{"id":2009,"date":"2022-01-13T19:07:48","date_gmt":"2022-01-13T11:07:48","guid":{"rendered":"http:\/\/139.9.1.231\/?p=2009"},"modified":"2023-02-11T10:49:06","modified_gmt":"2023-02-11T02:49:06","slug":"transformer","status":"publish","type":"post","link":"http:\/\/139.9.1.231\/index.php\/2022\/01\/13\/transformer\/","title":{"rendered":"Transformer"},"content":{"rendered":"\n

Transformer \u6700\u521d\u4e3b\u8981\u5e94\u7528\u4e8e\u4e00\u4e9b\u81ea\u7136\u8bed\u8a00\u5904\u7406\u573a\u666f\uff0c\u6bd4\u5982\u7ffb\u8bd1\u3001\u6587\u672c\u5206\u7c7b\u3001\u5199\u5c0f\u8bf4\u3001\u5199\u6b4c\u7b49\u3002\u968f\u7740\u6280\u672f\u7684\u53d1\u5c55\uff0cTransformer \u5f00\u59cb\u5f81\u6218\u89c6\u89c9\u9886\u57df\uff0c\u5206\u7c7b\u3001\u68c0\u6d4b\u7b49\u4efb\u52a1\u5747\u4e0d\u5728\u8bdd\u4e0b\uff0c\u9010\u6e10\u8d70\u4e0a\u4e86\u591a\u6a21\u6001<\/strong>\u7684\u9053\u8def\u3002<\/p>\n\n\n\n

Transformer \u662f Google \u5728 2017 \u5e74\u63d0\u51fa\u7684\u7528\u4e8e\u673a\u5668\u7ffb\u8bd1\u7684\u6a21\u578b\u3002<\/p>\n\n\n\n

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

Transformer \u7684\u5185\u90e8\uff0c\u5728\u672c\u8d28\u4e0a\u662f\u4e00\u4e2a Encoder-Decoder \u7684\u7ed3\u6784\uff0c\u5373 \u7f16\u7801\u5668-\u89e3\u7801\u5668\u3002<\/p>\n\n\n\n

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

Transformer \u4e2d\u629b\u5f03\u4e86\u4f20\u7edf\u7684 CNN \u548c RNN\uff0c\u6574\u4e2a\u7f51\u7edc\u7ed3\u6784\u5b8c\u5168\u7531 Attention \u673a\u5236\u7ec4\u6210\uff0c\u5e76\u4e14\u91c7\u7528\u4e86 6 \u5c42 Encoder-Decoder \u7ed3\u6784\u3002<\/p>\n\n\n\n

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

\u663e\u7136\uff0cTransformer \u4e3b\u8981\u5206\u4e3a\u4e24\u5927\u90e8\u5206<\/strong>\uff0c\u5206\u522b\u662f\u7f16\u7801\u5668<\/strong>\u548c\u89e3\u7801\u5668<\/strong>\u3002\u6574\u4e2a Transformer \u662f\u7531 6 \u4e2a\u8fd9\u6837\u7684\u7ed3\u6784\u7ec4\u6210\uff0c\u4e3a\u4e86\u65b9\u4fbf\u7406\u89e3\uff0c\u6211\u4eec\u53ea\u770b\u5176\u4e2d\u4e00\u4e2aEncoder-Decoder \u7ed3\u6784\u3002<\/p>\n\n\n\n

\u4ee5\u4e00\u4e2a\u7b80\u5355\u7684\u4f8b\u5b50\u8fdb\u884c\u8bf4\u660e\uff1a<\/p>\n\n\n\n

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

Why do we work?\uff0c\u6211\u4eec\u4e3a\u4ec0\u4e48\u5de5\u4f5c\uff1f\u5de6\u4fa7\u7ea2\u6846\u662f\u7f16\u7801\u5668<\/strong>\uff0c\u53f3\u4fa7\u7ea2\u6846\u662f\u89e3\u7801\u5668<\/strong>\uff0c\u7f16\u7801\u5668<\/strong>\u8d1f\u8d23\u628a\u81ea\u7136\u8bed\u8a00\u5e8f\u5217\u6620\u5c04\u6210\u4e3a\u9690\u85cf\u5c42\uff08\u4e0a\u56fe\u7b2c2\u6b65\uff09\uff0c\u5373\u542b\u6709\u81ea\u7136\u8bed\u8a00\u5e8f\u5217\u7684\u6570\u5b66\u8868\u8fbe\u3002\u89e3\u7801\u5668<\/strong>\u628a\u9690\u85cf\u5c42\u518d\u6620\u5c04\u4e3a\u81ea\u7136\u8bed\u8a00\u5e8f\u5217\uff0c\u4ece\u800c\u4f7f\u6211\u4eec\u53ef\u4ee5\u89e3\u51b3\u5404\u79cd\u95ee\u9898\uff0c\u5982\u60c5\u611f\u5206\u6790\u3001\u673a\u5668\u7ffb\u8bd1\u3001\u6458\u8981\u751f\u6210\u3001\u8bed\u4e49\u5173\u7cfb\u62bd\u53d6\u7b49\u3002\u7b80\u5355\u8bf4\u4e0b\uff0c\u4e0a\u56fe\u6bcf\u4e00\u6b65\u90fd\u505a\u4e86\u4ec0\u4e48\uff1a<\/p>\n\n\n\n

