{"id":2639,"date":"2025-05-05T20:02:23","date_gmt":"2025-05-05T12:02:23","guid":{"rendered":"https:\/\/www.lanxy.ink\/?p=2639"},"modified":"2025-05-09T20:18:58","modified_gmt":"2025-05-09T12:18:58","slug":"%e4%bb%8e-0-%e5%bc%80%e5%a7%8b%e7%9a%84-transformer-%e6%a2%b3%e7%90%86%ef%bc%883%ef%bc%89%e6%b3%a8%e6%84%8f%e5%8a%9b%e6%9c%ba%e5%88%b6%e4%b8%8e-transformer-%e5%8e%9f%e7%90%86","status":"publish","type":"post","link":"https:\/\/www.lanxy.ink\/?p=2639","title":{"rendered":"\u4ece 0 \u5f00\u59cb\u7684 Transformer \u68b3\u7406\uff083\uff09\u6ce8\u610f\u529b\u673a\u5236\u4e0e Transformer \u539f\u7406"},"content":{"rendered":"\n

\u666e\u901a\u6ce8\u610f\u529b\u673a\u5236\u7684\u539f\u7406<\/h2>\n\n\n\n

\u4eba\u7684\u6ce8\u610f\u529b\u5f80\u5f80\u96c6\u4e2d\u5728\u89c6\u7ebf\u7684\u6b63\u4e2d\u95f4\uff0c\u4e13\u6ce8\u4e8e\u90a3\u4e00\u5c0f\u90e8\u5206\u91cd\u8981\u7684\u5185\u5bb9\u3002\u6a21\u578b\u600e\u4e48\u53bb\u6ce8\u610f\u5230\u90a3\u4e9b\u91cd\u8981\u7684\u5185\u5bb9\u5462\uff1f\u6211\u4eec\u5582\u7ed9\u6a21\u578b\u7684\u6570\u636e\u662f\u4e00\u4e2a\u4e2a\u77e9\u9635\u6216\u5f20\u91cf\uff0c\u8981\u60f3\u8ba9\u6a21\u578b\u91cd\u70b9\u5173\u6ce8\u5230\u67d0\u4e9b\u4fe1\u606f\uff0c\u4e00\u4e2a\u76f4\u89c2\u7684\u60f3\u6cd5\u5c31\u662f\u8ba9\u8fd9\u4e9b\u5730\u65b9\u7684\u6570\u503c\u53d8\u5f97\u66f4\u5927\u3002\u6ce8\u610f\u529b\u673a\u5236\u5c31\u662f\u8fd9\u6837\u4e00\u4e2a\u8fc7\u7a0b\uff0c\u901a\u8fc7\u52a0\u6743\u6c42\u548c\u7684\u65b9\u6cd5\uff0c\u8d4b\u4e88\u4e0d\u540c\u7684\u8f93\u5165\u5143\u7d20\u4ee5\u4e0d\u540c\u7684\u6743\u91cd<\/strong>\uff0c\u91cd\u8981\u7684\u4fe1\u606f\u6743\u91cd\u5927\uff0c\u91cd\u70b9\u5173\u6ce8\uff0c\u4e0d\u91cd\u8981\u7684\u4fe1\u606f\u6743\u91cd\u5c0f\uff0c\u88ab\u5ffd\u7565\u3002<\/p>\n\n\n\n

\u6ce8\u610f\u529b\u673a\u5236\u91cc\u6709 \u952e Key\u3001\u503c Value \u548c \u67e5\u8be2 Query\uff0c\u76f4\u63a5\u770b\u8fd9\u4e09\u4e2a\u4e1c\u897f\u4f1a\u6bd4\u8f83\u62bd\u8c61\uff0c\u6240\u4ee5\u6211\u7528\u4e00\u4e2a\u4f8b\u5b50\u6765\u6253\u6bd4\u65b9\u3002<\/p>\n\n\n\n

\u5047\u8bbe\u6211\u4eec\u534a\u591c\u56de\u5bb6\uff0c\u5bb6\u91cc\u4e4c\u6f06\u561b\u9ed1\u7684\uff0c\u800c\u6211\u4eec\u624b\u91cc\u6709\u624b\u7535\u7b52<\/strong>\uff0c\u7528\u624b\u7535\u7b52\u626b\u6765\u626b\u53bb\u5730\u627e\u8981\u6362\u7684\u978b<\/strong>\u3002<\/p>\n\n\n\n

\u5bb6\u91cc\u4f1a\u6709\u5f88\u591a\u67dc\u5b50\uff08Key\uff09\uff0c\u67dc\u5b50\u91cc\u4f1a\u653e\u6709\u4e00\u4e9b\u4e1c\u897f\uff08Value\uff09\uff0c\u800c\u6211\u4eec\u8981\u6362\u978b\uff0c\u5c31\u5f97\u53bb\u67e5<\/strong>\u4e00\u4e0b\uff0c\u54ea\u4e2a\u5730\u65b9\u66f4\u6709\u53ef\u80fd\u4f1a\u653e\u7740\u978b\u5462\uff1f\u6211\u4eec\u7684\u978b\u5c31\u662f Query \u4e86\u3002<\/p>\n\n\n\n

\u4f60\u53ef\u80fd\u4f1a\u60f3\uff0c\u978b\u67dc\u5440\uff0c\u56e0\u4e3a\u978b\u4e0d\u90fd\u662f\u653e\u978b\u67dc\u91cc\u7684\u561b\uff01\u5f53\u6211\u4eec\u5728\u8111\u8865\u8fd9\u4e2a\u60f3\u6cd5\u7684\u65f6\u5019\uff0c\u5c31\u5f00\u59cb\u5728\u505a\u7c7b\u4f3c\u4e8e\u201c\u76f8\u4f3c\u5ea6\u5339\u914d\u201d\u7684\u4e8b\u60c5\u4e86\uff0c\u978b\u548c\u978b\u67dc\u4e4b\u95f4<\/strong>\u6bd4\u978b\u548c\u5e8a\u5934\u67dc\u6216\u8005\u5176\u4ed6\u5bb6\u5177\u4e4b\u95f4\u6709\u7740\u66f4\u9ad8\u7684\u5339\u914d\u5ea6\u3002<\/p>\n\n\n\n

\u7136\u540e\u6211\u4eec\u628a\u624b\u7535\u7b52\u6253\u5f00\uff0c\u624b\u7535\u7b52\u628a\u5149\u675f\u805a\u7126\u5230\u6700\u6709\u5e0c\u671b\u7684\u4f4d\u7f6e\u2014\u2014\u624b\u7535\u7b52\u7684\u5149\u675f\u5c31\u76f8\u5f53\u4e8e\u6743\u91cd<\/strong>\uff0c\u4e5f\u5c31\u662f\u6ce8\u00b7\u610f\u00b7\u529b<\/strong>\u3002\u4f60\u4f1a\u628a\u6ce8\u610f\u529b\u96c6\u4e2d\u5728\u624b\u7535\u7b52\u7167\u4eae\u7684\u5730\u65b9\uff08\u6743\u91cd\u5927\uff09\uff0c\u800c\u4e0d\u4f1a\u53bb\u770b\u6697\u7684\u5730\u65b9\uff08\u6743\u91cd\u5c0f\uff09\u3002<\/p>\n\n\n\n

