{"id":37551,"date":"2025-05-16T09:33:42","date_gmt":"2025-05-16T01:33:42","guid":{"rendered":"https:\/\/www.wsisp.com\/helps\/37551.html"},"modified":"2025-05-16T09:33:42","modified_gmt":"2025-05-16T01:33:42","slug":"%e3%80%90%e7%8e%b0%e4%bb%a3%e6%b7%b1%e5%ba%a6%e5%ad%a6%e4%b9%a0%e6%8a%80%e6%9c%af%e3%80%91%e6%b3%a8%e6%84%8f%e5%8a%9b%e6%9c%ba%e5%88%b607%ef%bc%9atransformer","status":"publish","type":"post","link":"https:\/\/www.wsisp.com\/helps\/37551.html","title":{"rendered":"\u3010\u73b0\u4ee3\u6df1\u5ea6\u5b66\u4e60\u6280\u672f\u3011\u6ce8\u610f\u529b\u673a\u523607\uff1aTransformer"},"content":{"rendered":"<p><img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2025\/05\/20250516013340-682695f417dab.gif\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" width=\"500\" \/><\/p>\n<p>\u3010\u4f5c\u8005\u4e3b\u9875\u3011Francek Chen \u3010\u4e13\u680f\u4ecb\u7ecd\u3011<span class=\"katex--inline\"><span class=\"katex\"><span class=\"katex-mathml\"> <\/p>\n<p>           \u2308 <\/p>\n<p>          \u2308 <\/p>\n<p>      <\/span><span class=\"katex-html\"><span class=\"base\"><span class=\"strut\" style=\"height: 1em;vertical-align: -0.25em\"><\/span><span class=\"mopen\">\u2308<\/span><\/span><\/span><\/span><\/span>PyTorch\u6df1\u5ea6\u5b66\u4e60<span class=\"katex--inline\"><span class=\"katex\"><span class=\"katex-mathml\"> <\/p>\n<p>           \u230b <\/p>\n<p>          \u230b <\/p>\n<p>      <\/span><span class=\"katex-html\"><span class=\"base\"><span class=\"strut\" style=\"height: 1em;vertical-align: -0.25em\"><\/span><span class=\"mclose\">\u230b<\/span><\/span><\/span><\/span><\/span> \u6df1\u5ea6\u5b66\u4e60 (DL, Deep Learning) \u7279\u6307\u57fa\u4e8e\u6df1\u5c42\u795e\u7ecf\u7f51\u7edc\u6a21\u578b\u548c\u65b9\u6cd5\u7684\u673a\u5668\u5b66\u4e60\u3002\u5b83\u662f\u5728\u7edf\u8ba1\u673a\u5668\u5b66\u4e60\u3001\u4eba\u5de5\u795e\u7ecf\u7f51\u7edc\u7b49\u7b97\u6cd5\u6a21\u578b\u57fa\u7840\u4e0a&#xff0c;\u7ed3\u5408\u5f53\u4ee3\u5927\u6570\u636e\u548c\u5927\u7b97\u529b\u7684\u53d1\u5c55\u800c\u53d1\u5c55\u51fa\u6765\u7684\u3002\u6df1\u5ea6\u5b66\u4e60\u6700\u91cd\u8981\u7684\u6280\u672f\u7279\u5f81\u662f\u5177\u6709\u81ea\u52a8\u63d0\u53d6\u7279\u5f81\u7684\u80fd\u529b\u3002\u795e\u7ecf\u7f51\u7edc\u7b97\u6cd5\u3001\u7b97\u529b\u548c\u6570\u636e\u662f\u5f00\u5c55\u6df1\u5ea6\u5b66\u4e60\u7684\u4e09\u8981\u7d20\u3002\u6df1\u5ea6\u5b66\u4e60\u5728\u8ba1\u7b97\u673a\u89c6\u89c9\u3001\u81ea\u7136\u8bed\u8a00\u5904\u7406\u3001\u591a\u6a21\u6001\u6570\u636e\u5206\u6790\u3001\u79d1\u5b66\u63a2\u7d22\u7b49\u9886\u57df\u90fd\u53d6\u5f97\u4e86\u5f88\u591a\u6210\u679c\u3002\u672c\u4e13\u680f\u4ecb\u7ecd\u57fa\u4e8ePyTorch\u7684\u6df1\u5ea6\u5b66\u4e60\u7b97\u6cd5\u5b9e\u73b0\u3002 \u3010GitCode\u3011\u4e13\u680f\u8d44\u6e90\u4fdd\u5b58\u5728\u6211\u7684GitCode\u4ed3\u5e93&#xff1a;https:\/\/gitcode.com\/Morse_Chen\/PyTorch_deep_learning\u3002<\/p>\n<\/p>\n<h4>\u6587\u7ae0\u76ee\u5f55<\/h4>\n<ul>\n<li>\n<ul>\n<li>\u4e00\u3001\u6a21\u578b<\/li>\n<li>\u4e8c\u3001\u57fa\u4e8e\u4f4d\u7f6e\u7684\u524d\u9988\u7f51\u7edc<\/li>\n<li>\u4e09\u3001\u6b8b\u5dee\u8fde\u63a5\u548c\u5c42\u89c4\u8303\u5316<\/li>\n<li>\u56db\u3001\u7f16\u7801\u5668<\/li>\n<li>\u4e94\u3001\u89e3\u7801\u5668<\/li>\n<li>\u516d\u3001\u8bad\u7ec3<\/li>\n<li>\u5c0f\u7ed3<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<hr \/>\n<p>\u2003\u2003\u81ea\u6ce8\u610f\u529b\u548c\u4f4d\u7f6e\u7f16\u7801\u4e2d\u6bd4\u8f83\u4e86\u5377\u79ef\u795e\u7ecf\u7f51\u7edc&#xff08;CNN&#xff09;\u3001\u5faa\u73af\u795e\u7ecf\u7f51\u7edc&#xff08;RNN&#xff09;\u548c\u81ea\u6ce8\u610f\u529b&#xff08;self-attention&#xff09;\u3002\u503c\u5f97\u6ce8\u610f\u7684\u662f&#xff0c;\u81ea\u6ce8\u610f\u529b\u540c\u65f6\u5177\u6709\u5e76\u884c\u8ba1\u7b97\u548c\u6700\u77ed\u7684\u6700\u5927\u8def\u5f84\u957f\u5ea6\u8fd9\u4e24\u4e2a\u4f18\u52bf\u3002\u56e0\u6b64&#xff0c;\u4f7f\u7528\u81ea\u6ce8\u610f\u529b\u6765\u8bbe\u8ba1\u6df1\u5ea6\u67b6\u6784\u662f\u5f88\u6709\u5438\u5f15\u529b\u7684\u3002\u5bf9\u6bd4\u4e4b\u524d\u4ecd\u7136\u4f9d\u8d56\u5faa\u73af\u795e\u7ecf\u7f51\u7edc\u5b9e\u73b0\u8f93\u5165\u8868\u793a\u7684\u81ea\u6ce8\u610f\u529b\u6a21\u578b&#xff0c;Transformer\u6a21\u578b\u5b8c\u5168\u57fa\u4e8e\u6ce8\u610f\u529b\u673a\u5236&#xff0c;\u6ca1\u6709\u4efb\u4f55\u5377\u79ef\u5c42\u6216\u5faa\u73af\u795e\u7ecf\u7f51\u7edc\u5c42\u3002\u5c3d\u7ba1Transformer\u6700\u521d\u662f\u5e94\u7528\u4e8e\u5728\u6587\u672c\u6570\u636e\u4e0a\u7684\u5e8f\u5217\u5230\u5e8f\u5217\u5b66\u4e60&#xff0c;\u4f46\u73b0\u5728\u5df2\u7ecf\u63a8\u5e7f\u5230\u5404\u79cd\u73b0\u4ee3\u7684\u6df1\u5ea6\u5b66\u4e60\u4e2d&#xff0c;\u4f8b\u5982\u8bed\u8a00\u3001\u89c6\u89c9\u3001\u8bed\u97f3\u548c\u5f3a\u5316\u5b66\u4e60\u9886\u57df\u3002<\/p>\n<h3>\u4e00\u3001\u6a21\u578b<\/h3>\n<p>\u2003\u2003Transformer\u4f5c\u4e3a\u7f16\u7801\u5668-\u89e3\u7801\u5668\u67b6\u6784\u7684\u4e00\u4e2a\u5b9e\u4f8b&#xff0c;\u5176\u6574\u4f53\u67b6\u6784\u56fe\u5728\u56fe1\u4e2d\u5c55\u793a\u3002\u6b63\u5982\u6240\u89c1\u5230\u7684&#xff0c;Transformer\u662f\u7531\u7f16\u7801\u5668\u548c\u89e3\u7801\u5668\u7ec4\u6210\u7684\u3002\u4e0e\u57fa\u4e8eBahdanau\u6ce8\u610f\u529b\u5b9e\u73b0\u7684\u5e8f\u5217\u5230\u5e8f\u5217\u7684\u5b66\u4e60\u76f8\u6bd4&#xff0c;Transformer\u7684\u7f16\u7801\u5668\u548c\u89e3\u7801\u5668\u662f\u57fa\u4e8e\u81ea\u6ce8\u610f\u529b\u7684\u6a21\u5757\u53e0\u52a0\u800c\u6210\u7684&#xff0c;\u6e90&#xff08;\u8f93\u5165&#xff09;\u5e8f\u5217\u548c\u76ee\u6807&#xff08;\u8f93\u51fa&#xff09;\u5e8f\u5217\u7684\u5d4c\u5165&#xff08;embedding&#xff09;\u8868\u793a\u5c06\u52a0\u4e0a\u4f4d\u7f6e\u7f16\u7801&#xff08;positional encoding&#xff09;&#xff0c;\u518d\u5206\u522b\u8f93\u5165\u5230\u7f16\u7801\u5668\u548c\u89e3\u7801\u5668\u4e2d\u3002<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2025\/05\/20250516013340-682695f424dd0.png\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" width=\"450\" \/><\/p>\n<p> <font face=\"\u4eff\u5b8b\">\u56fe1 Transformer\u67b6\u6784<\/font> <\/p>\n<p>\u2003\u2003\u56fe1\u4e2d\u6982\u8ff0\u4e86Transformer\u7684\u67b6\u6784\u3002\u4ece\u5b8f\u89c2\u89d2\u5ea6\u6765\u770b&#xff0c;Transformer\u7684\u7f16\u7801\u5668\u662f\u7531\u591a\u4e2a\u76f8\u540c\u7684\u5c42\u53e0\u52a0\u800c\u6210\u7684&#xff0c;\u6bcf\u4e2a\u5c42\u90fd\u6709\u4e24\u4e2a\u5b50\u5c42&#xff08;\u5b50\u5c42\u8868\u793a\u4e3a<span class=\"katex--inline\"><span class=\"katex\"><span class=\"katex-mathml\"> <\/p>\n<p>         s <\/p>\n<p>         u <\/p>\n<p>         b <\/p>\n<p>         l <\/p>\n<p>         a <\/p>\n<p>         y <\/p>\n<p>         e <\/p>\n<p>         r <\/p>\n<p>        \\\\mathrm{sublayer} <\/p>\n<p>    <\/span><span class=\"katex-html\"><span class=\"base\"><span class=\"strut\" style=\"height: 0.8889em;vertical-align: -0.1944em\"><\/span><span class=\"mord\"><span class=\"mord mathrm\">sublayer<\/span><\/span><\/span><\/span><\/span><\/span>&#xff09;\u3002\u7b2c\u4e00\u4e2a\u5b50\u5c42\u662f\u591a\u5934\u81ea\u6ce8\u610f\u529b&#xff08;multi-head self-attention&#xff09;\u6c47\u805a&#xff1b;\u7b2c\u4e8c\u4e2a\u5b50\u5c42\u662f\u57fa\u4e8e\u4f4d\u7f6e\u7684\u524d\u9988\u7f51\u7edc&#xff08;positionwise feed-forward network&#xff09;\u3002\u5177\u4f53\u6765\u8bf4&#xff0c;\u5728\u8ba1\u7b97\u7f16\u7801\u5668\u7684\u81ea\u6ce8\u610f\u529b\u65f6&#xff0c;\u67e5\u8be2\u3001\u952e\u548c\u503c\u90fd\u6765\u81ea\u524d\u4e00\u4e2a\u7f16\u7801\u5668\u5c42\u7684\u8f93\u51fa\u3002\u53d7\u6b8b\u5dee\u7f51\u7edc\u7684\u542f\u53d1&#xff0c;\u6bcf\u4e2a\u5b50\u5c42\u90fd\u91c7\u7528\u4e86\u6b8b\u5dee\u8fde\u63a5&#xff08;residual connection&#xff09;\u3002\u5728Transformer\u4e2d&#xff0c;\u5bf9\u4e8e\u5e8f\u5217\u4e2d\u4efb\u4f55\u4f4d\u7f6e\u7684\u4efb\u4f55\u8f93\u5165<span class=\"katex--inline\"><span class=\"katex\"><span class=\"katex-mathml\"> <\/p>\n<p>         x <\/p>\n<p>         \u2208 <\/p>\n<p>          R <\/p>\n<p>          d <\/p>\n<p>        \\\\mathbf{x} \\\\in \\\\mathbb{R}^d <\/p>\n<p>    <\/span><span class=\"katex-html\"><span class=\"base\"><span class=\"strut\" style=\"height: 0.5782em;vertical-align: -0.0391em\"><\/span><span class=\"mord mathbf\">x<\/span><span class=\"mspace\" style=\"margin-right: 0.2778em\"><\/span><span class=\"mrel\">\u2208<\/span><span class=\"mspace\" style=\"margin-right: 0.2778em\"><\/span><\/span><span class=\"base\"><span class=\"strut\" style=\"height: 0.8491em\"><\/span><span class=\"mord\"><span class=\"mord mathbb\">R<\/span><span class=\"msupsub\"><span class=\"vlist-t\"><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.8491em\"><span class=\"\" style=\"top: -3.063em;margin-right: 0.05em\"><span class=\"pstrut\" style=\"height: 2.7em\"><\/span><span class=\"sizing reset-size6 size3 mtight\"><span class=\"mord mathnormal mtight\">d<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>&#xff0c;\u90fd\u8981\u6c42\u6ee1\u8db3<span class=\"katex--inline\"><span class=\"katex\"><span class=\"katex-mathml\"> <\/p>\n<p>          s <\/p>\n<p>          u <\/p>\n<p>          b <\/p>\n<p>          l <\/p>\n<p>          a <\/p>\n<p>          y <\/p>\n<p>          e <\/p>\n<p>          r <\/p>\n<p>         ( <\/p>\n<p>         x <\/p>\n<p>         ) <\/p>\n<p>         \u2208 <\/p>\n<p>          R <\/p>\n<p>          d <\/p>\n<p>        \\\\mathrm{sublayer}(\\\\mathbf{x}) \\\\in \\\\mathbb{R}^d <\/p>\n<p>    <\/span><span class=\"katex-html\"><span class=\"base\"><span class=\"strut\" style=\"height: 1em;vertical-align: -0.25em\"><\/span><span class=\"mord\"><span class=\"mord mathrm\">sublayer<\/span><\/span><span class=\"mopen\">(<\/span><span class=\"mord mathbf\">x<\/span><span class=\"mclose\">)<\/span><span class=\"mspace\" style=\"margin-right: 0.2778em\"><\/span><span class=\"mrel\">\u2208<\/span><span class=\"mspace\" style=\"margin-right: 0.2778em\"><\/span><\/span><span class=\"base\"><span class=\"strut\" style=\"height: 0.8491em\"><\/span><span class=\"mord\"><span class=\"mord mathbb\">R<\/span><span class=\"msupsub\"><span class=\"vlist-t\"><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.8491em\"><span class=\"\" style=\"top: -3.063em;margin-right: 0.05em\"><span class=\"pstrut\" style=\"height: 2.7em\"><\/span><span class=\"sizing reset-size6 size3 mtight\"><span class=\"mord mathnormal mtight\">d<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>&#xff0c;\u4ee5\u4fbf\u6b8b\u5dee\u8fde\u63a5\u6ee1\u8db3<span class=\"katex--inline\"><span class=\"katex\"><span class=\"katex-mathml\"> <\/p>\n<p>         x <\/p>\n<p>         &#043; <\/p>\n<p>          s <\/p>\n<p>          u <\/p>\n<p>          b <\/p>\n<p>          l <\/p>\n<p>          a <\/p>\n<p>          y <\/p>\n<p>          e <\/p>\n<p>          r <\/p>\n<p>         ( <\/p>\n<p>         x <\/p>\n<p>         ) <\/p>\n<p>         \u2208 <\/p>\n<p>          R <\/p>\n<p>          d <\/p>\n<p>        \\\\mathbf{x} &#043; \\\\mathrm{sublayer}(\\\\mathbf{x}) \\\\in \\\\mathbb{R}^d <\/p>\n<p>    <\/span><span class=\"katex-html\"><span class=\"base\"><span class=\"strut\" style=\"height: 0.6667em;vertical-align: -0.0833em\"><\/span><span class=\"mord mathbf\">x<\/span><span class=\"mspace\" style=\"margin-right: 0.2222em\"><\/span><span class=\"mbin\">&#043;<\/span><span class=\"mspace\" style=\"margin-right: 0.2222em\"><\/span><\/span><span class=\"base\"><span class=\"strut\" style=\"height: 1em;vertical-align: -0.25em\"><\/span><span class=\"mord\"><span class=\"mord mathrm\">sublayer<\/span><\/span><span class=\"mopen\">(<\/span><span class=\"mord mathbf\">x<\/span><span class=\"mclose\">)<\/span><span class=\"mspace\" style=\"margin-right: 0.2778em\"><\/span><span class=\"mrel\">\u2208<\/span><span class=\"mspace\" style=\"margin-right: 0.2778em\"><\/span><\/span><span class=\"base\"><span class=\"strut\" style=\"height: 0.8491em\"><\/span><span class=\"mord\"><span class=\"mord mathbb\">R<\/span><span class=\"msupsub\"><span class=\"vlist-t\"><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.8491em\"><span class=\"\" style=\"top: -3.063em;margin-right: 0.05em\"><span class=\"pstrut\" style=\"height: 2.7em\"><\/span><span class=\"sizing reset-size6 size3 mtight\"><span class=\"mord mathnormal mtight\">d<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u3002\u5728\u6b8b\u5dee\u8fde\u63a5\u7684\u52a0\u6cd5\u8ba1\u7b97\u4e4b\u540e&#xff0c;\u7d27\u63a5\u7740\u5e94\u7528\u5c42\u89c4\u8303\u5316&#xff08;layer normalization&#xff09;\u3002\u56e0\u6b64&#xff0c;\u8f93\u5165\u5e8f\u5217\u5bf9\u5e94\u7684\u6bcf\u4e2a\u4f4d\u7f6e&#xff0c;Transformer\u7f16\u7801\u5668\u90fd\u5c06\u8f93\u51fa\u4e00\u4e2a<span class=\"katex--inline\"><span class=\"katex\"><span class=\"katex-mathml\"> <\/p>\n<p>         d <\/p>\n<p>        