  • \u8f93\u5165\u81ea\u7136\u8bed\u8a00\u5e8f\u5217\u5230\u7f16\u7801\u5668: Why do we work?(\u4e3a\u4ec0\u4e48\u8981\u5de5\u4f5c)\uff1b<\/li>
  • \u7f16\u7801\u5668\u8f93\u51fa\u7684\u9690\u85cf\u5c42\uff0c\u518d\u8f93\u5165\u5230\u89e3\u7801\u5668\uff1b<\/li>
  • \u8f93\u5165 <\ud835\udc60\ud835\udc61\ud835\udc4e\ud835\udc5f\ud835\udc61> (\u8d77\u59cb)\u7b26\u53f7\u5230\u89e3\u7801\u5668\uff1b<\/li>
  • \u89e3\u7801\u5668\u5f97\u5230\u7b2c\u4e00\u4e2a\u5b57”\u4e3a”\uff1b<\/li>
  • \u5c06\u5f97\u5230\u7684\u7b2c\u4e00\u4e2a\u5b57”\u4e3a”\u843d\u4e0b\u6765\u518d\u8f93\u5165\u5230\u89e3\u7801\u5668\uff1b<\/li>
  • \u89e3\u7801\u5668\u5f97\u5230\u7b2c\u4e8c\u4e2a\u5b57”\u4ec0”\uff1b<\/li>
  • \u5c06\u5f97\u5230\u7684\u7b2c\u4e8c\u5b57\u518d\u843d\u4e0b\u6765\uff0c\u76f4\u5230\u89e3\u7801\u5668\u8f93\u51fa <\ud835\udc52\ud835\udc5b\ud835\udc51> (\u7ec8\u6b62\u7b26)\uff0c\u5373\u5e8f\u5217\u751f\u6210\u5b8c\u6210\u3002<\/li><\/ul>\n\n\n\n

    \u89e3\u7801\u5668\u548c\u7f16\u7801\u5668\u7684\u7ed3\u6784\u7c7b\u4f3c\uff0c\u672c\u6587\u4ee5\u7f16\u7801\u5668\u90e8\u5206\u8fdb\u884c\u8bb2\u89e3\u3002\u5373\u628a\u81ea\u7136\u8bed\u8a00\u5e8f\u5217\u6620\u5c04\u4e3a\u9690\u85cf\u5c42\u7684\u6570\u5b66\u8868\u8fbe<\/strong>\u7684\u8fc7\u7a0b\uff0c\u56e0\u4e3a\u7406\u89e3\u4e86\u7f16\u7801\u5668\u4e2d\u7684\u7ed3\u6784\uff0c\u7406\u89e3\u89e3\u7801\u5668\u5c31\u975e\u5e38\u7b80\u5355\u4e86\u3002\u4e3a\u4e86\u65b9\u4fbf\u5b66\u4e60\uff0c\u6211\u5c06\u7f16\u7801\u5668\u5206\u4e3a 4 \u4e2a\u90e8\u5206\uff0c\u4f9d\u6b21\u8bb2\u89e3\u3002<\/p>\n\n\n\n

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

    1\u3001\u4f4d\u7f6e\u5d4c\u5165\uff08\ud835\udc5d\ud835\udc5c\ud835\udc60\ud835\udc56\ud835\udc61\ud835\udc56\ud835\udc5c\ud835\udc5b\ud835\udc4e\ud835\udc59 \ud835\udc52\ud835\udc5b\ud835\udc50\ud835\udc5c\ud835\udc51\ud835\udc56\ud835\udc5b\ud835\udc54\uff09<\/strong><\/h3>\n\n\n\n

    \u6211\u4eec\u8f93\u5165\u6570\u636e X \u7ef4\u5ea6\u4e3a[batch size, sequence length]\u7684\u6570\u636e\uff0c\u6bd4\u5982\u6211\u4eec\u4e3a\u4ec0\u4e48\u5de5\u4f5c\u3002batch size \u5c31\u662f batch \u7684\u5927\u5c0f\uff0c\u8fd9\u91cc\u53ea\u6709\u4e00\u53e5\u8bdd\uff0c\u6240\u4ee5 batch size \u4e3a 1\uff0csequence length \u662f\u53e5\u5b50\u7684\u957f\u5ea6\uff0c\u4e00\u5171 7 \u4e2a\u5b57\uff0c\u6240\u4ee5\u8f93\u5165\u7684\u6570\u636e\u7ef4\u5ea6\u662f [1, 7]\u3002\u6211\u4eec\u4e0d\u80fd\u76f4\u63a5\u5c06\u8fd9\u53e5\u8bdd\u8f93\u5165\u5230\u7f16\u7801\u5668<\/strong>\u4e2d\uff0c\u56e0\u4e3a Tranformer \u4e0d\u8ba4\u8bc6\uff0c\u6211\u4eec\u9700\u8981\u5148\u8fdb\u884c\u5b57\u5d4c\u5165<\/strong>\uff0c\u5373\u5f97\u5230\u56fe\u4e2d\u7684 \u3002\u7b80\u5355\u70b9\u8bf4\uff0c\u5c31\u662f\u6587\u5b57->\u5b57\u5411\u91cf\u7684\u8f6c\u6362\uff0c\u8fd9\u79cd\u8f6c\u6362\u662f\u5c06\u6587\u5b57\u8f6c\u6362\u4e3a\u8ba1\u7b97\u673a\u8ba4\u8bc6\u7684\u6570\u5b66\u8868\u793a\uff0c\u7528\u5230\u7684\u65b9\u6cd5\u5c31\u662f Word2Vec\uff0cWord2Vec \u7684\u5177\u4f53\u7ec6\u8282\uff0c\u5bf9\u4e8e\u521d\u5b66\u8005\u6682\u4e14\u4e0d\u7528\u4e86\u89e3\uff0c\u8fd9\u4e2a\u662f\u53ef\u4ee5\u76f4\u63a5\u4f7f\u7528\u7684\u3002\u5f97\u5230\u7ef4\u5ea6\u662f [batch size, sequence length, embedding dimension]\uff0cembedding dimension \u7684\u5927\u5c0f\u7531 Word2Vec \u7b97\u6cd5\u51b3\u5b9a\uff0cTranformer \u91c7\u7528 512 \u957f\u5ea6\u7684\u5b57\u5411\u91cf\u3002\u6240\u4ee5 \u7684\u7ef4\u5ea6\u662f [1, 7, 512]\u3002\u81f3\u6b64\uff0c\u8f93\u5165\u7684\u6211\u4eec\u4e3a\u4ec0\u4e48\u5de5\u4f5c\uff0c\u53ef\u4ee5\u7528\u4e00\u4e2a\u77e9\u9635\u6765\u7b80\u5316\u8868\u793a\u3002<\/p>\n\n\n\n

