{"id":4744,"date":"2026-03-05T15:42:36","date_gmt":"2026-03-05T07:42:36","guid":{"rendered":"http:\/\/www.puchao.com\/?p=4744"},"modified":"2026-03-05T15:55:57","modified_gmt":"2026-03-05T07:55:57","slug":"%e6%99%ae%e8%b6%85insight-%e8%ae%ba%e6%96%87%e6%8b%86%e8%a7%a3%e7%b3%bb%e5%88%971%ef%bc%9a%e5%a4%a7%e7%99%bd%e8%af%9d%e8%ae%b2%e8%ae%b2transformer","status":"publish","type":"post","link":"http:\/\/www.puchao.com\/index.php\/2026\/03\/05\/%e6%99%ae%e8%b6%85insight-%e8%ae%ba%e6%96%87%e6%8b%86%e8%a7%a3%e7%b3%bb%e5%88%971%ef%bc%9a%e5%a4%a7%e7%99%bd%e8%af%9d%e8%ae%b2%e8%ae%b2transformer\/","title":{"rendered":"\u666e\u8d85Insight | \u8bba\u6587\u62c6\u89e3\u7cfb\u52171\uff1a\u5927\u767d\u8bdd\u8bb2\u8bb2Transformer"},"content":{"rendered":"<p>2017\u5e746\u6708\uff0cGoogle Brain\uff08\u73b0\u5728\u5df2\u548cDeep Mind\u5408\u5e76\uff09\u7684Illia Polosukhin\u3001Ashish Vaswani \u7b49\u516b\u4f4d\u4f5c\u8005\uff0c\u53d1\u8868\u4e86\u8bba\u6587\u300aAttention Is All You Need\u300b\uff0c\u6b63\u5f0f\u63d0\u51faTransformer\u67b6\u6784\u3002<\/p>\n<p>\u5f7c\u65f6\uff0c\u8fd9\u7bc7\u8bba\u6587\u88ab\u5f52\u7c7b\u4e3a\u673a\u5668\u7ffb\u8bd1\u9886\u57df\u7684\u4e00\u6b21\u6280\u672f\u8fed\u4ee3\uff0c\u6ca1\u4eba\u610f\u8bc6\u5230\u5b83\u662f\u4eba\u5de5\u667a\u80fd\u5386\u53f2\u4e0a\u7684\u4e07\u5386\u5341\u4e94\u5e74\u3002<\/p>\n<p><span style=\"color: #3366ff;\"><strong>\u4e00\u3001Transformer\u4e3a\u4e86\u89e3\u51b3\u4ec0\u4e48\u95ee\u9898<\/strong><\/span><\/p>\n<p>\u5728\u8fd9\u7bc7\u8bba\u6587\u4e4b\u524d\uff0c\u7edf\u6cbb\u673a\u5668\u5b66\u4e60\u754c\u7684\u662f<strong>RNN<\/strong>\uff08\u5faa\u73af\u795e\u7ecf\u7f51\u7edc\uff09\u53ca\u5176\u53d8\u4f53LSTM\uff08\u957f\u77ed\u671f\u8bb0\u5fc6\u7f51\u7edc\uff09\u3002\u800cRNN\u6700\u663e\u8457\u7684\u7279\u5f81\u662f<strong>\u7ebf\u6027\u3001\u4e32\u884c<\/strong>\u3002\u901a\u8fc7\u65f6\u95f4\u6b65\u957f\u7684\u5f15\u5165\uff0cRNN\u5728\u751f\u6210\u4e0b\u4e00\u4e2atoken\u65f6\uff0c\u53ef\u4ee5\u628a\u524d\u5e8f\u6240\u6709\u65f6\u95f4\u6b65\u957f\u7684\u4fe1\u606f\u88c5\u8fdb\u9690\u85cf\u72b6\u6001\u91cc\u5e26\u8fc7\u6765\uff0c\u56e0\u6b64\u7406\u8bba\u4e0aRNN\u53ef\u4ee5\u628a\u4efb\u610f\u957f\u5ea6\u7684\u5e8f\u5217\u4f5c\u4e3a\u8f93\u5165\u9879\u3002<\/p>\n<p>\u4f46\u5b9e\u9645\u60c5\u51b5\u5e76\u4e0d\u7406\u60f3\u3002<\/p>\n<p>RNN\u5728\u5904\u7406\u957f\u5e8f\u5217\u65f6\uff0c\u9700\u8981\u5c06\u4fe1\u606f\u901a\u8fc7\u65f6\u95f4\u6b65\u9010\u5c42\u4f20\u9012\u3002\u7531\u4e8e\u53cd\u5411\u4f20\u64ad\u6d89\u53ca\u6743\u91cd\u77e9\u9635\u7684\u8fde\u4e58\uff0c\u5982\u679c\u6fc0\u6d3b\u51fd\u6570\u7684\u5bfc\u6570\u5c0f\u4e8e1\uff0c\u68af\u5ea6\u4f1a\u968f\u7740\u65f6\u95f4\u6b65\u7684\u589e\u52a0\u5448\u6307\u6570\u7ea7\u8870\u51cf\uff08\u68af\u5ea6\u6d88\u5931\uff09\uff1b\u5982\u679c\u6743\u91cd\u77e9\u9635\u8f83\u5927\uff0c\u68af\u5ea6\u4f1a\u5448\u6307\u6570\u7ea7\u589e\u957f\uff0c\u5bfc\u81f4\u6a21\u578b\u8bad\u7ec3\u6781\u5ea6\u4e0d\u7a33\u5b9a\uff0c\u751a\u81f3\u76f4\u63a5\u62a5\u9519\uff08\u68af\u5ea6\u7206\u70b8\uff09\u3002\u6240\u4ee5\u5b9e\u9645\u4e0aRNN\u53ea\u80fd\u8bb0\u4f4f10-20\u6b65\u7684\u4fe1\u606f\u3002<\/p>\n<p>\u5c31\u50cf\u5f53\u4f60\u8bfb\u4e00\u672c\u957f\u7bc7\u5c0f\u8bf4\uff0c\u8bfb\u5230\u7b2c500\u9875\u65f6\uff0c\u4f60\u5df2\u7ecf\u5ffd\u7565\u4e86\u7b2c1\u9875\u57cb\u4e0b\u7684\u4f0f\u7b14\u3002RNN\u4e5f\u662f\u5982\u6b64\uff0c\u968f\u7740\u5e8f\u5217\u7684\u4e0d\u65ad\u62c9\u957f\uff0c\u65e9\u671f\u7684\u4fe1\u606f\u5728\u4f20\u9012\u8fc7\u7a0b\u4e2d\u4f1a\u9010\u6e10\u8870\u51cf\uff0c\u76f4\u81f3\u6d88\u5931\uff0c\u8fd9\u79cd\u60c5\u51b5\u4e0bRNN\u5728\u7528\u6700\u8fd1\u7684\u77ed\u671f\u8bb0\u5fc6\u53bb\u8986\u76d6\u957f\u671f\u7684\u6587\u672c\u4fe1\u606f\uff0c\u81ea\u7136\u4e0d\u51c6\u786e\u3002<\/p>\n<p>\u9664\u6b64\u4e4b\u5916\uff0cRNN\u4e5f\u65e0\u6cd5\u5e76\u884c\u8ba1\u7b97\u3002\u5b83\u7684\u6838\u5fc3\u903b\u8f91\u662f\u4e0b\u4e00\u4e2a\u6b65\u957f\u7684\u8ba1\u7b97\u4f9d\u8d56\u4e8e\u8fd9\u4e00\u6b65\u957f\u7684\u7ed3\u679c\u3002\u6a21\u578b\u5fc5\u987b\u5148\u7b97\u51fa\u7b2c1\u4e2a\u8bcd\uff0c\u624d\u80fd\u7b97\u7b2c2\u4e2a\u8bcd\uff0c\u4ee5\u6b64\u7c7b\u63a8\u3002\u4e32\u884c\u8ba1\u7b97\u6a21\u5f0f\u5bfc\u81f4\u5b83\u65e0\u6cd5\u5145\u5206\u5229\u7528GPU\u7b49\u52a0\u901f\u786c\u4ef6\u7684\u5e76\u884c\u8ba1\u7b97\u80fd\u529b\uff0c\u8bad\u7ec3\u6548\u7387\u6781\u4f4e\u3002<\/p>\n<p>Transformer\u67b6\u6784\u5c31\u662f\u4e3a\u4e86\u89e3\u51b3\u4ee5\u4e0a\u95ee\u9898\u88ab\u63d0\u51fa\u7684\u3002<\/p>\n<p><strong><span style=\"color: #3366ff;\">\u4e8c\u3001Transformer\u662f\u600e\u4e48\u5de5\u4f5c\u7684<\/span><\/strong><\/p>\n<p>\u5728\u5f00\u59cb\u4e4b\u524d\uff0c\u6211\u4eec\u9605\u89c82\u4e2a\u6982\u5ff5\uff0c\u6709\u52a9\u4e8e\u6211\u4eec\u8fdb\u4e00\u6b65\u7406\u89e3Transformer\u7684\u5e95\u5c42\u539f\u7406\uff1a<\/p>\n<p><strong>\u5927\u6a21\u578b\uff1a<\/strong>\u751f\u6210\u5f0f\u5927\u6a21\u578b\u672c\u4f53\u5f88\u50cf\u4e00\u4e2a\u7531\u591a\u4e2a\u7ecf\u8fc7\u8bad\u7ec3\u540e\u751f\u6210\u7684\u3001\u590d\u6742\u4e14\u9ad8\u7ef4\u77e9\u9635\u7ec4\u5408\u5728\u4e00\u8d77\u7684\u52a8\u6001\u6982\u7387\u5b57\u5178\u3002\u4e00\u822c\u7684\u5927\u6a21\u578b\uff08\u5982Llama 3\u6216 GPT-3\uff09\u5305\u542b\u7684\u6743\u91cd\u77e9\u9635\u6570\u91cf\u901a\u5e38\u5728\u51e0\u767e\u4e2a\u6570\u91cf\u7ea7\u3002\u5176\u4e2d\u5305\u62ec\u5168\u5c40\u77e9\u9635\u548c\u6bcf\u4e00\u5c42\u7684\u77e9\u9635\uff1a<\/p>\n<ul>\n<li>\u5168\u5c40\u77e9\u9635\u4e00\u822c\u53ea\u67092\u4e2a\uff1a\u8bcd\u5d4c\u5165\u77e9\u9635\uff08<em>We<\/em>\uff0c\u659c\u4f53\u7684\u5927\u5199\u5b57\u6bcd\u4ee3\u8868\u5b83\u662f\u4e00\u4e2a\u77e9\u9635\uff0c\u4e0b\u540c\uff09\u548c\u8f93\u51fa\u5934\u77e9\u9635\uff08<em>Whead<\/em> \uff09\uff1b<\/li>\n<li>\u6bcf\u4e00\u5c42\u7684\u77e9\u9635\uff1a\u81ea\u6ce8\u610f\u529b\u673a\u5236\u6bcf\u4e00\u5c42\u90fd\u5b58\u57284\u4e2a\u6743\u91cd\u77e9\u9635\uff08<em>WQ\u3001WK\u3001WV\u3001WO<\/em>\uff09\uff0c\u524d\u9988\u795e\u7ecf\u7f51\u7edc\u6bcf\u4e00\u5c42\u9700\u89813\u4e2a\uff0c\u5206\u522b\u7528\u6765\u505a\u95e8\u63a7\u3001\u5347\u7ef4\u3001\u964d\u7ef4\uff08Transformer\u8bba\u6587\u4e2d\u53ea\u4f7f\u7528ReLU\u3001FNN 2\u4e2a\u77e9\u9635\uff09\u3002<\/li>\n<\/ul>\n<p><strong>\u6a21\u578b\u7684\u8bad\u7ec3\uff1a<\/strong>\u5728\u6240\u6709\u7684\u6a21\u578b\u77e9\u9635\u91cc\u586b\u5165\u968f\u673a\u6570\uff0c\u5f00\u59cb\u5f80\u524d\u9010\u5c42\u8f93\u5165\uff08\u524d\u5411\u4f20\u64ad\uff0cForward Pass\uff09\u3002\u968f\u5373\u5f00\u59cb\u590d\u76d8\uff0c\u6839\u636e\u5df2\u6709\u7684\u7b54\u6848\uff08\u4e8b\u5148\u51c6\u5907\u597d\u7684\uff09\uff0c\u544a\u8bc9\u6a21\u578b\u731c\u9519\u4e86\uff0c\u731c\u9519\u7684\u79bb\u8c31\u7a0b\u5ea6\u7528\u4ea4\u53c9\u71b5\u635f\u5931\u51fd\u6570\u5927\u5c0f\u6765\u8868\u793a\uff0cLoss=-log\uff08\u9884\u6d4b\u6b63\u786e\u5b57\u7684\u6982\u7387\uff09\u3002\u7136\u540e\u5f00\u59cb\u53cd\u8fc7\u6765\u627e\u9519\u8bef\u7684\u75c7\u7ed3\u5728\u54ea\uff08\u53cd\u5411\u4f20\u64ad\uff0cBackpropagation\uff09\uff0c\u5177\u4f53\u64cd\u4f5c\u5c31\u662f\u8ba1\u7b97\u51fa\u53c2\u6570W\u5bf9\u6700\u7ec8 Loss \u7684\u5f71\u54cd\u529b\uff0c\u5373\u68af\u5ea6=\u2202Loss\/\u2202W\u3002\u518d\u901a\u8fc7\u4f18\u5316\u5668\u5c1d\u8bd5\u66f4\u65b0\u77e9\u9635\u53c2\u6570\uff08<em>Wnew =Wold<\/em>\u00a0&#8211; \u5b66\u4e60\u7387\u00d7\u68af\u5ea6\uff09\u3002\u968f\u7740\u68af\u5ea6\u4e0b\u964d\uff0c\u7ed3\u679c\u4f1a\u9010\u6e10\u6536\u655b\u3002<\/p>\n<p><strong>\u4ee5\u4e0b\u8fdb\u5165\u6b63\u6587\uff1a<\/strong><\/p>\n<p>Transformer\u4f9d\u7136\u6cbf\u7528\u4e86RNN\u7f16\u7801\u5668-\u89e3\u7801\u5668\uff08Encoder-Decoder\uff09\u8fd9\u90e8\u5206\u7ed3\u6784\uff0c\u4f46\u662f\u6452\u5f03\u4e86\u4f20\u7edf\u7684\u5faa\u73af\u548c\u5377\u79ef\uff0c\u5b8c\u5168\u4f9d\u8d56<strong>\u6ce8\u610f\u529b\u673a\u5236<\/strong>\uff08Attention\uff09\u6765\u6355\u6349\u8f93\u5165\u548c\u8f93\u51fa\u4e4b\u95f4\u7684\u5168\u5c40\u4f9d\u8d56\u5173\u7cfb\u3002\u6211\u4eec\u5c06\u6839\u636e\u8bba\u6587\u91cc\u7684\u6d41\u7a0b\u56fe\uff0c\u4eceInputs\u5f00\u59cb\uff0c\u4f9d\u7167\u903b\u8f91\u987a\u5e8f\u6765\u89e3\u91caTransformer\u7684\u5177\u4f53\u5de5\u4f5c\u6d41\u7a0b\uff1a<\/p>\n<p><img decoding=\"async\" class=\"aligncenter size-full wp-image-4745\" src=\"http:\/\/www.puchao.com\/wp-content\/uploads\/640-4.png\" alt=\"\" width=\"472\" height=\"695\" srcset=\"http:\/\/www.puchao.com\/wp-content\/uploads\/640-4-200x294.png 200w, http:\/\/www.puchao.com\/wp-content\/uploads\/640-4-204x300.png 204w, http:\/\/www.puchao.com\/wp-content\/uploads\/640-4-400x589.png 400w, http:\/\/www.puchao.com\/wp-content\/uploads\/640-4.png 472w\" sizes=\"(max-width: 472px) 100vw, 472px\" \/><\/p>\n<p style=\"text-align: center;\"><span style=\"color: #808080;\">Transformer\u67b6\u6784\u793a\u610f<\/span><\/p>\n<p><span style=\"color: #808080;\"><strong>\u7b2c\u4e00\u9636\u6bb5\uff1a\u9884\u5904\u7406<\/strong><\/span><\/p>\n<p><strong>(1)\u8bcd\u5d4c\u5165\uff08Input Embedding\uff09<\/strong><\/p>\n<p><strong>\u8bcd\u5d4c\u5165<\/strong>\u8868\u53ef\u4ee5\u91c7\u7528Word2Vec\u3001Glove\u7b49\u7b97\u6cd5\u9884\u8bad\u7ec3\u5f97\u5230\uff0c\u4e5f\u53ef\u4ee5\u5728\u00a0Transformer \u4e2d\u8bad\u7ec3\u5f97\u5230\u3002\u67e5\u8be2\u8bcd\u5d4c\u5165\u8868\uff0c\u628a\u5305\u542bn\u4e2a\u5355\u8bcd\uff08\u5b57\uff09\u7684Inputs\u5e8f\u5217\u8f6c\u6362\u6210\u4e00\u4e2a\u56fa\u5b9a\u7684\u5411\u91cf<em>X<\/em>=[w1,w2,&#8230;,wn]\uff0c\u5176\u7ef4\u5ea6\u4e3a