d <\/p>\n<p>    <\/span><span class=\"katex-html\"><span class=\"base\"><span class=\"strut\" style=\"height: 0.6944em\"><\/span><span class=\"mord mathnormal\">d<\/span><\/span><\/span><\/span><\/span>\u7ef4\u8868\u793a\u5411\u91cf\u3002<\/p>\n<p>\u2003\u2003Transformer\u89e3\u7801\u5668\u4e5f\u662f\u7531\u591a\u4e2a\u76f8\u540c\u7684\u5c42\u53e0\u52a0\u800c\u6210\u7684&#xff0c;\u5e76\u4e14\u5c42\u4e2d\u4f7f\u7528\u4e86\u6b8b\u5dee\u8fde\u63a5\u548c\u5c42\u89c4\u8303\u5316\u3002\u9664\u4e86\u7f16\u7801\u5668\u4e2d\u63cf\u8ff0\u7684\u4e24\u4e2a\u5b50\u5c42\u4e4b\u5916&#xff0c;\u89e3\u7801\u5668\u8fd8\u5728\u8fd9\u4e24\u4e2a\u5b50\u5c42\u4e4b\u95f4\u63d2\u5165\u4e86\u7b2c\u4e09\u4e2a\u5b50\u5c42&#xff0c;\u79f0\u4e3a\u7f16\u7801\u5668&#xff0d;\u89e3\u7801\u5668\u6ce8\u610f\u529b&#xff08;encoder-decoder attention&#xff09;\u5c42\u3002\u5728\u7f16\u7801\u5668&#xff0d;\u89e3\u7801\u5668\u6ce8\u610f\u529b\u4e2d&#xff0c;\u67e5\u8be2\u6765\u81ea\u524d\u4e00\u4e2a\u89e3\u7801\u5668\u5c42\u7684\u8f93\u51fa&#xff0c;\u800c\u952e\u548c\u503c\u6765\u81ea\u6574\u4e2a\u7f16\u7801\u5668\u7684\u8f93\u51fa\u3002\u5728\u89e3\u7801\u5668\u81ea\u6ce8\u610f\u529b\u4e2d&#xff0c;\u67e5\u8be2\u3001\u952e\u548c\u503c\u90fd\u6765\u81ea\u4e0a\u4e00\u4e2a\u89e3\u7801\u5668\u5c42\u7684\u8f93\u51fa\u3002\u4f46\u662f&#xff0c;\u89e3\u7801\u5668\u4e2d\u7684\u6bcf\u4e2a\u4f4d\u7f6e\u53ea\u80fd\u8003\u8651\u8be5\u4f4d\u7f6e\u4e4b\u524d\u7684\u6240\u6709\u4f4d\u7f6e\u3002\u8fd9\u79cd\u63a9\u853d&#xff08;masked&#xff09;\u6ce8\u610f\u529b\u4fdd\u7559\u4e86\u81ea\u56de\u5f52&#xff08;auto-regressive&#xff09;\u5c5e\u6027&#xff0c;\u786e\u4fdd\u9884\u6d4b\u4ec5\u4f9d\u8d56\u4e8e\u5df2\u751f\u6210\u7684\u8f93\u51fa\u8bcd\u5143\u3002<\/p>\n<p>\u2003\u2003\u5728\u6b64\u4e4b\u524d\u5df2\u7ecf\u63cf\u8ff0\u5e76\u5b9e\u73b0\u4e86\u57fa\u4e8e\u7f29\u653e\u70b9\u79ef\u591a\u5934\u6ce8\u610f\u529b\u548c\u4f4d\u7f6e\u7f16\u7801\u3002\u63a5\u4e0b\u6765\u5c06\u5b9e\u73b0Transformer\u6a21\u578b\u7684\u5269\u4f59\u90e8\u5206\u3002<\/p>\n<p><span class=\"token keyword\">import<\/span> math<br \/>\n<span class=\"token keyword\">import<\/span> pandas <span class=\"token keyword\">as<\/span> pd<br \/>\n<span class=\"token keyword\">import<\/span> torch<br \/>\n<span class=\"token keyword\">from<\/span> torch <span class=\"token keyword\">import<\/span> nn<br \/>\n<span class=\"token keyword\">from<\/span> d2l <span class=\"token keyword\">import<\/span> torch <span class=\"token keyword\">as<\/span> d2l<\/p>\n<h3>\u4e8c\u3001\u57fa\u4e8e\u4f4d\u7f6e\u7684\u524d\u9988\u7f51\u7edc<\/h3>\n<p>\u2003\u2003\u57fa\u4e8e\u4f4d\u7f6e\u7684\u524d\u9988\u7f51\u7edc\u5bf9\u5e8f\u5217\u4e2d\u7684\u6240\u6709\u4f4d\u7f6e\u7684\u8868\u793a\u8fdb\u884c\u53d8\u6362\u65f6\u4f7f\u7528\u7684\u662f\u540c\u4e00\u4e2a\u591a\u5c42\u611f\u77e5\u673a&#xff08;MLP&#xff09;&#xff0c;\u8fd9\u5c31\u662f\u79f0\u524d\u9988\u7f51\u7edc\u662f\u57fa\u4e8e\u4f4d\u7f6e\u7684&#xff08;positionwise&#xff09;\u7684\u539f\u56e0\u3002\u5728\u4e0b\u9762\u7684\u5b9e\u73b0\u4e2d&#xff0c;\u8f93\u5165X\u7684\u5f62\u72b6&#xff08;\u6279\u91cf\u5927\u5c0f&#xff0c;\u65f6\u95f4\u6b65\u6570\u6216\u5e8f\u5217\u957f\u5ea6&#xff0c;\u9690\u5355\u5143\u6570\u6216\u7279\u5f81\u7ef4\u5ea6&#xff09;\u5c06\u88ab\u4e00\u4e2a\u4e24\u5c42\u7684\u611f\u77e5\u673a\u8f6c\u6362\u6210\u5f62\u72b6\u4e3a&#xff08;\u6279\u91cf\u5927\u5c0f&#xff0c;\u65f6\u95f4\u6b65\u6570&#xff0c;ffn_num_outputs&#xff09;\u7684\u8f93\u51fa\u5f20\u91cf\u3002<\/p>\n<p><span class=\"token comment\">#&#064;save<\/span><br \/>\n<span class=\"token keyword\">class<\/span> <span class=\"token class-name\">PositionWiseFFN<\/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 triple-quoted-string string\">&#034;&#034;&#034;\u57fa\u4e8e\u4f4d\u7f6e\u7684\u524d\u9988\u7f51\u7edc&#034;&#034;&#034;<\/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> ffn_num_input<span class=\"token punctuation\">,<\/span> ffn_num_hiddens<span class=\"token punctuation\">,<\/span> ffn_num_outputs<span class=\"token punctuation\">,<\/span> <span class=\"token operator\">**<\/span>kwargs<span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">:<\/span><br \/>\n        <span class=\"token builtin\">super<\/span><span class=\"token punctuation\">(<\/span>PositionWiseFFN<span class=\"token punctuation\">,<\/span> self<span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">.<\/span>__init__<span class=\"token punctuation\">(<\/span><span class=\"token operator\">**<\/span>kwargs<span class=\"token punctuation\">)<\/span><br \/>\n        self<span class=\"token punctuation\">.<\/span>dense1 <span class=\"token operator\">&#061;<\/span> nn<span class=\"token punctuation\">.<\/span>Linear<span class=\"token punctuation\">(<\/span>ffn_num_input<span class=\"token punctuation\">,<\/span> ffn_num_hiddens<span class=\"token punctuation\">)<\/span><br \/>\n        self<span class=\"token punctuation\">.<\/span>relu <span class=\"token operator\">&#061;<\/span> nn<span class=\"token punctuation\">.<\/span>ReLU<span class=\"token punctuation\">(<\/span><span class=\"token punctuation\">)<\/span><br \/>\n        self<span class=\"token punctuation\">.<\/span>dense2 <span class=\"token operator\">&#061;<\/span> nn<span class=\"token punctuation\">.<\/span>Linear<span class=\"token punctuation\">(<\/span>ffn_num_hiddens<span class=\"token punctuation\">,<\/span> ffn_num_outputs<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> X<span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">:<\/span><br \/>\n        <span class=\"token keyword\">return<\/span> self<span class=\"token punctuation\">.<\/span>dense2<span class=\"token punctuation\">(<\/span>self<span class=\"token punctuation\">.<\/span>relu<span class=\"token punctuation\">(<\/span>self<span class=\"token punctuation\">.<\/span>dense1<span class=\"token punctuation\">(<\/span>X<span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">)<\/span><\/p>\n<p>\u2003\u2003\u4e0b\u9762\u7684\u4f8b\u5b50\u663e\u793a&#xff0c;\u6539\u53d8\u5f20\u91cf\u7684\u6700\u91cc\u5c42\u7ef4\u5ea6\u7684\u5c3a\u5bf8&#xff0c;\u4f1a\u6539\u53d8\u6210\u57fa\u4e8e\u4f4d\u7f6e\u7684\u524d\u9988\u7f51\u7edc\u7684\u8f93\u51fa\u5c3a\u5bf8\u3002\u56e0\u4e3a\u7528\u540c\u4e00\u4e2a\u591a\u5c42\u611f\u77e5\u673a\u5bf9\u6240\u6709\u4f4d\u7f6e\u4e0a\u7684\u8f93\u5165\u8fdb\u884c\u53d8\u6362&#xff0c;\u6240\u4ee5\u5f53\u6240\u6709\u8fd9\u4e9b\u4f4d\u7f6e\u7684\u8f93\u5165\u76f8\u540c\u65f6&#xff0c;\u5b83\u4eec\u7684\u8f93\u51fa\u4e5f\u662f\u76f8\u540c\u7684\u3002<\/p>\n<p>ffn <span class=\"token operator\">&#061;<\/span> PositionWiseFFN<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> <span class=\"token number\">8<\/span><span class=\"token punctuation\">)<\/span><br \/>\nffn<span class=\"token punctuation\">.<\/span><span class=\"token builtin\">eval<\/span><span class=\"token punctuation\">(<\/span><span class=\"token punctuation\">)<\/span><br \/>\nffn<span class=\"token punctuation\">(<\/span>torch<span class=\"token punctuation\">.<\/span>ones<span class=\"token punctuation\">(<\/span><span class=\"token punctuation\">(<\/span><span class=\"token number\">2<\/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 punctuation\">)<\/span><span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">[<\/span><span class=\"token number\">0<\/span><span class=\"token punctuation\">]<\/span><\/p>\n<p><img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2025\/05\/20250516013340-682695f43f42a.png\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" width=\"600\" \/><\/p>\n<h3>\u4e09\u3001\u6b8b\u5dee\u8fde\u63a5\u548c\u5c42\u89c4\u8303\u5316<\/h3>\n<p>\u2003\u2003\u73b0\u5728\u8ba9\u6211\u4eec\u5173\u6ce8\u56fe1\u4e2d\u7684\u52a0\u6cd5\u548c\u89c4\u8303\u5316&#xff08;add&amp;norm&#xff09;\u7ec4\u4ef6\u3002\u6b63\u5982\u5728\u672c\u8282\u5f00\u5934\u6240\u8ff0&#xff0c;\u8fd9\u662f\u7531\u6b8b\u5dee\u8fde\u63a5\u548c\u7d27\u968f\u5176\u540e\u7684\u5c42\u89c4\u8303\u5316\u7ec4\u6210\u7684\u3002\u4e24\u8005\u90fd\u662f\u6784\u5efa\u6709\u6548\u7684\u6df1\u5ea6\u67b6\u6784\u7684\u5173\u952e\u3002<\/p>\n<p>\u2003\u2003\u6279\u91cf\u89c4\u8303\u5316\u4e2d\u89e3\u91ca\u4e86\u5728\u4e00\u4e2a\u5c0f\u6279\u91cf\u7684\u6837\u672c\u5185\u57fa\u4e8e\u6279\u91cf\u89c4\u8303\u5316\u5bf9\u6570\u636e\u8fdb\u884c\u91cd\u65b0\u4e2d\u5fc3\u5316\u548c\u91cd\u65b0\u7f29\u653e\u7684\u8c03\u6574\u3002\u5c42\u89c4\u8303\u5316\u548c\u6279\u91cf\u89c4\u8303\u5316\u7684\u76ee\u6807\u76f8\u540c&#xff0c;\u4f46\u5c42\u89c4\u8303\u5316\u662f\u57fa\u4e8e\u7279\u5f81\u7ef4\u5ea6\u8fdb\u884c\u89c4\u8303\u5316\u3002\u5c3d\u7ba1\u6279\u91cf\u89c4\u8303\u5316\u5728\u8ba1\u7b97\u673a\u89c6\u89c9\u4e2d\u88ab\u5e7f\u6cdb\u5e94\u7528&#xff0c;\u4f46\u5728\u81ea\u7136\u8bed\u8a00\u5904\u7406\u4efb\u52a1\u4e2d&#xff08;\u8f93\u5165\u901a\u5e38\u662f\u53d8\u957f\u5e8f\u5217&#xff09;\u6279\u91cf\u89c4\u8303\u5316\u901a\u5e38\u4e0d\u5982\u5c42\u89c4\u8303\u5316\u7684\u6548\u679c\u597d\u3002<\/p>\n<p>\u2003\u2003\u4ee5\u4e0b\u4ee3\u7801\u5bf9\u6bd4\u4e0d\u540c\u7ef4\u5ea6\u7684\u5c42\u89c4\u8303\u5316\u548c\u6279\u91cf\u89c4\u8303\u5316\u7684\u6548\u679c\u3002<\/p>\n<p>ln <span class=\"token operator\">&#061;<\/span> nn<span class=\"token punctuation\">.<\/span>LayerNorm<span class=\"token punctuation\">(<\/span><span class=\"token number\">2<\/span><span class=\"token punctuation\">)<\/span><br \/>\nbn <span class=\"token operator\">&#061;<\/span> nn<span class=\"token punctuation\">.<\/span>BatchNorm1d<span class=\"token punctuation\">(<\/span><span class=\"token number\">2<\/span><span class=\"token punctuation\">)<\/span><br \/>\nX <span class=\"token operator\">&#061;<\/span> torch<span class=\"token punctuation\">.<\/span>tensor<span class=\"token punctuation\">(<\/span><span class=\"token punctuation\">[<\/span><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> <span class=\"token punctuation\">[<\/span><span class=\"token number\">2<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token number\">3<\/span><span class=\"token punctuation\">]<\/span><span class=\"token punctuation\">]<\/span><span class=\"token punctuation\">,<\/span> dtype<span class=\"token operator\">&#061;<\/span>torch<span class=\"token punctuation\">.<\/span>float32<span class=\"token punctuation\">)<\/span><br \/>\n<span class=\"token comment\"># \u5728\u8bad\u7ec3\u6a21\u5f0f\u4e0b\u8ba1\u7b97X\u7684\u5747\u503c\u548c\u65b9\u5dee<\/span><br \/>\n<span class=\"token keyword\">print<\/span><span class=\"token punctuation\">(<\/span><span class=\"token string\">&#039;layer norm:&#039;<\/span><span class=\"token punctuation\">,<\/span> ln<span class=\"token punctuation\">(<\/span>X<span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token string\">&#039;\\\\nbatch norm:&#039;<\/span><span class=\"token punctuation\">,<\/span> bn<span class=\"token punctuation\">(<\/span>X<span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">)<\/span><\/p>\n<p><img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2025\/05\/20250516013340-682695f44f16b.png\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" width=\"500\" \/><\/p>\n<p>\u2003\u2003\u73b0\u5728\u53ef\u4ee5\u4f7f\u7528\u6b8b\u5dee\u8fde\u63a5\u548c\u5c42\u89c4\u8303\u5316\u6765\u5b9e\u73b0AddNorm\u7c7b\u3002\u6682\u9000\u6cd5\u4e5f\u88ab\u4f5c\u4e3a\u6b63\u5219\u5316\u65b9\u6cd5\u4f7f\u7528\u3002<\/p>\n<p><span class=\"token comment\">#&#064;save<\/span><br \/>\n<span class=\"token keyword\">class<\/span> <span class=\"token class-name\">AddNorm<\/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 triple-quoted-string string\">&#034;&#034;&#034;\u6b8b\u5dee\u8fde\u63a5\u540e\u8fdb\u884c\u5c42\u89c4\u8303\u5316&#034;&#034;&#034;<\/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> normalized_shape<span class=\"token punctuation\">,<\/span> dropout<span class=\"token punctuation\">,<\/span> <span class=\"token operator\">**<\/span>kwargs<span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">:<\/span><br \/>\n        <span class=\"token builtin\">super<\/span><span class=\"token punctuation\">(<\/span>AddNorm<span class=\"token punctuation\">,<\/span> self<span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">.<\/span>__init__<span class=\"token punctuation\">(<\/span><span class=\"token operator\">**<\/span>kwargs<span class=\"token punctuation\">)<\/span><br \/>\n        self<span class=\"token punctuation\">.<\/span>dropout <span class=\"token operator\">&#061;<\/span> nn<span class=\"token punctuation\">.<\/span>Dropout<span class=\"token punctuation\">(<\/span>dropout<span class=\"token punctuation\">)<\/span><br \/>\n        self<span class=\"token punctuation\">.