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

    \u6211\u4eec\u77e5\u9053\uff0c\u6587\u5b57\u7684\u5148\u540e\u987a\u5e8f\uff0c\u5f88\u91cd\u8981\u3002\u6bd4\u5982\u5403\u996d\u6ca1\u3001\u6ca1\u5403\u996d\u3001\u6ca1\u996d\u5403\u3001\u996d\u5403\u6ca1\u3001\u996d\u6ca1\u5403\uff0c\u540c\u6837\u4e09\u4e2a\u5b57\uff0c\u987a\u5e8f\u98a0\u5012\uff0c\u6240\u8868\u8fbe\u7684\u542b\u4e49\u5c31\u4e0d\u540c\u4e86\u3002\u6587\u5b57\u7684\u4f4d\u7f6e\u4fe1\u606f\u5f88\u91cd\u8981\uff0cTranformer \u6ca1\u6709\u7c7b\u4f3c RNN \u7684\u5faa\u73af\u7ed3\u6784\uff0c\u6ca1\u6709\u6355\u6349\u987a\u5e8f\u5e8f\u5217\u7684\u80fd\u529b\u3002\u4e3a\u4e86\u4fdd\u7559\u8fd9\u79cd\u4f4d\u7f6e\u4fe1\u606f\u4ea4\u7ed9 Tranformer \u5b66\u4e60\uff0c\u6211\u4eec\u9700\u8981\u7528\u5230\u4f4d\u7f6e\u5d4c\u5165\u3002<\/strong>\u52a0\u5165\u4f4d\u7f6e\u4fe1\u606f\u7684\u65b9\u5f0f\u975e\u5e38\u591a\uff0c\u6700\u7b80\u5355\u7684\u53ef\u4ee5\u662f\u76f4\u63a5\u5c06\u7edd\u5bf9\u5750\u6807 0,1,2 \u7f16\u7801\u3002Tranformer \u91c7\u7528\u7684\u662f sin-cos \u89c4\u5219\uff0c\u4f7f\u7528\u4e86 sin \u548c cos \u51fd\u6570\u7684\u7ebf\u6027\u53d8\u6362\u6765\u63d0\u4f9b\u7ed9\u6a21\u578b\u4f4d\u7f6e\u4fe1\u606f\uff1a<\/p>\n\n\n\n

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

    \u4e0a\u5f0f\u4e2d pos \u6307\u7684\u662f\u53e5\u4e2d\u5b57\u7684\u4f4d\u7f6e\uff0c\u53d6\u503c\u8303\u56f4\u662f [0, \ud835\udc5a\ud835\udc4e\ud835\udc65 \ud835\udc60\ud835\udc52\ud835\udc5e\ud835\udc62\ud835\udc52\ud835\udc5b\ud835\udc50\ud835\udc52 \ud835\udc59\ud835\udc52\ud835\udc5b\ud835\udc54\ud835\udc61\u210e)\uff0ci \u6307\u7684\u662f\u5b57\u5d4c\u5165\u7684\u7ef4\u5ea6, \u53d6\u503c\u8303\u56f4\u662f [0, \ud835\udc52\ud835\udc5a\ud835\udc4f\ud835\udc52\ud835\udc51\ud835\udc51\ud835\udc56\ud835\udc5b\ud835\udc54 \ud835\udc51\ud835\udc56\ud835\udc5a\ud835\udc52\ud835\udc5b\ud835\udc60\ud835\udc56\ud835\udc5c\ud835\udc5b)\u3002 \u5c31\u662f \ud835\udc52\ud835\udc5a\ud835\udc4f\ud835\udc52\ud835\udc51\ud835\udc51\ud835\udc56\ud835\udc5b\ud835\udc54 \ud835\udc51\ud835\udc56\ud835\udc5a\ud835\udc52\ud835\udc5b\ud835\udc60\ud835\udc56\ud835\udc5c\ud835\udc5b \u7684\u5927\u5c0f\u3002\u4e0a\u9762\u6709 sin \u548c cos \u4e00\u7ec4\u516c\u5f0f\uff0c\u4e5f\u5c31\u662f\u5bf9\u5e94\u7740 \ud835\udc52\ud835\udc5a\ud835\udc4f\ud835\udc52\ud835\udc51\ud835\udc51\ud835\udc56\ud835\udc5b\ud835\udc54 \ud835\udc51\ud835\udc56\ud835\udc5a\ud835\udc52\ud835\udc5b\ud835\udc60\ud835\udc56\ud835\udc5c\ud835\udc5b \u7ef4\u5ea6\u7684\u4e00\u7ec4\u5947\u6570\u548c\u5076\u6570\u7684\u5e8f\u53f7\u7684\u7ef4\u5ea6\uff0c\u4ece\u800c\u4ea7\u751f\u4e0d\u540c\u7684\u5468\u671f\u6027\u53d8\u5316\u3002<\/p>\n\n\n\n

