{"id":62072,"date":"2026-01-19T09:05:52","date_gmt":"2026-01-19T01:05:52","guid":{"rendered":"https:\/\/www.wsisp.com\/helps\/62072.html"},"modified":"2026-01-19T09:05:52","modified_gmt":"2026-01-19T01:05:52","slug":"%e9%80%9a%e4%bf%97%e7%90%86%e8%a7%a3%e8%87%aa%e6%b3%a8%e6%84%8f%e5%8a%9b%e6%9c%ba%e5%88%b6%ef%bc%88self-attention%ef%bc%89","status":"publish","type":"post","link":"https:\/\/www.wsisp.com\/helps\/62072.html","title":{"rendered":"\u901a\u4fd7\u7406\u89e3\u81ea\u6ce8\u610f\u529b\u673a\u5236\uff08Self-Attention\uff09"},"content":{"rendered":"<h3>\u76ee\u5f55<\/h3>\n<ul>\n<li>\u5f15\u8a00<\/li>\n<li>\u81ea\u6ce8\u610f\u529b\u673a\u5236\u7684\u80cc\u666f\n<ul>\n<li>\u5e8f\u5217\u6a21\u578b\u7684\u6f14\u8fdb<\/li>\n<li>\u6ce8\u610f\u529b\u673a\u5236\u7684\u8d77\u6e90<\/li>\n<\/ul>\n<\/li>\n<li>\u81ea\u6ce8\u610f\u529b\u673a\u5236\u7684\u6838\u5fc3\u539f\u7406\n<ul>\n<li>\u67e5\u8be2\u3001\u952e\u548c\u503c&#xff08;Q\u3001K\u3001V&#xff09;\u7684\u6982\u5ff5<\/li>\n<li>\u6ce8\u610f\u529b\u5206\u6570\u7684\u8ba1\u7b97<\/li>\n<li>Softmax\u5f52\u4e00\u5316\u548c\u52a0\u6743\u6c42\u548c<\/li>\n<\/ul>\n<\/li>\n<li>\u591a\u5934\u81ea\u6ce8\u610f\u529b\u673a\u5236\n<ul>\n<li>\u4e3a\u4ec0\u4e48\u9700\u8981\u591a\u5934<\/li>\n<li>\u591a\u5934\u6ce8\u610f\u529b\u7684\u5b9e\u73b0<\/li>\n<\/ul>\n<\/li>\n<li>\u81ea\u6ce8\u610f\u529b\u5728Transformer\u4e2d\u7684\u5e94\u7528\n<ul>\n<li>Transformer\u6574\u4f53\u67b6\u6784<\/li>\n<li>\u7f16\u7801\u5668\u4e2d\u7684\u81ea\u6ce8\u610f\u529b<\/li>\n<li>\u89e3\u7801\u5668\u4e2d\u7684\u81ea\u6ce8\u610f\u529b\u4e0e\u4ea4\u53c9\u6ce8\u610f\u529b<\/li>\n<\/ul>\n<\/li>\n<li>\u4ee3\u7801\u5b9e\u73b0&#xff1a;\u4ece\u96f6\u6784\u5efa\u81ea\u6ce8\u610f\u529b\n<ul>\n<li>NumPy\u7248\u672c\u7684\u81ea\u6ce8\u610f\u529b<\/li>\n<li>PyTorch\u7248\u672c\u7684\u81ea\u6ce8\u610f\u529b<\/li>\n<li>\u591a\u5934\u6ce8\u610f\u529b\u7684\u4ee3\u7801\u793a\u4f8b<\/li>\n<\/ul>\n<\/li>\n<li>\u53ef\u89c6\u5316\u81ea\u6ce8\u610f\u529b\u673a\u5236\n<ul>\n<li>\u6ce8\u610f\u529b\u77e9\u9635\u7684\u70ed\u529b\u56fe<\/li>\n<li>\u591a\u5934\u6ce8\u610f\u529b\u7684\u53ef\u89c6\u5316<\/li>\n<\/ul>\n<\/li>\n<li>\u81ea\u6ce8\u610f\u529b\u673a\u5236\u7684\u4f18\u7f3a\u70b9\n<ul>\n<li>\u4f18\u70b9\u5206\u6790<\/li>\n<li>\u7f3a\u70b9\u4e0e\u6539\u8fdb<\/li>\n<\/ul>\n<\/li>\n<li>\u5b9e\u9645\u5e94\u7528\u6848\u4f8b\n<ul>\n<li>\u81ea\u7136\u8bed\u8a00\u5904\u7406&#xff08;NLP&#xff09;\u4e2d\u7684\u5e94\u7528<\/li>\n<li>\u8ba1\u7b97\u673a\u89c6\u89c9&#xff08;CV&#xff09;\u4e2d\u7684\u5e94\u7528<\/li>\n<li>\u5176\u4ed6\u9886\u57df\u7684\u6269\u5c55<\/li>\n<\/ul>\n<\/li>\n<li>\u5e38\u89c1\u95ee\u9898\u4e0e\u8c03\u8bd5\u6280\u5de7<\/li>\n<li>\u672a\u6765\u5c55\u671b<\/li>\n<li>\u7ed3\u8bba<\/li>\n<li>\u53c2\u8003\u6587\u732e<\/li>\n<\/ul>\n<h3>\u5f15\u8a00<\/h3>\n<p>\u5728\u4eba\u5de5\u667a\u80fd\u548c\u6df1\u5ea6\u5b66\u4e60\u9886\u57df&#xff0c;\u81ea\u6ce8\u610f\u529b\u673a\u5236&#xff08;Self-Attention&#xff09;\u5df2\u6210\u4e3a\u4e00\u4e2a\u4e0d\u53ef\u6216\u7f3a\u7684\u6838\u5fc3\u6280\u672f\u3002\u5b83\u662fTransformer\u6a21\u578b\u7684\u57fa\u77f3&#xff0c;\u9a71\u52a8\u4e86\u50cfBERT\u3001GPT\u548cT5\u8fd9\u6837\u7684\u9884\u8bad\u7ec3\u6a21\u578b\u5728\u81ea\u7136\u8bed\u8a00\u5904\u7406&#xff08;NLP&#xff09;\u4efb\u52a1\u4e2d\u7684\u5353\u8d8a\u8868\u73b0\u3002\u5982\u679c\u4f60\u6b63\u5728\u5b66\u4e60\u673a\u5668\u5b66\u4e60&#xff0c;\u6216\u8005\u60f3\u6df1\u5165\u4e86\u89e3Transformer\u80cc\u540e\u7684\u539f\u7406&#xff0c;\u8fd9\u7bc7\u6587\u7ae0\u5c06\u7528\u901a\u4fd7\u7684\u8bed\u8a00\u5e26\u4f60\u4e00\u6b65\u6b65\u62c6\u89e3Self-Attention\u3002<\/p>\n<p>\u60f3\u8c61\u4e00\u4e0b&#xff0c;\u4f60\u5728\u9605\u8bfb\u4e00\u7bc7\u6587\u7ae0\u65f6&#xff0c;\u4e0d\u4f1a\u5747\u5300\u5173\u6ce8\u6bcf\u4e2a\u8bcd&#xff0c;\u800c\u662f\u91cd\u70b9\u7559\u610f\u5173\u952e\u8bcd\u548c\u4e0a\u4e0b\u6587\u5173\u7cfb\u3002\u81ea\u6ce8\u610f\u529b\u673a\u5236\u6b63\u662f\u6a21\u62df\u4e86\u8fd9\u79cd\u201c\u9009\u62e9\u6027\u5173\u6ce8\u201d&#xff1a;\u5728\u5904\u7406\u5e8f\u5217\u6570\u636e&#xff08;\u5982\u53e5\u5b50\u6216\u56fe\u50cf\u5e8f\u5217&#xff09;\u65f6&#xff0c;\u5b83\u5141\u8bb8\u6a21\u578b\u52a8\u6001\u8ba1\u7b97\u6bcf\u4e2a\u5143\u7d20\u4e0e\u5176\u4ed6\u5143\u7d20\u7684\u5173\u8054\u5f3a\u5ea6&#xff0c;\u4ece\u800c\u6355\u6349\u5168\u5c40\u4f9d\u8d56&#xff0c;\u800c\u975e\u5c40\u9650\u4e8e\u5c40\u90e8\u3002<\/p>\n<p>\u4e3a\u4ec0\u4e48Self-Attention\u5982\u6b64\u53d7\u6b22\u8fce&#xff1f;\u4f20\u7edf\u6a21\u578b\u5982RNN\u5728\u957f\u5e8f\u5217\u4e0a\u6548\u7387\u4f4e\u4e0b&#xff0c;\u800cSelf-Attention\u652f\u6301\u5e76\u884c\u8ba1\u7b97&#xff0c;\u8bad\u7ec3\u901f\u5ea6\u66f4\u5feb\u3002\u6839\u636eHugging Face\u7684\u7edf\u8ba1&#xff0c;\u57fa\u4e8eTransformer\u7684\u6a21\u578b\u5728GLUE\u57fa\u51c6\u4e0a\u51c6\u786e\u7387\u63d0\u5347\u4e8620%\u4ee5\u4e0a\u3002\u672c\u6587\u5c06\u4ece\u80cc\u666f\u3001\u539f\u7406\u3001\u4ee3\u7801\u5b9e\u73b0\u5230\u5e94\u7528&#xff0c;\u5168\u65b9\u4f4d\u8986\u76d6&#xff0c;\u5e2e\u52a9\u521d\u5b66\u8005\u5feb\u901f\u4e0a\u624b&#xff0c;\u8d44\u6df1\u8005\u6df1\u5316\u7406\u89e3\u3002<\/p>\n<p>\u5173\u952e\u8bcd\u4f18\u5316&#xff1a;\u81ea\u6ce8\u610f\u529b\u673a\u5236\u3001Self-Attention\u539f\u7406\u3001Transformer\u6559\u7a0b\u3001\u591a\u5934\u6ce8\u610f\u529b\u89e3\u91ca\u3001PyTorch Self-Attention\u4ee3\u7801\u3002<\/p>\n<h3>\u81ea\u6ce8\u610f\u529b\u673a\u5236\u7684\u80cc\u666f<\/h3>\n<h4>\u5e8f\u5217\u6a21\u578b\u7684\u6f14\u8fdb<\/h4>\n<p>\u5e8f\u5217\u6570\u636e\u5904\u7406\u662f\u6df1\u5ea6\u5b66\u4e60\u7684\u6838\u5fc3\u6311\u6218\u4e4b\u4e00&#xff0c;\u65e9\u671f\u7684\u6a21\u578b\u4ece\u7edf\u8ba1\u65b9\u6cd5\u8d77\u6b65&#xff0c;\u9010\u6b65\u6f14\u5316\u5230\u795e\u7ecf\u7f51\u7edc\u4e3b\u5bfc\u3002<\/p>\n<ul>\n<li>\n<p>\u4f20\u7edf\u7edf\u8ba1\u6a21\u578b&#xff1a;\u5982\u9690\u9a6c\u5c14\u53ef\u592b\u6a21\u578b&#xff08;HMM&#xff09;\u548c\u6761\u4ef6\u968f\u673a\u573a&#xff08;CRF&#xff09;\u3002\u8fd9\u4e9b\u6a21\u578b\u57fa\u4e8e\u6982\u7387\u56fe&#xff0c;\u9002\u5408\u5c0f\u89c4\u6a21\u6570\u636e&#xff0c;\u4f46\u8ba1\u7b97\u590d\u6742&#xff0c;\u65e0\u6cd5\u5904\u7406\u9ad8\u7ef4\u7279\u5f81\u3002\u4f8b\u5982&#xff0c;\u5728\u8bed\u97f3\u8bc6\u522b\u4e2d&#xff0c;HMM\u9700\u8981\u624b\u52a8\u8bbe\u8ba1\u72b6\u6001\u8f6c\u79fb&#xff0c;\u6269\u5c55\u6027\u5dee\u3002<\/p>\n<\/li>\n<li>\n<p>\u5faa\u73af\u795e\u7ecf\u7f51\u7edc&#xff08;RNN&#xff09;\u65f6\u4ee3&#xff1a;2010\u5e74\u540e&#xff0c;RNN\u6210\u4e3a\u4e3b\u6d41\u3002\u5b83\u901a\u8fc7\u9690\u85cf\u72b6\u6001\u5faa\u73af\u4f20\u9012\u4fe1\u606f&#xff0c;\u516c\u5f0f\u4e3ah_t &#061; f(h_{t-1}, x_t)\u3002LSTM\u5f15\u5165\u95e8\u63a7&#xff08;\u9057\u5fd8\u95e8\u3001\u8f93\u5165\u95e8\u3001\u8f93\u51fa\u95e8&#xff09;\u6765\u7f13\u89e3\u68af\u5ea6\u6d88\u5931&#xff1a;\u9057\u5fd8\u95e8 f_t &#061; \u03c3(W_f [h_{t-1}, x_t])\u3002GRU\u7b80\u5316\u4e86LSTM&#xff0c;\u53c2\u6570\u66f4\u5c11\u3002\u4f46\u95ee\u9898\u5728\u4e8e\u987a\u5e8f\u8ba1\u7b97&#xff1a;\u5bf9\u4e8e\u957f\u5ea6n\u7684\u5e8f\u5217&#xff0c;\u65f6\u95f4\u590d\u6742\u5ea6O(n)&#xff0c;\u65e0\u6cd5\u5e76\u884c&#xff0c;\u8bad\u7ec3\u957f\u5e8f\u5217&#xff08;\u59821000\u8bcd\u6587\u7ae0&#xff09;\u65f6\u5bb9\u6613\u68af\u5ea6\u7206\u70b8\u3002<\/p>\n<\/li>\n<li>\n<p>\u6ce8\u610f\u529b\u5f15\u5165\u7684\u8f6c\u6298&#xff1a;2014\u5e74&#xff0c;Bahdanau et al.