\u6536\u655b\u901f\u5ea6\uff0c\u5f71\u54cd\u6536\u655b\u7ed3\u679c\uff0c\u751a\u81f3\u9020\u6210Nan\u7b49\u4e00\u7cfb\u5217\u95ee\u9898\u3002 \u4e0b\u9762\u7684n_in\u4e3a\u7f51\u7edc\u7684\u8f93\u5165\u5927\u5c0f\uff0cn_out\u4e3a\u7f51\u7edc\u7684\u8f93\u51fa\u5927\u5c0f\uff0cn\u4e3an_in\u6216(n_in+n_out)*0.5<\/p>\n
Xavier\u521d\u59cb\u6cd5\u8bba\u6587\uff1a<a class=""url"" href=""http:\/\/jmlr.org\/proceedings\/papers\/v9\/glorot10a\/glorot10a.pdf"" target=""_blank"">http:\/\/jmlr.org\/proceedings\/papers\/v9\/glorot10a\/glorot10a.pdf<\/a><\/p>\n
He\u521d\u59cb\u5316\u8bba\u6587\uff1a<a class=""url"" href=""https:\/\/arxiv.org\/abs\/1502.01852"" target=""_blank"">https:\/\/arxiv.org\/abs\/1502.01852<\/a> uniform\u5747\u5300\u5206\u5e03\u521d\u59cb\u5316\uff1aw = np.random.uniform(low=-scale, high=scale, size=[n_in,n_out])<\/p>\n
Xavier\u521d\u59cb\u6cd5\uff0c\u9002\u7528\u4e8e\u666e\u901a\u6fc0\u6d3b\u51fd\u6570(tanh,sigmoid)\uff1ascale = np.sqrt(3\/n)<\/p>\n
He\u521d\u59cb\u5316\uff0c\u9002\u7528\u4e8eReLU\uff1ascale = np.sqrt(6\/n) normal\u9ad8\u65af\u5206\u5e03\u521d\u59cb\u5316\uff1aw = np.random.randn(n_in,n_out) * stdev # stdev\u4e3a\u9ad8\u65af\u5206\u5e03\u7684\u6807\u51c6\u5dee\uff0c\u5747\u503c\u8bbe\u4e3a0<\/p>\n
Xavier\u521d\u59cb\u6cd5\uff0c\u9002\u7528\u4e8e\u666e\u901a\u6fc0\u6d3b\u51fd\u6570 (tanh,sigmoid)\uff1astdev = np.sqrt(n)<\/p>\n
He\u521d\u59cb\u5316\uff0c\u9002\u7528\u4e8eReLU\uff1astdev = np.sqrt(2\/n)<\/p>\n
svd\u521d\u59cb\u5316\uff1a\u5bf9RNN\u6709\u6bd4\u8f83\u597d\u7684\u6548\u679c\u3002\u53c2\u8003\u8bba\u6587\uff1a<a class=""url"" href=""https:\/\/arxiv.org\/abs\/1312.6120"" target=""_blank"">https:\/\/arxiv.org\/abs\/1312.6120<\/a><\/p>\n
\u4e8c\u3001\u6570\u636e\u9884\u5904\u7406\u65b9\u5f0f zero-center ,\u8fd9\u4e2a\u633a\u5e38\u7528\u7684.X -= np.mean(X, axis = 0) # zero-centerX \/= np.std(X, axis = 0) # normalize PCA whitening,\u8fd9\u4e2a\u7528\u7684\u6bd4\u8f83\u5c11.<\/p>\n
\u4e09\u3001\u8bad\u7ec3\u6280\u5de7 \u8981\u505a\u68af\u5ea6\u5f52\u4e00\u5316,\u5373\u7b97\u51fa\u6765\u7684\u68af\u5ea6\u9664\u4ee5minibatch size clip c(\u68af\u5ea6\u88c1\u526a): \u9650\u5236\u6700\u5927\u68af\u5ea6,\u5176\u5b9e\u662fvalue = sqrt(w1^2+w2^2\u2026.),\u5982\u679cvalue\u8d85\u8fc7\u4e86\u9608\u503c,\u5c31\u7b97\u4e00\u4e2a\u8870\u51cf\u7cfb\u7cfb\u6570,\u8ba9value\u7684\u503c\u7b49\u4e8e\u9608\u503c: 5,10,15<\/p>\n