    # \u5bfc\u5165\u4f9d\u8d56\u5e93\nimport numpy as np\nimport matplotlib.pyplot as plt\nimport seaborn as sns\nimport math\n\ndef get_positional_encoding(max_seq_len, embed_dim):\n    # \u521d\u59cb\u5316\u4e00\u4e2apositional encoding\n    # embed_dim: \u5b57\u5d4c\u5165\u7684\u7ef4\u5ea6\n    # max_seq_len: \u6700\u5927\u7684\u5e8f\u5217\u957f\u5ea6\n    positional_encoding = np.array([\n        [pos \/ np.power(10000, 2 * i \/ embed_dim) for i in range(embed_dim)]\n        if pos != 0 else np.zeros(embed_dim) for pos in range(max_seq_len)])\n    positional_encoding[1:, 0::2] = np.sin(positional_encoding[1:, 0::2])  # dim 2i \u5076\u6570\n    positional_encoding[1:, 1::2] = np.cos(positional_encoding[1:, 1::2])  # dim 2i+1 \u5947\u6570\n    # \u5f52\u4e00\u5316, \u7528\u4f4d\u7f6e\u5d4c\u5165\u7684\u6bcf\u4e00\u884c\u9664\u4ee5\u5b83\u7684\u6a21\u957f\n    # denominator = np.sqrt(np.sum(position_enc**2, axis=1, keepdims=True))\n    # position_enc = position_enc \/ (denominator + 1e-8)\n    return positional_encoding\n    \npositional_encoding = get_positional_encoding(max_seq_len=100, embed_dim=16)\nplt.figure(figsize=(10,10))\nsns.heatmap(positional_encoding)\nplt.title(\"Sinusoidal Function\")\nplt.xlabel(\"hidden dimension\")\nplt.ylabel(\"sequence length\")<\/code><\/pre>\n\n\n\n
    \u53ef\u4ee5\u770b\u5230\uff0c\u4f4d\u7f6e\u5d4c\u5165\u5728 \ud835\udc52\ud835\udc5a\ud835\udc4f\ud835\udc52\ud835\udc51\ud835\udc51\ud835\udc56\ud835\udc5b\ud835\udc54 \ud835\udc51\ud835\udc56\ud835\udc5a\ud835\udc52\ud835\udc5b\ud835\udc60\ud835\udc56\ud835\udc5c\ud835\udc5b \uff08\u4e5f\u662fhidden dimension \uff09\u7ef4\u5ea6\u4e0a\u968f\u7740\u7ef4\u5ea6\u5e8f\u53f7\u589e\u5927\uff0c\u5468\u671f\u53d8\u5316\u4f1a\u8d8a\u6765\u8d8a\u6162\uff0c\u800c\u4ea7\u751f\u4e00\u79cd\u5305\u542b\u4f4d\u7f6e\u4fe1\u606f\u7684\u7eb9\u7406\u3002\n\n\u5c31\u8fd9\u6837\uff0c\u4ea7\u751f\u72ec\u4e00\u7684\u7eb9\u7406\u4f4d\u7f6e\u4fe1\u606f\uff0c\u6a21\u578b\u4ece\u800c\u5b66\u5230\u4f4d\u7f6e\u4e4b\u95f4\u7684\u4f9d\u8d56\u5173\u7cfb\u548c\u81ea\u7136\u8bed\u8a00\u7684\u65f6\u5e8f\u7279\u6027\u3002\n<\/code><\/pre>\n\n\n\n
    \"\"<\/figure><\/div>\n\n\n\n
    \u6700\u540e\uff0c\u5c06Xembedding  \u548c \u4f4d\u7f6e\u5d4c\u5165<\/code> \u76f8\u52a0\uff0c\u9001\u7ed9\u4e0b\u4e00\u5c42\u3002<\/p>\n\n\n\n
    2\u3001\u81ea\u6ce8\u610f\u529b\u5c42\uff08\ud835\udc60\ud835\udc52\ud835\udc59\ud835\udc53 \ud835\udc4e\ud835\udc61\ud835\udc61\ud835\udc52\ud835\udc5b\ud835\udc61\ud835\udc56\ud835\udc5c\ud835\udc5b \ud835\udc5a\ud835\udc52\ud835\udc50\u210e\ud835\udc4e\ud835\udc5b\ud835\udc56\ud835\udc60\ud835\udc5a\uff09<\/h3>\n\n\n\n
    \u76f4\u63a5\u770b\u4e0b\u56fe\u7b14\u8bb0\uff0c\u8bb2\u89e3\u7684\u975e\u5e38\u8be6\u7ec6\u3002<\/p>\n\n\n\n
    \"\"<\/figure>\n\n\n\n
    \u591a\u5934\u7684\u610f\u4e49\u5728\u4e8e\uff0c\\(QK^{T}\\) \u5f97\u5230\u7684\u77e9\u9635\u5c31\u53eb\u6ce8\u610f\u529b\u77e9\u9635\uff0c\u5b83\u53ef\u4ee5\u8868\u793a\u6bcf\u4e2a\u5b57\u4e0e\u5176\u4ed6\u5b57\u7684\u76f8\u4f3c\u7a0b\u5ea6\u3002\u56e0\u4e3a\uff0c\u5411\u91cf\u7684\u70b9\u79ef\u503c\u8d8a\u5927\uff0c\u8bf4\u660e\u4e24\u4e2a\u5411\u91cf\u8d8a\u63a5\u8fd1\u3002<\/p>\n\n\n\n
    \"\"<\/figure>\n\n\n\n
    \u6211\u4eec\u7684\u76ee\u7684\u662f\uff0c\u8ba9\u6bcf\u4e2a\u5b57\u90fd\u542b\u6709\u5f53\u524d\u8fd9\u4e2a\u53e5\u5b50\u4e2d\u7684\u6240\u6709\u5b57\u7684\u4fe1\u606f\uff0c\u7528\u6ce8\u610f\u529b\u5c42\uff0c\u6211\u4eec\u505a\u5230\u4e86\u3002<\/p>\n\n\n\n
    \u9700\u8981\u6ce8\u610f\u7684\u662f<\/strong>\uff0c\u5728\u4e0a\u9762 \ud835\udc60\ud835\udc52\ud835\udc59\ud835\udc53 \ud835\udc4e\ud835\udc61\ud835\udc61\ud835\udc52\ud835\udc5b\ud835\udc61\ud835\udc56\ud835\udc5c\ud835\udc5b<\/code> \u7684\u8ba1\u7b97\u8fc7\u7a0b\u4e2d\uff0c\u6211\u4eec\u901a\u5e38\u4f7f\u7528 \ud835\udc5a\ud835\udc56\ud835\udc5b\ud835\udc56 \ud835\udc4f\ud835\udc4e\ud835\udc61\ud835\udc50\u210e<\/code>\uff0c\u4e5f\u5c31\u662f\u4e00\u6b21\u8ba1\u7b97\u591a\u53e5\u8bdd\uff0c\u4e0a\u6587\u4e3e\u4f8b\u53ea\u7528\u4e86\u4e00\u4e2a\u53e5\u5b50\u3002<\/p>\n\n\n\n