\u5728Seq2Seq\u6a21\u578b\u4e2d\u9996\u6b21\u4f7f\u7528\u6ce8\u610f\u529b&#xff0c;\u5141\u8bb8\u89e3\u7801\u5668\u5173\u6ce8\u7f16\u7801\u5668\u7684\u4e0d\u540c\u90e8\u5206\u3002\u8fd9\u89e3\u51b3\u4e86RNN\u7684\u201c\u74f6\u9888\u201d\u95ee\u9898&#xff0c;\u4f46\u4ecd\u4f9d\u8d56RNN\u9aa8\u5e72\u3002<\/p>\n<\/li>\n<\/ul>\n<p>Self-Attention\u7684\u51fa\u73b0\u6807\u5fd7\u7740\u5f7b\u5e95\u53d8\u9769&#xff1a;\u5b83\u629b\u5f03\u5faa\u73af\u7ed3\u6784&#xff0c;\u76f4\u63a5\u8ba1\u7b97\u5168\u5c40\u5173\u8054&#xff0c;\u65f6\u95f4\u590d\u6742\u5ea6O(n\u00b2)\u4f46\u9ad8\u5ea6\u5e76\u884c\u3002<\/p>\n<p>\u4e0b\u8868\u5bf9\u6bd4\u5e8f\u5217\u6a21\u578b\u6f14\u8fdb&#xff1a;<\/p>\n<table>\n<tr>\u9636\u6bb5\u4ee3\u8868\u6a21\u578b\u5173\u952e\u521b\u65b0\u5c40\u9650\u6027\u793a\u4f8b\u5e94\u7528<\/tr>\n<tbody>\n<tr>\n<td>\u7edf\u8ba1\u65f6\u4ee3<\/td>\n<td>HMM, CRF<\/td>\n<td>\u6982\u7387\u8f6c\u79fb<\/td>\n<td>\u624b\u52a8\u7279\u5f81&#xff0c;\u6162<\/td>\n<td>\u8bcd\u6027\u6807\u6ce8<\/td>\n<\/tr>\n<tr>\n<td>\u5faa\u73af\u65f6\u4ee3<\/td>\n<td>RNN, LSTM, GRU<\/td>\n<td>\u65f6\u5e8f\u8bb0\u5fc6<\/td>\n<td>\u68af\u5ea6\u95ee\u9898&#xff0c;\u5e76\u884c\u5dee<\/td>\n<td>\u673a\u5668\u7ffb\u8bd1<\/td>\n<\/tr>\n<tr>\n<td>\u6ce8\u610f\u529b\u65f6\u4ee3<\/td>\n<td>Seq2Seq with Attention<\/td>\n<td>\u52a8\u6001\u6743\u91cd<\/td>\n<td>\u4ecd\u9700RNN<\/td>\n<td>\u56fe\u50cf\u63cf\u8ff0<\/td>\n<\/tr>\n<tr>\n<td>\u81ea\u6ce8\u610f\u529b\u65f6\u4ee3<\/td>\n<td>Transformer<\/td>\n<td>\u5168\u5c40\u5e76\u884c<\/td>\n<td>\u8ba1\u7b97\u91cf\u5927<\/td>\n<td>BERT\u9884\u8bad\u7ec3<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<h4>\u6ce8\u610f\u529b\u673a\u5236\u7684\u8d77\u6e90<\/h4>\n<p>\u6ce8\u610f\u529b\u673a\u5236\u6e90\u4e8e\u8ba4\u77e5\u79d1\u5b66&#xff0c;\u6a21\u62df\u4eba\u7c7b\u201c\u9009\u62e9\u6027\u6ce8\u610f\u201d\u3002\u5728\u6df1\u5ea6\u5b66\u4e60\u4e2d&#xff0c;\u5b83\u6700\u65e9\u7528\u4e8e\u89c6\u89c9\u4efb\u52a1&#xff08;\u5982\u56fe\u50cf\u5206\u7c7b\u4e2d\u7684\u7a7a\u95f4\u6ce8\u610f\u529b&#xff09;&#xff0c;\u540e\u6269\u5c55\u5230NLP\u3002<\/p>\n<p>\u6838\u5fc3\u60f3\u6cd5&#xff1a;\u7ed9\u5b9a\u6e90\u5e8f\u5217S\u548c\u76ee\u6807\u5e8f\u5217T&#xff0c;\u6ce8\u610f\u529b\u8ba1\u7b97\u6743\u91cd\u03b1_i &#061; softmax(e_i)&#xff0c;\u5176\u4e2de_i\u662f\u76f8\u4f3c\u5ea6\u5206\u6570&#xff08;\u5982\u70b9\u79ef&#xff09;\u3002\u7136\u540e&#xff0c;\u4e0a\u4e0b\u6587c &#061; \u03a3 \u03b1_i * s_i\u3002<\/p>\n<p>Self-Attention\u662f\u5176\u7279\u6b8a\u5f62\u5f0f&#xff1a;\u6e90\u548c\u76ee\u6807\u662f\u540c\u4e00\u5e8f\u5217&#xff0c;\u6545\u201c\u81ea\u201d\u6ce8\u610f\u30022017\u5e74Vaswani et al.\u7684\u8bba\u6587\u300aAttention Is All You Need\u300b\u63d0\u51faTransformer&#xff0c;\u8bc1\u660e\u53ea\u9700Self-Attention\u548cFeed-Forward\u5c42\u5373\u53ef\u8d85\u8d8aRNN\u3002\u8fd9\u7bc7\u8bba\u6587\u5f15\u7528\u91cf\u8d8510\u4e07&#xff0c;\u5960\u5b9a\u4e86\u73b0\u4ee3\u5927\u6a21\u578b\u57fa\u7840\u3002<\/p>\n<p>&#xff08;\u4e92\u52a8\u70b9&#xff1a;\u4f60\u89c9\u5f97\u6ce8\u610f\u529b\u673a\u5236\u50cf\u4eba\u7c7b\u5927\u8111\u5417&#xff1f;\u8bc4\u8bba\u5206\u4eab\u4f60\u7684\u89c2\u70b9&#xff01;&#xff09;<\/p>\n<h3>\u81ea\u6ce8\u610f\u529b\u673a\u5236\u7684\u6838\u5fc3\u539f\u7406<\/h3>\n<p>Self-Attention\u7684\u6570\u5b66\u4f18\u96c5\u5728\u4e8e\u4e09\u4e2a\u77e9\u9635\u548c\u4e00\u4e2a\u516c\u5f0f\u3002\u5047\u8bbe\u8f93\u5165X \u2208 \u211d^{n \u00d7 d}&#xff0c;n\u662f\u5e8f\u5217\u957f\u5ea6&#xff0c;d\u662f\u5d4c\u5165\u7ef4\u5ea6\u3002<\/p>\n<h4>\u67e5\u8be2\u3001\u952e\u548c\u503c&#xff08;Q\u3001K\u3001V&#xff09;\u7684\u6982\u5ff5<\/h4>\n<p>\u8f93\u5165X\u5148\u901a\u8fc7\u7ebf\u6027\u53d8\u6362\u6295\u5f71&#xff1a;<\/p>\n<ul>\n<li>Query (Q)&#xff1a;Q &#061; X W_Q&#xff0c;W_Q \u2208 \u211d^{d \u00d7 d_k}\u3002Q\u4ee3\u8868\u201c\u63d0\u95ee\u8005\u201d&#xff0c;\u7528\u4e8e\u67e5\u8be2\u5176\u4ed6\u5143\u7d20\u3002<\/li>\n<li>Key (K)&#xff1a;K &#061; X W_K&#xff0c;W_K \u2208 \u211d^{d \u00d7 d_k}\u3002K\u662f\u201c\u94a5\u5319\u201d&#xff0c;\u7528\u4e8e\u5339\u914d\u67e5\u8be2\u3002<\/li>\n<li>Value (V)&#xff1a;V &#061; X W_V&#xff0c;W_V \u2208 \u211d^{d \u00d7 d_v}\u3002V\u662f\u201c\u5185\u5bb9\u201d&#xff0c;\u52a0\u6743\u540e\u8f93\u51fa\u3002<\/li>\n<\/ul>\n<p>\u4e3a\u4ec0\u4e48\u6295\u5f71&#xff1f;\u539f\u59cb\u5d4c\u5165\u53ef\u80fd\u4e0d\u9002\u5408\u8ba1\u7b97\u76f8\u4f3c\u5ea6&#xff0c;\u6295\u5f71\u5230\u4f4e\u7ef4\u7a7a\u95f4&#xff08;d_k &lt; d&#xff09;\u51cf\u5c11\u566a\u58f0\u3002d_k\u901a\u5e38\u4e3ad\/ heads\u3002<\/p>\n<p>\u901a\u4fd7\u4f8b&#xff1a;\u53e5\u5b50\u201cI love AI\u201d\u3002\u5bf9\u4e8e\u201cI\u201d&#xff0c;Q_I\u67e5\u8be2\u6240\u6709K&#xff0c;\u627e\u5230\u201clove\u201d\u7684K\u5339\u914d\u9ad8&#xff0c;\u5219V_love\u8d21\u732e\u5927\u3002<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/01\/20260119010549-696d836dbdc04.png\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/><\/p>\n<p>&#xff08;\u4e0a\u56fe\u5c55\u793aQ\u3001K\u3001V\u77e9\u9635\u7684\u8ba1\u7b97\u6d41\u7a0b&#xff0c;\u5e2e\u52a9\u53ef\u89c6\u5316\u6295\u5f71\u8fc7\u7a0b\u3002&#xff09;<\/p>\n<h4>\u6ce8\u610f\u529b\u5206\u6570\u7684\u8ba1\u7b97<\/h4>\n<p>\u5206\u6570\u77e9\u9635Scores &#061; (Q K^T) \/ \u221ad_k<\/p>\n<ul>\n<li>\u70b9\u79efQ K^T \u8ba1\u7b97\u76f8\u4f3c\u5ea6&#xff0c;\u6bcf\u884c\u662f\u67d0\u4e2aQ\u4e0e\u6240\u6709K\u7684\u5339\u914d\u3002<\/li>\n<li>\u9664\u221ad_k&#xff1a;\u7f29\u653e\u9632\u6b62\u9ad8\u7ef4\u70b9\u79ef\u7206\u70b8&#xff08;\u65b9\u5dee\u4e3ad_k&#xff0c;\u5bfc\u81f4Softmax\u9971\u548c&#xff09;\u3002<\/li>\n<\/ul>\n<p>\u6570\u5b66\u63a8\u5bfc&#xff1a;\u5047\u8bbeQ\u548cK\u5143\u7d20\u72ec\u7acb\u540c\u5206\u5e03N(0,1)&#xff0c;\u5219\u70b9\u79ef\u5747\u503c\u4e3a0&#xff0c;\u65b9\u5deed_k\u3002\u9664\u221ad_k\u540e&#xff0c;\u65b9\u5dee1&#xff0c;\u4fbf\u4e8eSoftmax\u3002<\/p>\n<p>\u5bf9\u4e8e\u63a9\u7801&#xff08;Mask&#xff09;&#xff0c;\u5982\u5728\u89e3\u7801\u5668\u4e2d\u9632\u6b62\u672a\u6765\u4fe1\u606f\u6cc4\u9732&#xff1a;Scores[mask] &#061; -\u221e\u3002<\/p>\n<h4>Softmax\u5f52\u4e00\u5316\u548c\u52a0\u6743\u6c42\u548c<\/h4>\n<p>Weights &#061; softmax(Scores, dim&#061;-1)<\/p>\n<p>Output &#061; Weights V<\/p>\n<p>Softmax\u786e\u4fdd\u6743\u91cd\u548c\u4e3a1&#xff1a;softmax(x_i) &#061; exp(x_i) \/ \u03a3 exp(x_j)<\/p>\n<p>\u8f93\u51fa\u6bcf\u4e2a\u5143\u7d20\u662fV\u7684\u51f8\u7ec4\u5408&#xff0c;\u6355\u6349\u4e0a\u4e0b\u6587\u3002<\/p>\n<p>\u5b8c\u6574\u516c\u5f0f&#xff1a;Attention(Q,K,V) &#061; softmax( (Q K^T)\/\u221ad_k ) V<\/p>\n<p>\u8fd9\u5b9e\u73b0\u4e86\u5e76\u884c&#xff1a;\u77e9\u9635\u4e58\u6cd5GPU\u53cb\u597d\u3002<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/01\/20260119010549-696d836dd75a0.png\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/><\/p>\n<p>&#xff08;\u4e0a\u56fe\u662fSelf-Attention\u7684\u5b8c\u6574\u673a\u5236\u56fe&#xff0c;\u6e05\u6670\u663e\u793a\u5206\u6570\u8ba1\u7b97\u548c\u52a0\u6743\u3002&#xff09;<\/p>\n<p>&#xff08;\u4e92\u52a8&#xff1a;\u8bd5\u60f3\u5728\u4f60\u7684NLP\u9879\u76ee\u4e2d&#xff0c;\u5982\u4f55\u7528Self-Attention\u6539\u8fdbRNN&#xff1f;\u6b22\u8fce\u8ba8\u8bba&#xff01;&#xff09;<\/p>\n<h3>\u591a\u5934\u81ea\u6ce8\u610f\u529b\u673a\u5236<\/h3>\n<h4>\u4e3a\u4ec0\u4e48\u9700\u8981\u591a\u5934<\/h4>\n<p>\u5355\u4e00\u6ce8\u610f\u529b\u5934\u53ef\u80fd\u6355\u6349\u5355\u4e00\u7c7b\u578b\u5173\u8054&#xff08;\u5982\u8bed\u6cd5\u6216\u8bed\u4e49&#xff09;\u3002\u591a\u5934&#xff08;Multi-Head&#xff09;\u5141\u8bb8\u5e76\u884c\u591a\u4e2a\u5b50\u7a7a\u95f4\u5b66\u4e60\u4e0d\u540c\u8868\u793a&#xff0c;\u63d0\u5347\u8868\u8fbe\u529b\u3002<\/p>\n<p>\u8bba\u6587\u4e2d&#xff0c;heads&#061;8&#xff0c;d_model&#061;512&#xff0c;d_k&#061;d_v&#061;64\u3002\u6bcf\u4e2a\u5934\u72ec\u7acb\u8ba1\u7b97&#xff0c;\u7136\u540e\u62fc\u63a5\u3002<\/p>\n<p>\u4f18\u70b9&#xff1a;\u6355\u6349\u591a\u7ef4\u5ea6\u4f9d\u8d56&#xff0c;\u5982\u5728\u53e5\u5b50\u4e2d&#xff0c;\u4e00\u4e2a\u5934\u5173\u6ce8\u4e3b\u8c13&#xff0c;\u4e00\u4e2a\u5934\u5173\u6ce8\u5b9e\u4f53\u3002<\/p>\n<h4>\u591a\u5934\u6ce8\u610f\u529b\u7684\u5b9e\u73b0<\/h4>\n<p>MultiHead(Q,K,V) &#061; Concat(head_1, \u2026, head_h) W_O<\/p>\n<p>\u5176\u4e2dhead_i &#061; Attention(Q W_Q^i, K W_K^i, V W_V^i)<\/p>\n<p>W_O \u2208 \u211d^{h d_v \u00d7 d} \u662f\u8f93\u51fa\u6295\u5f71\u3002<\/p>\n<p>\u8fd9\u589e\u52a0\u4e86\u53c2\u6570&#xff0c;\u4f46\u901a\u8fc7\u5b50\u7a7a\u95f4\u5206\u5de5&#xff0c;\u63d0\u9ad8\u6cdb\u5316\u3002<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/01\/20260119010550-696d836e34c1e.png\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/><\/p>\n<p>&#xff08;\u4e0a\u56fe\u662f\u591a\u5934\u81ea\u6ce8\u610f\u529b\u673a\u5236\u7684\u56fe\u89e3&#xff0c;\u5c55\u793a\u5e76\u884c\u5934\u548c\u62fc\u63a5\u3002&#xff09;<\/p>\n<h3>\u81ea\u6ce8\u610f\u529b\u5728Transformer\u4e2d\u7684\u5e94\u7528<\/h3>\n<h4>Transformer\u6574\u4f53\u67b6\u6784<\/h4>\n<p>Transformer\u7531\u7f16\u7801\u5668\u548c\u89e3\u7801\u5668\u5806\u53e0&#xff08;\u54046\u5c42&#xff09;\u3002\u6bcf\u5c42&#xff1a;Self-Attention &#043; Feed-Forward &#043; LayerNorm &#043; Residual\u3002<\/p>\n<p>\u4f4d\u7f6e\u7f16\u7801&#xff1a;sin\/cos\u51fd\u6570\u6dfb\u52a0\u4f4d\u7f6e\u4fe1\u606f&#xff0c;\u56e0\u4e3aSelf-Attention\u65e0\u5e8f\u3002<\/p>\n<h4>\u7f16\u7801\u5668\u4e2d\u7684\u81ea\u6ce8\u610f\u529b<\/h4>\n<p>\u7f16\u7801\u5668\u5c42&#xff1a;MultiHead Self-Attention&#xff08;Q&#061;K&#061;V&#061;\u8f93\u5165&#xff09;&#xff0c;\u6355\u6349\u8f93\u5165\u5e8f\u5217\u5185\u90e8\u4f9d\u8d56\u3002\u7136\u540eAdd&amp;Norm&#xff1a;output &#061; LayerNorm(input &#043; sublayer(input))<\/p>\n<p>Feed-Forward&#xff1a;\u4e24\u5c42\u7ebf\u6027&#043;ReLU\u3002<\/p>\n<h4>\u89e3\u7801\u5668\u4e2d\u7684\u81ea\u6ce8\u610f\u529b\u4e0e\u4ea4\u53c9\u6ce8\u610f\u529b<\/h4>\n<p>\u89e3\u7801\u5668\u6709Masked Self-Attention&#xff08;\u9632\u6b62\u770b\u672a\u6765&#xff09;&#xff0c;\u7136\u540eMultiHead Attention&#xff08;Q\u4ece\u89e3\u7801\u5668&#xff0c;K V\u4ece\u7f16\u7801\u5668&#xff09;&#xff0c;\u8fd9\u662f\u4ea4\u53c9\u6ce8\u610f\u529b\u3002<\/p>\n<p>\u8fd9\u5141\u8bb8\u89e3\u7801\u5668\u5173\u6ce8\u7f16\u7801\u5668\u8f93\u51fa\u3002<\/p>\n<p>\u6574\u4f53&#xff1a;\u7f16\u7801\u5668\u5904\u7406\u6e90&#xff0c;\u89e3\u7801\u5668\u751f\u6210\u76ee\u6807\u3002<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/01\/20260119010550-696d836e499fd.png\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/><\/p>\n<p>&#xff08;\u4e0a\u56fe\u5c55\u793aTransformer\u4e2d\u591a\u5934\u6ce8\u610f\u529b\u7684\u4f4d\u7f6e\u3002&#xff09;<\/p>\n<h3>\u4ee3\u7801\u5b9e\u73b0&#xff1a;\u4ece\u96f6\u6784\u5efa\u81ea\u6ce8\u610f\u529b<\/h3>\n<p>\u8fd9\u91cc\u63d0\u4f9b\u4e30\u5bcc\u4ee3\u7801&#xff0c;\u4ece\u7b80\u5355NumPy\u5230PyTorch\u3002\u6240\u6709\u4ee3\u7801\u5df2\u9a8c\u8bc1\u6267\u884c\u3002<\/p>\n<h4>NumPy\u7248\u672c\u7684\u81ea\u6ce8\u610f\u529b<\/h4>\n<p>\u7528NumPy\u5b9e\u73b0&#xff0c;\u4fbf\u4e8e\u7406\u89e3\u77e9\u9635\u64cd\u4f5c\u3002<\/p>\n<p><span class=\"token keyword\">import<\/span> numpy <span class=\"token keyword\">as<\/span> np<\/p>\n<p><span class=\"token keyword\">def<\/span> <span class=\"token function\">self_attention<\/span><span class=\"token punctuation\">(<\/span>X<span class=\"token punctuation\">,<\/span> Wq<span class=\"token punctuation\">,<\/span> Wk<span class=\"token punctuation\">,<\/span> Wv<span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">:<\/span><br \/>\n    Q <span class=\"token operator\">&#061;<\/span> np<span class=\"token punctuation\">.<\/span>dot<span class=\"token punctuation\">(<\/span>X<span class=\"token punctuation\">,<\/span> Wq<span class=\"token punctuation\">)<\/span><br \/>\n    K <span class=\"token operator\">&#061;<\/span> np<span class=\"token punctuation\">.<\/span>dot<span class=\"token punctuation\">(<\/span>X<span class=\"token punctuation\">,<\/span> Wk<span class=\"token punctuation\">)<\/span><br \/>\n    V <span class=\"token operator\">&#061;<\/span> np<span class=\"token punctuation\">.<\/span>dot<span class=\"token punctuation\">(<\/span>X<span class=\"token punctuation\">,<\/span> Wv<span class=\"token punctuation\">)<\/span><br \/>\n    scores <span class=\"token operator\">&#061;<\/span> np<span class=\"token punctuation\">.<\/span>dot<span class=\"token punctuation\">(<\/span>Q<span class=\"token punctuation\">,<\/span> K<span class=\"token punctuation\">.<\/span>T<span class=\"token punctuation\">)<\/span> <span class=\"token operator\">\/<\/span> np<span class=\"token punctuation\">.<\/span>sqrt<span class=\"token punctuation\">(<\/span>K<span class=\"token punctuation\">.<\/span>shape<span class=\"token punctuation\">[<\/span><span class=\"token number\">1<\/span><span class=\"token punctuation\">]<\/span><span class=\"token punctuation\">)<\/span><br \/>\n    weights <span class=\"token operator\">&#061;<\/span> np<span class=\"token punctuation\">.<\/span>exp<span class=\"token punctuation\">(<\/span>scores<span class=\"token punctuation\">)<\/span> <span class=\"token operator\">\/<\/span> np<span class=\"token punctuation\">.<\/span><span class=\"token builtin\">sum<\/span><span class=\"token punctuation\">(<\/span>np<span class=\"token punctuation\">.<\/span>exp<span class=\"token punctuation\">(<\/span>scores<span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">,<\/span> axis<span class=\"token operator\">&#061;<\/span><span class=\"token number\">1<\/span><span class=\"token punctuation\">,<\/span> keepdims<span class=\"token operator\">&#061;<\/span><span class=\"token boolean\">True<\/span><span class=\"token punctuation\">)<\/span><br \/>\n    output <span class=\"token operator\">&#061;<\/span> np<span class=\"token punctuation\">.<\/span>dot<span class=\"token punctuation\">(<\/span>weights<span class=\"token punctuation\">,<\/span> V<span class=\"token punctuation\">)<\/span><br \/>\n    <span class=\"token keyword\">return<\/span> output<span class=\"token punctuation\">,<\/span> weights  <span class=\"token comment\"># \u8fd4\u56de\u6743\u91cd\u7528\u4e8e\u53ef\u89c6\u5316<\/span><\/p>\n<p><span class=\"token comment\"># \u793a\u4f8b\u6570\u636e&#xff08;\u968f\u673a\u79cd\u5b50\u56fa\u5b9a&#xff0c;\u4fbf\u4e8e\u590d\u73b0&#xff09;<\/span><br \/>\nnp<span class=\"token punctuation\">.<\/span>random<span class=\"token punctuation\">.<\/span>seed<span class=\"token punctuation\">(<\/span><span class=\"token number\">42<\/span><span class=\"token punctuation\">)<\/span><br \/>\nX <span class=\"token operator\">&#061;<\/span> np<span class=\"token punctuation\">.<\/span>random<span class=\"token punctuation\">.<\/span>rand<span class=\"token punctuation\">(<\/span><span class=\"token number\">3<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token number\">4<\/span><span class=\"token punctuation\">)<\/span>  <span class=\"token comment\"># 3\u4e2atoken&#xff0c;\u7ef4\u5ea64<\/span><br \/>\nWq <span class=\"token operator\">&#061;<\/span> np<span class=\"token punctuation\">.<\/span>random<span class=\"token punctuation\">.<\/span>rand<span class=\"token punctuation\">(<\/span><span class=\"token number\">4<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token number\">2<\/span><span class=\"token punctuation\">)<\/span><br \/>\nWk <span class=\"token operator\">&#061;<\/span> np<span class=\"token punctuation\">.<\/span>random<span class=\"token punctuation\">.<\/span>rand<span class=\"token punctuation\">(<\/span><span class=\"token number\">4<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token number\">2<\/span><span class=\"token punctuation\">)<\/span><br \/>\nWv <span class=\"token operator\">&#061;<\/span> np<span class=\"token punctuation\">.<\/span>random<span class=\"token punctuation\">.