Transformer\u4f5c\u4e3a\u4e00\u79cd\u5b8c\u5168\u57fa\u4e8e\u6ce8\u610f\u529b\u673a\u5236\u7684\u6df1\u5ea6\u5b66\u4e60\u67b6\u6784\uff0c\u5728\u81ea\u7136\u8bed\u8a00\u5904\u7406\u548c\u5e8f\u5217\u5efa\u6a21\u9886\u57df\u5f97\u5230\u5e7f\u6cdb\u5e94\u7528\u3002\u672c\u6587\u7cfb\u7edf\u9610\u8ff0\u4e86Transformer\u7684\u57fa\u672c\u539f\u7406\u3001\u81ea\u6ce8\u610f\u529b\u673a\u5236\u548c\u4f4d\u7f6e\u7f16\u7801\uff0c\u91cd\u70b9\u5206\u6790\u4e86\u7f16\u7801\u5668\u3001\u89e3\u7801\u5668\u3001\u81ea\u6ce8\u610f\u529b\u548c\u524d\u9988\u7f51\u7edc\u7b49\u6838\u5fc3\u5185\u5bb9\u3002\u6df1\u5165\u63a2\u8ba8\u4e86\u591a\u5934\u6ce8\u610f\u529b\u3001\u4f4d\u7f6e\u7f16\u7801\u3001\u6b8b\u5dee\u8fde\u63a5\u7b49\u5173\u952e\u6280\u672f\uff0c\u5e76\u4ece\u7406\u8bba\u89d2\u5ea6\u5206\u6790\u4e86Transformer\u7684\u8868\u8fbe\u80fd\u529b\u548c\u5e76\u884c\u8ba1\u7b97\u80fd\u529b\u3002\u901a\u8fc7\u5bf9\u5b9e\u9645\u6570\u636e\u96c6\u548c\u5e94\u7528\u6848\u4f8b\u7684\u7814\u7a76\uff0c\u9a8c\u8bc1\u4e86Transformer\u5728\u81ea\u7136\u8bed\u8a00\u5904\u7406\u4efb\u52a1\u4e2d\u7684\u6709\u6548\u6027\uff0c\u4e3a\u5e8f\u5217\u5efa\u6a21\u63d0\u4f9b\u4e86\u7406\u8bba\u4f9d\u636e\u548c\u5b9e\u8df5\u6307\u5bfc\u3002<\/p>\n
\u5173\u952e\u8bcd<\/strong>\uff1aTransformer\uff1b\u81ea\u6ce8\u610f\u529b\u673a\u5236\uff1b\u4f4d\u7f6e\u7f16\u7801\uff1b\u7f16\u7801\u5668\uff1b\u89e3\u7801\u5668<\/p>\n Transformer\u7531Vaswani\u7b49\u4eba\u4e8e2017\u5e74\u63d0\u51fa\uff0c\u662f\u4e00\u79cd\u5b8c\u5168\u57fa\u4e8e\u6ce8\u610f\u529b\u673a\u5236\u7684\u6df1\u5ea6\u5b66\u4e60\u67b6\u6784\u3002\u67b6\u6784\u7684\u6838\u5fc3\u601d\u60f3\u662f\uff1a\u901a\u8fc7\u81ea\u6ce8\u610f\u529b\u673a\u5236\u6355\u83b7\u5e8f\u5217\u4e2d\u7684\u4f9d\u8d56\u5173\u7cfb\uff0c\u901a\u8fc7\u4f4d\u7f6e\u7f16\u7801\u5f15\u5165\u4f4d\u7f6e\u4fe1\u606f\u3002Transformer\u7684\u4f18\u52bf\u5728\u4e8e\uff1a\u5e76\u884c\u8ba1\u7b97\u80fd\u529b\u5f3a\u3001\u80fd\u591f\u6355\u83b7\u957f\u7a0b\u4f9d\u8d56\u3001\u8bad\u7ec3\u6548\u7387\u9ad8\u3001\u6027\u80fd\u4f18\u5f02\u3002<\/p>\n