<\/span>ln <span class=\"token operator\">&#061;<\/span> nn<span class=\"token punctuation\">.<\/span>LayerNorm<span class=\"token punctuation\">(<\/span>normalized_shape<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> X<span class=\"token punctuation\">,<\/span> Y<span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">:<\/span><br \/>\n        <span class=\"token keyword\">return<\/span> self<span class=\"token punctuation\">.<\/span>ln<span class=\"token punctuation\">(<\/span>self<span class=\"token punctuation\">.<\/span>dropout<span class=\"token punctuation\">(<\/span>Y<span class=\"token punctuation\">)<\/span> <span class=\"token operator\">&#043;<\/span> X<span class=\"token punctuation\">)<\/span><\/p>\n<p>\u2003\u2003\u6b8b\u5dee\u8fde\u63a5\u8981\u6c42\u4e24\u4e2a\u8f93\u5165\u7684\u5f62\u72b6\u76f8\u540c&#xff0c;\u4ee5\u4fbf\u52a0\u6cd5\u64cd\u4f5c\u540e\u8f93\u51fa\u5f20\u91cf\u7684\u5f62\u72b6\u76f8\u540c\u3002<\/p>\n<p>add_norm <span class=\"token operator\">&#061;<\/span> AddNorm<span class=\"token punctuation\">(<\/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 punctuation\">,<\/span> <span class=\"token number\">0.5<\/span><span class=\"token punctuation\">)<\/span><br \/>\nadd_norm<span class=\"token punctuation\">.<\/span><span class=\"token builtin\">eval<\/span><span class=\"token punctuation\">(<\/span><span class=\"token punctuation\">)<\/span><br \/>\nadd_norm<span class=\"token punctuation\">(<\/span>torch<span class=\"token punctuation\">.<\/span>ones<span class=\"token punctuation\">(<\/span><span class=\"token punctuation\">(<\/span><span class=\"token number\">2<\/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 punctuation\">)<\/span><span class=\"token punctuation\">,<\/span> torch<span class=\"token punctuation\">.<\/span>ones<span class=\"token punctuation\">(<\/span><span class=\"token punctuation\">(<\/span><span class=\"token number\">2<\/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 punctuation\">)<\/span><span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">.<\/span>shape<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2025\/05\/20250516013340-682695f45a597.png\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" width=\"220\" \/><\/p>\n<h3>\u56db\u3001\u7f16\u7801\u5668<\/h3>\n<p>\u2003\u2003\u6709\u4e86\u7ec4\u6210Transformer\u7f16\u7801\u5668\u7684\u57fa\u7840\u7ec4\u4ef6&#xff0c;\u73b0\u5728\u53ef\u4ee5\u5148\u5b9e\u73b0\u7f16\u7801\u5668\u4e2d\u7684\u4e00\u4e2a\u5c42\u3002\u4e0b\u9762\u7684EncoderBlock\u7c7b\u5305\u542b\u4e24\u4e2a\u5b50\u5c42&#xff1a;\u591a\u5934\u81ea\u6ce8\u610f\u529b\u548c\u57fa\u4e8e\u4f4d\u7f6e\u7684\u524d\u9988\u7f51\u7edc&#xff0c;\u8fd9\u4e24\u4e2a\u5b50\u5c42\u90fd\u4f7f\u7528\u4e86\u6b8b\u5dee\u8fde\u63a5\u548c\u7d27\u968f\u7684\u5c42\u89c4\u8303\u5316\u3002<\/p>\n<p><span class=\"token comment\">#&#064;save<\/span><br \/>\n<span class=\"token keyword\">class<\/span> <span class=\"token class-name\">EncoderBlock<\/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 triple-quoted-string string\">&#034;&#034;&#034;Transformer\u7f16\u7801\u5668\u5757&#034;&#034;&#034;<\/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> key_size<span class=\"token punctuation\">,<\/span> query_size<span class=\"token punctuation\">,<\/span> value_size<span class=\"token punctuation\">,<\/span> num_hiddens<span class=\"token punctuation\">,<\/span><br \/>\n                 norm_shape<span class=\"token punctuation\">,<\/span> ffn_num_input<span class=\"token punctuation\">,<\/span> ffn_num_hiddens<span class=\"token punctuation\">,<\/span> num_heads<span class=\"token punctuation\">,<\/span><br \/>\n                 dropout<span class=\"token punctuation\">,<\/span> use_bias<span class=\"token operator\">&#061;<\/span><span class=\"token boolean\">False<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token operator\">**<\/span>kwargs<span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">:<\/span><br \/>\n        <span class=\"token builtin\">super<\/span><span class=\"token punctuation\">(<\/span>EncoderBlock<span class=\"token punctuation\">,<\/span> self<span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">.<\/span>__init__<span class=\"token punctuation\">(<\/span><span class=\"token operator\">**<\/span>kwargs<span class=\"token punctuation\">)<\/span><br \/>\n        self<span class=\"token punctuation\">.<\/span>attention <span class=\"token operator\">&#061;<\/span> d2l<span class=\"token punctuation\">.<\/span>MultiHeadAttention<span class=\"token punctuation\">(<\/span><br \/>\n            key_size<span class=\"token punctuation\">,<\/span> query_size<span class=\"token punctuation\">,<\/span> value_size<span class=\"token punctuation\">,<\/span> num_hiddens<span class=\"token punctuation\">,<\/span> num_heads<span class=\"token punctuation\">,<\/span> dropout<span class=\"token punctuation\">,<\/span> use_bias<span class=\"token punctuation\">)<\/span><br \/>\n        self<span class=\"token punctuation\">.<\/span>addnorm1 <span class=\"token operator\">&#061;<\/span> AddNorm<span class=\"token punctuation\">(<\/span>norm_shape<span class=\"token punctuation\">,<\/span> dropout<span class=\"token punctuation\">)<\/span><br \/>\n        self<span class=\"token punctuation\">.<\/span>ffn <span class=\"token operator\">&#061;<\/span> PositionWiseFFN<span class=\"token punctuation\">(<\/span>ffn_num_input<span class=\"token punctuation\">,<\/span> ffn_num_hiddens<span class=\"token punctuation\">,<\/span> num_hiddens<span class=\"token punctuation\">)<\/span><br \/>\n        self<span class=\"token punctuation\">.<\/span>addnorm2 <span class=\"token operator\">&#061;<\/span> AddNorm<span class=\"token punctuation\">(<\/span>norm_shape<span class=\"token punctuation\">,<\/span> dropout<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> X<span class=\"token punctuation\">,<\/span> valid_lens<span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">:<\/span><br \/>\n        Y <span class=\"token operator\">&#061;<\/span> self<span class=\"token punctuation\">.<\/span>addnorm1<span class=\"token punctuation\">(<\/span>X<span class=\"token punctuation\">,<\/span> self<span class=\"token punctuation\">.<\/span>attention<span class=\"token punctuation\">(<\/span>X<span class=\"token punctuation\">,<\/span> X<span class=\"token punctuation\">,<\/span> X<span class=\"token punctuation\">,<\/span> valid_lens<span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">)<\/span><br \/>\n        <span class=\"token keyword\">return<\/span> self<span class=\"token punctuation\">.<\/span>addnorm2<span class=\"token punctuation\">(<\/span>Y<span class=\"token punctuation\">,<\/span> self<span class=\"token punctuation\">.<\/span>ffn<span class=\"token punctuation\">(<\/span>Y<span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">)<\/span><\/p>\n<p>\u2003\u2003\u6b63\u5982\u4ece\u4ee3\u7801\u4e2d\u6240\u770b\u5230\u7684&#xff0c;Transformer\u7f16\u7801\u5668\u4e2d\u7684\u4efb\u4f55\u5c42\u90fd\u4e0d\u4f1a\u6539\u53d8\u5176\u8f93\u5165\u7684\u5f62\u72b6\u3002<\/p>\n<p>X <span class=\"token operator\">&#061;<\/span> torch<span class=\"token punctuation\">.<\/span>ones<span class=\"token punctuation\">(<\/span><span class=\"token punctuation\">(<\/span><span class=\"token number\">2<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token number\">100<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token number\">24<\/span><span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">)<\/span><br \/>\nvalid_lens <span class=\"token operator\">&#061;<\/span> torch<span class=\"token punctuation\">.<\/span>tensor<span class=\"token punctuation\">(<\/span><span class=\"token punctuation\">[<\/span><span class=\"token number\">3<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token number\">2<\/span><span class=\"token punctuation\">]<\/span><span class=\"token punctuation\">)<\/span><br \/>\nencoder_blk <span class=\"token operator\">&#061;<\/span> EncoderBlock<span class=\"token punctuation\">(<\/span><span class=\"token number\">24<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token number\">24<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token number\">24<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token number\">24<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token punctuation\">[<\/span><span class=\"token number\">100<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token number\">24<\/span><span class=\"token punctuation\">]<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token number\">24<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token number\">48<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token number\">8<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token number\">0.5<\/span><span class=\"token punctuation\">)<\/span><br \/>\nencoder_blk<span class=\"token punctuation\">.<\/span><span class=\"token builtin\">eval<\/span><span class=\"token punctuation\">(<\/span><span class=\"token punctuation\">)<\/span><br \/>\nencoder_blk<span class=\"token punctuation\">(<\/span>X<span class=\"token punctuation\">,<\/span> valid_lens<span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">.<\/span>shape<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2025\/05\/20250516013340-682695f462ead.png\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" width=\"240\" \/><\/p>\n<p>\u2003\u2003\u4e0b\u9762\u5b9e\u73b0\u7684Transformer\u7f16\u7801\u5668\u7684\u4ee3\u7801\u4e2d&#xff0c;\u5806\u53e0\u4e86num_layers\u4e2aEncoderBlock\u7c7b\u7684\u5b9e\u4f8b\u3002\u7531\u4e8e\u8fd9\u91cc\u4f7f\u7528\u7684\u662f\u503c\u8303\u56f4\u5728<span class=\"katex--inline\"><span class=\"katex\"><span class=\"katex-mathml\"> <\/p>\n<p>         \u2212 <\/p>\n<p>         1 <\/p>\n<p>        -1 <\/p>\n<p>    <\/span><span class=\"katex-html\"><span class=\"base\"><span class=\"strut\" style=\"height: 0.7278em;vertical-align: -0.0833em\"><\/span><span class=\"mord\">\u2212<\/span><span class=\"mord\">1<\/span><\/span><\/span><\/span><\/span>\u548c<span class=\"katex--inline\"><span class=\"katex\"><span class=\"katex-mathml\"> <\/p>\n<p>         1 <\/p>\n<p>        1 <\/p>\n<p>    <\/span><span class=\"katex-html\"><span class=\"base\"><span class=\"strut\" style=\"height: 0.6444em\"><\/span><span class=\"mord\">1<\/span><\/span><\/span><\/span><\/span>\u4e4b\u95f4\u7684\u56fa\u5b9a\u4f4d\u7f6e\u7f16\u7801&#xff0c;\u56e0\u6b64\u901a\u8fc7\u5b66\u4e60\u5f97\u5230\u7684\u8f93\u5165\u7684\u5d4c\u5165\u8868\u793a\u7684\u503c\u9700\u8981\u5148\u4e58\u4ee5\u5d4c\u5165\u7ef4\u5ea6\u7684\u5e73\u65b9\u6839\u8fdb\u884c\u91cd\u65b0\u7f29\u653e&#xff0c;\u7136\u540e\u518d\u4e0e\u4f4d\u7f6e\u7f16\u7801\u76f8\u52a0\u3002<\/p>\n<p><span class=\"token comment\">#&#064;save<\/span><br \/>\n<span class=\"token keyword\">class<\/span> <span class=\"token class-name\">TransformerEncoder<\/span><span class=\"token punctuation\">(<\/span>d2l<span class=\"token punctuation\">.<\/span>Encoder<span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">:<\/span><br \/>\n    <span class=\"token triple-quoted-string string\">&#034;&#034;&#034;Transformer\u7f16\u7801\u5668&#034;&#034;&#034;<\/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> vocab_size<span class=\"token punctuation\">,<\/span> key_size<span class=\"token punctuation\">,<\/span> query_size<span class=\"token punctuation\">,<\/span> value_size<span class=\"token punctuation\">,<\/span><br \/>\n                 num_hiddens<span class=\"token punctuation\">,<\/span> norm_shape<span class=\"token punctuation\">,<\/span> ffn_num_input<span class=\"token punctuation\">,<\/span> ffn_num_hiddens<span class=\"token punctuation\">,<\/span><br \/>\n                 num_heads<span class=\"token punctuation\">,<\/span> num_layers<span class=\"token punctuation\">,<\/span> dropout<span class=\"token punctuation\">,<\/span> use_bias<span class=\"token operator\">&#061;<\/span><span class=\"token boolean\">False<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token operator\">**<\/span>kwargs<span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">:<\/span><br \/>\n        <span class=\"token builtin\">super<\/span><span class=\"token punctuation\">(<\/span>TransformerEncoder<span class=\"token punctuation\">,<\/span> self<span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">.<\/span>__init__<span class=\"token punctuation\">(<\/span><span class=\"token operator\">**<\/span>kwargs<span class=\"token punctuation\">)<\/span><br \/>\n        self<span class=\"token punctuation\">.