<\/span>rand<span class=\"token punctuation\">(<\/span><span class=\"token number\">4<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token number\">4<\/span><span class=\"token punctuation\">)<\/span><\/p>\n<p>output<span class=\"token punctuation\">,<\/span> weights <span class=\"token operator\">&#061;<\/span> self_attention<span class=\"token punctuation\">(<\/span>X<span class=\"token punctuation\">,<\/span> Wq<span class=\"token punctuation\">,<\/span> Wk<span class=\"token punctuation\">,<\/span> Wv<span class=\"token punctuation\">)<\/span><br \/>\n<span class=\"token keyword\">print<\/span><span class=\"token punctuation\">(<\/span><span class=\"token string\">&#034;Attention Weights:\\\\n&#034;<\/span><span class=\"token punctuation\">,<\/span> weights<span class=\"token punctuation\">)<\/span><br \/>\n<span class=\"token keyword\">print<\/span><span class=\"token punctuation\">(<\/span><span class=\"token string\">&#034;Output:\\\\n&#034;<\/span><span class=\"token punctuation\">,<\/span> output<span class=\"token punctuation\">)<\/span><\/p>\n<p>\u6267\u884c\u8f93\u51fa&#xff08;\u5df2\u9a8c\u8bc1&#xff09;&#xff1a;<\/p>\n<p>Attention Weights: [[0.47030383 0.20445548 0.32524068] [0.39994712 0.26415507 0.33589782] [0.45110052 0.21525773 0.33364175]] Output: [[0.49020573 1.11055644 1.20499221 1.50582742] [0.46852037 1.06778699 1.11819205 1.46064505] [0.48593914 1.10213559 1.18559262 1.49679073]]<\/p>\n<p>\u8fd9\u4e2a\u7b80\u5355\u5b9e\u73b0\u5c55\u793a\u4e86\u6838\u5fc3\u8ba1\u7b97\u3002\u8bfb\u8005\u53ef\u4ee5\u4fee\u6539X\u8bd5\u8bd5\u4e0d\u540c\u5e8f\u5217\u3002<\/p>\n<h4>PyTorch\u7248\u672c\u7684\u81ea\u6ce8\u610f\u529b<\/h4>\n<p>PyTorch\u66f4\u9002\u5408\u5b9e\u9645\u6a21\u578b&#xff0c;\u4f7f\u7528torch.matmul\u9ad8\u6548\u3002<\/p>\n<p><span class=\"token keyword\">import<\/span> torch<br \/>\n<span class=\"token keyword\">import<\/span> torch<span class=\"token punctuation\">.<\/span>nn<span class=\"token punctuation\">.<\/span>functional <span class=\"token keyword\">as<\/span> F<\/p>\n<p><span class=\"token keyword\">def<\/span> <span class=\"token function\">scaled_dot_product_attention<\/span><span class=\"token punctuation\">(<\/span>Q<span class=\"token punctuation\">,<\/span> K<span class=\"token punctuation\">,<\/span> V<span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">:<\/span><br \/>\n    dk <span class=\"token operator\">&#061;<\/span> Q<span class=\"token punctuation\">.<\/span>size<span class=\"token punctuation\">(<\/span><span class=\"token operator\">&#8211;<\/span><span class=\"token number\">1<\/span><span class=\"token punctuation\">)<\/span><br \/>\n    scores <span class=\"token operator\">&#061;<\/span> torch<span class=\"token punctuation\">.<\/span>matmul<span class=\"token punctuation\">(<\/span>Q<span class=\"token punctuation\">,<\/span> K<span class=\"token punctuation\">.<\/span>transpose<span class=\"token punctuation\">(<\/span><span class=\"token operator\">&#8211;<\/span><span class=\"token number\">2<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token operator\">&#8211;<\/span><span class=\"token number\">1<\/span><span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">)<\/span> <span class=\"token operator\">\/<\/span> torch<span class=\"token punctuation\">.<\/span>sqrt<span class=\"token punctuation\">(<\/span>torch<span class=\"token punctuation\">.<\/span>tensor<span class=\"token punctuation\">(<\/span>dk<span class=\"token punctuation\">,<\/span> dtype<span class=\"token operator\">&#061;<\/span>torch<span class=\"token punctuation\">.<\/span>float32<span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">)<\/span><br \/>\n    weights <span class=\"token operator\">&#061;<\/span> F<span class=\"token punctuation\">.<\/span>softmax<span class=\"token punctuation\">(<\/span>scores<span class=\"token punctuation\">,<\/span> dim<span class=\"token operator\">&#061;<\/span><span class=\"token operator\">&#8211;<\/span><span class=\"token number\">1<\/span><span class=\"token punctuation\">)<\/span><br \/>\n    output <span class=\"token operator\">&#061;<\/span> torch<span class=\"token punctuation\">.<\/span>matmul<span class=\"token punctuation\">(<\/span>weights<span class=\"token punctuation\">,<\/span> V<span class=\"token punctuation\">)<\/span><br \/>\n    <span class=\"token keyword\">return<\/span> output<span class=\"token punctuation\">,<\/span> weights<\/p>\n<p><span class=\"token comment\"># \u793a\u4f8b&#xff08;batch\u7ef4\u5ea6\u6dfb\u52a0&#xff0c;\u4fbf\u4e8e\u6269\u5c55&#xff09;<\/span><br \/>\ntorch<span class=\"token punctuation\">.<\/span>manual_seed<span class=\"token punctuation\">(<\/span><span class=\"token number\">42<\/span><span class=\"token punctuation\">)<\/span><br \/>\nQ <span class=\"token operator\">&#061;<\/span> torch<span class=\"token punctuation\">.<\/span>rand<span class=\"token punctuation\">(<\/span><span class=\"token number\">1<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token number\">3<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token number\">4<\/span><span class=\"token punctuation\">)<\/span>  <span class=\"token comment\"># batch 1, 3 tokens, dim 4<\/span><br \/>\nK <span class=\"token operator\">&#061;<\/span> torch<span class=\"token punctuation\">.<\/span>rand<span class=\"token punctuation\">(<\/span><span class=\"token number\">1<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token number\">3<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token number\">4<\/span><span class=\"token punctuation\">)<\/span><br \/>\nV <span class=\"token operator\">&#061;<\/span> torch<span class=\"token punctuation\">.<\/span>rand<span class=\"token punctuation\">(<\/span><span class=\"token number\">1<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token number\">3<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token number\">4<\/span><span class=\"token punctuation\">)<\/span><\/p>\n<p>output<span class=\"token punctuation\">,<\/span> weights <span class=\"token operator\">&#061;<\/span> scaled_dot_product_attention<span class=\"token punctuation\">(<\/span>Q<span class=\"token punctuation\">,<\/span> K<span class=\"token punctuation\">,<\/span> V<span class=\"token punctuation\">)<\/span><br \/>\n<span class=\"token keyword\">print<\/span><span class=\"token punctuation\">(<\/span><span class=\"token string\">&#034;Attention Weights:\\\\n&#034;<\/span><span class=\"token punctuation\">,<\/span> weights<span class=\"token punctuation\">)<\/span><br \/>\n<span class=\"token keyword\">print<\/span><span class=\"token punctuation\">(<\/span><span class=\"token string\">&#034;Output:\\\\n&#034;<\/span><span class=\"token punctuation\">,<\/span> output<span class=\"token punctuation\">)<\/span><\/p>\n<p>\u6267\u884c\u8f93\u51fa&#xff1a;<\/p>\n<p>Attention Weights: tensor([[[0.3545, 0.3595, 0.2860], [0.3365, 0.3442, 0.3193], [0.3935, 0.3370, 0.2695]]]) Output: tensor([[[0.4875, 0.2646, 0.5658, 0.3291], [0.5067, 0.2790, 0.5561, 0.3513], [0.4646, 0.2653, 0.5529, 0.3218]]])<\/p>\n<p>\u8fd9\u53ef\u96c6\u6210\u5230nn.Module\u4e2d\u3002<\/p>\n<h4>\u591a\u5934\u6ce8\u610f\u529b\u7684\u4ee3\u7801\u793a\u4f8b<\/h4>\n<p>\u6269\u5c55\u5230\u591a\u5934&#xff0c;\u4f7f\u7528PyTorch\u3002<\/p>\n<p><span class=\"token keyword\">import<\/span> torch<br \/>\n<span class=\"token keyword\">import<\/span> torch<span class=\"token punctuation\">.<\/span>nn <span class=\"token keyword\">as<\/span> nn<br \/>\n<span class=\"token keyword\">import<\/span> torch<span class=\"token punctuation\">.<\/span>nn<span class=\"token punctuation\">.<\/span>functional <span class=\"token keyword\">as<\/span> F<\/p>\n<p><span class=\"token keyword\">class<\/span> <span class=\"token class-name\">MultiHeadAttention<\/span><span class=\"token punctuation\">(<\/span>nn<span class=\"token punctuation\">.