n\u00d7d model \uff08\u8bba\u6587\u91cc\u6a21\u578b\u7ef4\u5ea6d model =512\uff09\uff0c<em>X<\/em>\u7684\u6bcf\u4e00\u884c\u90fd\u4ee3\u8868\u4e00\u4e2a\u8bcd\u3002\u8fd9\u4e00\u6b65\u662f\u4e3a\u4e86\u628a\u5e8f\u5217\u53d8\u6210\u80fd\u7528\u8ba1\u7b97\u673a\u8ba1\u7b97\u7684\u5f62\u5f0f\u3002\u57fa\u4e8e\u8fd9\u79cd\u53d8\u6362\uff0c\u540e\u7eed\u8bad\u7ec3\u8fc7\u7a0b\u53ef\u4ee5\u8fdb\u884c\u5927\u91cf\u5e76\u884c\u7684\u77e9\u9635\u8fd0\u7b97\uff0c\u53ef\u88ab\u5e76\u884c\u8ba1\u7b97\u786c\u4ef6\u52a0\u901f\u3002<\/p>\n<p>\u8ba1\u7b97\u8fc7\u7a0b\uff1a<\/p>\n<ul>\n<li>\u8f93\u5165\u4e3a<em>X<\/em>=[w1,w2,&#8230;,wn]<\/li>\n<li>\u521d\u59cb\u5316\u4e00\u4e2a\u8bcd\u5d4c\u5165\u77e9\u9635\uff0cV\u4e3a\u8bcd\u8868\u5927\u5c0f\u00a0 \u00a0 <img decoding=\"async\" class=\"wp-image-4747 alignnone\" src=\"http:\/\/www.puchao.com\/wp-content\/uploads\/edf040018b92abe3ebbdf4c9028980cb-1.png\" alt=\"\" width=\"140\" height=\"49\" srcset=\"http:\/\/www.puchao.com\/wp-content\/uploads\/edf040018b92abe3ebbdf4c9028980cb-1-200x70.png 200w, http:\/\/www.puchao.com\/wp-content\/uploads\/edf040018b92abe3ebbdf4c9028980cb-1.png 214w\" sizes=\"(max-width: 140px) 100vw, 140px\" \/><\/li>\n<\/ul>\n<ul>\n<li>\u5bf9\u4e8e\u6bcf\u4e2a<em>Wi<\/em>\uff0c\u4ece<em>WE<\/em>\u4e2d\u53d6\u51fa\u5bf9\u5e94\u7684\u7b2ci\u884c\u5411\u91cf<\/li>\n<li>\u901a\u8fc7\u8bcd\u5d4c\u5165\u77e9\u9635\u53d8\u6362\u540e\uff0c<em>X=Xembed<\/em><\/li>\n<\/ul>\n<p><strong>(2)\u4f4d\u7f6e\u7f16\u7801\uff08Positional Encoding\uff09<\/strong><\/p>\n<p>\u56e0\u4e3aTransformer\u5185\u90e8\u6ca1\u6709RNN\u7ed3\u6784\uff0c\u5b83\u5e76\u884c\u5904\u7406\u6240\u6709\u8bcd\uff0c\u5982\u679c\u4e0d\u52a0\u4f4d\u7f6e\u7f16\u7801\uff0c\u6a21\u578b\u65e0\u6cd5\u7406\u89e3\u8bed\u5e8f\u3002\u6240\u4ee5\u5728\u9884\u5904\u7406\u65f6\uff0c\u9700\u8981\u52a0\u5165\u6bcf\u4e2a\u8bcd\u7684\u4f4d\u7f6e\u4fe1\u606f\u3002\u5177\u4f53\u64cd\u4f5c\u4e0a\uff0cTransformer\u4f1a\u5c06\u4e00\u6574\u53e5\u8bdd\u4e2d\u6bcf\u4e2a\u8bcd\u7684\u4f4d\u7f6e\u4fe1\u606f\u901a\u8fc7\u6b63\u5f26\u6216\u4f59\u5f26\u51fd\u6570\u7f16\u7801\u6210\u4e0eX\u7ef4\u5ea6\u4e00\u81f4\u7684\u5411\u91cf\uff0c\u76f4\u63a5\u52a0\u5230<em>Xembed<\/em>\u4e2d\u3002\u8fd9\u6837\u53d8\u6362\u4ee5\u540e\uff0c\u6bcf\u4e2a\u8bcd\u90fd\u5e26\u6709\u4e86\u5176\u4ed6\u8bcd\u7684\u4f4d\u7f6e\u4fe1\u606f\u3002<\/p>\n<p>\u8ba1\u7b97\u8fc7\u7a0b\uff1a<\/p>\n<ul>\n<li>\u4f7f\u7528\u6b63\u5f26\u548c\u4f59\u5f26\u51fd\u6570\u751f\u6210\u4f4d\u7f6e\u77e9\u9635PE\uff1a<img decoding=\"async\" class=\"wp-image-4748 alignnone\" src=\"http:\/\/www.puchao.com\/wp-content\/uploads\/0246a63f566558ef11faccaa86501f84.png\" alt=\"\" width=\"248\" height=\"57\" srcset=\"http:\/\/www.puchao.com\/wp-content\/uploads\/0246a63f566558ef11faccaa86501f84-200x46.png 200w, http:\/\/www.puchao.com\/wp-content\/uploads\/0246a63f566558ef11faccaa86501f84-300x69.png 300w, http:\/\/www.puchao.com\/wp-content\/uploads\/0246a63f566558ef11faccaa86501f84-400x92.png 400w, http:\/\/www.puchao.com\/wp-content\/uploads\/0246a63f566558ef11faccaa86501f84.png 500w\" sizes=\"(max-width: 248px) 100vw, 248px\" \/><\/li>\n<li>\u76f8\u52a0\uff1a<em>Xinput=Xembed+PE<\/em><\/li>\n<\/ul>\n<p><strong><span style=\"color: #808080;\">\u7b2c\u4e8c\u9636\u6bb5\uff1a\u7f16\u7801\u5668\uff08The Encoder\uff09<\/span><\/strong><\/p>\n<p>Nx\u6846\u5185\u90e8\u7684\u7f16\u7801\u5668\u6d41\u7a0b\u4f1a\u91cd\u590dx\u6b21\uff0c\u4ee5\u4e0b\u662f\u5355\u6b21\u7684\u5904\u7406\u6d41\u7a0b\uff1a<\/p>\n<p><strong>(3)\u591a\u5934\u81ea\u6ce8\u610f\u529b\u673a\u5236\uff08Multi-Head Self-Attention\uff09<\/strong><\/p>\n<p><em><strong>\u8fd9\u662fTransformer\u7684\u6838\u5fc3\u673a\u5236\u3002<\/strong><\/em><\/p>\n<p><em>Xinput<\/em>\u00a0\u5206\u522b\u4e0e\u4e09\u4e2a\u6743\u91cd\u77e9\u9635<em>WQ\u3001WK\u3001WV\u00a0<\/em>\u76f8\u4e58\uff0c\u5f97\u5230<em>Q\u3001K\u3001V<\/em>\u77e9\u9635\u3002<em>Q<\/em>\u4e0e<em>KT<\/em>\u76f8\u4e58\uff0c\u5f97\u5230<strong>\u6ce8\u610f\u529b\u6743\u91cd\u77e9\u9635<em>Score<\/em><\/strong>\uff0c\u77e9\u9635\u4e2d\u7684\u6bcf\u4e00\u4e2a\u6570\u5b57\u4ee3\u8868\u4e86\u4e24\u4e2a\u8bcd\u4e4b\u95f4\u7684\u76f8\u5173\u6027\u3002\u8ba1\u7b97\u5e8f\u5217\u4e2d\u6bcf\u4e00\u4e2a\u8bcd\u4e4b\u95f4\u7684\u5173\u8054\u5ea6\uff0c\u53ef\u4ee5\u89e3\u51b3\u4e00\u8bcd\u591a\u4e49\u95ee\u9898\u3002\u5982bank\u4ee3\u8868\u94f6\u884c\u8fd8\u662f\u6cb3\u5cb8\uff0c\u53d6\u51b3\u4e8e\u5b83\u548c\u5468\u56f4\u54ea\u4e2a\u8bcd\u5728\u6570\u5b66\u610f\u4e49\u4e0a\u5173\u8054\u66f4\u7d27\u5bc6\u3002<\/p>\n<p><img decoding=\"async\" class=\"aligncenter size-full