<\/span>num_hiddens <span class=\"token operator\">&#061;<\/span> num_hiddens<br \/>\n        self<span class=\"token punctuation\">.<\/span>embedding <span class=\"token operator\">&#061;<\/span> nn<span class=\"token punctuation\">.<\/span>Embedding<span class=\"token punctuation\">(<\/span>vocab_size<span class=\"token punctuation\">,<\/span> num_hiddens<span class=\"token punctuation\">)<\/span><br \/>\n        self<span class=\"token punctuation\">.<\/span>pos_encoding <span class=\"token operator\">&#061;<\/span> d2l<span class=\"token punctuation\">.<\/span>PositionalEncoding<span class=\"token punctuation\">(<\/span>num_hiddens<span class=\"token punctuation\">,<\/span> dropout<span class=\"token punctuation\">)<\/span><br \/>\n        self<span class=\"token punctuation\">.<\/span>blks <span class=\"token operator\">&#061;<\/span> nn<span class=\"token punctuation\">.<\/span>Sequential<span class=\"token punctuation\">(<\/span><span class=\"token punctuation\">)<\/span><br \/>\n        <span class=\"token keyword\">for<\/span> i <span class=\"token keyword\">in<\/span> <span class=\"token builtin\">range<\/span><span class=\"token punctuation\">(<\/span>num_layers<span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">:<\/span><br \/>\n            self<span class=\"token punctuation\">.<\/span>blks<span class=\"token punctuation\">.<\/span>add_module<span class=\"token punctuation\">(<\/span><span class=\"token string\">&#034;block&#034;<\/span><span class=\"token operator\">&#043;<\/span><span class=\"token builtin\">str<\/span><span class=\"token punctuation\">(<\/span>i<span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">,<\/span><br \/>\n                EncoderBlock<span class=\"token punctuation\">(<\/span>key_size<span class=\"token punctuation\">,<\/span> query_size<span class=\"token punctuation\">,<\/span> value_size<span class=\"token punctuation\">,<\/span> num_hiddens<span class=\"token punctuation\">,<\/span><br \/>\n                             norm_shape<span class=\"token punctuation\">,<\/span> ffn_num_input<span class=\"token punctuation\">,<\/span> ffn_num_hiddens<span class=\"token punctuation\">,<\/span> num_heads<span class=\"token punctuation\">,<\/span> dropout<span class=\"token punctuation\">,<\/span> use_bias<span class=\"token punctuation\">)<\/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> X<span class=\"token punctuation\">,<\/span> valid_lens<span class=\"token punctuation\">,<\/span> <span class=\"token operator\">*<\/span>args<span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">:<\/span><br \/>\n        <span class=\"token comment\"># \u56e0\u4e3a\u4f4d\u7f6e\u7f16\u7801\u503c\u5728-1\u548c1\u4e4b\u95f4&#xff0c;<\/span><br \/>\n        <span class=\"token comment\"># \u56e0\u6b64\u5d4c\u5165\u503c\u4e58\u4ee5\u5d4c\u5165\u7ef4\u5ea6\u7684\u5e73\u65b9\u6839\u8fdb\u884c\u7f29\u653e&#xff0c;<\/span><br \/>\n        <span class=\"token comment\"># \u7136\u540e\u518d\u4e0e\u4f4d\u7f6e\u7f16\u7801\u76f8\u52a0\u3002<\/span><br \/>\n        X <span class=\"token operator\">&#061;<\/span> self<span class=\"token punctuation\">.<\/span>pos_encoding<span class=\"token punctuation\">(<\/span>self<span class=\"token punctuation\">.<\/span>embedding<span class=\"token punctuation\">(<\/span>X<span class=\"token punctuation\">)<\/span> <span class=\"token operator\">*<\/span> math<span class=\"token punctuation\">.<\/span>sqrt<span class=\"token punctuation\">(<\/span>self<span class=\"token punctuation\">.<\/span>num_hiddens<span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">)<\/span><br \/>\n        self<span class=\"token punctuation\">.<\/span>attention_weights <span class=\"token operator\">&#061;<\/span> <span class=\"token punctuation\">[<\/span><span class=\"token boolean\">None<\/span><span class=\"token punctuation\">]<\/span> <span class=\"token operator\">*<\/span> <span class=\"token builtin\">len<\/span><span class=\"token punctuation\">(<\/span>self<span class=\"token punctuation\">.<\/span>blks<span class=\"token punctuation\">)<\/span><br \/>\n        <span class=\"token keyword\">for<\/span> i<span class=\"token punctuation\">,<\/span> blk <span class=\"token keyword\">in<\/span> <span class=\"token builtin\">enumerate<\/span><span class=\"token punctuation\">(<\/span>self<span class=\"token punctuation\">.<\/span>blks<span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">:<\/span><br \/>\n            X <span class=\"token operator\">&#061;<\/span> blk<span class=\"token punctuation\">(<\/span>X<span class=\"token punctuation\">,<\/span> valid_lens<span class=\"token punctuation\">)<\/span><br \/>\n            self<span class=\"token punctuation\">.<\/span>attention_weights<span class=\"token punctuation\">[<\/span>i<span class=\"token punctuation\">]<\/span> <span class=\"token operator\">&#061;<\/span> blk<span class=\"token punctuation\">.<\/span>attention<span class=\"token punctuation\">.<\/span>attention<span class=\"token punctuation\">.<\/span>attention_weights<br \/>\n        <span class=\"token keyword\">return<\/span> X<\/p>\n<p>\u2003\u2003\u4e0b\u9762\u6211\u4eec\u6307\u5b9a\u4e86\u8d85\u53c2\u6570\u6765\u521b\u5efa\u4e00\u4e2a\u4e24\u5c42\u7684Transformer\u7f16\u7801\u5668\u3002Transformer\u7f16\u7801\u5668\u8f93\u51fa\u7684\u5f62\u72b6\u662f&#xff08;\u6279\u91cf\u5927\u5c0f&#xff0c;\u65f6\u95f4\u6b65\u6570\u76ee&#xff0c;num_hiddens&#xff09;\u3002<\/p>\n<p>encoder <span class=\"token operator\">&#061;<\/span> TransformerEncoder<span class=\"token punctuation\">(<\/span><span class=\"token number\">200<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token number\">24<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token number\">24<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token number\">24<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token number\">24<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token punctuation\">[<\/span><span class=\"token number\">100<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token number\">24<\/span><span class=\"token punctuation\">]<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token number\">24<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token number\">48<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token number\">8<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token number\">2<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token number\">0.5<\/span><span class=\"token punctuation\">)<\/span><br \/>\nencoder<span class=\"token punctuation\">.<\/span><span class=\"token builtin\">eval<\/span><span class=\"token punctuation\">(<\/span><span class=\"token punctuation\">)<\/span><br \/>\nencoder<span class=\"token punctuation\">(<\/span>torch<span class=\"token punctuation\">.<\/span>ones<span class=\"token punctuation\">(<\/span><span class=\"token punctuation\">(<\/span><span class=\"token number\">2<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token number\">100<\/span><span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">,<\/span> dtype<span class=\"token operator\">&#061;<\/span>torch<span class=\"token punctuation\">.<\/span><span class=\"token builtin\">long<\/span><span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">,<\/span> valid_lens<span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">.<\/span>shape<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2025\/05\/20250516013340-682695f46b6b7.png\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" width=\"240\" \/><\/p>\n<h3>\u4e94\u3001\u89e3\u7801\u5668<\/h3>\n<p>\u2003\u2003\u5982\u56fe1\u6240\u793a&#xff0c;Transformer\u89e3\u7801\u5668\u4e5f\u662f\u7531\u591a\u4e2a\u76f8\u540c\u7684\u5c42\u7ec4\u6210\u3002\u5728DecoderBlock\u7c7b\u4e2d\u5b9e\u73b0\u7684\u6bcf\u4e2a\u5c42\u5305\u542b\u4e86\u4e09\u4e2a\u5b50\u5c42&#xff1a;\u89e3\u7801\u5668\u81ea\u6ce8\u610f\u529b\u3001\u201c\u7f16\u7801\u5668-\u89e3\u7801\u5668\u201d\u6ce8\u610f\u529b\u548c\u57fa\u4e8e\u4f4d\u7f6e\u7684\u524d\u9988\u7f51\u7edc\u3002\u8fd9\u4e9b\u5b50\u5c42\u4e5f\u90fd\u88ab\u6b8b\u5dee\u8fde\u63a5\u548c\u7d27\u968f\u7684\u5c42\u89c4\u8303\u5316\u56f4\u7ed5\u3002<\/p>\n<p>\u2003\u2003\u6b63\u5982\u5728\u672c\u8282\u524d\u9762\u6240\u8ff0&#xff0c;\u5728\u63a9\u853d\u591a\u5934\u89e3\u7801\u5668\u81ea\u6ce8\u610f\u529b\u5c42&#xff08;\u7b2c\u4e00\u4e2a\u5b50\u5c42&#xff09;\u4e2d&#xff0c;\u67e5\u8be2\u3001\u952e\u548c\u503c\u90fd\u6765\u81ea\u4e0a\u4e00\u4e2a\u89e3\u7801\u5668\u5c42\u7684\u8f93\u51fa\u3002\u5173\u4e8e\u5e8f\u5217\u5230\u5e8f\u5217\u6a21\u578b&#xff08;sequence-to-sequence model&#xff09;&#xff0c;\u5728\u8bad\u7ec3\u9636\u6bb5&#xff0c;\u5176\u8f93\u51fa\u5e8f\u5217\u7684\u6240\u6709\u4f4d\u7f6e&#xff08;\u65f6\u95f4\u6b65&#xff09;\u7684\u8bcd\u5143\u90fd\u662f\u5df2\u77e5\u7684&#xff1b;\u7136\u800c&#xff0c;\u5728\u9884\u6d4b\u9636\u6bb5&#xff0c;\u5176\u8f93\u51fa\u5e8f\u5217\u7684\u8bcd\u5143\u662f\u9010\u4e2a\u751f\u6210\u7684\u3002\u56e0\u6b64&#xff0c;\u5728\u4efb\u4f55\u89e3\u7801\u5668\u65f6\u95f4\u6b65\u4e2d&#xff0c;\u53ea\u6709\u751f\u6210\u7684\u8bcd\u5143\u624d\u80fd\u7528\u4e8e\u89e3\u7801\u5668\u7684\u81ea\u6ce8\u610f\u529b\u8ba1\u7b97\u4e2d\u3002\u4e3a\u4e86\u5728\u89e3\u7801\u5668\u4e2d\u4fdd\u7559\u81ea\u56de\u5f52\u7684\u5c5e\u6027&#xff0c;\u5176\u63a9\u853d\u81ea\u6ce8\u610f\u529b\u8bbe\u5b9a\u4e86\u53c2\u6570dec_valid_lens&#xff0c;\u4ee5\u4fbf\u4efb\u4f55\u67e5\u8be2\u90fd\u53ea\u4f1a\u4e0e\u89e3\u7801\u5668\u4e2d\u6240\u6709\u5df2\u7ecf\u751f\u6210\u8bcd\u5143\u7684\u4f4d\u7f6e&#xff08;\u5373\u76f4\u5230\u8be5\u67e5\u8be2\u4f4d\u7f6e\u4e3a\u6b62&#xff09;\u8fdb\u884c\u6ce8\u610f\u529b\u8ba1\u7b97\u3002<\/p>\n<p><span class=\"token keyword\">class<\/span> <span class=\"token class-name\">DecoderBlock<\/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 triple-quoted-string string\">&#034;&#034;&#034;\u89e3\u7801\u5668\u4e2d\u7b2ci\u4e2a\u5757&#034;&#034;&#034;<\/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> key_size<span class=\"token punctuation\">,<\/span> query_size<span class=\"token punctuation\">,<\/span> value_size<span class=\"token punctuation\">,<\/span> num_hiddens<span class=\"token punctuation\">,<\/span><br \/>\n                 norm_shape<span class=\"token punctuation\">,<\/span> ffn_num_input<span class=\"token punctuation\">,<\/span> ffn_num_hiddens<span class=\"token punctuation\">,<\/span> num_heads<span class=\"token punctuation\">,<\/span><br \/>\n                 dropout<span class=\"token punctuation\">,<\/span> i<span class=\"token punctuation\">,<\/span> <span class=\"token operator\">**<\/span>kwargs<span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">:<\/span><br \/>\n        <span class=\"token builtin\">super<\/span><span class=\"token punctuation\">(<\/span>DecoderBlock<span class=\"token punctuation\">,<\/span> self<span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">.<\/span>__init__<span class=\"token punctuation\">(<\/span><span class=\"token operator\">**<\/span>kwargs<span class=\"token punctuation\">)<\/span><br \/>\n        self<span class=\"token punctuation\">.<\/span>i <span class=\"token operator\">&#061;<\/span> i<br \/>\n        self<span class=\"token punctuation\">.<\/span>attention1 <span class=\"token operator\">&#061;<\/span> d2l<span class=\"token punctuation\">.<\/span>MultiHeadAttention<span class=\"token punctuation\">(<\/span>key_size<span class=\"token punctuation\">,<\/span> query_size<span class=\"token punctuation\">,<\/span> value_size<span class=\"token punctuation\">,<\/span> num_hiddens<span class=\"token punctuation\">,<\/span> num_heads<span class=\"token punctuation\">,<\/span> dropout<span class=\"token punctuation\">)<\/span><br \/>\n        self<span class=\"token punctuation\">.<\/span>addnorm1 <span class=\"token operator\">&#061;<\/span> AddNorm<span class=\"token punctuation\">(<\/span>norm_shape<span class=\"token punctuation\">,<\/span> dropout<span class=\"token punctuation\">)<\/span><br \/>\n        self<span class=\"token punctuation\">.<\/span>attention2 <span class=\"token operator\">&#061;<\/span> d2l<span class=\"token punctuation\">.<\/span>MultiHeadAttention<span class=\"token punctuation\">(<\/span>key_size<span class=\"token punctuation\">,<\/span> query_size<span class=\"token punctuation\">,<\/span> value_size<span class=\"token punctuation\">,<\/span> num_hiddens<span class=\"token punctuation\">,<\/span> num_heads<span class=\"token punctuation\">,<\/span> dropout<span class=\"token punctuation\">)<\/span><br \/>\n        self<span class=\"token punctuation\">.