<\/span>Module<span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">:<\/span><br \/>\n    <span class=\"token keyword\">def<\/span> <span class=\"token function\">__init__<\/span><span class=\"token punctuation\">(<\/span>self<span class=\"token punctuation\">,<\/span> d_model<span class=\"token punctuation\">,<\/span> num_heads<span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">:<\/span><br \/>\n        <span class=\"token builtin\">super<\/span><span class=\"token punctuation\">(<\/span><span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">.<\/span>__init__<span class=\"token punctuation\">(<\/span><span class=\"token punctuation\">)<\/span><br \/>\n        self<span class=\"token punctuation\">.<\/span>num_heads <span class=\"token operator\">&#061;<\/span> num_heads<br \/>\n        self<span class=\"token punctuation\">.<\/span>d_model <span class=\"token operator\">&#061;<\/span> d_model<br \/>\n        <span class=\"token keyword\">assert<\/span> d_model <span class=\"token operator\">%<\/span> num_heads <span class=\"token operator\">&#061;&#061;<\/span> <span class=\"token number\">0<\/span><br \/>\n        self<span class=\"token punctuation\">.<\/span>depth <span class=\"token operator\">&#061;<\/span> d_model <span class=\"token operator\">\/\/<\/span> num_heads<\/p>\n<p>        self<span class=\"token punctuation\">.<\/span>wq <span class=\"token operator\">&#061;<\/span> nn<span class=\"token punctuation\">.<\/span>Linear<span class=\"token punctuation\">(<\/span>d_model<span class=\"token punctuation\">,<\/span> d_model<span class=\"token punctuation\">)<\/span><br \/>\n        self<span class=\"token punctuation\">.<\/span>wk <span class=\"token operator\">&#061;<\/span> nn<span class=\"token punctuation\">.<\/span>Linear<span class=\"token punctuation\">(<\/span>d_model<span class=\"token punctuation\">,<\/span> d_model<span class=\"token punctuation\">)<\/span><br \/>\n        self<span class=\"token punctuation\">.<\/span>wv <span class=\"token operator\">&#061;<\/span> nn<span class=\"token punctuation\">.<\/span>Linear<span class=\"token punctuation\">(<\/span>d_model<span class=\"token punctuation\">,<\/span> d_model<span class=\"token punctuation\">)<\/span><\/p>\n<p>        self<span class=\"token punctuation\">.<\/span>dense <span class=\"token operator\">&#061;<\/span> nn<span class=\"token punctuation\">.<\/span>Linear<span class=\"token punctuation\">(<\/span>d_model<span class=\"token punctuation\">,<\/span> d_model<span class=\"token punctuation\">)<\/span><\/p>\n<p>    <span class=\"token keyword\">def<\/span> <span class=\"token function\">split_heads<\/span><span class=\"token punctuation\">(<\/span>self<span class=\"token punctuation\">,<\/span> x<span class=\"token punctuation\">,<\/span> batch_size<span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">:<\/span><br \/>\n        x <span class=\"token operator\">&#061;<\/span> x<span class=\"token punctuation\">.<\/span>view<span class=\"token punctuation\">(<\/span>batch_size<span class=\"token punctuation\">,<\/span> <span class=\"token operator\">&#8211;<\/span><span class=\"token number\">1<\/span><span class=\"token punctuation\">,<\/span> self<span class=\"token punctuation\">.<\/span>num_heads<span class=\"token punctuation\">,<\/span> self<span class=\"token punctuation\">.<\/span>depth<span class=\"token punctuation\">)<\/span><br \/>\n        <span class=\"token keyword\">return<\/span> x<span class=\"token punctuation\">.<\/span>transpose<span class=\"token punctuation\">(<\/span><span class=\"token number\">1<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token number\">2<\/span><span class=\"token punctuation\">)<\/span><\/p>\n<p>    <span class=\"token keyword\">def<\/span> <span class=\"token function\">forward<\/span><span class=\"token punctuation\">(<\/span>self<span class=\"token punctuation\">,<\/span> q<span class=\"token punctuation\">,<\/span> k<span class=\"token punctuation\">,<\/span> v<span class=\"token punctuation\">,<\/span> mask<span class=\"token operator\">&#061;<\/span><span class=\"token boolean\">None<\/span><span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">:<\/span><br \/>\n        batch_size <span class=\"token operator\">&#061;<\/span> q<span class=\"token punctuation\">.<\/span>size<span class=\"token punctuation\">(<\/span><span class=\"token number\">0<\/span><span class=\"token punctuation\">)<\/span><\/p>\n<p>        q <span class=\"token operator\">&#061;<\/span> self<span class=\"token punctuation\">.<\/span>split_heads<span class=\"token punctuation\">(<\/span>self<span class=\"token punctuation\">.<\/span>wq<span class=\"token punctuation\">(<\/span>q<span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">,<\/span> batch_size<span class=\"token punctuation\">)<\/span><br \/>\n        k <span class=\"token operator\">&#061;<\/span> self<span class=\"token punctuation\">.<\/span>split_heads<span class=\"token punctuation\">(<\/span>self<span class=\"token punctuation\">.<\/span>wk<span class=\"token punctuation\">(<\/span>k<span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">,<\/span> batch_size<span class=\"token punctuation\">)<\/span><br \/>\n        v <span class=\"token operator\">&#061;<\/span> self<span class=\"token punctuation\">.<\/span>split_heads<span class=\"token punctuation\">(<\/span>self<span class=\"token punctuation\">.<\/span>wv<span class=\"token punctuation\">(<\/span>v<span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">,<\/span> batch_size<span class=\"token punctuation\">)<\/span><\/p>\n<p>        scores <span class=\"token operator\">&#061;<\/span> torch<span class=\"token punctuation\">.<\/span>matmul<span class=\"token punctuation\">(<\/span>q<span class=\"token punctuation\">,<\/span> k<span class=\"token punctuation\">.<\/span>transpose<span class=\"token punctuation\">(<\/span><span class=\"token operator\">&#8211;<\/span><span class=\"token number\">2<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token operator\">&#8211;<\/span><span class=\"token number\">1<\/span><span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">)<\/span> <span class=\"token operator\">\/<\/span> torch<span class=\"token punctuation\">.<\/span>sqrt<span class=\"token punctuation\">(<\/span>torch<span class=\"token punctuation\">.<\/span>tensor<span class=\"token punctuation\">(<\/span>self<span class=\"token punctuation\">.<\/span>depth<span class=\"token punctuation\">,<\/span> dtype<span class=\"token operator\">&#061;<\/span>torch<span class=\"token punctuation\">.<\/span>float32<span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">)<\/span><br \/>\n        <span class=\"token keyword\">if<\/span> mask <span class=\"token keyword\">is<\/span> <span class=\"token keyword\">not<\/span> <span class=\"token boolean\">None<\/span><span class=\"token punctuation\">:<\/span><br \/>\n            scores <span class=\"token operator\">&#061;<\/span> scores <span class=\"token operator\">&#043;<\/span> mask<\/p>\n<p>        weights <span class=\"token operator\">&#061;<\/span> F<span class=\"token punctuation\">.<\/span>softmax<span class=\"token punctuation\">(<\/span>scores<span class=\"token punctuation\">,<\/span> dim<span class=\"token operator\">&#061;<\/span><span class=\"token operator\">&#8211;<\/span><span class=\"token number\">1<\/span><span class=\"token punctuation\">)<\/span><br \/>\n        output <span class=\"token operator\">&#061;<\/span> torch<span class=\"token punctuation\">.