\u5728\u4eba\u5de5\u667a\u80fd\u7684\u53d1\u5c55\u5386\u7a0b\u4e2d，Transformer \u67b6\u6784\u65e0\u7591\u662f\u4e00\u5ea7\u5177\u6709\u91cc\u7a0b\u7891\u610f\u4e49\u7684\u4e30\u7891\u3002\u81ea\u4ece 2017 \u5e74 Google \u56e2\u961f\u5728\u8bba\u6587\u300aAttention Is All You Need\u300b\u4e2d\u9996\u6b21\u63d0\u51fa Transformer \u67b6\u6784\u4ee5\u6765，\u5b83\u4fbf\u8fc5\u901f\u5728\u81ea\u7136\u8bed\u8a00\u5904\u7406（NLP）\u9886\u57df\u5f15\u53d1\u4e86\u9769\u547d\u6027\u7684\u53d8\u9769，\u5e76\u9010\u6e10\u62d3\u5c55\u5230\u8ba1\u7b97\u673a\u89c6\u89c9\u3001\u8bed\u97f3\u8bc6\u522b\u7b49\u591a\u4e2a\u9886\u57df\u3002Transformer \u67b6\u6784\u4ee5\u5176\u5353\u8d8a\u7684\u5e76\u884c\u8ba1\u7b97\u80fd\u529b\u3001\u5f3a\u5927\u7684\u957f\u5e8f\u5217\u5904\u7406\u80fd\u529b\u4ee5\u53ca\u51fa\u8272\u7684\u5efa\u6a21\u6548\u679c，\u6210\u4e3a\u4e86\u4f17\u591a\u5148\u8fdb\u5927\u6a21\u578b\u7684\u6838\u5fc3\u57fa\u7840，\u5982 GPT \u7cfb\u5217\u3001BERT \u7b49\u3002<\/p>\n
\u672c\u6587\u5c06\u6df1\u5165\u5256\u6790 Transformer \u67b6\u6784，\u4ece\u5176\u6838\u5fc3\u539f\u7406\u3001\u7ec4\u4ef6\u6784\u6210\u5230\u6e90\u7801\u5b9e\u73b0，\u8fdb\u884c\u5168\u65b9\u4f4d\u3001\u7ec6\u81f4\u5165\u5fae\u7684\u5206\u6790\u3002\u901a\u8fc7\u5bf9\u6e90\u7801\u7684\u9010\u884c\u89e3\u8bfb，\u5e2e\u52a9\u8bfb\u8005\u6df1\u5165\u7406\u89e3 Transformer \u67b6\u6784\u7684\u5de5\u4f5c\u673a\u5236，\u4e3a\u8fdb\u4e00\u6b65\u7814\u7a76\u548c\u5e94\u7528\u57fa\u4e8e Transformer \u7684\u5927\u6a21\u578b\u5960\u5b9a\u575a\u5b9e\u7684\u57fa\u7840\u3002<\/p>\n
\u5728 Transformer \u67b6\u6784\u51fa\u73b0\u4e4b\u524d，\u5faa\u73af\u795e\u7ecf\u7f51\u7edc（RNN）\u53ca\u5176\u53d8\u4f53，\u5982\u957f\u77ed\u671f\u8bb0\u5fc6\u7f51\u7edc（LSTM）\u548c\u95e8\u63a7\u5faa\u73af\u5355\u5143（GRU），\u662f\u5904\u7406\u5e8f\u5217\u6570\u636e\u7684\u4e3b\u6d41\u6a21\u578b\u3002\u7136\u800c，\u8fd9\u4e9b\u6a21\u578b\u5b58\u5728\u4e00\u4e9b\u56fa\u6709\u7684\u5c40\u9650\u6027：<\/p>\n
\u4e3a\u4e86\u89e3\u51b3\u4f20\u7edf\u5e8f\u5217\u5904\u7406\u6a21\u578b\u7684\u5c40\u9650\u6027，Google \u56e2\u961f\u63d0\u51fa\u4e86 Transformer \u67b6\u6784\u3002Transformer \u67b6\u6784\u6452\u5f03\u4e86\u4f20\u7edf\u7684\u5faa\u73af\u7ed3\u6784，\u5b8c\u5168\u57fa\u4e8e\u6ce8\u610f\u529b\u673a\u5236（Attention Mechanism）\u6784\u5efa，\u4ece\u800c\u5b9e\u73b0\u4e86\u5e76\u884c\u8ba1\u7b97，\u5927\u5927\u63d0\u9ad8\u4e86\u6a21\u578b\u7684\u8bad\u7ec3\u6548\u7387\u548c\u5904\u7406\u957f\u5e8f\u5217\u7684\u80fd\u529b\u3002<\/p>\n