<\/span>addnorm2 <span class=\"token operator\">&#061;<\/span> AddNorm<span class=\"token punctuation\">(<\/span>norm_shape<span class=\"token punctuation\">,<\/span> dropout<span class=\"token punctuation\">)<\/span><br \/>\n        self<span class=\"token punctuation\">.<\/span>ffn <span class=\"token operator\">&#061;<\/span> PositionWiseFFN<span class=\"token punctuation\">(<\/span>ffn_num_input<span class=\"token punctuation\">,<\/span> ffn_num_hiddens<span class=\"token punctuation\">,<\/span> num_hiddens<span class=\"token punctuation\">)<\/span><br \/>\n        self<span class=\"token punctuation\">.<\/span>addnorm3 <span class=\"token operator\">&#061;<\/span> AddNorm<span class=\"token punctuation\">(<\/span>norm_shape<span class=\"token punctuation\">,<\/span> dropout<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> X<span class=\"token punctuation\">,<\/span> state<span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">:<\/span><br \/>\n        enc_outputs<span class=\"token punctuation\">,<\/span> enc_valid_lens <span class=\"token operator\">&#061;<\/span> state<span class=\"token punctuation\">[<\/span><span class=\"token number\">0<\/span><span class=\"token punctuation\">]<\/span><span class=\"token punctuation\">,<\/span> state<span class=\"token punctuation\">[<\/span><span class=\"token number\">1<\/span><span class=\"token punctuation\">]<\/span><br \/>\n        <span class=\"token comment\"># \u8bad\u7ec3\u9636\u6bb5&#xff0c;\u8f93\u51fa\u5e8f\u5217\u7684\u6240\u6709\u8bcd\u5143\u90fd\u5728\u540c\u4e00\u65f6\u95f4\u5904\u7406&#xff0c;<\/span><br \/>\n        <span class=\"token comment\"># \u56e0\u6b64state[2][self.i]\u521d\u59cb\u5316\u4e3aNone\u3002<\/span><br \/>\n        <span class=\"token comment\"># \u9884\u6d4b\u9636\u6bb5&#xff0c;\u8f93\u51fa\u5e8f\u5217\u662f\u901a\u8fc7\u8bcd\u5143\u4e00\u4e2a\u63a5\u7740\u4e00\u4e2a\u89e3\u7801\u7684&#xff0c;<\/span><br \/>\n        <span class=\"token comment\"># \u56e0\u6b64state[2][self.i]\u5305\u542b\u7740\u76f4\u5230\u5f53\u524d\u65f6\u95f4\u6b65\u7b2ci\u4e2a\u5757\u89e3\u7801\u7684\u8f93\u51fa\u8868\u793a<\/span><br \/>\n        <span class=\"token keyword\">if<\/span> state<span class=\"token punctuation\">[<\/span><span class=\"token number\">2<\/span><span class=\"token punctuation\">]<\/span><span class=\"token punctuation\">[<\/span>self<span class=\"token punctuation\">.<\/span>i<span class=\"token punctuation\">]<\/span> <span class=\"token keyword\">is<\/span> <span class=\"token boolean\">None<\/span><span class=\"token punctuation\">:<\/span><br \/>\n            key_values <span class=\"token operator\">&#061;<\/span> X<br \/>\n        <span class=\"token keyword\">else<\/span><span class=\"token punctuation\">:<\/span><br \/>\n            key_values <span class=\"token operator\">&#061;<\/span> torch<span class=\"token punctuation\">.<\/span>cat<span class=\"token punctuation\">(<\/span><span class=\"token punctuation\">(<\/span>state<span class=\"token punctuation\">[<\/span><span class=\"token number\">2<\/span><span class=\"token punctuation\">]<\/span><span class=\"token punctuation\">[<\/span>self<span class=\"token punctuation\">.<\/span>i<span class=\"token punctuation\">]<\/span><span class=\"token punctuation\">,<\/span> X<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><br \/>\n        state<span class=\"token punctuation\">[<\/span><span class=\"token number\">2<\/span><span class=\"token punctuation\">]<\/span><span class=\"token punctuation\">[<\/span>self<span class=\"token punctuation\">.<\/span>i<span class=\"token punctuation\">]<\/span> <span class=\"token operator\">&#061;<\/span> key_values<br \/>\n        <span class=\"token keyword\">if<\/span> self<span class=\"token punctuation\">.<\/span>training<span class=\"token punctuation\">:<\/span><br \/>\n            batch_size<span class=\"token punctuation\">,<\/span> num_steps<span class=\"token punctuation\">,<\/span> _ <span class=\"token operator\">&#061;<\/span> X<span class=\"token punctuation\">.<\/span>shape<br \/>\n            <span class=\"token comment\"># dec_valid_lens\u7684\u5f00\u5934:(batch_size,num_steps),<\/span><br \/>\n            <span class=\"token comment\"># \u5176\u4e2d\u6bcf\u4e00\u884c\u662f[1,2,&#8230;,num_steps]<\/span><br \/>\n            dec_valid_lens <span class=\"token operator\">&#061;<\/span> torch<span class=\"token punctuation\">.<\/span>arange<span class=\"token punctuation\">(<\/span><span class=\"token number\">1<\/span><span class=\"token punctuation\">,<\/span> num_steps <span class=\"token operator\">&#043;<\/span> <span class=\"token number\">1<\/span><span class=\"token punctuation\">,<\/span> device<span class=\"token operator\">&#061;<\/span>X<span class=\"token punctuation\">.<\/span>device<span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">.<\/span>repeat<span class=\"token punctuation\">(<\/span>batch_size<span class=\"token punctuation\">,<\/span> <span class=\"token number\">1<\/span><span class=\"token punctuation\">)<\/span><br \/>\n        <span class=\"token keyword\">else<\/span><span class=\"token punctuation\">:<\/span><br \/>\n            dec_valid_lens <span class=\"token operator\">&#061;<\/span> <span class=\"token boolean\">None<\/span><\/p>\n<p>        <span class=\"token comment\"># \u81ea\u6ce8\u610f\u529b<\/span><br \/>\n        X2 <span class=\"token operator\">&#061;<\/span> self<span class=\"token punctuation\">.<\/span>attention1<span class=\"token punctuation\">(<\/span>X<span class=\"token punctuation\">,<\/span> key_values<span class=\"token punctuation\">,<\/span> key_values<span class=\"token punctuation\">,<\/span> dec_valid_lens<span class=\"token punctuation\">)<\/span><br \/>\n        Y <span class=\"token operator\">&#061;<\/span> self<span class=\"token punctuation\">.<\/span>addnorm1<span class=\"token punctuation\">(<\/span>X<span class=\"token punctuation\">,<\/span> X2<span class=\"token punctuation\">)<\/span><br \/>\n        <span class=\"token comment\"># \u7f16\u7801\u5668&#xff0d;\u89e3\u7801\u5668\u6ce8\u610f\u529b\u3002<\/span><br \/>\n        <span class=\"token comment\"># enc_outputs\u7684\u5f00\u5934:(batch_size,num_steps,num_hiddens)<\/span><br \/>\n        Y2 <span class=\"token operator\">&#061;<\/span> self<span class=\"token punctuation\">.<\/span>attention2<span class=\"token punctuation\">(<\/span>Y<span class=\"token punctuation\">,<\/span> enc_outputs<span class=\"token punctuation\">,<\/span> enc_outputs<span class=\"token punctuation\">,<\/span> enc_valid_lens<span class=\"token punctuation\">)<\/span><br \/>\n        Z <span class=\"token operator\">&#061;<\/span> self<span class=\"token punctuation\">.<\/span>addnorm2<span class=\"token punctuation\">(<\/span>Y<span class=\"token punctuation\">,<\/span> Y2<span class=\"token punctuation\">)<\/span><br \/>\n        <span class=\"token keyword\">return<\/span> self<span class=\"token punctuation\">.<\/span>addnorm3<span class=\"token punctuation\">(<\/span>Z<span class=\"token punctuation\">,<\/span> self<span class=\"token punctuation\">.<\/span>ffn<span class=\"token punctuation\">(<\/span>Z<span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">,<\/span> state<\/p>\n<p>\u2003\u2003\u4e3a\u4e86\u4fbf\u4e8e\u5728\u201c\u7f16\u7801\u5668-\u89e3\u7801\u5668\u201d\u6ce8\u610f\u529b\u4e2d\u8fdb\u884c\u7f29\u653e\u70b9\u79ef\u8ba1\u7b97\u548c\u6b8b\u5dee\u8fde\u63a5\u4e2d\u8fdb\u884c\u52a0\u6cd5\u8ba1\u7b97&#xff0c;\u7f16\u7801\u5668\u548c\u89e3\u7801\u5668\u7684\u7279\u5f81\u7ef4\u5ea6\u90fd\u662fnum_hiddens\u3002<\/p>\n<p>decoder_blk <span class=\"token operator\">&#061;<\/span> DecoderBlock<span class=\"token punctuation\">(<\/span><span class=\"token number\">24<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token number\">24<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token number\">24<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token number\">24<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token punctuation\">[<\/span><span class=\"token number\">100<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token number\">24<\/span><span class=\"token punctuation\">]<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token number\">24<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token number\">48<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token number\">8<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token number\">0.5<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token number\">0<\/span><span class=\"token punctuation\">)<\/span><br \/>\ndecoder_blk<span class=\"token punctuation\">.<\/span><span class=\"token builtin\">eval<\/span><span class=\"token punctuation\">(<\/span><span class=\"token punctuation\">)<\/span><br \/>\nX <span class=\"token operator\">&#061;<\/span> torch<span class=\"token punctuation\">.<\/span>ones<span class=\"token punctuation\">(<\/span><span class=\"token punctuation\">(<\/span><span class=\"token number\">2<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token number\">100<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token number\">24<\/span><span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">)<\/span><br \/>\nstate <span class=\"token operator\">&#061;<\/span> <span class=\"token punctuation\">[<\/span>encoder_blk<span class=\"token punctuation\">(<\/span>X<span class=\"token punctuation\">,<\/span> valid_lens<span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">,<\/span> valid_lens<span class=\"token punctuation\">,<\/span> <span class=\"token punctuation\">[<\/span><span class=\"token boolean\">None<\/span><span class=\"token punctuation\">]<\/span><span class=\"token punctuation\">]<\/span><br \/>\ndecoder_blk<span class=\"token punctuation\">(<\/span>X<span class=\"token punctuation\">,<\/span> state<span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">[<\/span><span class=\"token number\">0<\/span><span class=\"token punctuation\">]<\/span><span class=\"token punctuation\">.<\/span>shape<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2025\/05\/20250516013340-682695f474d79.png\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" width=\"220\" \/><\/p>\n<p>\u2003\u2003\u73b0\u5728\u6211\u4eec\u6784\u5efa\u4e86\u7531num_layers\u4e2aDecoderBlock\u5b9e\u4f8b\u7ec4\u6210\u7684\u5b8c\u6574\u7684Transformer\u89e3\u7801\u5668\u3002\u6700\u540e&#xff0c;\u901a\u8fc7\u4e00\u4e2a\u5168\u8fde\u63a5\u5c42\u8ba1\u7b97\u6240\u6709vocab_size\u4e2a\u53ef\u80fd\u7684\u8f93\u51fa\u8bcd\u5143\u7684\u9884\u6d4b\u503c\u3002\u89e3\u7801\u5668\u7684\u81ea\u6ce8\u610f\u529b\u6743\u91cd\u548c\u7f16\u7801\u5668\u89e3\u7801\u5668\u6ce8\u610f\u529b\u6743\u91cd\u90fd\u88ab\u5b58\u50a8\u4e0b\u6765&#xff0c;\u65b9\u4fbf\u65e5\u540e\u53ef\u89c6\u5316\u7684\u9700\u8981\u3002<\/p>\n<p><span class=\"token keyword\">class<\/span> <span class=\"token class-name\">TransformerDecoder<\/span><span class=\"token punctuation\">(<\/span>d2l<span class=\"token punctuation\">.<\/span>AttentionDecoder<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> vocab_size<span class=\"token punctuation\">,<\/span> key_size<span class=\"token punctuation\">,<\/span> query_size<span class=\"token punctuation\">,<\/span> value_size<span class=\"token punctuation\">,<\/span><br \/>\n                 num_hiddens<span class=\"token punctuation\">,<\/span> norm_shape<span class=\"token punctuation\">,<\/span> ffn_num_input<span class=\"token punctuation\">,<\/span> ffn_num_hiddens<span class=\"token punctuation\">,<\/span><br \/>\n                 num_heads<span class=\"token punctuation\">,<\/span> num_layers<span class=\"token punctuation\">,<\/span> dropout<span class=\"token punctuation\">,<\/span> <span class=\"token operator\">**<\/span>kwargs<span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">:<\/span><br \/>\n        <span class=\"token builtin\">super<\/span><span class=\"token punctuation\">(<\/span>TransformerDecoder<span class=\"token punctuation\">,<\/span> self<span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">.<\/span>__init__<span class=\"token punctuation\">(<\/span><span class=\"token operator\">**<\/span>kwargs<span class=\"token punctuation\">)<\/span><br \/>\n        self<span class=\"token punctuation\">.<\/span>num_hiddens <span class=\"token operator\">&#061;<\/span> num_hiddens<br \/>\n        self<span class=\"token punctuation\">.<\/span>num_layers <span class=\"token operator\">&#061;<\/span> num_layers<br \/>\n        self<span class=\"token punctuation\">.<\/span>embedding <span class=\"token operator\">&#061;<\/span> nn<span class=\"token punctuation\">.