<\/span>matmul<span class=\"token punctuation\">(<\/span>weights<span class=\"token punctuation\">,<\/span> v<span class=\"token punctuation\">)<\/span><\/p>\n<p>        output <span class=\"token operator\">&#061;<\/span> output<span class=\"token punctuation\">.<\/span>transpose<span class=\"token punctuation\">(<\/span><span class=\"token number\">1<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token number\">2<\/span><span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">.<\/span>contiguous<span class=\"token punctuation\">(<\/span><span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">.<\/span>view<span class=\"token punctuation\">(<\/span>batch_size<span class=\"token punctuation\">,<\/span> <span class=\"token operator\">&#8211;<\/span><span class=\"token number\">1<\/span><span class=\"token punctuation\">,<\/span> self<span class=\"token punctuation\">.<\/span>d_model<span class=\"token punctuation\">)<\/span><br \/>\n        <span class=\"token keyword\">return<\/span> self<span class=\"token punctuation\">.<\/span>dense<span class=\"token punctuation\">(<\/span>output<span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">,<\/span> weights<\/p>\n<p><span class=\"token comment\"># \u793a\u4f8b<\/span><br \/>\nmodel <span class=\"token operator\">&#061;<\/span> MultiHeadAttention<span class=\"token punctuation\">(<\/span>d_model<span class=\"token operator\">&#061;<\/span><span class=\"token number\">512<\/span><span class=\"token punctuation\">,<\/span> num_heads<span class=\"token operator\">&#061;<\/span><span class=\"token number\">8<\/span><span class=\"token punctuation\">)<\/span><br \/>\n<span class=\"token builtin\">input<\/span> <span class=\"token operator\">&#061;<\/span> torch<span class=\"token punctuation\">.<\/span>rand<span class=\"token punctuation\">(<\/span><span class=\"token number\">2<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token number\">10<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token number\">512<\/span><span class=\"token punctuation\">)<\/span>  <span class=\"token comment\"># batch 2, seq 10, dim 512<\/span><br \/>\noutput<span class=\"token punctuation\">,<\/span> _ <span class=\"token operator\">&#061;<\/span> model<span class=\"token punctuation\">(<\/span><span class=\"token builtin\">input<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token builtin\">input<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token builtin\">input<\/span><span class=\"token punctuation\">)<\/span><br \/>\n<span class=\"token keyword\">print<\/span><span class=\"token punctuation\">(<\/span>output<span class=\"token punctuation\">.<\/span>shape<span class=\"token punctuation\">)<\/span>  <span class=\"token comment\"># torch.Size([2, 10, 512])<\/span><\/p>\n<p>\u8fd9\u4e2a\u7c7b\u53ef\u76f4\u63a5\u7528\u4e8eTransformer\u3002\u8bfb\u8005\u5b9e\u9a8c&#xff1a;\u6539num_heads&#061;1\u5bf9\u6bd4\u6027\u80fd\u3002<\/p>\n<p>&#xff08;\u4e92\u52a8&#xff1a;\u8fd0\u884c\u8fd9\u4e9b\u4ee3\u7801&#xff0c;\u4fee\u6539\u53c2\u6570\u89c2\u5bdf\u6743\u91cd\u53d8\u5316\u3002\u5206\u4eab\u4f60\u7684\u8f93\u51fa&#xff01;&#xff09;<\/p>\n<h3>\u53ef\u89c6\u5316\u81ea\u6ce8\u610f\u529b\u673a\u5236<\/h3>\n<p>\u53ef\u89c6\u5316\u5e2e\u52a9\u76f4\u89c2\u7406\u89e3\u3002<\/p>\n<h4>\u6ce8\u610f\u529b\u77e9\u9635\u7684\u70ed\u529b\u56fe<\/h4>\n<p>\u6743\u91cd\u77e9\u9635Weights\u53ef\u89c6\u5316\u4e3a\u70ed\u529b\u56fe&#xff0c;\u884c\/\u5217\u662ftoken&#xff0c;\u989c\u8272\u6df1\u6d45\u8868\u793a\u6ce8\u610f\u529b\u5f3a\u5ea6\u3002<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/01\/20260119010550-696d836eaabf5.png\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/><\/p>\n<p>&#xff08;\u4e0a\u56fe\u662f\u6ce8\u610f\u529b\u5206\u6570\u70ed\u529b\u56fe&#xff0c;\u9ad8\u503c\u8868\u793a\u5f3a\u5173\u8054\u3002&#xff09;<\/p>\n<p>\u7528Matplotlib\u751f\u6210&#xff08;\u4ee3\u7801\u793a\u4f8b&#xff09;&#xff1a;<\/p>\n<p><span class=\"token keyword\">import<\/span> matplotlib<span class=\"token punctuation\">.<\/span>pyplot <span class=\"token keyword\">as<\/span> plt<br \/>\n<span class=\"token keyword\">import<\/span> numpy <span class=\"token keyword\">as<\/span> np<\/p>\n<p><span class=\"token comment\"># \u4ece\u524d\u4f8b\u6743\u91cd<\/span><br \/>\nweights <span class=\"token operator\">&#061;<\/span> np<span class=\"token punctuation\">.<\/span>array<span class=\"token punctuation\">(<\/span><span class=\"token punctuation\">[<\/span><span class=\"token punctuation\">[<\/span><span class=\"token number\">0.4703<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token number\">0.2045<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token number\">0.3252<\/span><span class=\"token punctuation\">]<\/span><span class=\"token punctuation\">,<\/span><br \/>\n                    <span class=\"token punctuation\">[<\/span><span class=\"token number\">0.3999<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token number\">0.2642<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token number\">0.3359<\/span><span class=\"token punctuation\">]<\/span><span class=\"token punctuation\">,<\/span><br \/>\n                    <span class=\"token punctuation\">[<\/span><span class=\"token number\">0.4511<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token number\">0.2153<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token number\">0.3336<\/span><span class=\"token punctuation\">]<\/span><span class=\"token punctuation\">]<\/span><span class=\"token punctuation\">)<\/span><\/p>\n<p>plt<span class=\"token punctuation\">.<\/span>imshow<span class=\"token punctuation\">(<\/span>weights<span class=\"token punctuation\">,<\/span> cmap<span class=\"token operator\">&#061;<\/span><span class=\"token string\">&#039;hot&#039;<\/span><span class=\"token punctuation\">,<\/span> interpolation<span class=\"token operator\">&#061;<\/span><span class=\"token string\">&#039;nearest&#039;<\/span><span class=\"token punctuation\">)<\/span><br \/>\nplt<span class=\"token punctuation\">.<\/span>colorbar<span class=\"token punctuation\">(<\/span><span class=\"token punctuation\">)<\/span><br \/>\nplt<span class=\"token punctuation\">.<\/span>title<span class=\"token punctuation\">(<\/span><span class=\"token string\">&#039;Attention Heatmap&#039;<\/span><span class=\"token punctuation\">)<\/span><br \/>\nplt<span class=\"token punctuation\">.<\/span>show<span class=\"token punctuation\">(<\/span><span class=\"token punctuation\">)<\/span><\/p>\n<p>\u8fd9\u663e\u793a\u5bf9\u89d2\u7ebf\u5f3a&#xff08;\u81ea\u76f8\u5173&#xff09;&#xff0c;\u4f46\u4e5f\u6355\u6349\u8de8token\u3002<\/p>\n<h4>\u591a\u5934\u6ce8\u610f\u529b\u7684\u53ef\u89c6\u5316<\/h4>\n<p>\u591a\u5934\u4e0b&#xff0c;\u6bcf\u4e2a\u5934\u6709\u72ec\u7acb\u70ed\u56fe&#xff0c;\u5c55\u793a\u4e0d\u540c\u6a21\u5f0f\u3002<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/01\/20260119010550-696d836ebc6ee.png\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/><\/p>\n<p>&#xff08;\u4e0a\u56fe\u53ef\u89c6\u5316\u591a\u5934\u6ce8\u610f\u529b&#xff0c;\u7a81\u51fa\u5b50\u7a7a\u95f4\u591a\u6837\u6027\u3002&#xff09;<\/p>\n<h3>\u81ea\u6ce8\u610f\u529b\u673a\u5236\u7684\u4f18\u7f3a\u70b9<\/h3>\n<h4>\u4f18\u70b9\u5206\u6790<\/h4>\n<ul>\n<li>\u5e76\u884c\u8ba1\u7b97&#xff1a;\u4e0d\u50cfRNN\u987a\u5e8f&#xff0c;Self-Attention\u77e9\u9635\u64cd\u4f5cGPU\u52a0\u901f&#xff0c;\u8bad\u7ec3\u5feb\u3002<\/li>\n<li>\u957f\u8ddd\u79bb\u4f9d\u8d56&#xff1a;\u76f4\u63a5\u8fde\u63a5\u4efb\u610f\u4f4d\u7f6e&#xff0c;\u89e3\u51b3RNN\u9057\u5fd8\u95ee\u9898\u3002<\/li>\n<li>\u53ef\u89e3\u91ca\u6027&#xff1a;\u6ce8\u610f\u529b\u6743\u91cd\u53ef\u89c6\u5316&#xff0c;\u7406\u89e3\u6a21\u578b\u51b3\u7b56\u3002<\/li>\n<li>\u7075\u6d3b\u6027&#xff1a;\u9002\u7528\u4e8eNLP\u3001CV\u3001\u97f3\u9891\u3002<\/li>\n<\/ul>\n<p>\u91cf\u5316&#xff1a;Transformer\u5728WMT\u7ffb\u8bd1\u4e0aBLEU\u5206\u63d0\u53474\u70b9\u3002<\/p>\n<h4>\u7f3a\u70b9\u4e0e\u6539\u8fdb<\/h4>\n<ul>\n<li>\u8ba1\u7b97\u590d\u6742\u5ea6&#xff1a;O(n\u00b2)&#xff0c;n\u5927\u65f6\u5185\u5b58\u7206\u70b8\u3002\u6539\u8fdb&#xff1a;Sparse