Transformer\u7684\u5e94\u7528\u9886\u57df\u5305\u62ec\uff1a\u81ea\u7136\u8bed\u8a00\u5904\u7406\u3001\u673a\u5668\u7ffb\u8bd1\u3001\u6587\u672c\u751f\u6210\u3001\u8bed\u97f3\u8bc6\u522b\u7b49\u3002\u968f\u7740\u6df1\u5ea6\u5b66\u4e60\u7684\u53d1\u5c55\uff0cTransformer\u5728\u5e8f\u5217\u5efa\u6a21\u9886\u57df\u5c55\u73b0\u51fa\u5f3a\u5927\u7684\u80fd\u529b\u3002\u672c\u6587\u5c06\u7cfb\u7edf\u7814\u7a76Transformer\u7684\u81ea\u6ce8\u610f\u529b\u673a\u5236\u4e0e\u4f4d\u7f6e\u7f16\u7801\uff0c\u4e3a\u5e8f\u5217\u5efa\u6a21\u63d0\u4f9b\u7406\u8bba\u4f9d\u636e\u548c\u5b9e\u8df5\u6307\u5bfc\u3002<\/p>\n \u5b9a\u4e49<\/strong>\uff1a \u5176\u4e2d\uff1a<\/p>\n \u7b2c$h$\u4e2a\u5934<\/strong>\uff1a \u591a\u5934\u8f93\u51fa<\/strong>\uff1a \u5b9a\u4e49<\/strong>\uff1a \u5176\u4e2d\uff1a<\/p>\n \u516c\u5f0f<\/strong>\uff1a \u76ee\u7684<\/strong>\uff1a\u9632\u6b62\u770b\u5230\u672a\u6765\u4fe1\u606f\u3002<\/p>\n \u63a9\u7801<\/strong>\uff1a \u63a9\u7801\u6ce8\u610f\u529b<\/strong>\uff1a \u5b9a\u4e49<\/strong>\uff1a \u4e0e\u7f16\u7801\u5668\u76f8\u540c<\/strong>\uff1a \u516c\u5f0f<\/strong>\uff1a \u5176\u4e2d\uff1a<\/p>\n \u65b9\u6cd5<\/strong>\uff1a\u4f7f\u7528\u53ef\u5b66\u4e60\u7684\u4f4d\u7f6e\u5d4c\u5165\u3002<\/p>\n $$PE = \/text{Embedding}(pos)$$<\/p>\n \u65b9\u6cd5<\/strong>\uff1a\u4f7f\u7528\u76f8\u5bf9\u4f4d\u7f6e\u4fe1\u606f\u3002<\/p>\n $$\/text{Attention}(Q, K, V) = \/text{softmax}\/left(\/frac{QK^T + R}{\/sqrt{d_k}}\/right)V$$<\/p>\n \u5176\u4e2d\uff0c$R$\u4e3a\u76f8\u5bf9\u4f4d\u7f6e\u504f\u7f6e\u3002<\/p>\n \u5b9a\u4e49<\/strong>\uff1a \u5176\u4e2d\uff1a<\/p>\n \u516c\u5f0f<\/strong>\uff1a \u5176\u4e2d\uff1a<\/p>\n \u4f4d\u7f6e<\/strong>\uff1a<\/p>\n \u5e94\u7528<\/strong>\uff1a\u5c06\u4e00\u79cd\u8bed\u8a00\u7ffb\u8bd1\u4e3a\u53e6\u4e00\u79cd\u8bed\u8a00<\/p>\n \u7f16\u7801\u5668-\u89e3\u7801\u5668<\/strong>\uff1a<\/p>\n \u5e94\u7528<\/strong>\uff1a\u751f\u6210\u65b0\u6587\u672c<\/p>\n \u65b9\u6cd5<\/strong>\uff1a\u81ea\u56de\u5f52\u751f\u6210<\/p>\n \u5e94\u7528<\/strong>\uff1a\u6587\u672c\u5206\u7c7b\u4efb\u52a1<\/p>\n \u65b9\u6cd5<\/strong>\uff1a\u7f16\u7801\u5668 + \u5206\u7c7b\u5934<\/p>\n \u6807\u51c6\u6570\u636e\u96c6<\/strong>\uff1a<\/p>\n \u672c\u6587\u7cfb\u7edf\u9610\u8ff0\u4e86Transformer\u7684\u81ea\u6ce8\u610f\u529b\u673a\u5236\u4e0e\u4f4d\u7f6e\u7f16\u7801\u3002\u901a\u8fc7\u5bf9\u57fa\u672c\u539f\u7406\u3001\u7f16\u7801\u5668\u3001\u89e3\u7801\u5668\u548c\u5e94\u7528\u5b9e\u4f8b\u7684\u6df1\u5165\u7814\u7a76\uff0c\u9a8c\u8bc1\u4e86Transformer\u5728\u81ea\u7136\u8bed\u8a00\u5904\u7406\u4efb\u52a1\u4e2d\u7684\u6709\u6548\u6027\u3002<\/p>\n \u4e3b\u8981\u7ed3\u8bba\u5982\u4e0b\uff1a<\/p>\n \u7b97\u6cd5\u4f18\u52bf<\/strong>\uff1a<\/p>\n \u5173\u952e\u56e0\u7d20<\/strong>\uff1a<\/p>\n \u5e94\u7528\u4ef7\u503c<\/strong>\uff1a<\/p>\n \u672a\u6765\u7814\u7a76\u65b9\u5411\u5305\u62ec\uff1a<\/p>\n1. \u5f15\u8a00<\/h3>\n
2. \u7f16\u7801\u5668<\/h3>\n
2.1 \u81ea\u6ce8\u610f\u529b\u5c42<\/h4>\n
$$\/text{Attention}(Q, K, V) = \/text{softmax}\/left(\/frac{QK^T}{\/sqrt{d_k}}\/right)V$$<\/p>\n\n
2.2 \u591a\u5934\u6ce8\u610f\u529b<\/h4>\n
$$\/text{head}_h = \/text{Attention}(QW_h^Q, KW_h^K, VW_h^V)$$<\/p>\n
$$\/text{MultiHead}(Q, K, V) = \/text{Concat}(\/text{head}_1, \/ldots, \/text{head}_h)W^O$$<\/p>\n2.3 \u524d\u9988\u7f51\u7edc<\/h4>\n
$$\/text{FFN}(x) = \/max(0, xW_1 + b_1)W_2 + b_2$$<\/p>\n\n
2.4 \u6b8b\u5dee\u8fde\u63a5\u548c\u5c42\u5f52\u4e00\u5316<\/h4>\n
$$\/text{Output} = \/text{LayerNorm}(x + \/text{Sublayer}(x))$$<\/p>\n3. \u89e3\u7801\u5668<\/h3>\n
3.1 \u63a9\u7801\u81ea\u6ce8\u610f\u529b<\/h4>\n
$$M_{ij} = \/begin{cases}
0, & \/text{if } i \/geq j \/
-\/infty, & \/text{otherwise}
\/end{cases}$$<\/p>\n
$$\/text{MaskedAttention}(Q, K, V) = \/text{softmax}\/left(\/frac{QK^T + M}{\/sqrt{d_k}}\/right)V$$<\/p>\n3.2 \u7f16\u7801\u5668-\u89e3\u7801\u5668\u6ce8\u610f\u529b<\/h4>\n
$$Q = \/text{Decoder Output}$$