\u6700\u540e\u6211\u4eec\u7167\u4eae\u5e76\u6253\u5f00\u4e86\u978b\u67dc\uff0c\u5728\u91cc\u9762\u53bb\u63d0\u53d6\u6211\u4eec\u771f\u6b63\u9700\u8981\u7684 Value \uff08\u978b\uff09\uff0c\u8fd9\u5c31\u662f\u6ce8\u610f\u529b\u673a\u5236\u7684\u539f\u7406\u3002<\/p>\n\n\n\n

\n

Attention \u673a\u5236\u6700\u65e9\u662f\u5728\u8ba1\u7b97\u673a\u89c6\u89c9\u91cc\u5e94\u7528\u7684\uff0c\u968f\u540e\u5728 NLP \u9886\u57df\u4e5f\u5f00\u59cb\u5e94\u7528\u4e86\uff0c\u771f\u6b63\u53d1\u626c\u5149\u5927\u662f\u5728 NLP \u9886\u57df\uff0c\u56e0\u4e3a 2018 \u5e74 BERT<\/a> \u548c GPT \u7684\u6548\u679c\u51fa\u5947\u7684\u597d\uff0c\u8fdb\u800c\u8d70\u7ea2\u3002\u800c Transformer<\/a> \u548c Attention \u8fd9\u4e9b\u6838\u5fc3\u5f00\u59cb\u88ab\u5927\u5bb6\u91cd\u70b9\u5173\u6ce8\u3002[1]<\/p>\n\n\n\n

\u5982\u679c\u7528\u56fe\u6765\u8868\u8fbe Attention \u7684\u4f4d\u7f6e\u5927\u81f4\u662f\u4e0b\u9762\u7684\u6837\u5b50\uff1a<\/p>\n\n\n

\n
\"\"<\/div><\/figure>\n<\/div><\/blockquote>\n\n\n\n

\u6ce8\u610f\u529b\u539f\u7406\u7684\u4e09\u6b65\u5206\u89e3<\/h3>\n\n\n\n

\u5206\u4e3a\u4e09\u4e2a\u9636\u6bb5\uff1a<\/p>\n\n\n

\n
\"\"<\/div><\/figure>\n<\/div>\n\n\n

 \u56e0\u4e3a\u662f\u666e\u901a\u7684\u6ce8\u610f\u529b\u673a\u5236\uff0c\u6240\u4ee5\u5148\u4e0d\u53bb\u7ba1 Query\u3001Key\u3001Value \u662f\u600e\u4e48\u6765\u7684\uff0c\u6211\u4eec\u76f4\u63a5\u7ed9\u5b9a\u3002<\/p>\n\n\n\n

\u7b2c\u4e00\u6b65\uff1a Query \u548c Key \u8fdb\u884c\u76f8\u4f3c\u5ea6\u5f97\u5206\u8ba1\u7b97\uff0c\u5f97\u5230\u6743\u503c
\u7b2c\u4e8c\u6b65\uff1a\u5c06\u6743\u503c\u8fdb\u884c\u5f52\u4e00\u5316\uff0c\u5f97\u5230\u76f4\u63a5\u53ef\u7528\u7684\u6743\u91cd
\u7b2c\u4e09\u6b65\uff1a\u5c06\u6743\u91cd\u548c Value \u8fdb\u884c\u52a0\u6743\u6c42\u548c\uff08\u77e9\u9635\u76f8\u4e58\uff09<\/p>\n\n\n\n

\u4e4b\u524d\u627e\u978b\u7684\u4f8b\u5b50\u91cc\u6211\u4eec\u7684\u6743\u91cd\u5c31\u662f 1 \uff08\u4eae\u7684\u5730\u65b9\uff09\u548c 0\uff08\u6697\u7684\u5730\u65b9\uff09\uff0c\u8fd9\u91cc\u6743\u91cd\u662f 0~1 \u4e4b\u95f4\u7684\u5c0f\u6570\u4e86\uff0c\u6240\u4ee5\u662f\u52a0\u6743\u6c42\u548c\u3002\u8fd9\u4e09\u6b65\u7528\u4e00\u4e2a\u516c\u5f0f\u6765\u8868\u793a\u5c31\u662f\uff1a\\(\\)<\/p>\n\n\n\n

$$ \\mathrm{Attention}(Q,K,V)=f_{\u5f52\u4e00\u5316} \\left( \\mathrm{score} (Q,K) \\right) V $$<\/p>\n\n\n\n

\u6709\u4e86\u6ce8\u610f\u529b\u673a\u5236\u540e\uff0c\u6a21\u578b\u5c31\u4ece\u8fc7\u53bb\u7684\u6b7b\u8bb0\u786c\u80cc\u9636\u6bb5\u8f6c\u4e3a\u4e86\u878d\u4f1a\u8d2f\u901a\u9636\u6bb5\uff0c\u8ba9\u6a21\u578b\u7684\u6027\u80fd\u8fde\u4e0a\u597d\u51e0\u4e2a\u9ad8\u5ea6\u3002<\/p>\n\n\n\n

\u4ee3\u8868\u6027\u7684\u6ce8\u610f\u529b\u673a\u5236\u6a21\u578b<\/h3>\n\n\n\n

\u4e0a\u4e00\u8282\u91cc\u7684 \\(\\mathrm{score}(\\cdot)\\) \u53ea\u662f\u4e00\u4e2a\u7b3c\u7edf\u7684\u5199\u6cd5\uff0c\u56e0\u4e3a\u5b83\u672c\u8eab\u6709\u5f88\u591a\u79cd\u53d8\u4f53\uff0c\u5177\u6709\u4ee3\u8868\u6027\u7684\u76f8\u4f3c\u5ea6\u5f97\u5206\u8ba1\u7b97\u65b9\u5f0f\u6709\u5982\u4e0b\u51e0\u79cd\u3002<\/p>\n\n\n\n

Bahdanau Attention\uff08\u52a0\u6027\u6ce8\u610f\u529b\uff09<\/h4>\n\n\n\n

$$ \\mathrm{score} (s_{t-1},h_i) = v_{a}^{T} \\mathrm{Tanh} \\left( W_as_{t-1} + U_ah_i \\right) $$<\/p>\n\n\n\n

\u8fd9\u4e2a\u6ce8\u610f\u529b\u91cc\u7528\u4e86\u4e00\u5c42\u524d\u9988\u795e\u7ecf\u7f51\u7edc\uff0c\u800c\u4e14\u8be5\u6ce8\u610f\u529b\u9700\u8981\u7ed3\u5408 encoder \u548c decoder \u8003\u8651\uff0c\u63a5\u53d7\u4e0a\u4e00\u4e2a decoder \u7684\u8f93\u51fa \\( s_{t-1} \\) \u548c\u5f53\u524d encoder \u7684\u8f93\u51fa \\( h_{t} \\)\uff0c\u6fc0\u6d3b\u51fd\u6570\u662f Tanh\uff0c\u518d\u4e0e\u53ef\u5b66\u4e60\u7684\u5411\u91cf \\(v\\) \u505a\u70b9\u4e58\u3002<\/p>\n\n\n\n