wp-image-4749\" src=\"http:\/\/www.puchao.com\/wp-content\/uploads\/27ac518a05b4bb39b6123427d880a20f.png\" alt=\"\" width=\"153\" height=\"303\" srcset=\"http:\/\/www.puchao.com\/wp-content\/uploads\/27ac518a05b4bb39b6123427d880a20f-151x300.png 151w, http:\/\/www.puchao.com\/wp-content\/uploads\/27ac518a05b4bb39b6123427d880a20f.png 153w\" sizes=\"(max-width: 153px) 100vw, 153px\" \/><\/p>\n<p style=\"text-align: center;\"><span style=\"color: #808080;\">\u7f29\u653e\u70b9\u79ef\u6ce8\u610f\u529bScaled Dot-Product Attention<\/span><\/p>\n<p>\u591a\u5934\uff08Multi-Head\uff09\u5c31\u662f\u8ba9<em>Q\u3001K\u3001V<\/em>\u77e9\u9635\uff0c\u4e58\u4ee5\u4e0d\u540c\u7684\u6743\u91cd\u77e9\u9635W\uff0c\u901a\u8fc7\u7ebf\u6027\u53d8\u6362\u540e\uff0c\u518d\u8fd0\u884c\u5355\u4e2a\u7684\u6ce8\u610f\u529b\u8fd0\u7b97\uff0c\u6700\u540e\u5c06\u591a\u4e2a\u5934\u4ea7\u751f\u7684\u7ed3\u679c\u8fdb\u884c\u62fc\u63a5\u878d\u5408\u3002\u5176\u610f\u4e49\u5728\u4e8e\uff0c\u4e0d\u540c\u7684\u6743\u91cd\u77e9\u9635W\u53ef\u4ee5\u5bf9\u5e94\u8bed\u6cd5\u7684\u4e0d\u540c\u65b9\u9762\uff0c\u6bd4\u5982\u4e00\u4e2a\u5934\u5173\u6ce8\u8bed\u4e49\u6307\u4ee3\uff0c\u4e00\u4e2a\u5934\u5173\u6ce8\u52a8\u5bbe\u8bed\u6cd5\u7b49\u3002\u4e3e\u4e2a\u4f8b\u5b50\uff0c\u5728\u5904\u7406&#8221;The animal didn&#8217;t cross the street because it was too tired&#8221;\u8fd9\u53e5\u8bdd\u65f6\uff0c\u81ea\u6ce8\u610f\u529b\u673a\u5236\u4f1a\u8ba9&#8221;it&#8221; \u8fd9\u4e2a\u8bcd\u4e0e &#8220;animal&#8221;\u4ea7\u751f\u6781\u9ad8\u7684\u5173\u8054\u5ea6\uff0c\u4ece\u800c\u201c\u7406\u89e3\u201dit\u6307\u7684\u662f\u52a8\u7269\u800c\u4e0d\u662f\u8857\u9053\u3002\u8bba\u6587\u91cc\u662f8\u5934\uff0c\u610f\u5473\u7740\u5b83\u4f1a\u751f\u62108\u7ec4\u4e0d\u540c\u7684<em>Q\u3001K\u3001V<\/em>\u00a0\u77e9\u9635\u3002<\/p>\n<p><img decoding=\"async\" class=\"aligncenter size-full wp-image-4750\" src=\"http:\/\/www.puchao.com\/wp-content\/uploads\/b55f8f0e272f3c93aa8a62b647cc2b2b.png\" alt=\"\" width=\"230\" height=\"312\" srcset=\"http:\/\/www.puchao.com\/wp-content\/uploads\/b55f8f0e272f3c93aa8a62b647cc2b2b-200x271.png 200w, http:\/\/www.puchao.com\/wp-content\/uploads\/b55f8f0e272f3c93aa8a62b647cc2b2b-221x300.png 221w, http:\/\/www.puchao.com\/wp-content\/uploads\/b55f8f0e272f3c93aa8a62b647cc2b2b.png 230w\" sizes=\"(max-width: 230px) 100vw, 230px\" \/><\/p>\n<p style=\"text-align: center;\"><span style=\"color: #808080;\">\u591a\u5934\u6ce8\u610f\u529bMulti-Head Attention<\/span><\/p>\n<p>\u4ee5\u8bba\u6587\u76848\u5934\u4e3a\u4f8b\uff0c\u5177\u4f53\u8ba1\u7b97\u8fc7\u7a0b\uff1a<\/p>\n<ul>\n<li>\u7ebf\u6027\u53d8\u6362\uff1a\u5bf9\u4e8e8\u4e2a\u5934\u4e2d\u7684\u6bcf\u4e00\u4e2a\u5934i(i=1,2,\u2026,8)\uff0c\u90fd\u6709\u5bf9\u5e94\u7684\u6743\u91cd\u77e9\u9635<em>WiQ\u3001WiK\u3001WiV<\/em>\uff0c\u7ef4\u5ea6\u4e3a<em>dmodel\u00d7dk<\/em>\uff08\u5176\u4e2ddk=dmodel\/8\uff09\uff0c\u5c06<em>Xinput<\/em>\u5206\u522b\u4e58\u4ee5\u4e09\u4e2a\u6743\u91cd\u77e9\u9635\uff1a<img decoding=\"async\" class=\"size-full wp-image-4751 alignnone\" src=\"http:\/\/www.puchao.com\/wp-content\/uploads\/f67413012216ff2a6f71aafcfe6fdad5.png\" alt=\"\" width=\"185\" height=\"181\" srcset=\"http:\/\/www.puchao.com\/wp-content\/uploads\/f67413012216ff2a6f71aafcfe6fdad5-66x66.png 66w, http:\/\/www.puchao.com\/wp-content\/uploads\/f67413012216ff2a6f71aafcfe6fdad5.png 185w\" sizes=\"(max-width: 185px) 100vw, 185px\" \/><\/li>\n<li>\u8ba1\u7b97\u5355\u4e2a\u5934\u7684\u6ce8\u610f\u529b\u6743\u91cd\u77e9\u9635\uff1a<em>Score=Qi\u00a0\u00b7 KiT<\/em><\/li>\n<li>\u5bf9\u6ce8\u610f\u529b\u6743\u91cd\u8fdb\u884c\u7f29\u653e\uff0c\u9632\u6b62\u70b9\u79ef\u8fc7\u5927\u5bfc\u81f4Softmax\u8fdb\u5165\u9971\u548c\u533a\uff0c\u4f7f\u5f97\u68af\u5ea6\u8d8b\u8fd1\u4e8e0\u00a0 <img decoding=\"async\" class=\" wp-image-4752 alignnone\" src=\"http:\/\/www.puchao.com\/wp-content\/uploads\/c45baface41113351bf3f12d7e7335c0.png\" alt=\"\" width=\"212\" height=\"42\" srcset=\"http:\/\/www.puchao.com\/wp-content\/uploads\/c45baface41113351bf3f12d7e7335c0-200x40.png 200w, http:\/\/www.puchao.com\/wp-content\/uploads\/c45baface41113351bf3f12d7e7335c0.png 278w\" sizes=\"(max-width: 212px) 100vw, 212px\" \/><\/li>\n<li>\u8f6c\u5316\u4e3a\u6982\u7387\u5206\u5e03\uff08\u6bcf\u4e00\u884c\u4e4b\u548c\u90fd\u4e3a1\uff09\uff1a<em>Attention_Weights=Softmax(Score)<\/em>\u3002\u7b2ci\u884c\u3001j\u5217\u7684\u6570\u5b57\u5927\u5c0f\u8868\u793a\u4e86\u7b2ci\u4e2a\u8bcd\u548c\u7b2cj\u4e2a\u8bcd\u7684\u5173\u8054\u5ea6<img decoding=\"async\" class=\" wp-image-4753 alignnone\" src=\"http:\/\/www.puchao.com\/wp-content\/uploads\/0da5f0c920feb7bb20ffbf2c32cbfe85.png\" alt=\"\" width=\"280\" height=\"108\" srcset=\"http:\/\/www.puchao.com\/wp-content\/uploads\/0da5f0c920feb7bb20ffbf2c32cbfe85-200x77.png 