dropout\u5bf9\u5c0f\u6570\u636e\u9632\u6b62\u8fc7\u62df\u5408\u6709\u5f88\u597d\u7684\u6548\u679c,\u503c\u4e00\u822c\u8bbe\u4e3a0.5,\u5c0f\u6570\u636e\u4e0adropout+sgd\u5728\u6211\u7684\u5927\u90e8\u5206\u5b9e\u9a8c\u4e2d\uff0c\u6548\u679c\u63d0\u5347\u90fd\u975e\u5e38\u660e\u663e.\u56e0\u6b64\u53ef\u80fd\u7684\u8bdd\uff0c\u5efa\u8bae\u4e00\u5b9a\u8981\u5c1d\u8bd5\u4e00\u4e0b\u3002<\/p>\n
dropout\u7684\u4f4d\u7f6e\u6bd4\u8f83\u6709\u8bb2\u7a76, \u5bf9\u4e8eRNN,\u5efa\u8bae\u653e\u5230\u8f93\u5165->RNN\u4e0eRNN->\u8f93\u51fa\u7684\u4f4d\u7f6e.\u5173\u4e8eRNN\u5982\u4f55\u7528dropout,\u53ef\u4ee5\u53c2\u8003\u8fd9\u7bc7\u8bba\u6587:<a class=""url"" href=""http:\/\/arxiv.org\/abs\/1409.2329"" target=""_blank"">http:\/\/arxiv.org\/abs\/1409.2329<\/a> adam,adadelta\u7b49,\u5728\u5c0f\u6570\u636e\u4e0a,\u6211\u8fd9\u91cc\u5b9e\u9a8c\u7684\u6548\u679c\u4e0d\u5982sgd, sgd\u6536\u655b\u901f\u5ea6\u4f1a\u6162\u4e00\u4e9b\uff0c\u4f46\u662f\u6700\u7ec8\u6536\u655b\u540e\u7684\u7ed3\u679c\uff0c\u4e00\u822c\u90fd\u6bd4\u8f83\u597d\u3002<\/p>\n
\u5982\u679c\u4f7f\u7528sgd\u7684\u8bdd,\u53ef\u4ee5\u9009\u62e9\u4ece1.0\u6216\u80050.1\u7684\u5b66\u4e60\u7387\u5f00\u59cb,\u9694\u4e00\u6bb5\u65f6\u95f4,\u5728\u9a8c\u8bc1\u96c6\u4e0a\u68c0\u67e5\u4e00\u4e0b,\u5982\u679ccost\u6ca1\u6709\u4e0b\u964d,\u5c31\u5bf9\u5b66\u4e60\u7387\u51cf\u534a. \u6211\u770b\u8fc7\u5f88\u591a\u8bba\u6587\u90fd\u8fd9\u4e48\u641e,\u6211\u81ea\u5df1\u5b9e\u9a8c\u7684\u7ed3\u679c\u4e5f\u5f88\u597d. \u5f53\u7136,\u4e5f\u53ef\u4ee5\u5148\u7528ada\u7cfb\u5217\u5148\u8dd1,\u6700\u540e\u5feb\u6536\u655b\u7684\u65f6\u5019,\u66f4\u6362\u6210sgd\u7ee7\u7eed\u8bad\u7ec3.\u540c\u6837\u4e5f\u4f1a\u6709\u63d0\u5347.\u636e\u8bf4adadelta\u4e00\u822c\u5728\u5206\u7c7b\u95ee\u9898\u4e0a\u6548\u679c\u6bd4\u8f83\u597d\uff0cadam\u5728\u751f\u6210\u95ee\u9898\u4e0a\u6548\u679c\u6bd4\u8f83\u597d\u3002<\/p>\n
\u9664\u4e86gate\u4e4b\u7c7b\u7684\u5730\u65b9,\u9700\u8981\u628a\u8f93\u51fa\u9650\u5236\u62100-1\u4e4b\u5916,\u5c3d\u91cf\u4e0d\u8981\u7528sigmoid,\u53ef\u4ee5\u7528tanh\u6216\u8005relu\u4e4b\u7c7b\u7684\u6fc0\u6d3b\u51fd\u6570.1. sigmoid\u51fd\u6570\u5728-4\u52304\u7684\u533a\u95f4\u91cc\uff0c\u624d\u6709\u8f83\u5927\u7684\u68af\u5ea6\u3002\u4e4b\u5916\u7684\u533a\u95f4\uff0c\u68af\u5ea6\u63a5\u8fd10\uff0c\u5f88\u5bb9\u6613\u9020\u6210\u68af\u5ea6\u6d88\u5931\u95ee\u9898\u30022. \u8f93\u51650\u5747\u503c\uff0csigmoid\u51fd\u6570\u7684\u8f93\u51fa\u4e0d\u662f0\u5747\u503c\u7684\u3002 