Luong Attention\uff08\u70b9\u79ef\u6ce8\u610f\u529b\uff09<\/h4>\n\n\n\n

\u76f4\u63a5\u70b9\u79ef\uff1a<\/p>\n\n\n\n

$$ \\mathrm{score} (q,k) = q^Tk $$<\/p>\n\n\n\n

\u5e26\u7f29\u653e\u7684\u70b9\u79ef\uff1a<\/p>\n\n\n\n

$$ \\mathrm{score} (q,k) =\\frac {qk^T} {\\sqrt{d_k}} $$<\/p>\n\n\n\n

\u7f29\u653e\u70b9\u79ef\u662f Transformer \u91c7\u7528\u7684\u6ce8\u610f\u529b\uff0c\u91cc\u9762\u7684 \\(d_k\\) \u4e3a \\(K\\) \u7684\u7ef4\u5ea6\uff0c\u76f8\u5f53\u4e8e\u5f52\u4e00\u5316\u64cd\u4f5c\uff0c\u9632\u6b62\u68af\u5ea6\u6d88\u5931\u6216\u7206\u70b8\u3002<\/p>\n\n\n

\n
\"\"<\/div><\/figure>\n<\/div>\n\n\n

\u5e26\u7f29\u653e\u7684\u70b9\u79ef\u6ce8\u610f\u529b\uff08Scaled Dot-Product Attention\uff09\u6709\u65f6\u4f1a\u5728\u5c3a\u5ea6\u7f29\u653e\uff08Scale\uff09\u548c\u5f52\u4e00\u5316\uff08Softmax\uff09\u64cd\u4f5c\u4e4b\u95f4\u6dfb\u52a0\u4e00\u4e2a Mask \u64cd\u4f5c\uff0c\u7528\u6765\u5c4f\u853d\u4e00\u90e8\u5206\u6ce8\u610f\u529b\uff0cMask \u5728\u9700\u8981\u5c4f\u853d\u7684\u4f4d\u7f6e\u53d6 -\u221e\uff08\u5f88\u5927\u7684\u8d1f\u6570\uff09\uff0csoftmax\u202f\u4e4b\u540e\u7684\u6743\u91cd\u5c31\u4f1a\u53d8\u4e3a\u202f0\uff0c\u4e0d\u53c2\u4e0e\u52a0\u6743\u6c42\u548c\u3002<\/p>\n\n\n\n

\u5e38\u7528\u7684 Mask \u79cd\u7c7b\u6709\u4e24\u4e2a\uff0cPadding Mask \u548c Look\u2011Ahead\u202fMask\u3002
Padding Mask \u5c06\u4e0d\u540c\u957f\u5ea6\u7684\u5e8f\u5217\u5bf9\u9f50\u4e3a\u76f8\u540c\u957f\u5ea6\uff08\u6bd4\u5982\u4e0d\u540c\u7684\u53e5\u5b50\u6709\u4e0d\u540c\u7684\u957f\u5ea6\uff09\uff1b
Look-Ahead Mask \u5c4f\u853d\u5f53\u524d\u65f6\u95f4\u6b65\u4e4b\u540e\u7684\u4fe1\u606f\uff0c\u4e14\u53ea\u5728 Decoder \u7aef\u8fdb\u884c\uff0c\u6bd4\u5982\u5728\u751f\u6210\u53e5\u5b50\u65f6\u4fdd\u8bc1\u9884\u6d4b\u7b2c \u202ft\u202f \u4e2a\u8bcd\u65f6\u4e0d\u80fd\u5077\u770b t\u202f+\u202f1\u3001t\u202f+\u202f2\u2026 \u7684\u8bcd\uff0c\u53ea\u80fd\u4f9d\u8d56\u8fc7\u53bb\u7684\u8bcd\u3002<\/div><\/div>\n\n\n\n

\u6ce8\u610f\u529b\u673a\u5236\u53d8\u79cd<\/h2>\n\n\n\n

\u8f6f\u6ce8\u610f\u529b\u4e0e\u786c\u6ce8\u610f\u529b\uff08Soft and Hard Attention\uff09<\/h3>\n\n\n\n

\u4e0a\u9762\u627e\u978b\u7684\u4f8b\u5b50\u5373\u786c\u6ce8\u610f\u529b\u7684\u539f\u7406\uff1a\u9009\u62e9\u6ce8\u610f\u529b\u5206\u6570\u6700\u9ad8\u7684 value\uff0c\u800c\u8f6f\u6ce8\u610f\u529b\u662f\u5bf9\u6240\u6709 value \u52a0\u6743\u6c42\u548c\u3002<\/p>\n\n\n\n

\u786c\u6ce8\u610f\u529b\u7684\u597d\u5904\u5c31\u662f\u6bcf\u4e00\u6b65\u90fd\u660e\u786e\u5730\u201c\u6307\u5411\u201d\u67d0\u4e00\u4e2a\u8f93\u5165\u4f4d\u7f6e\uff0c\u5f88\u5bb9\u6613\u53ef\u89c6\u5316\u5bf9\u9f50\u7ed3\u679c\uff0c\u4f46\u574f\u5904\u5c31\u662f\u4e0d\u80fd\u53cd\u5411\u4f20\u64ad\u4e86\uff08\u79bb\u6563\u53d6\u503c\uff0c\u4e0d\u53ef\u5fae\uff09\uff0c\u5f97\u7528\u5176\u4ed6\u7684\u65b9\u6cd5\u8fdb\u884c\u68af\u5ea6\u4f30\u8ba1\uff08\u5f3a\u5316\u5b66\u4e60\u6216\u8499\u7279\u5361\u7f57\u65b9\u6cd5\uff0c\u5f85\u5b66\u4e60<\/em>\uff09\u3002<\/p>\n\n\n\n

\u81ea\u6ce8\u610f\u529b\uff08Self-Attention\uff09<\/h3>\n\n\n\n

\u5373\u6ce8\u610f\u81ea\u8eab\u4e0a\u4e0b\u6587\u5173\u8054\u7684\u673a\u5236\uff0cQuery\u3001Key\u3001Value \u90fd\u6765\u81ea\u540c\u4e00\u4e2a\u8f93\u5165\u5e8f\u5217 \\( X \\)\u3002\u53c2\u7167\u4e0a\u9762 Scaled Dot-Product Attention \u7684\u56fe\uff0c\u81ea\u6ce8\u610f\u529b\u7684\u8868\u8fbe\u5f0f\u53ef\u4ee5\u5199\u4e3a\uff1a<\/p>\n\n\n\n

$$ \\mathrm{SelfAttention}(X) = \\mathrm{softmax} \\left ( \\frac {XW_Q X W_K} {\\sqrt{d_k}} \\right ) XW_V $$<\/p>\n\n\n\n

Transformer \u4e2d\u7528\u5230\u7684\u5c31\u662f\u8fd9\u4e00\u6b3e\u6ce8\u610f\u529b\u3002<\/p>\n\n\n\n