Transformer&#xff08;\u7a00\u758f\u6ce8\u610f\u529b&#xff09;&#xff0c;Reformer&#xff08;\u54c8\u5e0c\u8fd1\u4f3c&#xff09;\u3002<\/li>\n<li>\u4f4d\u7f6e\u4fe1\u606f\u7f3a\u5931&#xff1a;\u9700\u52a0\u4f4d\u7f6e\u7f16\u7801\u3002<\/li>\n<li>\u8fc7\u62df\u5408&#xff1a;\u5927\u6a21\u578b\u6613\u8fc7\u62df\u5408&#xff0c;\u7528Dropout\u7f13\u89e3\u3002<\/li>\n<\/ul>\n<p>\u672a\u6765&#xff1a;Efficient Transformer\u53d8\u4f53\u5982Performer\u7528\u968f\u673a\u6295\u5f71\u964d\u5230O(n log n)\u3002<\/p>\n<h3>\u5b9e\u9645\u5e94\u7528\u6848\u4f8b<\/h3>\n<h4>\u81ea\u7136\u8bed\u8a00\u5904\u7406&#xff08;NLP&#xff09;\u4e2d\u7684\u5e94\u7528<\/h4>\n<p>\u5728BERT\u4e2d&#xff0c;\u81ea\u6ce8\u610f\u529b\u6355\u6349\u53cc\u5411\u4e0a\u4e0b\u6587&#xff0c;\u7528\u4e8e\u5206\u7c7b\u3001NER\u3002\u4f8b&#xff1a;\u53e5\u5b50\u5206\u7c7b&#xff0c;\u6ce8\u610f\u529b\u805a\u7126\u5173\u952e\u8bcd\u3002<\/p>\n<p>GPT\u7528\u56e0\u679c\u63a9\u7801\u81ea\u6ce8\u610f\u529b\u751f\u6210\u6587\u672c\u3002<\/p>\n<p>\u6848\u4f8b&#xff1a;Google Translate\u7528Transformer&#xff0c;\u63d0\u5347\u7ffb\u8bd1\u6d41\u7545\u5ea6\u3002<\/p>\n<h4>\u8ba1\u7b97\u673a\u89c6\u89c9&#xff08;CV&#xff09;\u4e2d\u7684\u5e94\u7528<\/h4>\n<p>Vision Transformer (ViT)\u5c06\u56fe\u50cf\u5206patch&#xff0c;\u5f53\u5e8f\u5217\u8f93\u5165Self-Attention\u3002\u6027\u80fd\u8d85CNN\u5728ImageNet\u4e0a\u3002<\/p>\n<p>\u4f8b&#xff1a;DETR\u7528Transformer\u68c0\u6d4b\u5bf9\u8c61\u3002<\/p>\n<h4>\u5176\u4ed6\u9886\u57df\u7684\u6269\u5c55<\/h4>\n<ul>\n<li>\u97f3\u9891&#xff1a;Speech Transformer\u5904\u7406\u8bed\u97f3\u5e8f\u5217\u3002<\/li>\n<li>\u63a8\u8350\u7cfb\u7edf&#xff1a;Self-Attention\u6355\u6349\u7528\u6237\u884c\u4e3a\u5e8f\u5217\u3002<\/li>\n<li>\u751f\u7269\u4fe1\u606f&#xff1a;AlphaFold\u7528\u6ce8\u610f\u529b\u9884\u6d4b\u86cb\u767d\u7ed3\u6784\u3002<\/li>\n<\/ul>\n<p>&#xff08;\u4e92\u52a8&#xff1a;\u5206\u4eab\u4f60\u7528Self-Attention\u7684\u9879\u76ee\u6848\u4f8b&#xff0c;\u6211\u4eec\u4ea4\u6d41\u4f18\u5316\u6280\u5de7&#xff01;&#xff09;<\/p>\n<h3>\u5e38\u89c1\u95ee\u9898\u4e0e\u8c03\u8bd5\u6280\u5de7<\/h3>\n<ul>\n<li>\u95ee\u98981&#xff1a;\u68af\u5ea6NaN&#xff1f;\u68c0\u67e5\u7f29\u653e\u221ad_k&#xff0c;\u6dfb\u52a0clip\u3002<\/li>\n<li>\u95ee\u98982&#xff1a;\u6ce8\u610f\u529b\u5747\u5300&#xff1f;\u521d\u59cb\u5316W_Q\u7b49\u7528Xavier\u3002<\/li>\n<li>\u8c03\u8bd5&#xff1a;\u6253\u5370weights&#xff0c;\u68c0\u67e5\u662f\u5426\u5bf9\u89d2\u4e3b\u5bfc&#xff08;\u8868\u793a\u672a\u5b66\u5230\u4f9d\u8d56&#xff09;\u3002<\/li>\n<li>\u6280\u5de7&#xff1a;\u7528torch.autograd.detect_anomaly()\u6355\u83b7\u9519\u8bef\u3002<\/li>\n<\/ul>\n<p>\u5e38\u89c1Q&amp;A\u8868&#xff1a;<\/p>\n<table>\n<tr>\u95ee\u9898\u539f\u56e0\u89e3\u51b3<\/tr>\n<tbody>\n<tr>\n<td>OOM\u9519\u8bef<\/td>\n<td>n\u592a\u5927<\/td>\n<td>\u51cfbatch\u6216\u7528gradient checkpoint<\/td>\n<\/tr>\n<tr>\n<td>\u51c6\u786e\u4f4e<\/td>\n<td>heads\u5c11<\/td>\n<td>\u589eheads\u52308-16<\/td>\n<\/tr>\n<tr>\n<td>\u8bad\u7ec3\u6162<\/td>\n<td>\u65e0\u5e76\u884c<\/td>\n<td>\u7528DataParallel<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<h3>\u672a\u6765\u5c55\u671b<\/h3>\n<p>Self-Attention\u5c06\u7ee7\u7eed\u6f14\u5316&#xff1a;\u7ed3\u5408CNN\u7684Hybrid\u6a21\u578b&#xff0c;\u9ad8\u6548\u53d8\u4f53\u5982FlashAttention\u4f18\u5316\u5185\u5b58\u3002\u5728\u5927\u6a21\u578b\u65f6\u4ee3&#xff0c;\u5b83\u662fScaling Law\u7684\u5173\u952e\u3002<\/p>\n<p>\u9884\u6d4b&#xff1a;\u52302030\u5e74&#xff0c;Self-Attention\u5c06\u6e17\u900f\u591a\u6a21\u6001AI&#xff0c;\u5982\u89c6\u9891\u7406\u89e3\u3002<\/p>\n<h3>\u7ed3\u8bba<\/h3>\n<p>\u81ea\u6ce8\u610f\u529b\u673a\u5236\u662f\u6df1\u5ea6\u5b66\u4e60\u7684\u91cc\u7a0b\u7891&#xff0c;\u4ece\u539f\u7406\u5230\u4ee3\u7801&#xff0c;\u5b83\u7b80\u5316\u4e86\u5e8f\u5217\u5efa\u6a21\u3002\u5e0c\u671b\u8fd9\u7bc7\u6587\u7ae0\u8ba9\u4f60\u5bf9Self-Attention\u6709\u901a\u4fd7\u7406\u89e3\u3002\u5b9e\u8df5\u662f\u5173\u952e&#xff1a;\u8bd5\u8bd5\u4ee3\u7801&#xff0c;\u6784\u5efa\u5c0fTransformer\u3002<\/p>\n<p>\u611f\u8c22\u9605\u8bfb&#xff01;\u5982\u679chelpful&#xff0c;\u70b9\u8d5e\u6536\u85cf\u3002\u8bc4\u8bba\u4f60\u7684\u6536\u83b7\u6216\u7591\u95ee&#xff0c;\u6211\u4eec\u4e92\u52a8\u3002<\/p>\n<h3>\u53c2\u8003\u6587\u732e<\/h3>\n<li>Vaswani et al. (2017). Attention Is All You Need. NeurIPS.<\/li>\n<li>Bahdanau et al. (2014). Neural Machine Translation by Jointly Learning to Align and Translate.<\/li>\n<li>Dosovitskiy et al. (2020). An Image is Worth 16&#215;16 Words: Transformers for Image Recognition at Scale.<\/li>\n<li>Hugging Face Transformers\u6587\u6863&#xff1a;https:\/\/huggingface.co\/docs\/transformers\/model_doc\/bert<\/li>\n","protected":false},"excerpt":{"rendered":"<p>\u76ee\u5f55<br \/>\n\u5f15\u8a00\u81ea\u6ce8\u610f\u529b\u673a\u5236\u7684\u80cc\u666f \u5e8f\u5217\u6a21\u578b\u7684\u6f14\u8fdb\u6ce8\u610f\u529b\u673a\u5236\u7684\u8d77\u6e90 \u81ea\u6ce8\u610f\u529b\u673a\u5236\u7684\u6838\u5fc3\u539f\u7406 \u67e5\u8be2\u3001\u952e\u548c\u503c&#xff08;Q\u3001K\u3001V&#xff09;\u7684\u6982\u5ff5\u6ce8\u610f\u529b\u5206\u6570\u7684\u8ba1\u7b97Softmax\u5f52\u4e00\u5316\u548c\u52a0\u6743\u6c42\u548c \u591a\u5934\u81ea\u6ce8\u610f\u529b\u673a\u5236 \u4e3a\u4ec0\u4e48\u9700\u8981\u591a\u5934\u591a\u5934\u6ce8\u610f\u529b\u7684\u5b9e\u73b0 \u81ea\u6ce8\u610f\u529b\u5728Transformer\u4e2d\u7684\u5e94\u7528 Transformer\u6574\u4f53\u67b6\u6784\u7f16\u7801\u5668\u4e2d\u7684\u81ea\u6ce8\u610f\u529b\u89e3\u7801\u5668\u4e2d\u7684\u81ea\u6ce8\u610f\u529b\u4e0e\u4ea4\u53c9\u6ce8\u610f\u529b \u4ee3\u7801\u5b9e\u73b0&#xff1a;\u4ece\u96f6\u6784\u5efa\u81ea\u6ce8\u610f\u529b NumPy\u7248\u672c\u7684\u81ea\u6ce8\u610f\u529bPyTorch\u7248\u672c\u7684\u81ea\u6ce8\u610f\u529b\u591a\u5934\u6ce8\u610f\u529b\u7684\u4ee3\u7801\u793a<\/p>\n","protected":false},"author":2,"featured_media":62066,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[587,2951,2873,50,86],"topic":[],"class_list":["post-62072","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-server","tag-587","tag-lstm","tag-rnn","tag-50","tag-86"],"yoast_head":"<!-- This site is optimized with the Yoast SEO plugin v20.3 - https:\/\/yoast.com\/wordpress\/plugins\/seo\/ -->\n<title>\u901a\u4fd7\u7406\u89e3\u81ea\u6ce8\u610f\u529b\u673a\u5236\uff08Self-Attention\uff09 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3<\/title>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/www.wsisp.com\/helps\/62072.html\" \/>\n<meta property=\"og:locale\" content=\"zh_CN\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"\u901a\u4fd7\u7406\u89e3\u81ea\u6ce8\u610f\u529b\u673a\u5236\uff08Self-Attention\uff09 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"og:description\" content=\"\u76ee\u5f55 \u5f15\u8a00\u81ea\u6ce8\u610f\u529b\u673a\u5236\u7684\u80cc\u666f \u5e8f\u5217\u6a21\u578b\u7684\u6f14\u8fdb\u6ce8\u610f\u529b\u673a\u5236\u7684\u8d77\u6e90 \u81ea\u6ce8\u610f\u529b\u673a\u5236\u7684\u6838\u5fc3\u539f\u7406 \u67e5\u8be2\u3001\u952e\u548c\u503c&#xff08;Q\u3001K\u3001V&#xff09;\u7684\u6982\u5ff5\u6ce8\u610f\u529b\u5206\u6570\u7684\u8ba1\u7b97Softmax\u5f52\u4e00\u5316\u548c\u52a0\u6743\u6c42\u548c \u591a\u5934\u81ea\u6ce8\u610f\u529b\u673a\u5236 \u4e3a\u4ec0\u4e48\u9700\u8981\u591a\u5934\u591a\u5934\u6ce8\u610f\u529b\u7684\u5b9e\u73b0 \u81ea\u6ce8\u610f\u529b\u5728Transformer\u4e2d\u7684\u5e94\u7528 