\u6ce8\u610f\u529b\u673a\u5236\u662f\u4e00\u79cd\u6a21\u62df\u4eba\u7c7b\u6ce8\u610f\u529b\u7684\u673a\u5236，\u5b83\u53ef\u4ee5\u8ba9\u6a21\u578b\u5728\u5904\u7406\u5e8f\u5217\u6570\u636e\u65f6，\u81ea\u52a8\u5730\u5173\u6ce8\u5e8f\u5217\u4e2d\u7684\u91cd\u8981\u90e8\u5206\u3002\u5728\u81ea\u7136\u8bed\u8a00\u5904\u7406\u4e2d，\u6ce8\u610f\u529b\u673a\u5236\u53ef\u4ee5\u5e2e\u52a9\u6a21\u578b\u66f4\u597d\u5730\u6355\u6349\u4e0a\u4e0b\u6587\u4fe1\u606f，\u4ece\u800c\u63d0\u9ad8\u6a21\u578b\u7684\u6027\u80fd\u3002<\/p>\n
\u7f29\u653e\u70b9\u79ef\u6ce8\u610f\u529b\u662f Transformer \u67b6\u6784\u4e2d\u4f7f\u7528\u7684\u4e00\u79cd\u6ce8\u610f\u529b\u673a\u5236，\u5176\u8ba1\u7b97\u516c\u5f0f\u5982\u4e0b： (\\\\text{Attention}(Q, K, V) = \\\\text{softmax}(\\\\frac{QK^T}{\\\\sqrt{d_k}})V) \u5176\u4e2d，Q \u662f\u67e5\u8be2\u77e9\u9635（Query Matrix），K \u662f\u952e\u77e9\u9635（Key Matrix），V \u662f\u503c\u77e9\u9635（Value Matrix），(d_k) \u662f\u952e\u5411\u91cf\u7684\u7ef4\u5ea6\u3002<\/p>\n
\u4ee5\u4e0b\u662f\u7f29\u653e\u70b9\u79ef\u6ce8\u610f\u529b\u7684 Python \u4ee3\u7801\u5b9e\u73b0：<\/p>\n
python<\/p>\n
import<\/span> torch # \u5b9a\u4e49\u7f29\u653e\u70b9\u79ef\u6ce8\u610f\u529b\u7c7b<\/span> def<\/span> forward<\/span>(<\/span>self,<\/span> q,<\/span> k,<\/span> v,<\/span> mask=<\/span>None<\/span>)<\/span>:<\/span> if<\/span> mask is<\/span> not<\/span> None<\/span>:<\/span> # \u5bf9\u6ce8\u610f\u529b\u5206\u6570\u8fdb\u884c softmax \u64cd\u4f5c，\u5f97\u5230\u6ce8\u610f\u529b\u6743\u91cd<\/span> \u591a\u5934\u6ce8\u610f\u529b\u662f\u7f29\u653e\u70b9\u79ef\u6ce8\u610f\u529b\u7684\u6269\u5c55，\u5b83\u901a\u8fc7\u5c06\u67e5\u8be2\u3001\u952e\u548c\u503c\u5206\u522b\u6295\u5f71\u5230\u591a\u4e2a\u4f4e\u7ef4\u5b50\u7a7a\u95f4\u4e2d，\u7136\u540e\u5728\u6bcf\u4e2a\u5b50\u7a7a\u95f4\u4e2d\u72ec\u7acb\u5730\u8ba1\u7b97\u6ce8\u610f\u529b，\u6700\u540e\u5c06\u6240\u6709\u5b50\u7a7a\u95f4\u7684\u6ce8\u610f\u529b\u8f93\u51fa\u62fc\u63a5\u8d77\u6765\u5e76\u8fdb\u884c\u7ebf\u6027\u53d8\u6362，\u5f97\u5230\u6700\u7ec8\u7684\u6ce8\u610f\u529b\u8f93\u51fa\u3002<\/p>\n \u591a\u5934\u6ce8\u610f\u529b\u7684\u8ba1\u7b97\u516c\u5f0f\u5982\u4e0b： (\\\\text{MultiHead}(Q, K, V) = \\\\text{Concat}(\\\\text{head}_1, \\\\ldots, \\\\text{head}_h)W^O) \u5176\u4e2d，(\\\\text{head}_i = \\\\text{Attention}(QW_i^Q, KW_i^K, VW_iV))，(W_i<\/sup>Q)\u3001(W_iK)\u3001(W_i<\/sup>V) \u662f\u6295\u5f71\u77e9\u9635，(W^O) \u662f\u8f93\u51fa\u77e9\u9635\u3002<\/p>\n \u4ee5\u4e0b\u662f\u591a\u5934\u6ce8\u610f\u529b\u7684 Python \u4ee3\u7801\u5b9e\u73b0：<\/p>\n python<\/p>\n # \u5b9a\u4e49\u591a\u5934\u6ce8\u610f\u529b\u7c7b<\/span> # \u5b9a\u4e49\u67e5\u8be2\u3001\u952e\u548c\u503c\u7684\u7ebf\u6027\u53d8\u6362\u5c42<\/span> # \u5b9a\u4e49\u7f29\u653e\u70b9\u79ef\u6ce8\u610f\u529b\u5c42<\/span> def<\/span> forward<\/span>(<\/span>self,<\/span> q,<\/span> k,<\/span> v,<\/span> mask=<\/span>None<\/span>)<\/span>:<\/span> # \u5bf9\u67e5\u8be2\u3001\u952e\u548c\u503c\u8fdb\u884c\u7ebf\u6027\u53d8\u6362<\/span> # \u5c06\u67e5\u8be2\u3001\u952e\u548c\u503c\u5206\u5272\u6210\u591a\u4e2a\u5934<\/span>
\nimport<\/span> torch.<\/span>nn as<\/span> nn
\nimport<\/span> torch.<\/span>nn.<\/span>functional as<\/span> F<\/p>\n
\nclass<\/span> ScaledDotProductAttention<\/span>(<\/span>nn.<\/span>Module)<\/span>:<\/span>
\n def<\/span> __init__<\/span>(<\/span>self,<\/span> d_k)<\/span>:<\/span>
\n super<\/span>(<\/span>ScaledDotProductAttention,<\/span> self)<\/span>.<\/span>__init__(<\/span>)<\/span>
\n # \u952e\u5411\u91cf\u7684\u7ef4\u5ea6<\/span>
\n self.<\/span>d_k =<\/span> d_k <\/p>\n
\n # \u8ba1\u7b97 Q \u548c K \u7684\u8f6c\u7f6e\u7684\u70b9\u79ef<\/span>
\n attn_scores =<\/span> torch.<\/span>matmul(<\/span>q,<\/span> k.<\/span>transpose(<\/span>–<\/span>2<\/span>,<\/span> –<\/span>1<\/span>)<\/span>)<\/span>
\n # \u7f29\u653e\u70b9\u79ef<\/span>
\n attn_scores =<\/span> attn_scores \/<\/span> torch.<\/span>sqrt(<\/span>torch.<\/span>tensor(<\/span>self.<\/span>d_k,<\/span> dtype=<\/span>torch.<\/span>float32)<\/span>)<\/span> <\/p>\n
\n # \u5982\u679c\u6709\u63a9\u7801，\u5c06\u63a9\u7801\u4f4d\u7f6e\u7684\u6ce8\u610f\u529b\u5206\u6570\u8bbe\u4e3a\u8d1f\u65e0\u7a77<\/span>
\n attn_scores =<\/span> attn_scores.<\/span>masked_fill(<\/span>mask ==<\/span> 0<\/span>,<\/span> –<\/span>1e9<\/span>)<\/span> <\/p>\n