    \u6bcf\u4e2a\u53e5\u5b50\u7684\u957f\u5ea6\u662f\u4e0d\u4e00\u6837\u7684\uff0c\u9700\u8981\u6309\u7167\u6700\u957f\u7684\u53e5\u5b50\u7684\u957f\u5ea6\u7edf\u4e00\u5904\u7406\u3002\u5bf9\u4e8e\u77ed\u7684\u53e5\u5b50\uff0c\u8fdb\u884c Padding<\/code> \u64cd\u4f5c\uff0c\u4e00\u822c\u6211\u4eec\u7528 0<\/code> \u6765\u8fdb\u884c\u586b\u5145\u3002<\/p>\n\n\n\n
    \"\"<\/figure>\n\n\n\n
    3\u3001\u6b8b\u5dee\u94fe\u63a5\u548c\u5c42\u5f52\u4e00\u5316<\/h3>\n\n\n\n
    \u52a0\u5165\u4e86\u6b8b\u5dee\u8bbe\u8ba1\u548c\u5c42\u5f52\u4e00\u5316\u64cd\u4f5c\uff0c\u76ee\u7684\u662f\u4e3a\u4e86\u9632\u6b62\u68af\u5ea6\u6d88\u5931\uff0c\u52a0\u5feb\u6536\u655b\u3002<\/p>\n\n\n\n
    1) \u6b8b\u5dee\u8bbe\u8ba1<\/h4>\n\n\n\n
    \u6211\u4eec\u5728\u4e0a\u4e00\u6b65\u5f97\u5230\u4e86\u7ecf\u8fc7\u6ce8\u610f\u529b\u77e9\u9635\u52a0\u6743\u4e4b\u540e\u7684 \ud835\udc49<\/code>\uff0c \u4e5f\u5c31\u662f \ud835\udc34\ud835\udc61\ud835\udc61\ud835\udc52\ud835\udc5b\ud835\udc61\ud835\udc56\ud835\udc5c\ud835\udc5b(\ud835\udc44, \ud835\udc3e, \ud835\udc49)<\/code>\uff0c\u6211\u4eec\u5bf9\u5b83\u8fdb\u884c\u4e00\u4e0b\u8f6c\u7f6e\uff0c\u4f7f\u5176\u548c \ud835\udc4b\ud835\udc52\ud835\udc5a\ud835\udc4f\ud835\udc52\ud835\udc51\ud835\udc51\ud835\udc56\ud835\udc5b\ud835\udc54<\/code> \u7684\u7ef4\u5ea6\u4e00\u81f4, \u4e5f\u5c31\u662f [\ud835\udc4f\ud835\udc4e\ud835\udc61\ud835\udc50\u210e \ud835\udc60\ud835\udc56\ud835\udc67\ud835\udc52, \ud835\udc60\ud835\udc52\ud835\udc5e\ud835\udc62\ud835\udc52\ud835\udc5b\ud835\udc50\ud835\udc52 \ud835\udc59\ud835\udc52\ud835\udc5b\ud835\udc54\ud835\udc61\u210e, \ud835\udc52\ud835\udc5a\ud835\udc4f\ud835\udc52\ud835\udc51\ud835\udc51\ud835\udc56\ud835\udc5b\ud835\udc54 \ud835\udc51\ud835\udc56\ud835\udc5a\ud835\udc52\ud835\udc5b\ud835\udc60\ud835\udc56\ud835\udc5c\ud835\udc5b]<\/code> \uff0c\u7136\u540e\u628a\u4ed6\u4eec\u52a0\u8d77\u6765\u505a\u6b8b\u5dee\u8fde\u63a5\uff0c\u76f4\u63a5\u8fdb\u884c\u5143\u7d20\u76f8\u52a0\uff0c\u56e0\u4e3a\u4ed6\u4eec\u7684\u7ef4\u5ea6\u4e00\u81f4:<\/p>\n\n\n\n
    $$
    X_{\\text {embedding }}+\\text { Attention }(Q, K, V)
    $$<\/p>\n\n\n\n
    \u5728\u4e4b\u540e\u7684\u8fd0\u7b97\u91cc\uff0c\u6bcf\u7ecf\u8fc7\u4e00\u4e2a\u6a21\u5757\u7684\u8fd0\u7b97\uff0c\u90fd\u8981\u628a\u8fd0\u7b97\u4e4b\u524d\u7684\u503c\u548c\u8fd0\u7b97\u4e4b\u540e\u7684\u503c\u76f8\u52a0\uff0c\u4ece\u800c\u5f97\u5230\u6b8b\u5dee\u8fde\u63a5\uff0c\u8bad\u7ec3\u7684\u65f6\u5019\u53ef\u4ee5\u4f7f\u68af\u5ea6\u76f4\u63a5\u8d70\u6377\u5f84\u53cd\u4f20\u5230\u6700\u521d\u59cb\u5c42\uff1a<\/p>\n\n\n\n
    $$
    X+\\text { SubLayer }(X)
    $$<\/p>\n\n\n\n
    2) \u5c42\u5f52\u4e00\u5316<\/h4>\n\n\n\n
    \u4f5c\u7528\u662f\u628a\u795e\u7ecf\u7f51\u7edc\u4e2d\u9690\u85cf\u5c42\u5f52\u4e00\u4e3a\u6807\u51c6\u6b63\u6001\u5206\u5e03\uff0c\u4e5f\u5c31\u662f \ud835\udc56.\ud835\udc56.\ud835\udc51<\/code> \u72ec\u7acb\u540c\u5206\u5e03\uff0c \u4ee5\u8d77\u5230\u52a0\u5feb\u8bad\u7ec3\u901f\u5ea6\uff0c \u52a0\u901f\u6536\u655b\u7684\u4f5c\u7528\u3002<\/p>\n\n\n\n
    $$
    \\mu_{i}=\\frac{1}{m} \\sum_{i=1}^{m} x_{i j}
    $$<\/p>\n\n\n\n
    \u4e0a\u5f0f\u4e2d\u4ee5\u77e9\u9635\u7684\u884c (\ud835\udc5f\ud835\udc5c\ud835\udc64) \u4e3a\u5355\u4f4d\u6c42\u5747\u503c\uff1a<\/p>\n\n\n\n
    $$
    \\sigma_{j}^{2}=\\frac{1}{m} \\sum_{i=1}^{m}\\left(x_{i j}-\\mu_{j}\\right)^{2}
    $$<\/p>\n\n\n\n
    \u4e0a\u5f0f\u4e2d\u4ee5\u77e9\u9635\u7684\u884c (\ud835\udc5f\ud835\udc5c\ud835\udc64) \u4e3a\u5355\u4f4d\u6c42\u65b9\u5dee\uff1a<\/p>\n\n\n\n
    $$
    \\operatorname{LayerNorm}(x)=\\alpha \\odot \\frac{x_{i j}-\\mu_{i}}{\\sqrt{\\sigma_{i}^{2}+\\epsilon}}+\\beta
    $$<\/p>\n\n\n\n