\u4ea4\u53c9\u6ce8\u610f\u529b\uff08Cross\u2011Attention\uff09<\/h3>\n\n\n\n

\u548c\u81ea\u6ce8\u610f\u529b\u51e0\u4e4e\u4e00\u81f4\uff0c\u5c31\u662f Query \u6765\u81ea\u5e8f\u5217 A\uff0c\u800c Key-Value \u6765\u81ea\u5e8f\u5217 B\uff0c\u6a21\u578b\u5728\u5bf9\u5e8f\u5217 A \u4e2d\u7684\u6bcf\u4e2a\u5143\u7d20\u505a\u6ce8\u610f\u529b\u8ba1\u7b97\u65f6\uff0c\u4f1a\u62ac\u5934\u53bb\u770b\u5e8f\u5217 B\uff0c\u4ece\u4e2d\u6311\u9009\u4e0e\u5f53\u524d\u67e5\u8be2\u6700\u76f8\u5173\u7684\u4fe1\u606f\uff0c\u518d\u628a\u6311\u9009\u5230\u7684\u4fe1\u606f\uff08\u52a0\u6743\u540e\u7684 V\uff09\u53cd\u9988\u7ed9\u5e8f\u5217 A\u3002\u53c2\u8003\u81ea\u6ce8\u610f\u529b\uff0c\u4ea4\u53c9\u6ce8\u610f\u529b\u7684\u8868\u8fbe\u5f0f\u53ef\u4ee5\u5199\u4e3a\uff1a<\/p>\n\n\n\n

$$ \\mathrm{CrossAttention}(Q_A, K_B, V_B) = \\mathrm{softmax} \\left ( \\frac {Q_A K_{B}^{T}} {\\sqrt{d_k}} \\right ) V_B $$<\/p>\n\n\n\n

Transformer \u4e2d\u4e5f\u7528\u5230\u4e86\u8fd9\u6b3e\u6ce8\u610f\u529b\u3002<\/p>\n\n\n\n

\u591a\u5934\u6ce8\u610f\u529b\uff08Multi-Head Attention\uff09<\/h3>\n\n\n\n

\u5c06\u6ce8\u610f\u529b\u673a\u5236\u5e76\u884c\u5730\u591a\u6b21\u6267\u884c\uff0c\u5373\u6709\u591a\u4e2a\u6ce8\u610f\u529b\u5934\uff0c\u518d\u5c06\u7ed3\u679c\u62fc\u63a5\u878d\u5408\u3002\u5982\u4e0b\u56fe\u6240\u793a\u3002<\/p>\n\n\n

\n
\"\"<\/div><\/figure>\n<\/div>\n\n\n

\u591a\u5934\u6ce8\u610f\u529b\u548c\u81ea\u6ce8\u610f\u529b\u7ecf\u5e38\u4e00\u8d77\u7528\uff0c\u53d8\u6210\u5f3a\u529b\u7684\u591a\u5934\u81ea\u6ce8\u610f\u529b\u673a\u5236\uff08Transformer \u4e2d\u4e5f\u7528\u5230\u4e86\u8fd9\u6b3e\u6ce8\u610f\u529b\uff09\u3002<\/p>\n\n\n\n

\u901a\u9053\u6ce8\u610f\u529b\uff08Channel Attention\uff09<\/h3>\n\n\n\n

\u901a\u9053\u53ef\u4ee5\u4ece\u7b80\u5355\u7684\u4e00\u5f20\u56fe\u7247\u5f00\u59cb\u7406\u89e3\uff0c\u4e00\u5f20\u5f69\u8272\u56fe\u7247\u5206\u4e3a R\u3001G\u3001B \u4e09\u4e2a\u989c\u8272\u901a\u9053\uff0c\u53ef\u4ee5\u60f3\u8c61\u6210\u4e09\u7247\u5f69\u8272\u5851\u6599\u7247\u53e0\u5728\u4e00\u8d77\u3002\u901a\u9053\u6ce8\u610f\u529b\u5c31\u662f\u5728\u8c03\u8282\u6bcf\u4e2a\u5851\u6599\u7247\u7684\u900f\u660e\u5ea6<\/strong>\uff0c\u5b83\u7684\u7ecf\u5178\u5b9e\u73b0\u51fa\u81ea SE-Net\uff08Squeeze-and-Excitation Networks\uff09\u4e2d\u3002<\/p>\n\n\n

\n
\"\"<\/div><\/figure>\n<\/div>\n\n\n

\u5982\u4e0a\u56fe\u6240\u793a\uff0c\u6709\u4e24\u4e2a\u5206\u652f\uff0c\u4e0a\u90e8\u7684\u5206\u652f\u5c31\u662f\u901a\u9053\u6ce8\u610f\u529b\u5206\u652f\u3002\u6a21\u578b\u5148\u770b\u770b\u6574\u5f20\u56fe\u91cc\uff0c\u6bcf\u4e00\u79cd\u901a\u9053\u5bf9\u4efb\u52a1\u6709\u591a\u91cd\u8981\uff0c\u518d\u628a\u91cd\u8981\u901a\u9053”\u900f\u660e\u5ea6”\u8c03\u5927\uff0c\u4e0d\u91cd\u8981\u7684\u901a\u9053”\u900f\u660e\u5ea6”\u8c03\u5c0f\u3002\u5177\u4f53\u7684\u5b9e\u73b0\u5206\u4e3a\u4e09\u6b65\uff0c\u5373 Squeeze \u3001 Excitation \u548c Scale\u3002<\/p>\n\n\n\n

Squeeze \u6b65\u9aa4 \\( F_{sq}(\\cdot) \\) \u4f1a\u5bf9\u6bcf\u4e2a\u901a\u9053\u7684 \\( w \\times h \\) \u7ef4\u7279\u5f81\u56fe\u505a\u5168\u5c40\u5e73\u5747\u6c60\u5316\uff08\u5168\u90e8\u6c42\u548c\u518d\u53d6\u5e73\u5747\uff09\u3002Excitation \u6b65\u9aa4 \\( F_{ex}(\\cdot, w) \\) \u4e3a\u4e24\u5c42\u202fFC\u202f+\u202fReLU\u202f+\u202fsigmoid\uff0c\u8f93\u51fa\u6743\u91cd\u5411\u91cf\u3002Scale \u6b65\u9aa4 \\( F_{scale}(\\cdot) \\) \u505a\u6ce8\u610f\u529b\u5f97\u5206\u52a0\u6743\u7684\u64cd\u4f5c\uff0c\u8fbe\u5230\u8c03\u6574\u6bcf\u4e2a\u901a\u9053\u91cd\u8981\u6027\u7684\u76ee\u7684\u3002<\/p>\n\n\n\n

\u7a7a\u95f4\u6ce8\u610f\u529b\uff08Spatial Attention\uff09<\/h3>\n\n\n\n

\u8fd8\u662f\u90a3\u4e00\u53e0\u5f69\u8272\u5851\u6599\u7247\uff0c\u7a7a\u95f4\u6ce8\u610f\u529b\u76f8\u5f53\u4e8e\u7ed9\u6574\u53e0\u5851\u6599\u7247\u76d6\u4e0a\u4e00\u5f20\u534a\u900f\u660e\u7684\u8d34\u7eb8<\/strong>\uff0c\u5b83\u7684\u7ecf\u5178\u5b9e\u73b0\u51fa\u81ea CBAM\uff08Convolutional Block Attention Module\uff09<\/strong>\u3002\u5047\u8bbe\u539f\u8f93\u5165\u7279\u5f81 \\( F \\) \u5df2\u7ecf\u7ecf\u8fc7\u4e86\u901a\u9053\u6ce8\u610f\u529b\u673a\u5236\u6a21\u5757\u5f97\u5230 \\( F\u2018 \\)<\/p>\n\n\n