200w, http:\/\/www.puchao.com\/wp-content\/uploads\/0da5f0c920feb7bb20ffbf2c32cbfe85-300x116.png 300w, http:\/\/www.puchao.com\/wp-content\/uploads\/0da5f0c920feb7bb20ffbf2c32cbfe85-400x154.png 400w, http:\/\/www.puchao.com\/wp-content\/uploads\/0da5f0c920feb7bb20ffbf2c32cbfe85.png 415w\" sizes=\"(max-width: 280px) 100vw, 280px\" \/><\/li>\n<li>\u52a0\u6743\u6c42\u548c,\u7528\u6743\u91cd\u77e9\u9635\u4e58Vi\uff1a<em>Headi=Attention_Weights \u00b7 Vi<\/em><\/li>\n<li>\u62fc\u63a5\u4e0e\u878d\u5408\uff1a\u5c068\u4e2a\u5934\u7684\u8f93\u51fa\u62fc\u63a5\uff0c\u518d\u7ecf\u8fc7\u4e00\u4e2a\u6743\u91cd\u77e9\u9635WO\u878d\u5408\uff1a<\/li>\n<\/ul>\n<p style=\"padding-left: 40px;\"><em>MultiHead=Concat(Head1,\u2026,Head8\uff09<\/em><br \/>\n<em>Xoutput=MultiHead \u00b7 Wo<\/em><\/p>\n<p>\u4ee5\u4e0a\u8ba1\u7b97\u8fc7\u7a0b\u53ef\u4ee5\u7efc\u5408\u4e3a\u4e00\u4e2aTransformer\u7ecf\u5178\u516c\u5f0f\uff1a<\/p>\n<p><img decoding=\"async\" class=\"aligncenter wp-image-4754\" src=\"http:\/\/www.puchao.com\/wp-content\/uploads\/87391f769aa72748adf78bf72ebba7f1.png\" alt=\"\" width=\"291\" height=\"57\" srcset=\"http:\/\/www.puchao.com\/wp-content\/uploads\/87391f769aa72748adf78bf72ebba7f1-200x39.png 200w, http:\/\/www.puchao.com\/wp-content\/uploads\/87391f769aa72748adf78bf72ebba7f1-300x59.png 300w, http:\/\/www.puchao.com\/wp-content\/uploads\/87391f769aa72748adf78bf72ebba7f1-400x78.png 400w, http:\/\/www.puchao.com\/wp-content\/uploads\/87391f769aa72748adf78bf72ebba7f1.png 594w\" sizes=\"(max-width: 291px) 100vw, 291px\" \/><\/p>\n<p><strong>(4)\u6b8b\u5dee\u8fde\u63a5\u4e0e\u5f52\u4e00\u5316\uff08Add &amp; Norm\uff09<\/strong><\/p>\n<p>\u5728\u4f20\u7edf\u7684RNN\u91cc\uff0c\u68af\u5ea6\u5728\u5c42\u5c42\u4f20\u9012\u4e2d\u4f1a\u9010\u6e10\u4e22\u5931\u3002Add\u5c31\u662f\u4e3a\u4e86\u89e3\u51b3\u68af\u5ea6\u6d88\u5931\u7684\u95ee\u9898\u3002\u5177\u4f53\u662f\u600e\u4e48\u505a\u5230\u7684\uff1f\u5047\u8bbe\u5728Add\u673a\u5236\u5f15\u5165\u4e4b\u524d\uff0cAttention\u7684\u8f93\u51fa\u662fMultiHeadAttention(X)\uff08\u7b80\u79f0M(X))\uff1bAdd\u673a\u5236\u7684\u52a0\u5165\uff0c\u4f1a\u5c06Attention\u7684\u8f93\u5165X\u548c\u8f93\u51faM(X)\u76f8\u52a0\uff0c\u7136\u540e\u6807\u51c6\u5316\u3002\u5982\u679c\u5728\u8fd9\u4e00\u5c42\u6ca1\u5b66\u5230\u4ec0\u4e48\u6709\u7528\u7684\u4e1c\u897f\uff0cM(X)\u8d8b\u8fd1\u4e8e0\uff0c\u90a3\u4e48\u539f\u59cb\u8f93\u5165X\u4f9d\u7136\u80fd\u6beb\u53d1\u65e0\u635f\u5730\u4f20\u5230\u4e0b\u4e00\u5c42\uff0c\u5728\u8bad\u7ec3\u56de\u4f20\u8bef\u5dee\u68af\u5ea6\u65f6\uff0cAdd\u4e5f\u53ef\u4ee5\u786e\u4fdd\u68af\u5ea6\u53ef\u4ee5\u76f4\u63a5\u8df3\u56de\u524d\u9762\u7684\u5c42\uff0c<strong>\u89e3\u51b3\u4e86\u68af\u5ea6\u6d88\u5931\u7684\u95ee\u9898<\/strong>\u3002<\/p>\n<p>\u518d\u8bf4Norm\u3002\u5728\u6a21\u578b\u5185\u90e8\uff0c\u6bcf\u4e00\u4e2a\u795e\u7ecf\u5143\u8f93\u51fa\u7684\u7edd\u5bf9\u503c\u53ef\u80fd\u4f1a\u6709\u5929\u58e4\u4e4b\u522b\uff0c\u6709\u7684\u7279\u522b\u5927\uff0c\u6709\u7684\u7279\u522b\u5c0f\u3002\u8fd9\u79cd\u6ce2\u52a8\u4f1a\u8ba9\u4e0b\u4e00\u5c42\u7684\u5b66\u4e60\u53d8\u5f97\u975e\u5e38\u5403\u529b\uff1a\u6ce8\u610f\u529b\u673a\u5236\u5929\u7136\u4f1a\u8ba9\u6a21\u578b\u66f4\u5173\u6ce8\u5927\u7684\u6570\u503c\uff0c\u5ffd\u7565\u5c0f\u7684\u6570\u503c\uff0c\u6216\u8005\u67d0\u4e00\u4e2a\u795e\u7ecf\u5143\u7684\u6570\u5b57\u9010\u6e10\u5927\u5230\u8ba1\u7b97\u673a\u7b97\u4e0d\u51fa\u6765\u3002Norm\u5c31\u662f\u5bf9\u8fd9\u4e9b\u6570\u5b57\u505a\u4e00\u4e9b\u6574\u7406\u3002\u5c06\u795e\u7ecf\u5143\u7684\u6570\u503c\u91cd\u7f6e\uff0c\u4f7f\u5176\u6574\u4f53\u7684<strong>\u5747\u503c\u4e3a0\u3001\u65b9\u5dee\u4e3a1<\/strong>\uff0c\u4f46\u662f\u4fdd\u7559\u4e86\u6240\u6709\u795e\u7ecf\u5143\u4e4b\u95f4\u7684\u533a\u522b\u7279\u5f81\uff0c\u4fdd\u8bc1\u4fe1\u606f\u65e0\u635f\u3002Norm\u53ef\u4ee5\u8ba9\u6a21\u578b\u4e0d\u9700\u8981\u56e0\u4e3a\u8f93\u5165\u6570\u636e\u7684\u5267\u70c8\u6ce2\u52a8\u800c\u53cd\u590d\u8c03\u6574\uff0c\u8bad\u7ec3\u901f\u5ea6\u66f4\u5feb\u3002<\/p>\n<p><strong>(5)\u524d\u9988\u795e\u7ecf\u7f51\u7edc\uff08Feed-Forward