    \u7136\u540e\u7528\u6bcf\u4e00\u884c\u7684\u6bcf\u4e00\u4e2a\u5143\u7d20\u51cf\u53bb\u8fd9\u884c\u7684\u5747\u503c\uff0c\u518d\u9664\u4ee5\u8fd9\u884c\u7684\u6807\u51c6\u5dee\uff0c\u4ece\u800c\u5f97\u5230\u5f52 \u4e00\u5316\u540e\u7684\u6570\u503c\uff0c \\(\\epsilon\\) \u662f\u4e3a\u4e86\u9632\u6b62\u9664 0 \uff1b\u4e4b\u540e\u5f15\u5165\u4e24\u4e2a\u53ef\u8bad\u7ec3\u53c2\u6570 \\(\\alpha, \\beta\\) \u6765\u5f25\u8865\u5f52\u4e00\u5316\u7684\u8fc7\u7a0b\u4e2d\u635f\u5931\u6389\u7684\u4fe1\u606f\uff0c\u6ce8\u610f \\(\\odot\\) \u8868\u793a \u5143\u7d20\u76f8\u4e58\u800c\u4e0d\u662f\u70b9\u79ef\uff0c\u6211\u4eec\u4e00\u822c\u521d\u59cb\u5316 \\(\\alpha\\) \u4e3a\u5168 1 \uff0c\u800c\\(\\beta\\) \u4e3a\u5168 0 \u3002<\/p>\n\n\n\n
    \u4ee3\u7801\u5c42\u9762\u975e\u5e38\u7b80\u5355\uff0c\u5355\u5934 attention<\/code> \u64cd\u4f5c\u5982\u4e0b\uff1a<\/p>\n\n\n\n
    class ScaledDotProductAttention(nn.Module):\n    ''' Scaled Dot-Product Attention '''\n\n    def __init__(self, temperature, attn_dropout=0.1):\n        super().__init__()\n        self.temperature = temperature\n        self.dropout = nn.Dropout(attn_dropout)\n\n    def forward(self, q, k, v, mask=None):\n        # self.temperature\u662f\u8bba\u6587\u4e2d\u7684d_k ** 0.5\uff0c\u9632\u6b62\u68af\u5ea6\u8fc7\u5927\n        # QxK\/sqrt(dk)\n        attn = torch.matmul(q \/ self.temperature, k.transpose(2, 3))\n\n        if mask is not None:\n            # \u5c4f\u853d\u4e0d\u60f3\u8981\u7684\u8f93\u51fa\n            attn = attn.masked_fill(mask == 0, -1e9)\n        # softmax+dropout\n        attn = self.dropout(F.softmax(attn, dim=-1))\n        # \u6982\u7387\u5206\u5e03xV\n        output = torch.matmul(attn, v)\n\n        return output, attn<\/code><\/pre>\n\n\n\n
    Multi-Head Attention<\/code> \u5b9e\u73b0\u5728 ScaledDotProductAttention<\/code> \u57fa\u7840\u4e0a\u6784\u5efa\uff1a<\/p>\n\n\n\n
    class MultiHeadAttention(nn.Module):\n    ''' Multi-Head Attention module '''\n\n    # n_head\u5934\u7684\u4e2a\u6570\uff0c\u9ed8\u8ba4\u662f8\n    # d_model\u7f16\u7801\u5411\u91cf\u957f\u5ea6\uff0c\u4f8b\u5982\u672c\u6587\u8bf4\u7684512\n    # d_k, d_v\u7684\u503c\u4e00\u822c\u4f1a\u8bbe\u7f6e\u4e3a n_head * d_k=d_model\uff0c\n    # \u6b64\u65f6concat\u540e\u6b63\u597d\u548c\u539f\u59cb\u8f93\u5165\u4e00\u6837\uff0c\u5f53\u7136\u4e0d\u76f8\u540c\u4e5f\u53ef\u4ee5\uff0c\u56e0\u4e3a\u540e\u9762\u6709fc\u5c42\n    # \u76f8\u5f53\u4e8e\u5c06\u53ef\u5b66\u4e60\u77e9\u9635\u5206\u6210\u72ec\u7acb\u7684n_head\u4efd\n    def __init__(self, n_head, d_model, d_k, d_v, dropout=0.1):\n        super().__init__()\n        # \u5047\u8bben_head=8\uff0cd_k=64\n        self.n_head = n_head\n        self.d_k = d_k\n        self.d_v = d_v\n        # d_model\u8f93\u5165\u5411\u91cf\uff0cn_head * d_k\u8f93\u51fa\u5411\u91cf\n        # \u53ef\u5b66\u4e60W^Q\uff0cW^K,W^V\u77e9\u9635\u53c2\u6570\u521d\u59cb\u5316\n        self.w_qs = nn.Linear(d_model, n_head * d_k, bias=False)\n        self.w_ks = nn.Linear(d_model, n_head * d_k, bias=False)\n        self.w_vs = nn.Linear(d_model, n_head * d_v, bias=False)\n        # \u6700\u540e\u7684\u8f93\u51fa\u7ef4\u5ea6\u53d8\u6362\u64cd\u4f5c\n        self.fc = nn.Linear(n_head * d_v, d_model, bias=False)\n        # \u5355\u5934\u81ea\u6ce8\u610f\u529b\n        self.attention = ScaledDotProductAttention(temperature=d_k ** 0.5)\n        self.dropout = nn.Dropout(dropout)\n        # \u5c42\u5f52\u4e00\u5316\n        self.layer_norm = nn.LayerNorm(d_model, eps=1e-6)\n\n    def forward(self, q, k, v, mask=None):\n        # \u5047\u8bbeqkv\u8f93\u5165\u662f(b,100,512),100\u662f\u8bad\u7ec3\u6bcf\u4e2a\u6837\u672c\u6700\u5927\u5355\u8bcd\u4e2a\u6570\n        # \u4e00\u822cqkv\u76f8\u7b49\uff0c\u5373\u81ea\u6ce8\u610f\u529b\n        residual = q\n        # \u5c06\u8f93\u5165x\u548c\u53ef\u5b66\u4e60\u77e9\u9635\u76f8\u4e58\uff0c\u5f97\u5230(b,100,512)\u8f93\u51fa\n        # \u5176\u4e2d512\u7684\u542b\u4e49\u5176\u5b9e\u662f8x64\uff0c8\u4e2ahead\uff0c\u6bcf\u4e2ahead\u7684\u53ef\u5b66\u4e60\u77e9\u9635\u4e3a64\u7ef4\u5ea6\n        # q\u7684\u8f93\u51fa\u662f(b,100,8,64),kv\u4e5f\u662f\u4e00\u6837\n        q = self.w_qs(q).view(sz_b, len_q, n_head, d_k)\n        k = self.w_ks(k).view(sz_b, len_k, n_head, d_k)\n        v = self.w_vs(v).view(sz_b, len_v, n_head, d_v)\n\n        # \u53d8\u6210(b,8,100,64)\uff0c\u65b9\u4fbf\u540e\u9762\u8ba1\u7b97\uff0c\u4e5f\u5c31\u662f8\u4e2a\u5934\u5355\u72ec\u8ba1\u7b97\n        q, k, v = q.transpose(1, 2), k.transpose(1, 2), v.transpose(1, 2)\n\n        if mask is not None:\n            mask = mask.unsqueeze(1)   # For head axis broadcasting.