<\/span>Embedding<span class=\"token punctuation\">(<\/span>vocab_size<span class=\"token punctuation\">,<\/span> num_hiddens<span class=\"token punctuation\">)<\/span><br \/>\n        self<span class=\"token punctuation\">.<\/span>pos_encoding <span class=\"token operator\">&#061;<\/span> d2l<span class=\"token punctuation\">.<\/span>PositionalEncoding<span class=\"token punctuation\">(<\/span>num_hiddens<span class=\"token punctuation\">,<\/span> dropout<span class=\"token punctuation\">)<\/span><br \/>\n        self<span class=\"token punctuation\">.<\/span>blks <span class=\"token operator\">&#061;<\/span> nn<span class=\"token punctuation\">.<\/span>Sequential<span class=\"token punctuation\">(<\/span><span class=\"token punctuation\">)<\/span><br \/>\n        <span class=\"token keyword\">for<\/span> i <span class=\"token keyword\">in<\/span> <span class=\"token builtin\">range<\/span><span class=\"token punctuation\">(<\/span>num_layers<span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">:<\/span><br \/>\n            self<span class=\"token punctuation\">.<\/span>blks<span class=\"token punctuation\">.<\/span>add_module<span class=\"token punctuation\">(<\/span><span class=\"token string\">&#034;block&#034;<\/span><span class=\"token operator\">&#043;<\/span><span class=\"token builtin\">str<\/span><span class=\"token punctuation\">(<\/span>i<span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">,<\/span><br \/>\n                DecoderBlock<span class=\"token punctuation\">(<\/span>key_size<span class=\"token punctuation\">,<\/span> query_size<span class=\"token punctuation\">,<\/span> value_size<span class=\"token punctuation\">,<\/span> num_hiddens<span class=\"token punctuation\">,<\/span><br \/>\n                             norm_shape<span class=\"token punctuation\">,<\/span> ffn_num_input<span class=\"token punctuation\">,<\/span> ffn_num_hiddens<span class=\"token punctuation\">,<\/span> num_heads<span class=\"token punctuation\">,<\/span> dropout<span class=\"token punctuation\">,<\/span> i<span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">)<\/span><br \/>\n        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>num_hiddens<span class=\"token punctuation\">,<\/span> vocab_size<span class=\"token punctuation\">)<\/span><\/p>\n<p>    <span class=\"token keyword\">def<\/span> <span class=\"token function\">init_state<\/span><span class=\"token punctuation\">(<\/span>self<span class=\"token punctuation\">,<\/span> enc_outputs<span class=\"token punctuation\">,<\/span> enc_valid_lens<span class=\"token punctuation\">,<\/span> <span class=\"token operator\">*<\/span>args<span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">:<\/span><br \/>\n        <span class=\"token keyword\">return<\/span> <span class=\"token punctuation\">[<\/span>enc_outputs<span class=\"token punctuation\">,<\/span> enc_valid_lens<span class=\"token punctuation\">,<\/span> <span class=\"token punctuation\">[<\/span><span class=\"token boolean\">None<\/span><span class=\"token punctuation\">]<\/span> <span class=\"token operator\">*<\/span> self<span class=\"token punctuation\">.<\/span>num_layers<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> X<span class=\"token punctuation\">,<\/span> state<span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">:<\/span><br \/>\n        X <span class=\"token operator\">&#061;<\/span> self<span class=\"token punctuation\">.<\/span>pos_encoding<span class=\"token punctuation\">(<\/span>self<span class=\"token punctuation\">.<\/span>embedding<span class=\"token punctuation\">(<\/span>X<span class=\"token punctuation\">)<\/span> <span class=\"token operator\">*<\/span> math<span class=\"token punctuation\">.<\/span>sqrt<span class=\"token punctuation\">(<\/span>self<span class=\"token punctuation\">.<\/span>num_hiddens<span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">)<\/span><br \/>\n        self<span class=\"token punctuation\">.<\/span>_attention_weights <span class=\"token operator\">&#061;<\/span> <span class=\"token punctuation\">[<\/span><span class=\"token punctuation\">[<\/span><span class=\"token boolean\">None<\/span><span class=\"token punctuation\">]<\/span> <span class=\"token operator\">*<\/span> <span class=\"token builtin\">len<\/span><span class=\"token punctuation\">(<\/span>self<span class=\"token punctuation\">.<\/span>blks<span class=\"token punctuation\">)<\/span> <span class=\"token keyword\">for<\/span> _ <span class=\"token keyword\">in<\/span> <span class=\"token builtin\">range<\/span> <span class=\"token punctuation\">(<\/span><span class=\"token number\">2<\/span><span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">]<\/span><br \/>\n        <span class=\"token keyword\">for<\/span> i<span class=\"token punctuation\">,<\/span> blk <span class=\"token keyword\">in<\/span> <span class=\"token builtin\">enumerate<\/span><span class=\"token punctuation\">(<\/span>self<span class=\"token punctuation\">.<\/span>blks<span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">:<\/span><br \/>\n            X<span class=\"token punctuation\">,<\/span> state <span class=\"token operator\">&#061;<\/span> blk<span class=\"token punctuation\">(<\/span>X<span class=\"token punctuation\">,<\/span> state<span class=\"token punctuation\">)<\/span><br \/>\n            <span class=\"token comment\"># \u89e3\u7801\u5668\u81ea\u6ce8\u610f\u529b\u6743\u91cd<\/span><br \/>\n            self<span class=\"token punctuation\">.<\/span>_attention_weights<span class=\"token punctuation\">[<\/span><span class=\"token number\">0<\/span><span class=\"token punctuation\">]<\/span><span class=\"token punctuation\">[<\/span>i<span class=\"token punctuation\">]<\/span> <span class=\"token operator\">&#061;<\/span> blk<span class=\"token punctuation\">.<\/span>attention1<span class=\"token punctuation\">.<\/span>attention<span class=\"token punctuation\">.<\/span>attention_weights<br \/>\n            <span class=\"token comment\"># \u201c\u7f16\u7801\u5668&#xff0d;\u89e3\u7801\u5668\u201d\u81ea\u6ce8\u610f\u529b\u6743\u91cd<\/span><br \/>\n            self<span class=\"token punctuation\">.<\/span>_attention_weights<span class=\"token punctuation\">[<\/span><span class=\"token number\">1<\/span><span class=\"token punctuation\">]<\/span><span class=\"token punctuation\">[<\/span>i<span class=\"token punctuation\">]<\/span> <span class=\"token operator\">&#061;<\/span> blk<span class=\"token punctuation\">.<\/span>attention2<span class=\"token punctuation\">.<\/span>attention<span class=\"token punctuation\">.<\/span>attention_weights<br \/>\n        <span class=\"token keyword\">return<\/span> self<span class=\"token punctuation\">.<\/span>dense<span class=\"token punctuation\">(<\/span>X<span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">,<\/span> state<\/p>\n<p>    <span class=\"token decorator annotation punctuation\">&#064;property<\/span><br \/>\n    <span class=\"token keyword\">def<\/span> <span class=\"token function\">attention_weights<\/span><span class=\"token punctuation\">(<\/span>self<span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">:<\/span><br \/>\n        <span class=\"token keyword\">return<\/span> self<span class=\"token punctuation\">.<\/span>_attention_weights<\/p>\n<h3>\u516d\u3001\u8bad\u7ec3<\/h3>\n<p>\u2003\u2003\u4f9d\u7167Transformer\u67b6\u6784\u6765\u5b9e\u4f8b\u5316\u7f16\u7801\u5668-\u89e3\u7801\u5668\u6a21\u578b\u3002\u5728\u8fd9\u91cc&#xff0c;\u6307\u5b9aTransformer\u7684\u7f16\u7801\u5668\u548c\u89e3\u7801\u5668\u90fd\u662f2\u5c42&#xff0c;\u90fd\u4f7f\u75284\u5934\u6ce8\u610f\u529b\u3002\u4e0e\u5e8f\u5217\u5230\u5e8f\u5217\u5b66\u4e60&#xff08;seq2seq&#xff09;\u7c7b\u4f3c&#xff0c;\u4e3a\u4e86\u8fdb\u884c\u5e8f\u5217\u5230\u5e8f\u5217\u7684\u5b66\u4e60&#xff0c;\u4e0b\u9762\u5728\u201c\u82f1\u8bed-\u6cd5\u8bed\u201d\u673a\u5668\u7ffb\u8bd1\u6570\u636e\u96c6\u4e0a\u8bad\u7ec3Transformer\u6a21\u578b\u3002<\/p>\n<p>num_hiddens<span class=\"token punctuation\">,<\/span> num_layers<span class=\"token punctuation\">,<\/span> dropout<span class=\"token punctuation\">,<\/span> batch_size<span class=\"token punctuation\">,<\/span> num_steps <span class=\"token operator\">&#061;<\/span> <span class=\"token number\">32<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token number\">2<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token number\">0.1<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token number\">64<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token number\">10<\/span><br \/>\nlr<span class=\"token punctuation\">,<\/span> num_epochs<span class=\"token punctuation\">,<\/span> device <span class=\"token operator\">&#061;<\/span> <span class=\"token number\">0.005<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token number\">200<\/span><span class=\"token punctuation\">,<\/span> d2l<span class=\"token punctuation\">.<\/span>try_gpu<span class=\"token punctuation\">(<\/span><span class=\"token punctuation\">)<\/span><br \/>\nffn_num_input<span class=\"token punctuation\">,<\/span> ffn_num_hiddens<span class=\"token punctuation\">,<\/span> num_heads <span class=\"token operator\">&#061;<\/span> <span class=\"token number\">32<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token number\">64<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token number\">4<\/span><br \/>\nkey_size<span class=\"token punctuation\">,<\/span> query_size<span class=\"token punctuation\">,<\/span> value_size <span class=\"token operator\">&#061;<\/span> <span class=\"token number\">32<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token number\">32<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token number\">32<\/span><br \/>\nnorm_shape <span class=\"token operator\">&#061;<\/span> <span class=\"token punctuation\">[<\/span><span class=\"token number\">32<\/span><span class=\"token punctuation\">]<\/span><\/p>\n<p>train_iter<span class=\"token punctuation\">,<\/span> src_vocab<span class=\"token punctuation\">,<\/span> tgt_vocab <span class=\"token operator\">&#061;<\/span> d2l<span class=\"token punctuation\">.<\/span>load_data_nmt<span class=\"token punctuation\">(<\/span>batch_size<span class=\"token punctuation\">,<\/span> num_steps<span class=\"token punctuation\">)<\/span><\/p>\n<p>encoder <span class=\"token operator\">&#061;<\/span> TransformerEncoder<span class=\"token punctuation\">(<\/span><br \/>\n    <span class=\"token builtin\">len<\/span><span class=\"token punctuation\">(<\/span>src_vocab<span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">,<\/span> key_size<span class=\"token punctuation\">,<\/span> query_size<span class=\"token punctuation\">,<\/span> value_size<span class=\"token punctuation\">,<\/span> num_hiddens<span class=\"token punctuation\">,<\/span><br \/>\n    norm_shape<span class=\"token punctuation\">,<\/span> ffn_num_input<span class=\"token punctuation\">,<\/span> ffn_num_hiddens<span class=\"token punctuation\">,<\/span> num_heads<span class=\"token punctuation\">,<\/span> num_layers<span class=\"token punctuation\">,<\/span> dropout<span class=\"token punctuation\">)<\/span><br \/>\ndecoder <span class=\"token operator\">&#061;<\/span> TransformerDecoder<span class=\"token punctuation\">(<\/span><br \/>\n    <span class=\"token builtin\">len<\/span><span class=\"token punctuation\">(<\/span>tgt_vocab<span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">,<\/span> key_size<span class=\"token punctuation\">,<\/span> query_size<span class=\"token punctuation\">,<\/span> value_size<span class=\"token punctuation\">,<\/span> num_hiddens<span class=\"token punctuation\">,<\/span><br \/>\n    norm_shape<span class=\"token punctuation\">,<\/span> ffn_num_input<span class=\"token punctuation\">,<\/span> ffn_num_hiddens<span class=\"token punctuation\">,<\/span> num_heads<span class=\"token punctuation\">,<\/span> num_layers<span class=\"token punctuation\">,<\/span> dropout<span class=\"token punctuation\">)<\/span><br \/>\nnet <span class=\"token operator\">&#061;<\/span> d2l<span class=\"token punctuation\">.<\/span>EncoderDecoder<span class=\"token punctuation\">(<\/span>encoder<span class=\"token punctuation\">,<\/span> decoder<span class=\"token punctuation\">)<\/span><br \/>\nd2l<span class=\"token punctuation\">.