Networks\uff0cFNN\uff09<\/strong><\/p>\n<p>FNN\u662f\u4eba\u5de5\u795e\u7ecf\u7f51\u7edc\u4e2d\u6700\u57fa\u7840\u3001\u4e5f\u662f\u6700\u65e9\u671f\u7684\u67b6\u6784\uff0c\u672c\u8d28\u4e0a\u662f\u4e00\u4e2a\u591a\u5c42\u51fd\u6570\u5d4c\u5957\u5728\u4e00\u8d77\u6784\u6210\u7684\u6a21\u578b\uff0c\u8f93\u5165\u7684\u6570\u636e\u4f1a\u9010\u5c42\u6620\u5c04\uff0c\u6700\u7ec8\u8f93\u51fa\u3002\u7406\u8bba\u4e0a\u53ea\u8981\u6709\u8db3\u591f\u7684\u9690\u85cf\u5c42\u548c\u8282\u70b9\uff0cFNN\u53ef\u4ee5\u62df\u5408\u4efb\u4f55\u590d\u6742\u7684\u7ebf\u6027\u6216\u975e\u7ebf\u6027\u51fd\u6570\u5173\u7cfb\uff0c\u4f46\u5b83\u4e0d\u8003\u8651\u4e0a\u4e0b\u6587\u6570\u636e\u7684\u5173\u8054\u3002<\/p>\n<p>\u505a\u4e2a\u4e0d\u90a3\u4e48\u6070\u5f53\u4f46\u662f\u53ef\u4ee5\u5e2e\u52a9\u7406\u89e3\u7684\u6bd4\u55bb\uff1aFNN\u50cf\u4e00\u4e2a\u786c\u76d8\uff0c\u8d1f\u8d23\u5b58\u50a8\u6a21\u578b\u5728\u8bad\u7ec3\u6570\u636e\u4e2d\u5b66\u5230\u7684\u9759\u6001\u4e8b\u5b9e\uff1bAttention\u5219\u50cf\u4e00\u4e2aCPU\uff0c\u8d1f\u8d23\u5904\u7406\u5f53\u524d\u7684\u52a8\u6001\u4e0a\u4e0b\u6587\uff0c\u628a\u76f8\u5173\u7684\u8bcd\u62c9\u5230\u4e00\u8d77\u3002\u6ca1\u6709Attention\uff0c\u6a21\u578b\u65e0\u6cd5\u7406\u89e3\u4e0a\u4e0b\u6587\uff1b\u6ca1\u6709FNN\uff0c\u6a21\u578b\u7f3a\u4e4f\u8db3\u591f\u7684\u975e\u7ebf\u6027\u53d8\u6362\u80fd\u529b\u6765\u5904\u7406\u590d\u6742\u7684\u6587\u672c\u7279\u5f81\u3002<\/p>\n<p>\u5177\u4f53\u8ba1\u7b97\u8fc7\u7a0b\uff1a<\/p>\n<ul>\n<li>\u591a\u5c42\u7ebf\u6027\u53d8\u6362+1\u4e2aReLU\u6fc0\u6d3b\u51fd\u6570\uff1aFFN(x\uff09=max(0,xW1+b1)W2+b2<\/li>\n<li>lFNN\u4e4b\u540e\u4f1a\u518d\u91cd\u590d\u505a\u4e00\u6b21Add &amp; Norm<\/li>\n<\/ul>\n<p><strong>\u4ee5\u4e0a\u6b65\u9aa4\u91cd\u590dX\u6b21\uff0c\u7f16\u7801\u5668\u4f1a\u8f93\u51fa\u4e00\u4e2a\u5305\u542b\u6df1\u5ea6\u4e0a\u4e0b\u6587\u4fe1\u606f\u7684\u77e9\u9635\u3002\u8fd9\u7ec4\u77e9\u9635\u5c06\u4f5c\u4e3aKey\u548cValue\u4f20\u7ed9Decoder\u4f7f\u7528\u3002<\/strong><\/p>\n<p><strong><span style=\"color: #808080;\">\u7b2c\u4e09\u9636\u6bb5\uff1a\u89e3\u7801\u5668(The Decoder)<\/span><\/strong><\/p>\n<p>\u8fd9\u4e00\u90e8\u5206\u53ea\u505a\u7b80\u5355\u8bf4\u660e\u3002<\/p>\n<p>\u9996\u5148\u662f\u5df2\u51c6\u5907\u597d\u7684Outputs\u88ab\u9884\u5904\u7406\uff0c\u9884\u5904\u7406\u7684\u65b9\u5f0f\u4e0eEncoder\u9636\u6bb5\u7c7b\u4f3c\u3002\u5904\u7406\u5b8c\u540e\uff0c\u628a\u5b83\u5582\u7ed9Decoder\u3002\u5f00\u59cb\u4ee5\u4e0b\u6b65\u9aa4\uff1a<\/p>\n<p><strong>(6)\u63a9\u7801\u591a\u5934\u6ce8\u610f\u529b\u673a\u5236\uff08Masked Multi-Head Attention\uff09<\/strong><\/p>\n<p>\u548cEncoder\u7684\u591a\u5934\u81ea\u6ce8\u610f\u529b\u673a\u5236\u7c7b\u4f3c\uff0c\u4e0d\u540c\u7684\u662f\u4f46\u52a0\u4e86\u4e00\u4e2aMask\uff08\u63a9\u7801\uff09\u3002\u5f53\u6a21\u578b\u5728\u9884\u6d4b\u7b2ci\u4e2a\u8bcd\u65f6\uff0c\u5b83\u4e0d\u80fd\u770b\u5230\u7b2ci+n\u4e2a\u8bcd\uff0c\u9632\u6b62\u53cd\u5411\u4f20\u64ad\u5931\u6548\u3002<\/p>\n<p><strong>(7)\u4ea4\u53c9\u6ce8\u610f\u529b\u673a\u5236\uff08Encoder-Decoder Attention\uff09<\/strong><\/p>\n<p>\u4e0eEncoder\u9636\u6bb5\u7684\u591a\u5934\u81ea\u6ce8\u610f\u529b\u673a\u5236\u7c7b\u4f3c\uff0c\u8fd9\u91cc\u4e5f\u6709<em>Q\u3001K\u3001V<\/em>\u77e9\u9635\uff0c\u4f46\u5b83\u4eec\u7684<strong>\u6765\u6e90\u4e0d\u540c<\/strong>\u3002\u4ee5\u7ffb\u8bd1\u5de5\u4f5c\u4e3a\u4f8b\uff1aQ\u6765\u81eaDecoder\u4e0a\u4e00\u5c42\u7684\u8f93\u51fa\uff0c\u662f\u5df2\u7ecf\u5b8c\u6210\u7ffb\u8bd1\u540e\u7684\u53e5\u5b50\u5e8f\u5217\uff0c\u4e5f\u5c31\u662f\u5582\u7ed9\u6a21\u578b\u7684\u7b54\u6848\uff1b<em>K<\/em>\u548c<em>V<\/em>\u6765\u81eaEncoder\u7684\u6700\u7ec8\u8f93\u51fa\uff0c\u662f\u7ffb\u8bd1\u524d\u7684\u539f\u59cb\u53e5\u5b50\u5e8f\u5217\u3002\u7136\u540e\u8fd0\u884c\u6bd4\u8f83\u72ec\u7279\u7684\u4ea4\u53c9\u6ce8\u610f\u529b\u673a\u5236\uff1aDecoder\u7528\u5b83\u7684<em>Q<\/em>\u53bb\u626b\u63cfEncoder\u7684<em>K<\/em>\uff0c\u627e\u5230\u5339\u914d\u5ea6\u6700\u9ad8\u7684\u90e8\u5206\uff0c\u53d6\u51fa\u5bf9\u5e94\u7684<em>V<\/em>\u3002<\/p>\n<p><strong>\u901a\u8fc7\u8fd9\u4e2a\u8fc7\u7a0b\uff0cDecoder\u5f97\u5230\u4e86\u4e00\u4e2a\u878d\u5408\u4e86\u6e90\u53e5\u76f8\u5173\u4fe1\u606f\u7684\u5411\u91cf\u3002<\/strong>\u5177\u4f53\u8ba1\u7b97\u8fc7\u7a0b\uff1a<\/p>\n<ul>\n<li>\u5148\u7ecf\u8fc7\u5404\u81ea\u7684\u7ebf\u6027\u5c42\u8fdb\u884c\u7279\u5f81\u8f6c\u6362\uff1a<img decoding=\"async\" class=\" wp-image-4755 alignnone\" src=\"http:\/\/www.puchao.com\/wp-content\/uploads\/f37b1846a1306420c3db5e710bf873a9.png\" alt=\"\" width=\"216\" height=\"94\" srcset=\"http:\/\/www.puchao.com\/wp-content\/uploads\/f37b1846a1306420c3db5e710bf873a9-200x87.png 200w, http:\/\/www.puchao.com\/wp-content\/uploads\/f37b1846a1306420c3db5e710bf873a9-300x131.png 300w, http:\/\/www.puchao.com\/wp-content\/uploads\/f37b1846a1306420c3db5e710bf873a9.png 340w\" sizes=\"(max-width: 216px) 100vw, 216px\" \/><\/li>\n<li>\u8ba1\u7b97\u5355\u4e2a\u5934\u7684\u6ce8\u610f\u529b\u6743\u91cd\u77e9\u9635\uff1aScore=Qi\u00a0\u00b7 KiT<\/li>\n<li>\u5f52\u4e00\u5316\uff1a<img decoding=\"async\" class=\" wp-image-4756 alignnone\" src=\"http:\/\/www.puchao.com\/wp-content\/uploads\/f72f6937fd2694ff884aa57b7a7fa36c.png\" alt=\"\" width=\"259\" height=\"56\" srcset=\"http:\/\/www.puchao.com\/wp-content\/uploads\/f72f6937fd2694ff884aa57b7a7fa36c-200x43.png 