\n        # \u8f93\u51faq\u662f(b,8,100,64),\u7ef4\u6301\u4e0d\u53d8,\u5185\u90e8\u8ba1\u7b97\u6d41\u7a0b\u662f\uff1a\n        # q*k\u8f6c\u7f6e\uff0c\u9664\u4ee5d_k ** 0.5\uff0c\u8f93\u51fa\u7ef4\u5ea6\u662fb,8,100,100\u5373\u5355\u8bcd\u548c\u5355\u8bcd\u76f4\u63a5\u7684\u76f8\u4f3c\u6027\n        # \u5bf9\u6700\u540e\u4e00\u4e2a\u7ef4\u5ea6\u8fdb\u884csoftmax\u64cd\u4f5c\u5f97\u5230b,8,100,100\n        # \u6700\u540e\u4e58\u4e0aV\uff0c\u5f97\u5230b,8,100,64\u8f93\u51fa\n        q, attn = self.attention(q, k, v, mask=mask)\n\n        # b,100,8,64-->b,100,512\n        q = q.transpose(1, 2).contiguous().view(sz_b, len_q, -1)\n        q = self.dropout(self.fc(q))\n        # \u6b8b\u5dee\u8ba1\u7b97\n        q += residual\n        # \u5c42\u5f52\u4e00\u5316\uff0c\u5728512\u7ef4\u5ea6\u8ba1\u7b97\u5747\u503c\u548c\u65b9\u5dee\uff0c\u8fdb\u884c\u5c42\u5f52\u4e00\u5316\n        q = self.layer_norm(q)\n\n        return q, attn<\/code><\/pre>\n\n\n\n
    4\u3001\u524d\u9988\u7f51\u7edc<\/h3>\n\n\n\n
    \u8fd9\u4e2a\u5c42\u5c31\u6ca1\u5565\u8bf4\u7684\u4e86\uff0c\u975e\u5e38\u7b80\u5355\uff0c\u76f4\u63a5\u770b\u4ee3\u7801\u5427\uff1a<\/p>\n\n\n\n
    class PositionwiseFeedForward(nn.Module):\n    ''' A two-feed-forward-layer module '''\n\n    def __init__(self, d_in, d_hid, dropout=0.1):\n        super().__init__()\n        # \u4e24\u4e2afc\u5c42\uff0c\u5bf9\u6700\u540e\u7684512\u7ef4\u5ea6\u8fdb\u884c\u53d8\u6362\n        self.w_1 = nn.Linear(d_in, d_hid) # position-wise\n        self.w_2 = nn.Linear(d_hid, d_in) # position-wise\n        self.layer_norm = nn.LayerNorm(d_in, eps=1e-6)\n        self.dropout = nn.Dropout(dropout)\n\n    def forward(self, x):\n        residual = x\n\n        x = self.w_2(F.relu(self.w_1(x)))\n        x = self.dropout(x)\n        x += residual\n\n        x = self.layer_norm(x)\n\n        return x<\/code><\/pre>\n\n\n\n
    \u6700\u540e\uff0c\u56de\u987e\u4e0b \ud835\udc61\ud835\udc5f\ud835\udc4e\ud835\udc5b\ud835\udc60\ud835\udc53\ud835\udc5c\ud835\udc5f\ud835\udc5a\ud835\udc52\ud835\udc5f \ud835\udc52\ud835\udc5b\ud835\udc50\ud835\udc5c\ud835\udc51\ud835\udc52\ud835\udc5f<\/code> \u7684\u6574\u4f53\u7ed3\u6784\u3002<\/p>\n\n\n\n
    \u7ecf\u8fc7\u4e0a\u6587\u7684\u68b3\u7406\uff0c\u6211\u4eec\u5df2\u7ecf\u57fa\u672c\u4e86\u89e3\u4e86 \ud835\udc61\ud835\udc5f\ud835\udc4e\ud835\udc5b\ud835\udc60\ud835\udc53\ud835\udc5c\ud835\udc5f\ud835\udc5a\ud835\udc52\ud835\udc5f<\/code> \u7f16\u7801\u5668\u7684\u4e3b\u8981\u6784\u6210\u90e8\u5206\uff0c\u6211\u4eec\u4e0b\u9762\u7528\u516c\u5f0f\u628a\u4e00\u4e2a \ud835\udc61\ud835\udc5f\ud835\udc4e\ud835\udc5b\ud835\udc60\ud835\udc53\ud835\udc5c\ud835\udc5f\ud835\udc5a\ud835\udc52\ud835\udc5f \ud835\udc4f\ud835\udc59\ud835\udc5c\ud835\udc50\ud835\udc58<\/code> \u7684\u8ba1\u7b97\u8fc7\u7a0b\u6574\u7406\u4e00\u4e0b\uff1a<\/p>\n\n\n\n
    1) \u5b57\u5411\u91cf\u4e0e\u4f4d\u7f6e\u7f16\u7801
    $$X= EmbeddingLookup (X)+ PositionalEncoding$$  <\/p>\n\n\n\n
     $$X \\in \\mathbb{R}^{\\text {batch size } * \\text { seq. len. * embed.dim. }}$$
    2) \u81ea\u6ce8\u610f\u529b\u673a\u5236<\/p>\n\n\n\n
    $$
    Q=\\operatorname{Linear}(X)=X W_{Q}
    $$<\/p>\n\n\n\n
    $$
    K=\\operatorname{Linear}(X)=X W_{K}
    $$<\/p>\n\n\n\n
    $$
    V=\\operatorname{Linear}(X)=X W_{V}
    $$<\/p>\n\n\n\n
    $$
    X_{\\text {attention }}=\\text { SelfAttention }(Q, K, V)
    $$<\/p>\n\n\n\n