<\/span>train_seq2seq<span class=\"token punctuation\">(<\/span>net<span class=\"token punctuation\">,<\/span> train_iter<span class=\"token punctuation\">,<\/span> lr<span class=\"token punctuation\">,<\/span> num_epochs<span class=\"token punctuation\">,<\/span> tgt_vocab<span class=\"token punctuation\">,<\/span> device<span class=\"token punctuation\">)<\/span><\/p>\n<p><img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2025\/05\/20250516013340-682695f47d36a.png\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" width=\"380\" \/> <img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2025\/05\/20250516013340-682695f486b09.png\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" width=\"320\" \/><\/p>\n<p>\u2003\u2003\u8bad\u7ec3\u7ed3\u675f\u540e&#xff0c;\u4f7f\u7528Transformer\u6a21\u578b\u5c06\u4e00\u4e9b\u82f1\u8bed\u53e5\u5b50\u7ffb\u8bd1\u6210\u6cd5\u8bed&#xff0c;\u5e76\u4e14\u8ba1\u7b97\u5b83\u4eec\u7684BLEU\u5206\u6570\u3002<\/p>\n<p>engs <span class=\"token operator\">&#061;<\/span> <span class=\"token punctuation\">[<\/span><span class=\"token string\">&#039;go .&#039;<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token string\">&#034;i lost .&#034;<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token string\">&#039;he\\\\&#039;s calm .&#039;<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token string\">&#039;i\\\\&#039;m home .&#039;<\/span><span class=\"token punctuation\">]<\/span><br \/>\nfras <span class=\"token operator\">&#061;<\/span> <span class=\"token punctuation\">[<\/span><span class=\"token string\">&#039;va !&#039;<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token string\">&#039;j\\\\&#039;ai perdu .&#039;<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token string\">&#039;il est calme .&#039;<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token string\">&#039;je suis chez moi .&#039;<\/span><span class=\"token punctuation\">]<\/span><br \/>\n<span class=\"token keyword\">for<\/span> eng<span class=\"token punctuation\">,<\/span> fra <span class=\"token keyword\">in<\/span> <span class=\"token builtin\">zip<\/span><span class=\"token punctuation\">(<\/span>engs<span class=\"token punctuation\">,<\/span> fras<span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">:<\/span><br \/>\n    translation<span class=\"token punctuation\">,<\/span> dec_attention_weight_seq <span class=\"token operator\">&#061;<\/span> d2l<span class=\"token punctuation\">.<\/span>predict_seq2seq<span class=\"token punctuation\">(<\/span>net<span class=\"token punctuation\">,<\/span> eng<span class=\"token punctuation\">,<\/span> src_vocab<span class=\"token punctuation\">,<\/span> tgt_vocab<span class=\"token punctuation\">,<\/span> num_steps<span class=\"token punctuation\">,<\/span> device<span class=\"token punctuation\">,<\/span> <span class=\"token boolean\">True<\/span><span class=\"token punctuation\">)<\/span><br \/>\n    <span class=\"token keyword\">print<\/span><span class=\"token punctuation\">(<\/span><span class=\"token string-interpolation\"><span class=\"token string\">f&#039;<\/span><span class=\"token interpolation\"><span class=\"token punctuation\">{<\/span>eng<span class=\"token punctuation\">}<\/span><\/span><span class=\"token string\"> &#061;&gt; <\/span><span class=\"token interpolation\"><span class=\"token punctuation\">{<\/span>translation<span class=\"token punctuation\">}<\/span><\/span><span class=\"token string\">, &#039;<\/span><\/span><span class=\"token punctuation\">,<\/span> <span class=\"token string-interpolation\"><span class=\"token string\">f&#039;bleu <\/span><span class=\"token interpolation\"><span class=\"token punctuation\">{<\/span>d2l<span class=\"token punctuation\">.<\/span>bleu<span class=\"token punctuation\">(<\/span>translation<span class=\"token punctuation\">,<\/span> fra<span class=\"token punctuation\">,<\/span> k<span class=\"token operator\">&#061;<\/span><span class=\"token number\">2<\/span><span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">:<\/span><span class=\"token format-spec\">.3f<\/span><span class=\"token punctuation\">}<\/span><\/span><span class=\"token string\">&#039;<\/span><\/span><span class=\"token punctuation\">)<\/span><\/p>\n<p><img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2025\/05\/20250516013340-682695f497e53.png\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" width=\"400\" \/><\/p>\n<p>\u2003\u2003\u5f53\u8fdb\u884c\u6700\u540e\u4e00\u4e2a\u82f1\u8bed\u5230\u6cd5\u8bed\u7684\u53e5\u5b50\u7ffb\u8bd1\u5de5\u4f5c\u65f6&#xff0c;\u8ba9\u6211\u4eec\u53ef\u89c6\u5316Transformer\u7684\u6ce8\u610f\u529b\u6743\u91cd\u3002\u7f16\u7801\u5668\u81ea\u6ce8\u610f\u529b\u6743\u91cd\u7684\u5f62\u72b6\u4e3a&#xff08;\u7f16\u7801\u5668\u5c42\u6570&#xff0c;\u6ce8\u610f\u529b\u5934\u6570&#xff0c;num_steps\u6216\u67e5\u8be2\u7684\u6570\u76ee&#xff0c;num_steps\u6216\u201c\u952e-\u503c\u201d\u5bf9\u7684\u6570\u76ee&#xff09;\u3002<\/p>\n<p>enc_attention_weights <span class=\"token operator\">&#061;<\/span> torch<span class=\"token punctuation\">.<\/span>cat<span class=\"token punctuation\">(<\/span>net<span class=\"token punctuation\">.<\/span>encoder<span class=\"token punctuation\">.<\/span>attention_weights<span class=\"token punctuation\">,<\/span> <span class=\"token number\">0<\/span><span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">.<\/span>reshape<span class=\"token punctuation\">(<\/span><span class=\"token punctuation\">(<\/span>num_layers<span class=\"token punctuation\">,<\/span> num_heads<span class=\"token punctuation\">,<\/span> <span class=\"token operator\">&#8211;<\/span><span class=\"token number\">1<\/span><span class=\"token punctuation\">,<\/span> num_steps<span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">)<\/span><br \/>\nenc_attention_weights<span class=\"token punctuation\">.<\/span>shape<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2025\/05\/20250516013340-682695f4a39a9.png\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" width=\"260\" \/><\/p>\n<p>\u2003\u2003\u5728\u7f16\u7801\u5668\u7684\u81ea\u6ce8\u610f\u529b\u4e2d&#xff0c;\u67e5\u8be2\u548c\u952e\u90fd\u6765\u81ea\u76f8\u540c\u7684\u8f93\u5165\u5e8f\u5217\u3002\u56e0\u4e3a\u586b\u5145\u8bcd\u5143\u662f\u4e0d\u643a\u5e26\u4fe1\u606f\u7684&#xff0c;\u56e0\u6b64\u901a\u8fc7\u6307\u5b9a\u8f93\u5165\u5e8f\u5217\u7684\u6709\u6548\u957f\u5ea6\u53ef\u4ee5\u907f\u514d\u67e5\u8be2\u4e0e\u4f7f\u7528\u586b\u5145\u8bcd\u5143\u7684\u4f4d\u7f6e\u8ba1\u7b97\u6ce8\u610f\u529b\u3002\u63a5\u4e0b\u6765&#xff0c;\u5c06\u9010\u884c\u5448\u73b0\u4e24\u5c42\u591a\u5934\u6ce8\u610f\u529b\u7684\u6743\u91cd\u3002\u6bcf\u4e2a\u6ce8\u610f\u529b\u5934\u90fd\u6839\u636e\u67e5\u8be2\u3001\u952e\u548c\u503c\u7684\u4e0d\u540c\u7684\u8868\u793a\u5b50\u7a7a\u95f4\u6765\u8868\u793a\u4e0d\u540c\u7684\u6ce8\u610f\u529b\u3002<\/p>\n<p>d2l<span class=\"token punctuation\">.<\/span>show_heatmaps<span class=\"token punctuation\">(<\/span><br \/>\n    enc_attention_weights<span class=\"token punctuation\">.<\/span>cpu<span class=\"token punctuation\">(<\/span><span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">,<\/span> xlabel<span class=\"token operator\">&#061;<\/span><span class=\"token string\">&#039;Key positions&#039;<\/span><span class=\"token punctuation\">,<\/span><br \/>\n    ylabel<span class=\"token operator\">&#061;<\/span><span class=\"token string\">&#039;Query positions&#039;<\/span><span class=\"token punctuation\">,<\/span> titles<span class=\"token operator\">&#061;<\/span><span class=\"token punctuation\">[<\/span><span class=\"token string\">&#039;Head %d&#039;<\/span> <span class=\"token operator\">%<\/span> i <span class=\"token keyword\">for<\/span> i <span class=\"token keyword\">in<\/span> <span class=\"token builtin\">range<\/span><span class=\"token punctuation\">(<\/span><span class=\"token number\">1<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token number\">5<\/span><span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">]<\/span><span class=\"token punctuation\">,<\/span><br \/>\n    figsize<span class=\"token operator\">&#061;<\/span><span class=\"token punctuation\">(<\/span><span class=\"token number\">7<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token number\">3.5<\/span><span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">)<\/span><\/p>\n<p><img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2025\/05\/20250516013340-682695f4ac22a.png\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" width=\"500\" \/><\/p>\n<p>\u2003\u2003\u4e3a\u4e86\u53ef\u89c6\u5316\u89e3\u7801\u5668\u7684\u81ea\u6ce8\u610f\u529b\u6743\u91cd\u548c\u201c\u7f16\u7801\u5668-\u89e3\u7801\u5668\u201d\u7684\u6ce8\u610f\u529b\u6743\u91cd&#xff0c;\u6211\u4eec\u9700\u8981\u5b8c\u6210\u66f4\u591a\u7684\u6570\u636e\u64cd\u4f5c\u5de5\u4f5c\u3002\u4f8b\u5982\u7528\u96f6\u586b\u5145\u88ab\u63a9\u853d\u4f4f\u7684\u6ce8\u610f\u529b\u6743\u91cd\u3002\u503c\u5f97\u6ce8\u610f\u7684\u662f&#xff0c;\u89e3\u7801\u5668\u7684\u81ea\u6ce8\u610f\u529b\u6743\u91cd\u548c\u201c\u7f16\u7801\u5668-\u89e3\u7801\u5668\u201d\u7684\u6ce8\u610f\u529b\u6743\u91cd\u90fd\u6709\u76f8\u540c\u7684\u67e5\u8be2&#xff1a;\u5373\u4ee5\u5e8f\u5217\u5f00\u59cb\u8bcd\u5143&#xff08;beginning-of-sequence,BOS&#xff09;\u6253\u5934&#xff0c;\u518d\u4e0e\u540e\u7eed\u8f93\u51fa\u7684\u8bcd\u5143\u5171\u540c\u7ec4\u6210\u5e8f\u5217\u3002<\/p>\n<p>dec_attention_weights_2d <span class=\"token operator\">&#061;<\/span> <span class=\"token punctuation\">[<\/span>head<span class=\"token punctuation\">[<\/span><span class=\"token number\">0<\/span><span class=\"token punctuation\">]<\/span><span class=\"token punctuation\">.<\/span>tolist<span class=\"token punctuation\">(<\/span><span class=\"token punctuation\">)<\/span><br \/>\n                            <span class=\"token keyword\">for<\/span> step <span class=\"token keyword\">in<\/span> dec_attention_weight_seq<br \/>\n                            <span class=\"token keyword\">for<\/span> attn <span class=\"token keyword\">in<\/span> step <span class=\"token keyword\">for<\/span> blk <span class=\"token keyword\">in<\/span> attn <span class=\"token keyword\">for<\/span> head <span class=\"token keyword\">in<\/span> blk<span class=\"token punctuation\">]<\/span><br \/>\ndec_attention_weights_filled <span class=\"token operator\">&#061;<\/span> torch<span class=\"token punctuation\">.<\/span>tensor<span class=\"token punctuation\">(<\/span>pd<span class=\"token punctuation\">.<\/span>DataFrame<span class=\"token punctuation\">(<\/span>dec_attention_weights_2d<span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">.<\/span>fillna<span class=\"token punctuation\">(<\/span><span class=\"token number\">0.0<\/span><span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">.<\/span>values<span class=\"token punctuation\">)<\/span><br \/>\ndec_attention_weights <span class=\"token operator\">&#061;<\/span> dec_attention_weights_filled<span class=\"token punctuation\">.<\/span>reshape<span class=\"token punctuation\">(<\/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 number\">2<\/span><span class=\"token punctuation\">,<\/span> num_layers<span class=\"token punctuation\">,<\/span> num_heads<span class=\"token punctuation\">,<\/span> num_steps<span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">)<\/span><br \/>\ndec_self_attention_weights<span class=\"token punctuation\">,<\/span> dec_inter_attention_weights <span class=\"token operator\">&#061;<\/span> dec_attention_weights<span class=\"token punctuation\">.<\/span>permute<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 number\">3<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token number\">0<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token number\">4<\/span><span class=\"token punctuation\">)<\/span><br \/>\ndec_self_attention_weights<span class=\"token punctuation\">.<\/span>shape<span class=\"token punctuation\">,<\/span> dec_inter_attention_weights<span class=\"token punctuation\">.<\/span>shape<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2025\/05\/20250516013340-682695f4cf91e.png\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" width=\"450\" \/><\/p>\n<p>\u2003\u2003\u7531\u4e8e\u89e3\u7801\u5668\u81ea\u6ce8\u610f\u529b\u7684\u81ea\u56de\u5f52\u5c5e\u6027&#xff0c;\u67e5\u8be2\u4e0d\u4f1a\u5bf9\u5f53\u524d\u4f4d\u7f6e\u4e4b\u540e\u7684\u201c\u952e-\u503c\u201d\u5bf9\u8fdb\u884c\u6ce8\u610f\u529b\u8ba1\u7b97\u3002<\/p>\n<p><span class=\"token comment\"># Plusonetoincludethebeginning-of-sequencetoken<\/span><br \/>\nd2l<span class=\"token punctuation\">.