200w, http:\/\/www.puchao.com\/wp-content\/uploads\/f72f6937fd2694ff884aa57b7a7fa36c-300x65.png 300w, http:\/\/www.puchao.com\/wp-content\/uploads\/f72f6937fd2694ff884aa57b7a7fa36c.png 347w\" sizes=\"(max-width: 259px) 100vw, 259px\" \/><\/li>\n<li>\u52a0\u6743\u6c42\u548c\uff1a<img decoding=\"async\" class=\" wp-image-4758 alignnone\" src=\"http:\/\/www.puchao.com\/wp-content\/uploads\/17ed0c929c5b7a25d221701e1847b4ee.png\" alt=\"\" width=\"237\" height=\"45\" srcset=\"http:\/\/www.puchao.com\/wp-content\/uploads\/17ed0c929c5b7a25d221701e1847b4ee-200x38.png 200w, http:\/\/www.puchao.com\/wp-content\/uploads\/17ed0c929c5b7a25d221701e1847b4ee-300x57.png 300w, http:\/\/www.puchao.com\/wp-content\/uploads\/17ed0c929c5b7a25d221701e1847b4ee-320x62.png 320w, http:\/\/www.puchao.com\/wp-content\/uploads\/17ed0c929c5b7a25d221701e1847b4ee.png 326w\" sizes=\"(max-width: 237px) 100vw, 237px\" \/><\/li>\n<\/ul>\n<p><strong>(8)\u91cd\u590d FNN \u548c Add &amp; Norm<\/strong><\/p>\n<p><span style=\"color: #808080;\"><strong>\u7b2c\u56db\u9636\u6bb5\uff1a\u8f93\u51fa\u751f\u6210(Final Output)<\/strong><\/span><\/p>\n<p><strong>(9)\u7ebf\u6027\u5c42 (Linear)<\/strong><\/p>\n<p>\u628aDecoder \u6700\u7ec8\u8f93\u51fa\u7684\u5411\u91cf\uff08dmodel\u7ef4\u5ea6\uff09\u6295\u5f71\u5230\u4e00\u4e2a\u5de8\u5927\u7684\u8bcd\u8868\u7ef4\u5ea6\uff0c\u5c06\u62bd\u8c61\u7684\u8bed\u4e49\u5411\u91cf\u8f6c\u5316\u4e3a\u5bf9\u5177\u4f53\u5355\u8bcd\u7684\u6253\u5206\u3002<\/p>\n<p>\u5177\u4f53\u8ba1\u7b97\u8fc7\u7a0b\uff1a<\/p>\n<ul>\n<li><em>Logits=Outputdecoder\u00a0\u00b7 Wvocab\uff08Wvocab<\/em>\u5f62\u72b6\u4e3a<em>dmodel\u00a0\u00d7V\uff09<\/em><\/li>\n<\/ul>\n<p><strong>(10)Softmax\u5c42<\/strong><\/p>\n<p>\u628a\u4e0a\u9762\u7684\u6570\u503c\u8f6c\u5316\u6210\u6982\u7387\u5206\u5e03\u3002\u6982\u7387\u6700\u9ad8\u7684\u90a3\u4e2a\u8bcd\uff0c\u5c31\u662f\u6a21\u578b\u9884\u6d4b\u7684\u4e0b\u4e00\u4e2a\u8bcd\u3002<\/p>\n<ul>\n<li>\u8f93\u51fa\u4e00\u4e2a\u6982\u7387\u5206\u5e03\uff0c\u603b\u548c\u4e3a1<\/li>\n<\/ul>\n<p><img decoding=\"async\" class=\" wp-image-4759 alignnone\" src=\"http:\/\/www.puchao.com\/wp-content\/uploads\/9bf922e0421ab096bec4f3fa2f378ac4.png\" alt=\"\" width=\"176\" height=\"81\" srcset=\"http:\/\/www.puchao.com\/wp-content\/uploads\/9bf922e0421ab096bec4f3fa2f378ac4-200x92.png 200w, http:\/\/www.puchao.com\/wp-content\/uploads\/9bf922e0421ab096bec4f3fa2f378ac4.png 247w\" sizes=\"(max-width: 176px) 100vw, 176px\" \/><\/p>\n<p><strong>\u4ee5\u4e0a\uff0c\u5c31\u662f\u901a\u8fc7Transformer\u505a\u4e00\u6b21\u524d\u5411\u4f20\u64ad\u7684\u5b8c\u6574\u8fc7\u7a0b\u3002<\/strong><\/p>\n<p><strong><span style=\"color: #3366ff;\">\u4e09\u3001\u5173\u4e8eTransformer\u7684\u5ef6\u4f38\u601d\u8003<\/span><\/strong><\/p>\n<p>RNN\u50cf\u662f\u6211\u4eec\u5927\u591a\u6570\u65f6\u5019\u7684\u7ebf\u6027\u601d\u7ef4\u4e60\u60ef\uff0c\u7b80\u5355\u76f4\u63a5\uff0c\u800c\u4e14\u6709\u8ff9\u53ef\u5faa\u3002\u6211\u4eec\u5728\u65f6\u95f4\u7684\u7ebf\u6027\u6d41\u6dcc\u4e2d\uff0c\u57fa\u4e8e\u6628\u5929\u7684\u7ecf\u9a8c\u3001\u4fe1\u606f\u548c\u77e5\u8bc6\uff0c\u6765\u63a8\u5bfc\u4eca\u65e5\u3002\u518d\u57fa\u4e8e\u4eca\u65e5\u53d1\u751f\u7684\u4e8b\u3001\u5bf9\u8fc7\u5f80\u6b63\u786e\u4fe1\u606f\u7684\u9a8c\u771f\uff08\u5956\u52b1\uff09\u4ee5\u53ca\u5bf9\u8fc7\u5f80\u9519\u8bef\u4fe1\u606f\u7684\u52d8\u8bef\uff08\u53cd\u5411\u4f20\u64ad\uff09\uff0c\u6765\u4e0d\u65ad\u5f3a\u5316\u6b63\u786e\u7684\u8ba4\u77e5\u6216\u4fee\u6b63\u9519\u8bef\u7684\u8ba4\u77e5\uff0c\u4ee5\u6b64\u6765\u63a8\u6d4b\u660e\u65e5\u7684\u53ef\u80fd\u6027\uff0c\u5e76\u4e0d\u65ad\u63d0\u5347\u5176\u51c6\u786e\u7387\u3002<\/p>\n<p>\u8fd9\u79cd\u7ebf\u6027\u601d\u7ef4\u5e26\u6765\u4e86\u5f88\u591a\u597d\u5904\uff0c\u4e14\u88ab\u9a8c\u8bc1\u8fc7\uff1a<\/p>\n<ul>\n<li><strong>\u7269\u79cd\u7684\u5ef6\u7eed\uff1a<\/strong>\u65e9\u671f\u7684\u4eba\u7c7b\uff0c\u5728\u7f3a\u4e4f\u53c2\u8003\u8d44\u6599\u7684\u60c5\u51b5\u4e0b\uff0c\u53ea\u9700\u8981\u8bb0\u4f4f\u8fdb\u98df\u67d0\u79cd\u98df\u7269\u7684\u4f19\u4f34\u72b6\u6001\u4e0d\u597d\uff0c\u540e\u7eed\u5c31\u4f1a\u964d\u4f4e\u98df\u7528\u7684\u6982\u7387\u3002\u57fa\u4e8e\u8fc7\u53bb\u7684\u8d1f\u9762\u7ecf\u9a8c\u6765\u9884\u6d4b\u5371\u9669\uff0c\u662f\u7269\u79cd\u5ef6\u7eed\u7684\u4e00\u9879\u91cd\u8981\u6cd5\u5219\u3002<\/li>\n<li><strong>\u5efa\u7acb\u5b89\u5168\u611f\uff1a<\/strong>\u57fa\u4e8e\u71b5\u589e\uff0c\u4e00\u4e2a\u5b64\u7acb\u7cfb\u7edf\u603b\u662f\u8d8b\u5411\u4e8e\u4ece\u6709\u5e8f\u8d70\u5411\u65e0\u5e8f\uff0c\u4e16\u754c\u4e5f\u662f\u4e00\u6837\u3002\u7ebf\u6027\u601d