rnn\u7684dim\u548cembdding size,\u4e00\u822c\u4ece128\u4e0a\u4e0b\u5f00\u59cb\u8c03\u6574. batch size,\u4e00\u822c\u4ece128\u5de6\u53f3\u5f00\u59cb\u8c03\u6574.batch size\u5408\u9002\u6700\u91cd\u8981,\u5e76\u4e0d\u662f\u8d8a\u5927\u8d8a\u597d. word2vec\u521d\u59cb\u5316,\u5728\u5c0f\u6570\u636e\u4e0a,\u4e0d\u4ec5\u53ef\u4ee5\u6709\u6548\u63d0\u9ad8\u6536\u655b\u901f\u5ea6,\u4e5f\u53ef\u4ee5\u53ef\u4ee5\u63d0\u9ad8\u7ed3\u679c.<\/p>\n
\u56db\u3001\u5c3d\u91cf\u5bf9\u6570\u636e\u505ashuffle LSTM \u7684forget gate\u7684bias,\u75281.0\u6216\u8005\u66f4\u5927\u7684\u503c\u505a\u521d\u59cb\u5316,\u53ef\u4ee5\u53d6\u5f97\u66f4\u597d\u7684\u7ed3\u679c,\u6765\u81ea\u8fd9\u7bc7\u8bba\u6587:<a class=""url"" href=""http:\/\/jmlr.org\/proceedings\/papers\/v37\/jozefowicz15.pdf"" target=""_blank"">http:\/\/jmlr.org\/proceedings\/papers\/v37\/jozefowicz15.pdf<\/a>, \u6211\u8fd9\u91cc\u5b9e\u9a8c\u8bbe\u62101.0,\u53ef\u4ee5\u63d0\u9ad8\u6536\u655b\u901f\u5ea6.\u5b9e\u9645\u4f7f\u7528\u4e2d,\u4e0d\u540c\u7684\u4efb\u52a1,\u53ef\u80fd\u9700\u8981\u5c1d\u8bd5\u4e0d\u540c\u7684\u503c. Batch Normalization\u636e\u8bf4\u53ef\u4ee5\u63d0\u5347\u6548\u679c\uff0c\u4e0d\u8fc7\u6211\u6ca1\u6709\u5c1d\u8bd5\u8fc7\uff0c\u5efa\u8bae\u4f5c\u4e3a\u6700\u540e\u63d0\u5347\u6a21\u578b\u7684\u624b\u6bb5\uff0c\u53c2\u8003\u8bba\u6587\uff1aAccelerating Deep Network Training by Reducing Internal Covariate Shift<\/p>\n
\u5982\u679c\u4f60\u7684\u6a21\u578b\u5305\u542b\u5168\u8fde\u63a5\u5c42\uff08MLP\uff09\uff0c\u5e76\u4e14\u8f93\u5165\u548c\u8f93\u51fa\u5927\u5c0f\u4e00\u6837\uff0c\u53ef\u4ee5\u8003\u8651\u5c06MLP\u66ff\u6362\u6210Highway Network,\u6211\u5c1d\u8bd5\u5bf9\u7ed3\u679c\u6709\u4e00\u70b9\u63d0\u5347\uff0c\u5efa\u8bae\u4f5c\u4e3a\u6700\u540e\u63d0\u5347\u6a21\u578b\u7684\u624b\u6bb5\uff0c\u539f\u7406\u5f88\u7b80\u5355\uff0c\u5c31\u662f\u7ed9\u8f93\u51fa\u52a0\u4e86\u4e00\u4e2agate\u6765\u63a7\u5236\u4fe1\u606f\u7684\u6d41\u52a8\uff0c\u8be6\u7ec6\u4ecb\u7ecd\u8bf7\u53c2\u8003\u8bba\u6587: <a class=""url"" href=""http:\/\/arxiv.org\/abs\/1505.00387"" target=""_blank"">http:\/\/arxiv.org\/abs\/1505.00387<\/a> \u4e00\u8f6e\u52a0\u6b63\u5219\uff0c\u4e00\u8f6e\u4e0d\u52a0\u6b63\u5219\uff0c\u53cd\u590d\u8fdb\u884c\u3002<\/p>\n