\n
\"\"<\/div><\/figure>\n<\/div>\n\n\n

\u5982\u4e0a\u56fe\u6240\u793a\uff0c\u5148\u5bf9\u8f93\u5165\u202f\\( F\u2018 \\)\u202f\u5728\u901a\u9053\u7ef4\u5ea6\u505a\u6700\u5927\u6c60\u5316\uff08MaxPool\uff09\u548c\u5e73\u5747\u6c60\u5316\uff08AvgPool\uff09\uff0c\u6cbf\u901a\u9053\u7ef4\u5ea6\u62fc\u63a5\uff0c\u518d\u7528\u5927\u5c0f\u4e3a 7\u00d77 \u7684\u5377\u79ef\u6838\u8fdb\u884c\u5377\u79ef\uff0c\u6700\u540e\u7ecf\u8fc7\u6fc0\u6d3b\u51fd\u6570\uff0c\u5f97\u5230\u6ce8\u610f\u529b\u56fe \\( M_s \\)\u3002\u8fd9\u4e2a\u6ce8\u610f\u529b\u56fe\u4f1a\u544a\u8bc9\u6a21\u578b\uff0c\u7279\u5f81\u56fe\u91cc\u7684\u54ea\u4e9b\u5730\u65b9\u6700\u6709\u7528\uff0c\u8d34\u7eb8\u7684\u8fd9\u4e9b\u5730\u65b9\u900f\u660e\u5ea6\u8c03\u5927\uff0c\u5176\u4ed6\u5730\u65b9\u900f\u660e\u5ea6\u8c03\u5c0f\u3002<\/p>\n\n\n\n

CBAM<\/strong> \u8ba9\u7279\u5f81\u5148\u540e\u901a\u8fc7\u901a\u9053\u6ce8\u610f\u529b\u548c\u7a7a\u95f4\u6ce8\u610f\u529b\uff0c\u8ba9\u6a21\u578b\u5b66\u4e60\u5230\u201c\u4ec0\u4e48\u7279\u5f81\u6700\u6709\u7528\u201d\u548c\u201c\u6700\u6709\u7528\u7684\u4fe1\u606f\u5728\u54ea\u91cc”\uff0c\u505a\u5230\u6ce8\u610f\u529b\u7684\u7d2f\u79ef\u589e\u76ca\uff0c\u5168\u9762\u6355\u83b7\u7279\u5f81\u4e2d\u7684\u5173\u952e\u4fe1\u606f\u3002<\/p>\n\n\n\n

\"\"<\/div><\/div>\n\n\n\n

\u5173\u4e8e\u6c60\u5316\u5c42\uff1a
\u6c60\u5316\u7684\u82f1\u6587\u4e3a”Pooling”\uff0c\u672c\u610f\u4e3a\u201d\u6c60\u5858\u201c\u6216\u201d\u6c47\u96c6\u201c\u3002\u6c60\u5316\u7684\u540d\u5b57\u548c\u5176\u4f5c\u7528\u76f8\u5bf9\u5e94\uff0c\u628a\u4e00\u5927\u6ee9\u6c34\u201d\u6c47\u96c6\u5230\u4e00\u4e2a\u5c0f\u6c60\u51fc\u5b50\u91cc\u201c\u2192\u628a\u5c40\u90e8\u7279\u5f81\u6c47\u603b\u6210\u4e00\u4e2a\u4ee3\u8868\u503c\u3002<\/div><\/div>\n\n\n\n

\u5176\u4ed6\u6ce8\u610f\u529b\u673a\u5236\uff08\u5f85\u7eed\uff09<\/h3>\n\n\n\n

\u53cc\u91cd\u6ce8\u610f\u529b\uff08Dual\u202fAttention\uff09\u3001\u53ef\u53d8\u5f62\u6ce8\u610f\u529b\uff08Deformable\u202fAttention\uff09\u3001\u5c42\u6b21\u6ce8\u610f\u529b\uff08Hierarchical Attention\uff09……\u7b49\uff0c\u9488\u5bf9\u4e0d\u540c\u4efb\u52a1\u7684\u4e0d\u540c\u6ce8\u610f\u529b\u673a\u5236\u3002<\/p>\n\n\n\n

Attention Is All You Need<\/h2>\n\n\n\n

\u5230\u4e86\u6700\u7ec8\u7684\u5b66\u4e60\u76ee\u6807\uff1aTransformer \u67b6\u6784\u3002\u548c\u8bba\u6587\u300aAttention Is All You Need\u300b\u4e2d\u7684\u987a\u5e8f\u4e00\u6837\uff0c\u6309\u7167 Encoder stacks\u3001Decoder stacks\u3001Input\/Output Embeddings\u3001Positional Encoding \u7684\u987a\u5e8f\u8fdb\u884c\u5b66\u4e60\u3002<\/p>\n\n\n\n

Transformer \u7684\u6574\u4f53\u67b6\u6784<\/h3>\n\n\n
\n
\"\"<\/div><\/figure>\n<\/div>\n\n\n

Encoder Stacks<\/h3>\n\n\n\n

\u5728 Transformer \u91cc\uff0cencoder \u7531\u516d\u4e2a\u76f8\u4f3c\u7684 stack \u5806\u53e0\u800c\u6210\u3002\u5355\u4e2a stack \u7ed3\u6784\u662f\u6a21\u578b\u6846\u67b6\u5de6\u4fa7\u8f83\u77ee\u7684\u7070\u8272\u5706\u89d2\u77e9\u9635\u6846\u4f4f\u7684\u90e8\u5206\u3002\u5c31\u50cf\u4e50\u9ad8\u79ef\u6728\u4e00\u6837\uff0c\u4ece\u4e0b\u5f80\u4e0a\u7684\u56db\u4e2a\u79ef\u6728\u5757\u53ef\u4ee5\u5f52\u7eb3\u4e3a\uff1a<\/p>\n\n\n\n

$$ \u591a\u5934\u81ea\u6ce8\u610f\u529b \\to \u6b8b\u5dee\u5f52\u4e00\u5316 \\to \u524d\u9988 \\to \u6b8b\u5dee\u5f52\u4e00\u5316 $$<\/p>\n\n\n\n

\u591a\u5934\u81ea\u6ce8\u610f\u529b\uff08Multi-Head Attention\uff09<\/h4>\n\n\n\n