\u7ef4\u901a\u8fc7\u7ecf\u9a8c\u4e3a\u6211\u4eec\u4e34\u65f6\u5efa\u7acb\u4e86\u4e00\u79cd\u53ef\u4ee5\u9884\u6d4b\u7684\u79e9\u5e8f\u611f\u3002\u5982\u679c\u6211\u4eec\u4e0d\u80fd\u786e\u4fe1\uff1a\u592a\u9633\u660e\u5929\u4f1a\u50cf\u6628\u5929\u4e00\u6837\u5347\u8d77\u3001\u6211\u80fd\u7528\u6628\u5929\u8d5a\u6765\u7684\u94b1\u4fdd\u8bc1\u6211\u4eca\u5929\u53ef\u4ee5\u4e00\u65e5\u4e09\u9910\uff0c\u90a3\u611f\u89c9\u4f1a\u5f88\u60f6\u6050\u3002<\/li>\n<\/ul>\n<p>\u57fa\u4e8e\u7ebf\u6027\u601d\u7ef4\uff0c\u6211\u4eec\u5929\u7136\u53ea\u5173\u6ce8\u4e0e\u81ea\u5df1<strong>\u76f8\u4f3c\u7684\u4fe1\u606f<\/strong>\uff0c\u800c\u4e0d\u662f\u6240\u6709<strong>\u6709\u7528\u7684\u4fe1\u606f<\/strong>\u3002\u8fd9\u5728\u793e\u4f1a\u5b66\u4e0a\u88ab\u79f0\u4e3a\u540c\u8d28\u6027\u504f\u597d\u3002\u4f46\u5982\u679c\u4eba\u7c7b\u793e\u4f1a\u5931\u53bb\u4e86\u5bf9\u5176\u4ed6\u5e8f\u5217\u548c\u751f\u547d\u4f53\u7684\u5305\u5bb9\u673a\u5236\uff0c\u5c31\u4f1a\u9000\u5316\u4e3a\u4e00\u4e2a\u4e2a\u4e92\u4e0d\u7406\u89e3\u7684\u5b64\u5c9b\u3002<\/p>\n<p>Transformer\u5176\u5b9e\u5305\u542b\u4e86\u4e00\u79cd\u9690\u55bb\uff0c\u4e2a\u4f53\u3001\u793e\u4f1a\u5982\u679c\u53ef\u4ee5\u5bf9\u63a5\u89e6\u5230\u7684\u6240\u6709\u4fe1\u606f\u8fdb\u884c\u5168\u9762\u7684\u3001\u5e26\u6743\u91cd\u7684\u6ce8\u610f\u529b\u5206\u914d\uff0c\u518d\u7ed3\u5408\u5185\u90e8\u6d88\u5316\u5904\u7406\uff08FNN\uff09\uff0c\u79ef\u7d2f\u5230\u4e00\u5b9a\u53c2\u6570\u89c4\u6a21\u65f6\uff0c\u53ef\u80fd\u65b0\u7684\u6587\u660e\u5f62\u5f0f\u5c06\u4f1a\u6d8c\u73b0\u3002<\/p>\n<p>\u518d\u8bf4\u56de\u5230\u5927\u6a21\u578b\u76ee\u524d\u9762\u4e34\u7684\u9ed1\u7bb1\u95ee\u9898\u3002<\/p>\n<p>\u867d\u7136\u5927\u6a21\u578b\u7684\u9690\u85cf\u5c42\u65e0\u6cd5\u88ab\u7b26\u53f7\u4e3b\u4e49\u89e3\u91ca\uff0c\u5176\u751f\u6210\u80fd\u529b\u76ee\u524d\u8fd8\u662f\u9690\u6027\u4e14\u6982\u7387\u67d4\u6027\uff0c\u4f46\u6211\u8ba4\u4e3a\u5b83\u5728\u4e00\u5b9a\u7a0b\u5ea6\u4e0a\u4e5f\u548c\u4eba\u7c7b\u7684\u5b66\u4e60\u6a21\u5f0f\u5df2\u7ecf\u6709\u76f8\u4f3c\u6027\uff0c<strong>\u4eba\u7c7b\u5927\u8111\u4f55\u5c1d\u4e0d\u662f\u4e00\u79cd\u8fd0\u884c\u673a\u5236\u6700\u590d\u6742\u7684\u9ed1\u7bb1\uff1f<\/strong>\u5728\u4f20\u7edf\u7684\u77e5\u8bc6\u4f20\u9012\u8303\u5f0f\u4e2d\uff0c\u6211\u4eec\u7684\u77e5\u8bc6\u7ed3\u6784\u642d\u5efa\u5728\u5f88\u591a\u57fa\u7840\u516c\u7406\u4e0a\uff0c\u5373\u4f7f\u4f60\u5bf9\u4e16\u754c\u548c\u4e8b\u7269\u7684\u8bde\u751f\u3001\u6f14\u5316\u3001\u8fed\u4ee3\u7684\u6700\u5e95\u5c42\u903b\u8f91\u5e76\u4e0d\u5b8c\u5168\u6e05\u6670\uff0c\u4f46\u4f60\u4ecd\u80fd\u6839\u636e\u5df2\u7ecf\u88ab\u666e\u904d\u8ba4\u53ef\u7684\u516c\u7406\uff0c\u751a\u81f3\u662f\u51ed\u501f\u6570\u6b21\u8bd5\u9519\u5efa\u7acb\u7684\u76f4\u89c9\uff0c\u7ed3\u5408\u5df2\u7ecf\u88ab\u8bc1\u5b9e\u6b63\u786e\u4e8b\u7269\u7684\u5f3a\u5316\uff0c\u5c31\u53ef\u4ee5\u5f97\u51fa\u7ed3\u8bba\uff0c\u800c\u4e14\u8fd9\u4e2a\u7ed3\u8bba\u6700\u7ec8\u4e5f\u7b26\u5408\u6211\u4eec\u89c1\u8bc1\u5230\u7684\u8868\u89c2\u7ed3\u679c\u3002<\/p>\n<h6><span style=\"color: #999999;\">\uff08\u4e2a\u4eba\u89c2\u70b9\uff0c\u4ec5\u4f9b\u53c2\u8003\uff09<\/span><\/h6>\n","protected":false},"excerpt":{"rendered":"<p>2017\u5e746\u6708\uff0cGo  [&#8230;]<\/p>\n","protected":false},"author":1,"featured_media":4761,"comment_status":"closed","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[26],"tags":[],"class_list":["post-4744","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-category-2"],"_links":{"self":[{"href":"http:\/\/www.puchao.com\/index.php\/wp-json\/wp\/v2\/posts\/4744","targetHints":{"allow":["GET"]}}],"collection":[{"href":"http:\/\/www.puchao.com\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"http:\/\/www.puchao.com\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"http:\/\/www.puchao.com\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"http:\/\/www.puchao.com\/index.php\/wp-json\/wp\/v2\/comments?post=4744"}],"version-history":[{"count":2,"href":"http:\/\/www.puchao.com\/index.php\/wp-json\/wp\/v2\/posts\/4744\/revisions"}],"predecessor-version":[{"id":4760,"href":"http:\/\/www.puchao.com\/index.php\/wp-json\/wp\/v2\/posts\/4744\/revisions\/4760"}],"wp:featuredmedia":[{"embeddable":true,"href":"http:\/\/www.puchao.com\/index.php\/wp-json\/wp\/v2\/media\/4761"}],"wp:attachment":[{"href":"http:\/\/www.puchao.com\/index.php\/wp-json\/wp\/v2\/media?parent=4744"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/www.puchao.com\/index.php\/wp-json\/wp\/v2\/categories?post=4744"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/www.puchao.com\/index.php\/wp-json\/wp\/v2\/tags?post=4744"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}