\u4e94\u3001Ensemble Ensemble\u662f\u8bba\u6587\u5237\u7ed3\u679c\u7684\u7ec8\u6781\u6838\u6b66\u5668,\u6df1\u5ea6\u5b66\u4e60\u4e2d\u4e00\u822c\u6709\u4ee5\u4e0b\u51e0\u79cd\u65b9\u5f0f \u540c\u6837\u7684\u53c2\u6570,\u4e0d\u540c\u7684\u521d\u59cb\u5316\u65b9\u5f0f \u4e0d\u540c\u7684\u53c2\u6570,\u901a\u8fc7cross-validation,\u9009\u53d6\u6700\u597d\u7684\u51e0\u7ec4 \u540c\u6837\u7684\u53c2\u6570,\u6a21\u578b\u8bad\u7ec3\u7684\u4e0d\u540c\u9636\u6bb5\uff0c\u5373\u4e0d\u540c\u8fed\u4ee3\u6b21\u6570\u7684\u6a21\u578b\u3002 \u4e0d\u540c\u7684\u6a21\u578b,\u8fdb\u884c\u7ebf\u6027\u878d\u5408. \u4f8b\u5982RNN\u548c\u4f20\u7edf\u6a21\u578b.<\/p>\n
<pre><code> "“`
\n<\/code><\/pre>\n","protected":false},"excerpt":{"rendered":"
“`” \u4e00\u3001\u53c2\u6570\u521d\u59cb\u5316 \u4e0b\u9762\u51e0\u79cd\u65b9\u5f0f,\u968f\u4fbf\u9009\u4e00\u4e2a,\u7ed3\u679c\u57fa\u672c\u90fd\u5dee\u4e0d\u591a\u3002\u4f46\u662f\u4e00\u5b9a\u8981\u505a\u3002\u5426\u5219 […]<\/p>\n","protected":false},"author":7,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[101],"tags":[],"class_list":["post-44045","post","type-post","status-publish","format-standard","hentry","category-c"],"_links":{"self":[{"href":"http:\/\/wx.kaifamiao.info\/index.php\/wp-json\/wp\/v2\/posts\/44045","targetHints":{"allow":["GET"]}}],"collection":[{"href":"http:\/\/wx.kaifamiao.info\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"http:\/\/wx.kaifamiao.info\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"http:\/\/wx.kaifamiao.info\/index.php\/wp-json\/wp\/v2\/users\/7"}],"replies":[{"embeddable":true,"href":"http:\/\/wx.kaifamiao.info\/index.php\/wp-json\/wp\/v2\/comments?post=44045"}],"version-history":[{"count":1,"href":"http:\/\/wx.kaifamiao.info\/index.php\/wp-json\/wp\/v2\/posts\/44045\/revisions"}],"predecessor-version":[{"id":44046,"href":"http:\/\/wx.kaifamiao.info\/index.php\/wp-json\/wp\/v2\/posts\/44045\/revisions\/44046"}],"wp:attachment":[{"href":"http:\/\/wx.kaifamiao.info\/index.php\/wp-json\/wp\/v2\/media?parent=44045"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/wx.kaifamiao.info\/index.php\/wp-json\/wp\/v2\/categories?post=44045"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/wx.kaifamiao.info\/index.php\/wp-json\/wp\/v2\/tags?post=44045"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}