Transformer\u6574\u4f53\u67b6\u6784\u7f16\u7801\u5668\u4e2d\u7684\u81ea\u6ce8\u610f\u529b\u89e3\u7801\u5668\u4e2d\u7684\u81ea\u6ce8\u610f\u529b\u4e0e\u4ea4\u53c9\u6ce8\u610f\u529b \u4ee3\u7801\u5b9e\u73b0&#xff1a;\u4ece\u96f6\u6784\u5efa\u81ea\u6ce8\u610f\u529b NumPy\u7248\u672c\u7684\u81ea\u6ce8\u610f\u529bPyTorch\u7248\u672c\u7684\u81ea\u6ce8\u610f\u529b\u591a\u5934\u6ce8\u610f\u529b\u7684\u4ee3\u7801\u793a\" \/>\n<meta property=\"og:url\" content=\"https:\/\/www.wsisp.com\/helps\/62072.html\" \/>\n<meta property=\"og:site_name\" content=\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"article:published_time\" content=\"2026-01-19T01:05:52+00:00\" \/>\n<meta property=\"og:image\" content=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/01\/20260119010549-696d836dbdc04.png\" \/>\n<meta name=\"author\" content=\"admin\" \/>\n<meta name=\"twitter:card\" content=\"summary_large_image\" \/>\n<meta name=\"twitter:label1\" content=\"\u4f5c\u8005\" \/>\n\t<meta name=\"twitter:data1\" content=\"admin\" \/>\n\t<meta name=\"twitter:label2\" content=\"\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4\" \/>\n\t<meta name=\"twitter:data2\" content=\"7 \u5206\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"WebPage\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/62072.html\",\"url\":\"https:\/\/www.wsisp.com\/helps\/62072.html\",\"name\":\"\u901a\u4fd7\u7406\u89e3\u81ea\u6ce8\u610f\u529b\u673a\u5236\uff08Self-Attention\uff09 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"isPartOf\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\"},\"datePublished\":\"2026-01-19T01:05:52+00:00\",\"dateModified\":\"2026-01-19T01:05:52+00:00\",\"author\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\"},\"breadcrumb\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/62072.html#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/www.wsisp.com\/helps\/62072.html\"]}]},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/62072.html#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\/\/www.wsisp.com\/helps\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"\u901a\u4fd7\u7406\u89e3\u81ea\u6ce8\u610f\u529b\u673a\u5236\uff08Self-Attention\uff09\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\",\"url\":\"https:\/\/www.wsisp.com\/helps\/\",\"name\":\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"description\":\"\u9999\u6e2f\u670d\u52a1\u5668_\u9999\u6e2f\u4e91\u670d\u52a1\u5668\u8d44\u8baf_\u670d\u52a1\u5668\u5e2e\u52a9\u6587\u6863_\u670d\u52a1\u5668\u6559\u7a0b\",\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\/\/www.wsisp.com\/helps\/?s={search_term_string}\"},\"query-input\":\"required name=search_term_string\"}],\"inLanguage\":\"zh-Hans\"},{\"@type\":\"Person\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\",\"name\":\"admin\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/image\/\",\"url\":\"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery\",\"contentUrl\":\"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery\",\"caption\":\"admin\"},\"sameAs\":[\"http:\/\/wp.wsisp.com\"],\"url\":\"https:\/\/www.wsisp.com\/helps\/author\/admin\"}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"\u901a\u4fd7\u7406\u89e3\u81ea\u6ce8\u610f\u529b\u673a\u5236\uff08Self-Attention\uff09 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/www.wsisp.com\/helps\/62072.html","og_locale":"zh_CN","og_type":"article","og_title":"\u901a\u4fd7\u7406\u89e3\u81ea\u6ce8\u610f\u529b\u673a\u5236\uff08Self-Attention\uff09 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","og_description":"\u76ee\u5f55 \u5f15\u8a00\u81ea\u6ce8\u610f\u529b\u673a\u5236\u7684\u80cc\u666f \u5e8f\u5217\u6a21\u578b\u7684\u6f14\u8fdb\u6ce8\u610f\u529b\u673a\u5236\u7684\u8d77\u6e90 \u81ea\u6ce8\u610f\u529b\u673a\u5236\u7684\u6838\u5fc3\u539f\u7406 \u67e5\u8be2\u3001\u952e\u548c\u503c&#xff08;Q\u3001K\u3001V&#xff09;\u7684\u6982\u5ff5\u6ce8\u610f\u529b\u5206\u6570\u7684\u8ba1\u7b97Softmax\u5f52\u4e00\u5316\u548c\u52a0\u6743\u6c42\u548c \u591a\u5934\u81ea\u6ce8\u610f\u529b\u673a\u5236 \u4e3a\u4ec0\u4e48\u9700\u8981\u591a\u5934\u591a\u5934\u6ce8\u610f\u529b\u7684\u5b9e\u73b0 \u81ea\u6ce8\u610f\u529b\u5728Transformer\u4e2d\u7684\u5e94\u7528 Transformer\u6574\u4f53\u67b6\u6784\u7f16\u7801\u5668\u4e2d\u7684\u81ea\u6ce8\u610f\u529b\u89e3\u7801\u5668\u4e2d\u7684\u81ea\u6ce8\u610f\u529b\u4e0e\u4ea4\u53c9\u6ce8\u610f\u529b \u4ee3\u7801\u5b9e\u73b0&#xff1a;\u4ece\u96f6\u6784\u5efa\u81ea\u6ce8\u610f\u529b NumPy\u7248\u672c\u7684\u81ea\u6ce8\u610f\u529bPyTorch\u7248\u672c\u7684\u81ea\u6ce8\u610f\u529b\u591a\u5934\u6ce8\u610f\u529b\u7684\u4ee3\u7801\u793a","og_url":"https:\/\/www.wsisp.com\/helps\/62072.html","og_site_name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","article_published_time":"2026-01-19T01:05:52+00:00","og_image":[{"url":"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/01\/20260119010549-696d836dbdc04.png"}],"author":"admin","twitter_card":"summary_large_image","twitter_misc":{"\u4f5c\u8005":"admin","\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4":"7 \u5206"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/www.wsisp.com\/helps\/62072.html","url":"https:\/\/www.wsisp.com\/helps\/62072.html","name":"\u901a\u4fd7\u7406\u89e3\u81ea\u6ce8\u610f\u529b\u673a\u5236\uff08Self-Attention\uff09 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","isPartOf":{"@id":"https:\/\/www.wsisp.com\/helps\/#website"},"datePublished":"2026-01-19T01:05:52+00:00","dateModified":"2026-01-19T01:05:52+00:00","author":{"@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41"},"breadcrumb":{"@id":"https:\/\/www.wsisp.com\/helps\/62072.html#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/www.wsisp.com\/helps\/62072.html"]}]},{"@type":"BreadcrumbList","@id":"https:\/\/www.wsisp.com\/helps\/62072.html#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/www.wsisp.com\/helps"},{"@type":"ListItem","position":2,"name":"\u901a\u4fd7\u7406\u89e3\u81ea\u6ce8\u610f\u529b\u673a\u5236\uff08Self-Attention\uff09"}]},{"@type":"WebSite","@id":"https:\/\/www.wsisp.com\/helps\/#website","url":"https:\/\/www.wsisp.com\/helps\/","name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","description":"\u9999\u6e2f\u670d\u52a1\u5668_\u9999\u6e2f\u4e91\u670d\u52a1\u5668\u8d44\u8baf_\u670d\u52a1\u5668\u5e2e\u52a9\u6587\u6863_\u670d\u52a1\u5668\u6559\u7a0b","potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/www.wsisp.com\/helps\/?s={search_term_string}"},"query-input":"required name=search_term_string"}],"inLanguage":"zh-Hans"},{"@type":"Person","@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41","name":"admin","image":{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/image\/","url":"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery","contentUrl":"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery","caption":"admin"},"sameAs":["http:\/\/wp.wsisp.com"],"url":"https:\/\/www.wsisp.com\/helps\/author\/admin"}]}},"_links":{"self":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/62072","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/comments?post=62072"}],"version-history":[{"count":0,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/62072\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/media\/62066"}],"wp:attachment":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/media?parent=62072"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/categories?post=62072"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/tags?post=62072"},{"taxonomy":"topic","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/topic?post=62072"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}