\u5047\u5982\u8bf4 h = 8\uff08\u6709 8 \u4e2a\u5e76\u884c\u7684\u6ce8\u610f\u529b\u5934\uff09\uff0c\u8f93\u5165\u7684 Q\u3001K\u3001V \u7ef4\u5ea6\u662f \\( [B, L, D] \\)\uff0c\u6682\u65f6\u4e0d\u7ba1 Batch Size \\( B \\)\uff0cQ\u3001K\u3001V \u4f1a\u5728 \\( D \\) \u7ef4\u5ea6\u5207\u6210\u7b49\u957f\u7684 8 \u5c0f\u6bb5\uff0c\u5f97\u5230 8 \u7ec4\u8f93\u5165\uff0c\u5e76\u884c\u8f93\u5165\u5230\u591a\u5934\u81ea\u6ce8\u610f\u529b\u6a21\u5757\u4e2d\uff0c\u5f97\u5230 8 \u4e2a\u5934\u7684\u8f93\u51fa\uff0c\u518d\u62fc\u63a5\u8d77\u6765\u5c31\u5f97\u5230\u591a\u5934\u81ea\u6ce8\u610f\u529b\u7684\u603b\u8f93\u51fa\uff0c\u6355\u83b7\u4e86\u5e8f\u5217\u7684\u4e0a\u4e0b\u6587\u4f9d\u8d56\u3002<\/p>\n\n\n\n

$$ \\mathrm{Attention}(Q,K,V)=\\mathrm{softmax} \\left ( \\frac {QK^T} {\\sqrt{d_k}} \\right ) V $$<\/p>\n\n\n\n

\u8fd9\u4e2a\u662f Transformer \u4e2d\u7684\u7ecf\u5178\u6ce8\u610f\u529b\u8868\u8fbe\u5f0f\uff0c\u91cc\u9762\u7684 \\(d_k\\) \u4e3a \\(K\\) \u7684\u7ef4\u5ea6\uff0c\u76f8\u5f53\u4e8e\u5f52\u4e00\u5316\u64cd\u4f5c\uff0c\u9632\u6b62\u68af\u5ea6\u6d88\u5931\u6216\u7206\u70b8\u3002softmax \u5c31\u662f\u6fc0\u6d3b\u51fd\u6570\u4e86\uff0c\u5c06\u6743\u91cd\u5f52\u4e00\u5316\u5230 0~1 \u8303\u56f4\u5185\uff0c\u800c\u4e14\u603b\u548c\u662f 1\u3002<\/p>\n\n\n\n

\u6b8b\u5dee\u8fde\u63a5\u548c\u5c42\u5f52\u4e00\u5316\uff08Add & Norm\uff09<\/h4>\n\n\n\n

\u5c06\u4e00\u4e2a\u5b50\u5c42 Sub-layer \u7684\u8f93\u51fa\u518d\u548c\u8f93\u5165\u76f8\u52a0\uff0c\u5e76\u7ecf\u8fc7\u5c42\u5f52\u4e00\u5316\u5904\u7406\u3002\u8868\u8fbe\u5f0f\uff1a<\/p>\n\n\n\n

$$ \\mathrm{LayerNorm}(x + \\mathrm{Sublayer}(x)) $$<\/p>\n\n\n\n

\u4e3a\u4e86\u5229\u4e8e\u6b8b\u5dee\u8fde\u63a5\u548c\u5c42\u5f52\u4e00\u5316\uff0c\u4f5c\u8005\u5c06\u6a21\u578b\u7684\u8f93\u51fa\u7ef4\u5ea6\u63a7\u5236\u5728 \\( d_{\\mathrm{model}} = 512 \\)\u3002<\/p>\n\n\n\n

\u6b8b\u5dee\u8fde\u63a5\u548c\u8df3\u8dc3\u8fde\u63a5\uff1a
\u6b8b\u5dee\u8fde\u63a5\u662f\u8df3\u8dc3\u8fde\u63a5\u7684\u4e00\u4e2a\u7279\u5b9a\u79cd\u7c7b\uff0c\u628a\u8f93\u5165\u76f4\u63a5\u201c\u76f8\u52a0\u201d\u5230\u82e5\u5e72\u5c42\u4e4b\u540e\u7684\u8f93\u51fa\u4e0a\u3002
\u8df3\u8dc3\u8fde\u63a5\u7684\u6982\u5ff5\u66f4\u5bbd\u6cdb\uff0c\u53ef\u4ee5\u662f\u76f8\u52a0\uff08Residual\uff09\u3001\u62fc\u63a5\uff08Concat\uff09\u3001\u95e8\u63a7\uff08Gated skip\uff09\u3001\u81ea\u9002\u5e94\u7f29\u653e\u7b49\u7b49\u3002<\/div><\/div>\n\n\n\n

\u524d\u9988\u795e\u7ecf\u7f51\u7edc\uff08Feed Forward Networks\uff09<\/h4>\n\n\n\n

\u5176\u5b9e\u5c31\u662f\u591a\u5c42\u611f\u77e5\u673a MLP\uff0c\u8f93\u5165\u5c42\u548c\u8f93\u51fa\u5c42\u7ef4\u5ea6\u5747\u4e3a 512\uff0c\u9690\u85cf\u5c42\u7ef4\u5ea6\u4e3a 2048\u3002<\/p>\n\n\n

\n
\"\"<\/div><\/figure>\n<\/div>\n\n\n

\u5c06\u8f93\u5165\u901a\u8fc7 6 \u4e2a Encoder stacks \u540e\u5c31\u5f97\u5230\u4e86\u7f16\u7801\u540e\u7684\u7279\u5f81\uff0c\u53ef\u4ee5\u9001\u5230 Decoder \u4e2d\uff0c\u5982\u4e0b\u56fe\u6240\u793a\u3002<\/p>\n\n\n

\n
\"\"<\/div><\/figure>\n<\/div>\n\n\n

Decoder Stacks<\/h3>\n\n\n\n

decoder \u7684\u5355\u4e2a stack \u7ed3\u6784\u662f\u6a21\u578b\u6846\u67b6\u53f3\u4fa7\u8f83\u9ad8\u7684\u7070\u8272\u5706\u89d2\u77e9\u9635\u6846\u4f4f\u7684\u90e8\u5206\u3002\u7ee7\u7eed\u7528\u79ef\u6728\u5757\u505a\u6bd4\u55bb\uff0c\u4ece\u4e0b\u5f80\u4e0a\u7684\u516d\u4e2a\u79ef\u6728\u5757\u53ef\u4ee5\u5f52\u7eb3\u4e3a\uff1a<\/p>\n\n\n\n

$$ \u63a9\u819c\u591a\u5934\u81ea\u6ce8\u610f\u529b \\to \u6b8b\u5dee\u5f52\u4e00\u5316 \\to \u4ea4\u53c9\u6ce8\u610f\u529b \\to \u6b8b\u5dee\u5f52\u4e00\u5316 \\to \u524d\u9988 \\to \u6b8b\u5dee\u5f52\u4e00\u5316 $$<\/p>\n\n\n\n

\u5176\u4e2d Add\u202f&\u202fNorm\u3001Position\u2011wise\u202fFeed Forward \u4e0e Encoder \u4e00\u6a21\u4e00\u6837\uff0c\u53ef\u4ee5\u628a Decoder \u60f3\u6210 Encoder\u202f \u7684 2 \u5904\u5347\u7ea7\u7248\u3002\u5347\u7ea7\u8fc7\u7684\u5730\u65b9\u5c31\u662f\u4e0b\u9762\u7684\u4e24\u4e2a\u90e8\u5206\u3002<\/p>\n\n\n\n