\n attn_weights =<\/span> F.<\/span>softmax(<\/span>attn_scores,<\/span> dim=<\/span>–<\/span>1<\/span>)<\/span>
\n # \u8ba1\u7b97\u6ce8\u610f\u529b\u8f93\u51fa<\/span>
\n output =<\/span> torch.<\/span>matmul(<\/span>attn_weights,<\/span> v)<\/span>
\n return<\/span> output,<\/span> attn_weights<\/p>\n3.1.3 \u591a\u5934\u6ce8\u610f\u529b（Multi-Head Attention）<\/h5>\n
\nclass<\/span> MultiHeadAttention<\/span>(<\/span>nn.<\/span>Module)<\/span>:<\/span>
\n def<\/span> __init__<\/span>(<\/span>self,<\/span> d_model,<\/span> num_heads)<\/span>:<\/span>
\n super<\/span>(<\/span>MultiHeadAttention,<\/span> self)<\/span>.<\/span>__init__(<\/span>)<\/span>
\n # \u6a21\u578b\u7684\u7ef4\u5ea6<\/span>
\n self.<\/span>d_model =<\/span> d_model
\n # \u6ce8\u610f\u529b\u5934\u7684\u6570\u91cf<\/span>
\n self.<\/span>num_heads =<\/span> num_heads
\n # \u6bcf\u4e2a\u5934\u7684\u7ef4\u5ea6<\/span>
\n self.<\/span>d_k =<\/span> d_model \/\/<\/span> num_heads <\/p>\n
\n self.<\/span>W_q =<\/span> nn.<\/span>Linear(<\/span>d_model,<\/span> d_model)<\/span>
\n self.<\/span>W_k =<\/span> nn.<\/span>Linear(<\/span>d_model,<\/span> d_model)<\/span>
\n self.<\/span>W_v =<\/span> nn.<\/span>Linear(<\/span>d_model,<\/span> d_model)<\/span>
\n # \u5b9a\u4e49\u8f93\u51fa\u7684\u7ebf\u6027\u53d8\u6362\u5c42<\/span>
\n self.<\/span>W_o =<\/span> nn.<\/span>Linear(<\/span>d_model,<\/span> d_model)<\/span> <\/p>\n
\n self.<\/span>attention =<\/span> ScaledDotProductAttention(<\/span>self.<\/span>d_k)<\/span><\/p>\n
\n batch_size =<\/span> q.<\/span>size(<\/span>0<\/span>)<\/span><\/p>\n
\n Q =<\/span> self.<\/span>W_q(<\/span>q)<\/span>
\n K =<\/span> self.<\/span>W_k(<\/span>k)<\/span>
\n V =<\/span> self.<\/span>W_v(<\/span>v)<\/span><\/p>\n
\n Q =<\/span> Q.<\/span>view(<\/span>batch_size,<\/span> –<\/span>1<\/span>,<\/span> self.<\/span>num_heads,<\/span> self.<\/span>d_k)<\/span>.<\/span>transpose(<\/span>1<\/span>,<\/span> 2<\/span>)<\/span>
\n K =<\/span> K.<\/span>view(<\/span>batch_size,<\/span> –<\/span>1<\/span>,<\/span> self.<\/span>num_heads,<\/span> self.<\/span>d_k)<\/span>.<\/span>transpose(<\/span>1<\/span>,<\/span> 2<\/span>)<\/span>
\n V =<\/span> V.<\/span>view(<\/span>batch_size,<\/span> –<\/span>1<\/span>,<\/span> self