<\/span>show_heatmaps<span class=\"token punctuation\">(<\/span><br \/>\n    dec_self_attention_weights<span class=\"token punctuation\">[<\/span><span class=\"token punctuation\">:<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token punctuation\">:<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token punctuation\">:<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token punctuation\">:<\/span><span class=\"token builtin\">len<\/span><span class=\"token punctuation\">(<\/span>translation<span class=\"token punctuation\">.<\/span>split<span class=\"token punctuation\">(<\/span><span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">)<\/span> <span class=\"token operator\">&#043;<\/span> <span class=\"token number\">1<\/span><span class=\"token punctuation\">]<\/span><span class=\"token punctuation\">,<\/span><br \/>\n    xlabel<span class=\"token operator\">&#061;<\/span><span class=\"token string\">&#039;Key positions&#039;<\/span><span class=\"token punctuation\">,<\/span> ylabel<span class=\"token operator\">&#061;<\/span><span class=\"token string\">&#039;Query positions&#039;<\/span><span class=\"token punctuation\">,<\/span><br \/>\n    titles<span class=\"token operator\">&#061;<\/span><span class=\"token punctuation\">[<\/span><span class=\"token string\">&#039;Head %d&#039;<\/span> <span class=\"token operator\">%<\/span> i <span class=\"token keyword\">for<\/span> i <span class=\"token keyword\">in<\/span> <span class=\"token builtin\">range<\/span><span class=\"token punctuation\">(<\/span><span class=\"token number\">1<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token number\">5<\/span><span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">]<\/span><span class=\"token punctuation\">,<\/span> figsize<span class=\"token operator\">&#061;<\/span><span class=\"token punctuation\">(<\/span><span class=\"token number\">7<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token number\">3.5<\/span><span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">)<\/span><\/p>\n<p><img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2025\/05\/20250516013340-682695f4dc3ad.png\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" width=\"500\" \/><\/p>\n<p>\u2003\u2003\u4e0e\u7f16\u7801\u5668\u7684\u81ea\u6ce8\u610f\u529b\u7684\u60c5\u51b5\u7c7b\u4f3c&#xff0c;\u901a\u8fc7\u6307\u5b9a\u8f93\u5165\u5e8f\u5217\u7684\u6709\u6548\u957f\u5ea6&#xff0c;\u8f93\u51fa\u5e8f\u5217\u7684\u67e5\u8be2\u4e0d\u4f1a\u4e0e\u8f93\u5165\u5e8f\u5217\u4e2d\u586b\u5145\u4f4d\u7f6e\u7684\u8bcd\u5143\u8fdb\u884c\u6ce8\u610f\u529b\u8ba1\u7b97\u3002<\/p>\n<p>d2l<span class=\"token punctuation\">.<\/span>show_heatmaps<span class=\"token punctuation\">(<\/span><br \/>\n    dec_inter_attention_weights<span class=\"token punctuation\">,<\/span> xlabel<span class=\"token operator\">&#061;<\/span><span class=\"token string\">&#039;Key positions&#039;<\/span><span class=\"token punctuation\">,<\/span><br \/>\n    ylabel<span class=\"token operator\">&#061;<\/span><span class=\"token string\">&#039;Query positions&#039;<\/span><span class=\"token punctuation\">,<\/span> titles<span class=\"token operator\">&#061;<\/span><span class=\"token punctuation\">[<\/span><span class=\"token string\">&#039;Head %d&#039;<\/span> <span class=\"token operator\">%<\/span> i <span class=\"token keyword\">for<\/span> i <span class=\"token keyword\">in<\/span> <span class=\"token builtin\">range<\/span><span class=\"token punctuation\">(<\/span><span class=\"token number\">1<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token number\">5<\/span><span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">]<\/span><span class=\"token punctuation\">,<\/span><br \/>\n    figsize<span class=\"token operator\">&#061;<\/span><span class=\"token punctuation\">(<\/span><span class=\"token number\">7<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token number\">3.5<\/span><span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">)<\/span><\/p>\n<p><img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2025\/05\/20250516013340-682695f4ef9c7.png\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" width=\"500\" \/><\/p>\n<p>\u2003\u2003\u5c3d\u7ba1Transformer\u67b6\u6784\u662f\u4e3a\u4e86\u5e8f\u5217\u5230\u5e8f\u5217\u7684\u5b66\u4e60\u800c\u63d0\u51fa\u7684&#xff0c;\u4f46\u6b63\u5982\u540e\u9762\u5c06\u63d0\u53ca\u7684\u90a3\u6837&#xff0c;Transformer\u7f16\u7801\u5668\u6216Transformer\u89e3\u7801\u5668\u901a\u5e38\u88ab\u5355\u72ec\u7528\u4e8e\u4e0d\u540c\u7684\u6df1\u5ea6\u5b66\u4e60\u4efb\u52a1\u4e2d\u3002<\/p>\n<h3>\u5c0f\u7ed3<\/h3>\n<ul>\n<li>Transformer\u662f\u7f16\u7801\u5668-\u89e3\u7801\u5668\u67b6\u6784\u7684\u4e00\u4e2a\u5b9e\u8df5&#xff0c;\u5c3d\u7ba1\u5728\u5b9e\u9645\u60c5\u51b5\u4e2d\u7f16\u7801\u5668\u6216\u89e3\u7801\u5668\u53ef\u4ee5\u5355\u72ec\u4f7f\u7528\u3002<\/li>\n<li>\u5728Transformer\u4e2d&#xff0c;\u591a\u5934\u81ea\u6ce8\u610f\u529b\u7528\u4e8e\u8868\u793a\u8f93\u5165\u5e8f\u5217\u548c\u8f93\u51fa\u5e8f\u5217&#xff0c;\u4e0d\u8fc7\u89e3\u7801\u5668\u5fc5\u987b\u901a\u8fc7\u63a9\u853d\u673a\u5236\u6765\u4fdd\u7559\u81ea\u56de\u5f52\u5c5e\u6027\u3002<\/li>\n<li>Transformer\u4e2d\u7684\u6b8b\u5dee\u8fde\u63a5\u548c\u5c42\u89c4\u8303\u5316\u662f\u8bad\u7ec3\u975e\u5e38\u6df1\u5ea6\u6a21\u578b\u7684\u91cd\u8981\u5de5\u5177\u3002<\/li>\n<li>Transformer\u6a21\u578b\u4e2d\u57fa\u4e8e\u4f4d\u7f6e\u7684\u524d\u9988\u7f51\u7edc\u4f7f\u7528\u540c\u4e00\u4e2a\u591a\u5c42\u611f\u77e5\u673a&#xff0c;\u4f5c\u7528\u662f\u5bf9\u6240\u6709\u5e8f\u5217\u4f4d\u7f6e\u7684\u8868\u793a\u8fdb\u884c\u8f6c\u6362\u3002<\/li>\n<\/ul>\n","protected":false},"excerpt":{"rendered":"<p>\u6587\u7ae0\u6d4f\u89c8\u9605\u8bfb884\u6b21\uff0c\u70b9\u8d5e74\u6b21\uff0c\u6536\u85cf63\u6b21\u3002Transformer\u57fa\u4e8e\u81ea\u6ce8\u610f\u529b\u548c\u4f4d\u7f6e\u7f16\u7801\uff0c\u91c7\u7528\u7f16\u7801\u5668-\u89e3\u7801\u5668\u67b6\u6784\u3002\u7f16\u7801\u5668\u548c\u89e3\u7801\u5668\u901a\u8fc7\u5806\u53e0\u591a\u5934\u6ce8\u610f\u529b\u5c42\u548c\u524d\u9988\u7f51\u7edc\u6784\u6210\uff0c\u5229\u7528\u6b8b\u5dee\u8fde\u63a5\u548c\u5c42\u89c4\u8303\u5316\u63d0\u5347\u8bad\u7ec3\u6548\u679c\uff0c\u5728\u5e76\u884c\u8ba1\u7b97\u548c\u77ed\u4f9d\u8d56\u8def\u5f84\u4f18\u52bf\u4e0b\uff0c\u5e7f\u6cdb\u5e94\u7528\u4e8e\u5e8f\u5217\u4efb\u52a1\u5982\u673a\u5668\u7ffb\u8bd1\u3002<\/p>\n","protected":false},"author":2,"featured_media":37535,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[587,152,3259,841,50,86],"topic":[],"class_list":{"0":"post-37551","1":"post","2":"type-post","3":"status-publish","4":"format-standard","5":"has-post-thumbnail","6":"hentry","7":"category-server","8":"tag-587","9":"tag-pytorch","11":"tag-transformer","12":"tag-50","13":"tag-86"},"yoast_head":"<!-- This site is optimized with the Yoast SEO plugin v20.3 - https:\/\/yoast.com\/wordpress\/plugins\/seo\/ -->\n<title>\u3010\u73b0\u4ee3\u6df1\u5ea6\u5b66\u4e60\u6280\u672f\u3011\u6ce8\u610f\u529b\u673a\u523607\uff1aTransformer - \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\/37551.html\" \/>\n<meta property=\"og:locale\" content=\"zh_CN\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"\u3010\u73b0\u4ee3\u6df1\u5ea6\u5b66\u4e60\u6280\u672f\u3011\u6ce8\u610f\u529b\u673a\u523607\uff1aTransformer - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"og:description\" content=\"\u6587\u7ae0\u6d4f\u89c8\u9605\u8bfb884\u6b21\uff0c\u70b9\u8d5e74\u6b21\uff0c\u6536\u85cf63\u6b21\u3002Transformer\u57fa\u4e8e\u81ea\u6ce8\u610f\u529b\u548c\u4f4d\u7f6e\u7f16\u7801\uff0c\u91c7\u7528\u7f16\u7801\u5668-\u89e3\u7801\u5668\u67b6\u6784\u3002\u7f16\u7801\u5668\u548c\u89e3\u7801\u5668\u901a\u8fc7\u5806\u53e0\u591a\u5934\u6ce8\u610f\u529b\u5c42\u548c\u524d\u9988\u7f51\u7edc\u6784\u6210\uff0c\u5229\u7528\u6b8b\u5dee\u8fde\u63a5\u548c\u5c42\u89c4\u8303\u5316\u63d0\u5347\u8bad\u7ec3\u6548\u679c\uff0c\u5728\u5e76\u884c\u8ba1\u7b97\u548c\u77ed\u4f9d\u8d56\u8def\u5f84\u4f18\u52bf\u4e0b\uff0c\u5e7f\u6cdb\u5e94\u7528\u4e8e\u5e8f\u5217\u4efb\u52a1\u5982\u673a\u5668\u7ffb\u8bd1\u3002\" \/>\n<meta property=\"og:url\" content=\"https:\/\/www.wsisp.com\/helps\/37551.html\" \/>\n<meta property=\"og:site_name\" content=\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"article:published_time\" content=\"2025-05-16T01:33:42+00:00\" \/>\n<meta property=\"og:image\" content=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2025\/05\/20250516013340-682695f417dab.gif\" \/>\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=\"9 \u5206\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"WebPage\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/37551.html\",\"url\":\"https:\/\/www.wsisp.com\/helps\/37551.html\",\"name\":\"\u3010\u73b0\u4ee3\u6df1\u5ea6\u5b66\u4e60\u6280\u672f\u3011\u6ce8\u610f\u529b\u673a\u523607\uff1aTransformer - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"isPartOf\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\"},\"datePublished\":\"2025-05-16T01:33:42+00:00\",\"dateModified\":\"2025-05-16T01:33:42+00:00\",\"author\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\"},\"breadcrumb\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/37551.html#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/www.wsisp.com\/helps\/37551.html\"]}]},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/37551.html#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\/\/www.wsisp.com\/helps\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"\u3010\u73b0\u4ee3\u6df1\u5ea6\u5b66\u4e60\u6280\u672f\u3011\u6ce8\u610f\u529b\u673a\u523607\uff1aTransformer\"}]},{\"@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":"\u3010\u73b0\u4ee3\u6df1\u5ea6\u5b66\u4e60\u6280\u672f\u3011\u6ce8\u610f\u529b\u673a\u523607\uff1aTransformer - \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\/37551.html","og_locale":"zh_CN","og_type":"article","og_title":"\u3010\u73b0\u4ee3\u6df1\u5ea6\u5b66\u4e60\u6280\u672f\u3011\u6ce8\u610f\u529b\u673a\u523607\uff1aTransformer - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","og_description":"\u6587\u7ae0\u6d4f\u89c8\u9605\u8bfb884\u6b21\uff0c\u70b9\u8d5e74\u6b21\uff0c\u6536\u85cf63\u6b21\u3002Transformer\u57fa\u4e8e\u81ea\u6ce8\u610f\u529b\u548c\u4f4d\u7f6e\u7f16\u7801\uff0c\u91c7\u7528\u7f16\u7801\u5668-\u89e3\u7801\u5668\u67b6\u6784\u3002\u7f16\u7801\u5668\u548c\u89e3\u7801\u5668\u901a\u8fc7\u5806\u53e0\u591a\u5934\u6ce8\u610f\u529b\u5c42\u548c\u524d\u9988\u7f51\u7edc\u6784\u6210\uff0c\u5229\u7528\u6b8b\u5dee\u8fde\u63a5\u548c\u5c42\u89c4\u8303\u5316\u63d0\u5347\u8bad\u7ec3\u6548\u679c\uff0c\u5728\u5e76\u884c\u8ba1\u7b97\u548c\u77ed\u4f9d\u8d56\u8def\u5f84\u4f18\u52bf\u4e0b\uff0c\u5e7f\u6cdb\u5e94\u7528\u4e8e\u5e8f\u5217\u4efb\u52a1\u5982\u673a\u5668\u7ffb\u8bd1\u3002","og_url":"https:\/\/www.wsisp.com\/helps\/37551.html","og_site_name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","article_published_time":"2025-05-16T01:33:42+00:00","og_image":[{"url":"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2025\/05\/20250516013340-682695f417dab.gif"}],"author":"admin","twitter_card":"summary_large_image","twitter_misc":{"\u4f5c\u8005":"admin","\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4":"9 \u5206"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/www.wsisp.com\/helps\/37551.html","url":"https:\/\/www.wsisp.com\/helps\/37551.html","name":"\u3010\u73b0\u4ee3\u6df1\u5ea6\u5b66\u4e60\u6280\u672f\u3011\u6ce8\u610f\u529b\u673a\u523607\uff1aTransformer - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","isPartOf":{"@id":"https:\/\/www.wsisp.com\/helps\/#website"},"datePublished":"2025-05-16T01:33:42+00:00","dateModified":"2025-05-16T01:33:42+00:00","author":{"@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41"},"breadcrumb":{"@id":"https:\/\/www.wsisp.com\/helps\/37551.html#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/www.wsisp.com\/helps\/37551.html"]}]},{"@type":"BreadcrumbList","@id":"https:\/\/www.wsisp.com\/helps\/37551.html#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/www.wsisp.com\/helps"},{"@type":"ListItem","position":2,"name":"\u3010\u73b0\u4ee3\u6df1\u5ea6\u5b66\u4e60\u6280\u672f\u3011\u6ce8\u610f\u529b\u673a\u523607\uff1aTransformer"}]},{"@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\/37551","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=37551"}],"version-history":[{"count":0,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/37551\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/media\/37535"}],"wp:attachment":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/media?parent=37551"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/categories?post=37551"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/tags?post=37551"},{"taxonomy":"topic","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/topic?post=37551"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}