\u5e26\u906e\u6321\u7684\u591a\u5934\u81ea\u6ce8\u610f\u529b\uff08Masked Multi\u2011Head Self\u2011Attention\uff09<\/h4>\n\n\n\n

\u5728\u524d\u9762\u5df2\u7ecf\u5b66\u4e60\u8fc7\uff0c\u6240\u52a0\u7684 Mask \u4e3a Look-Ahead Mask\uff0c\u5728\u8868\u8fbe\u5f0f\u4e2d\u7684\u4f53\u73b0\u5982\u4e0b\uff1a<\/p>\n\n\n\n

$$ \\mathrm{Attention}(Q,K,V)=\\mathrm{softmax} \\left ( \\frac {QK^T} {\\sqrt{d_k}} + M_{\\mathrm{casual}} \\right ) V $$<\/p>\n\n\n\n

\u5176\u4e2d<\/p>\n\n\n\n

$$ M_{\\mathrm{casual}} = \\begin{cases} 0 &j \\le i \\\\ -\u221e &j > i \\end{cases}$$<\/p>\n\n\n\n

\u6709\u4e86 Look-Ahead Mask\uff0c\u6a21\u578b\u5c31\u4e0d\u80fd\u5077\u770b\u5230\u672a\u6765\u7684\u4fe1\u606f\u4e86\uff0c\u4fdd\u6301\u81ea\u56de\u5f52<\/em>\u7279\u6027\uff08\u9650\u5236\u6a21\u578b\u53ea\u80fd\u7528\u524d\u9762\u7684\u4fe1\u606f\u9884\u6d4b\u540e\u9762\u7684\u4fe1\u606f\uff09<\/p>\n\n\n\n

\u4ea4\u53c9\u6ce8\u610f\u529b\uff08Encoder\u2011Decoder Cross\u2011Attention\uff09<\/h4>\n\n\n\n

\u5728 Transformer \u91cc\u9762 Query \u6765\u81ea\u524d\u9762\u7684 Decoder \u5c42\uff08\u522b\u5fd8\u4e86 Decoder \u662f 6 \u4e2a\u4e00\u6837\u7684\u6a21\u5757\u5806\u53e0\u800c\u6210\u7684\uff0c\u4e0a\u4e00\u4e2adecoder \u7684\u8f93\u51fa\u518d\u9001\u5165\u4e0b\u4e00\u4e2a decoder\uff09\uff0c\u800c Key-Value \u6765\u81ea\u6765\u81ea Encoder \u7684\u8f93\u51fa\u3002<\/p>\n\n\n\n

Input Embeddings<\/h3>\n\n\n\n

\u5bf9\u4e8e\u8f93\u5165\u7684\u8bcd\u5d4c\u5165\uff0c\u76ee\u7684\u662f\u5c06\u4eba\u7c7b\u770b\u5f97\u61c2\u7684\u81ea\u7136\u8bed\u8a00\u8f6c\u6362\u4e3a\u673a\u5668\u770b\u5f97\u61c2\u7684\u201c\u5411\u91cf\u8bed\u201d\uff0c\u5927\u81f4\u7684\u64cd\u4f5c\u5206\u4e3a\u4e09\u6b65\uff1a\u2460\u67e5\u8868 \u2192 \u2461\u7f29\u653e \u2192 \u2462\u4f4d\u7f6e\u7f16\u7801\u3002<\/p>\n\n\n\n

\u4e00\u5f00\u59cb\u4f1a\u7ed9\u6a21\u578b\u5b89\u6392\u201c\u4e00\u672c\u8bcd\u5178\u201d\uff0c\u4e00\u4e2a \\(\\left | V \\right |\\)\uff08\u6709\u591a\u5c11\u4e2a\u8bcd\uff09\\(\\times d_{\\mathrm{model}}\\) \uff08\u6bcf\u4e2a\u8bcd\u7684\u89e3\u91ca\uff0c\u56fa\u5b9a\u957f\u5ea6\uff09\u7684\u53ef\u5b66\u4e60\u77e9\u9635 \\(E\\)\u3002\u2460\u6bcf\u4e2a\u8bcd\u6839\u636e\u81ea\u5df1\u5728\u53e5\u5b50\u4e2d\u7684\u4f4d\u7f6e token id \\(x_i\\) \uff08\u7b2c\u4e00\u4e2a\u8bcd\u5c31\u662f 0\uff0c\u7b2c\u4e8c\u4e2a\u8bcd\u5c31\u662f 1……\uff09\u67e5\u8bcd\u5178\uff0c\u5f97\u5230\u8bcd\u5d4c\u5165 \\( \\mathbf{e}_i \\)\u3002\u2461\u5bf9\u8bcd\u5d4c\u5165\u518d\u4e58\u4ee5 \\(\\sqrt{d_{\\mathrm{model}}}\\) \u7a33\u5b9a\u65b9\u5dee\u3002\u2462\u518d\u4e3a\u8bcd\u5d4c\u5165\u52a0\u4e00\u4e2a\u53ef\u5b66\u4e60\u7684\u4f4d\u7f6e\u7f16\u7801<\/strong> \\( \\mathbf{P}_i \\)\uff0c\u6700\u540e\u5f97\u5230\u8f93\u5165\u8bcd\u5d4c\u5165\uff1a<\/p>\n\n\n\n

$$ \\tilde{\\mathbf{e}}_i = \\mathbf{e}_i + \\mathbf{P}_i $$<\/p>\n\n\n\n

Output Embeddings<\/h3>\n\n\n\n

\u4ece\u67b6\u6784\u56fe\u4e0a\u53ef\u4ee5\u770b\u5230\uff0coutputs \u4e0b\u65b9\u6709\u4e00\u4e2a\u6ce8\u91ca\uff1ashifted rights\uff0c\u610f\u601d\u662f\u5728\u8bad\u7ec3\u65f6\u628a\u771f\u5b9e\u76ee\u6807\u5e8f\u5217\u202f\\(<s>,y_1,y_2,\u2026y_{T-1}\\) \u6574\u4f53\u5411\u53f3\u79fb\u52a8\u4e00\u4f4d\u9001\u8fdb\u89e3\u7801\u5668\uff0c\u53bb\u9884\u6d4b\u4e0b\u4e00\u4e2a\u8bcd\uff1b\u63a8\u7406\u65f6\u6bcf\u4e00\u6b65\u628a\u81ea\u5df1\u521a\u751f\u6210\u7684\u8bcd\u518d\u67e5\u8bcd\u5178\uff0c\u8f93\u9001\u56de\u89e3\u7801\u5668\u3002<\/p>\n\n\n\n

\u89e3\u7801\u5668\u8f93\u5165\u8ddf\u7f16\u7801\u5668\u8f93\u5165\u4e00\u6837\uff0c\u90fd\u901a\u8fc7\u4e00\u4e2a\u53ef\u5b66\u4e60\u7684\u77e9\u9635 \\(E\\)\uff0c\u628a token id \u53d8\u6210 \\(d_{\\mathrm{model}}\\) \u7ef4\u5411\u91cf\uff0c\u518d\u52a0\u4f4d\u7f6e\u7f16\u7801\u3002<\/p>\n\n\n\n