    3) \u6b8b\u5dee\u8fde\u63a5\u4e0e\u5c42\u5f52\u4e00\u5316<\/p>\n\n\n\n
    $$
    X_{\\text {attention }}=X+X_{\\text {attention }}
    $$<\/p>\n\n\n\n
    $$
    X_{\\text {attention }}=\\text { Layer Norm }\\left(X_{\\text {attention }}\\right)
    $$<\/p>\n\n\n\n
    4) \u524d\u5411\u7f51\u7edc
    \u5176\u5b9e\u5c31\u662f\u4e24\u5c42\u7ebf\u6027\u6620\u5c04\u5e76\u7528\u6fc0\u6d3b\u51fd\u6570\u6fc0\u6d3b\uff0c\u6bd4\u5982\u8bf4 ReLU :
    $$
    X_{\\text {hidden }}=\\operatorname{Activate}\\left(\\text { Linear }\\left(\\text { Linear }\\left(X_{\\text {attention }}\\right)\\right)\\right)
    $$
    5) \u91cd\u590d3)<\/p>\n\n\n\n
    $$
    X_{\\text {hidden }}=X_{\\text {attention }}+X_{\\text {hidden }}
    $$<\/p>\n\n\n\n
    $$
     X_{\\text {hidden }}=\\text { Layer } N \\text { orm }\\left(X_{\\text {hidden }}\\right) 
    $$<\/p>\n\n\n\n
    $$
      X_{\\text {hidden }} \\in \\mathbb{R}^{\\text {batch size } * \\text { seq. len. } * \\text { embed. dim. }} 
    $$<\/p>\n","protected":false},"excerpt":{"rendered":"
    Transformer \u6700\u521d\u4e3b\u8981\u5e94\u7528\u4e8e\u4e00\u4e9b\u81ea\u7136\u8bed\u8a00\u5904\u7406\u573a\u666f\uff0c\u6bd4\u5982\u7ffb\u8bd1\u3001\u6587\u672c\u5206\u7c7b\u3001\u5199\u5c0f\u8bf4\u3001\u5199\u6b4c\u7b49\u3002\u968f\u7740\u6280\u672f\u7684\u53d1\u5c55 … \u7ee7\u7eed\u9605\u8bfbTransformer<\/span><\/a><\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":[],"categories":[4,12],"tags":[],"_links":{"self":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/2009"}],"collection":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/comments?post=2009"}],"version-history":[{"count":10,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/2009\/revisions"}],"predecessor-version":[{"id":13806,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/2009\/revisions\/13806"}],"wp:attachment":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/media?parent=2009"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/categories?post=2009"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/tags?post=2009"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}