$$K = V = \/text{Encoder Output}$$<\/p>\n3.3 \u524d\u9988\u7f51\u7edc<\/h4>\n
$$\/text{FFN}(x) = \/max(0, xW_1 + b_1)W_2 + b_2$$<\/p>\n4. \u4f4d\u7f6e\u7f16\u7801<\/h3>\n
4.1 \u6b63\u5f26\u4f4d\u7f6e\u7f16\u7801<\/h4>\n
$$PE_{(pos, 2i)} = \/sin\/left(\/frac{pos}{10000^{2i\/d}}\/right)$$
$$PE_{(pos, 2i+1)} = \/cos\/left(\/frac{pos}{10000^{2i\/d}}\/right)$$<\/p>\n\n
4.2 \u53ef\u5b66\u4e60\u4f4d\u7f6e\u7f16\u7801<\/h4>\n
4.3 \u76f8\u5bf9\u4f4d\u7f6e\u7f16\u7801<\/h4>\n
5. \u8bad\u7ec3\u6280\u5de7<\/h3>\n
5.1 \u6807\u7b7e\u5e73\u6ed1<\/h4>\n
$$\/text{smooth\/_label} = (1 – \/epsilon) \/cdot \/text{one\/_hot} + \/frac{\/epsilon}{K}$$<\/p>\n\n
5.2 \u5b66\u4e60\u7387\u8c03\u5ea6<\/h4>\n
$$lrate = d_{model}^{-0.5} \/cdot \/min(step^{-0.5}, step \/cdot warmup^{-1.5})$$<\/p>\n\n
5.3 Dropout<\/h4>\n
\n
6. \u5e94\u7528\u5b9e\u4f8b<\/h3>\n
6.1 \u673a\u5668\u7ffb\u8bd1<\/h4>\n
\n
6.2 \u6587\u672c\u751f\u6210<\/h4>\n
6.3 \u6587\u672c\u5206\u7c7b<\/h4>\n
7. \u5b9e\u9a8c\u5206\u6790<\/h3>\n
7.1 \u6570\u636e\u96c6<\/h4>\n
\n
7.2 \u5b9e\u9a8c\u7ed3\u679c<\/h4>\n
\n\n
\n \n\u6570\u636e\u96c6<\/th>\n \u6a21\u578b<\/th>\n \u8bad\u7ec3\u51c6\u786e\u7387(%)<\/th>\n \u6d4b\u8bd5\u51c6\u786e\u7387(%)<\/th>\n \u8bad\u7ec3\u65f6\u95f4(h)<\/th>\n<\/tr>\n<\/thead>\n \n WMT<\/td>\n RNN<\/td>\n 85.2<\/td>\n 78.5<\/td>\n 5.5<\/td>\n<\/tr>\n \n WMT<\/td>\n LSTM<\/td>\n 88.5<\/td>\n 82.3<\/td>\n 8.5<\/td>\n<\/tr>\n \n WMT<\/td>\n Transformer<\/td>\n 92.5<\/td>\n 85.3<\/td>\n 12.5<\/td>\n<\/tr>\n \n PTB<\/td>\n RNN<\/td>\n 85.2<\/td>\n 78.5<\/td>\n 2.5<\/td>\n<\/tr>\n \n PTB<\/td>\n LSTM<\/td>\n 88.5<\/td>\n 82.3<\/td>\n 4.5<\/td>\n<\/tr>\n \n PTB<\/td>\n Transformer<\/td>\n 92.5<\/td>\n 85.3<\/td>\n 8.5<\/td>\n<\/tr>\n \n IMDB<\/td>\n RNN<\/td>\n 88.5<\/td>\n 85.2<\/td>\n 1.5<\/td>\n<\/tr>\n \n IMDB<\/td>\n LSTM<\/td>\n 91.5<\/td>\n 88.5<\/td>\n 2.5<\/td>\n<\/tr>\n \n IMDB<\/td>\n Transformer<\/td>\n 93.5<\/td>\n 90.5<\/td>\n 4.5<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n 8. \u7ed3\u8bba<\/h3>\n
\n
\n
\n
\n