\u9884\u6d4b\u5b8c\u6bd5\u540e\uff0c\u8981\u628a\u9690\u85cf\u5411\u91cf\uff08\u5411\u91cf\u8bed\uff09 \\(h_t\\) \u6620\u5c04\u56de\u8bcd\u8868\uff08\u81ea\u7136\u8bed\u8a00\uff09\u8fdb\u884c\u6253\u5206\uff0c\u6700\u81ea\u7136\u7684\u505a\u6cd5\u662f\u518d\u7528 \\(E^T\\) \u505a\u4e00\u6b21\u70b9\u79ef\uff0c\u518d\u52a0\u504f\u7f6e\uff0c\u7136\u540e\u8fc7 soft\u2011max\uff0c\u8fd9\u4e00\u6b65\u5c31\u662f\u8f93\u51fa\u5d4c\u5165\/LM\u202fhead\u3002<\/p>\n\n\n\n

\u540c\u4e00\u5757\u53ef\u5b66\u4e60\u7684\u77e9\u9635 \\(E\\) \u5c31\u76f8\u5f53\u4e8e\u4e00\u4e2a\u53cc\u5411\u7ffb\u8bd1\u7684\u8bcd\u5178\uff0c\u7f16\u7801\u65f6\u628a one\u2011hot -> dense\uff08\u8f93\u5165\u5d4c\u5165\uff09\uff0c\u89e3\u7801\u65f6\u628a dense -> vocabulary logits\uff08\u8f93\u51fa\u5d4c\u5165\uff0c\u8f6c\u7f6e\u540e\u4f7f\u7528\uff09<\/p>\n\n\n\n

\uff08\u5d4c\u5165\u8fd9\u4e00\u5757\u5df2\u7ecf\u5728Transformer \u6a21\u578b\u5916\u90e8\u4e86\uff09<\/p>\n\n\n\n

Positional Encoding<\/h3>\n\n\n\n

\u81ea\u6ce8\u610f\u529b\u673a\u5236\u672c\u8eab\u5e76\u4e0d\u80fd\u611f\u77e5\u5230\u987a\u5e8f\u5173\u7cfb\uff0c\u628a\u201c\u732b\u5bf9\u6211\u54c8\u6c14\u201d\u548c\u201c\u6211\u5bf9\u732b\u54c8\u6c14\u201d\u91cc\u7684\u8bcd\u5411\u91cf\u76f4\u63a5\u4e22\u8fdb\u53bb\uff0c\u6ce8\u610f\u529b\u53ea\u770b\u5230 4\u202f\u4e2a\u8bcd\uff08\u732b\uff0c\u6211\uff0c\u5bf9\uff0c\u54c8\u6c14\uff09\uff0c\u5b8c\u5168\u4e0d\u77e5\u9053\u8c01\u5148\u8c01\u540e\u3002\u6240\u4ee5\u9700\u8981\u7ed9\u6bcf\u4e2a\u8bcd\u8d34\u4e00\u4e2a\u5ea7\u6807\u6807\u7b7e\uff0c\u544a\u8bc9\u6a21\u578b\u201c\u8fd9\u662f\u7b2c\u202f1\u202f\u4e2a\u201d\u201c\u8fd9\u662f\u7b2c\u202f2\u202f\u4e2a\u201d\u2026\u2026\u8fd9\u6837\u5b83\u5c31\u80fd\u5206\u6e05\u5148\u6765\u540e\u5230\u3002<\/p>\n\n\n\n

Transformer \u4e2d\u7528\u5230\u7684\u6b63\u5f26\/\u4f59\u5f26\u4f4d\u7f6e\u7f16\u7801\u4e3a\u4e09\u89d2\u5f0f\u3002\u5bf9\u4e8e\u5e8f\u5217\u4e2d\u7684\u7b2c pos \u4e2a\u4f4d\u7f6e\uff0c\u7edd\u5bf9\u4f4d\u7f6e\u7f16\u7801\u5411\u91cf\u7684\u7b2c k \u4e2a\u7ef4\u5ea6\u7684\u503c<\/p>\n\n\n\n

$$ \\begin{align} \\mathrm{PE} (pos, 2k) = \\mathrm{sin} \\left( \\frac{pos} {1000^(\\frac{2k} {d_{\\mathrm{model}}}} \\right), \\\\ \\mathrm{PE} (pos, 2k + 1) = \\mathrm{cos} \\left( \\frac{pos} {1000^(\\frac{2k} {d_{\\mathrm{model}}}} \\right) \\end{align}$$<\/p>\n\n\n\n

\u4e0b\u4e00\u6b65\u5b66\u4e60&\u56de\u987e…\uff1f<\/p>\n\n\n\n

BERT\uff1fTransformer\u53d8\u79cd\u5bb6\u65cf\uff1fAutoEncoder\u53d8\u79cd\u5bb6\u65cf\uff1fGAN\uff1fDiffusion\uff1f<\/p>\n","protected":false},"excerpt":{"rendered":"

\u666e\u901a\u6ce8\u610f\u529b\u673a\u5236\u7684\u539f\u7406 \u4eba\u7684\u6ce8\u610f\u529b\u5f80\u5f80\u96c6\u4e2d\u5728\u89c6\u7ebf\u7684\u6b63\u4e2d\u95f4\uff0c\u4e13\u6ce8\u4e8e\u90a3\u4e00\u5c0f\u90e8\u5206\u91cd\u8981\u7684\u5185\u5bb9\u3002\u6a21\u578b\u600e\u4e48\u53bb\u6ce8\u610f\u5230\u90a3\u4e9b\u91cd\u8981\u7684 […]<\/p>\n","protected":false},"author":311,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"_bbp_topic_count":0,"_bbp_reply_count":0,"_bbp_total_topic_count":0,"_bbp_total_reply_count":0,"_bbp_voice_count":0,"_bbp_anonymous_reply_count":0,"_bbp_topic_count_hidden":0,"_bbp_reply_count_hidden":0,"_bbp_forum_subforum_count":0,"footnotes":""},"categories":[18],"tags":[],"class_list":["post-2639","post","type-post","status-publish","format-standard","hentry","category-aigc"],"_links":{"self":[{"href":"https:\/\/www.lanxy.ink\/index.php?rest_route=\/wp\/v2\/posts\/2639","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.lanxy.ink\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.lanxy.ink\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.lanxy.ink\/index.php?rest_route=\/wp\/v2\/users\/311"}],"replies":[{"embeddable":true,"href":"https:\/\/www.lanxy.ink\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=2639"}],"version-history":[{"count":37,"href":"https:\/\/www.lanxy.ink\/index.php?rest_route=\/wp\/v2\/posts\/2639\/revisions"}],"predecessor-version":[{"id":2702,"href":"https:\/\/www.lanxy.ink\/index.php?rest_route=\/wp\/v2\/posts\/2639\/revisions\/2702"}],"wp:attachment":[{"href":"https:\/\/www.lanxy.ink\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=2639"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.lanxy.ink\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=2639"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.lanxy.ink\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=2639"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}