{"id":1205,"date":"2025-07-20T01:36:24","date_gmt":"2025-07-19T17:36:24","guid":{"rendered":"https:\/\/lingbo.online\/?p=1205"},"modified":"2025-07-21T00:37:14","modified_gmt":"2025-07-20T16:37:14","slug":"%e3%80%8a%e5%a4%a7%e8%a7%84%e6%a8%a1%e8%af%ad%e8%a8%80%e6%a8%a1%e5%9e%8b%e7%90%86%e8%ae%ba%e4%b8%8e%e5%ae%9e%e8%b7%b5%e3%80%8b%e8%af%bb%e5%90%8e%e6%84%9f","status":"publish","type":"post","link":"https:\/\/lingbo.online\/index.php\/uncategorized\/%e3%80%8a%e5%a4%a7%e8%a7%84%e6%a8%a1%e8%af%ad%e8%a8%80%e6%a8%a1%e5%9e%8b%e7%90%86%e8%ae%ba%e4%b8%8e%e5%ae%9e%e8%b7%b5%e3%80%8b%e8%af%bb%e5%90%8e%e6%84%9f\/","title":{"rendered":"\u300a\u5927\u89c4\u6a21\u8bed\u8a00\u6a21\u578b\u7406\u8bba\u4e0e\u5b9e\u8df5\u300b\u8bfb\u540e\u611f"},"content":{"rendered":"<h3>\u5f15\u8a00<\/h3>\n<h4>\u8bed\u8a00\u6a21\u578b\u7684\u53d1\u5c55\u5386\u7a0b<\/h4>\n<ul>\n<li>\u7edf\u8ba1\u8bed\u8a00\u6a21\u578b\uff08Statistical Language Model,SLM\uff09\u3002\u7edf\u8ba1\u8bed\u8a00\u6a21\u578b\u4f7f\u7528\u9a6c\u5c14\u53ef\u592b\u5047\u8bbe\u6765\u5efa\u7acb\u8bed\u8a00\u5e8f\u5217\u7684\u9884\u6d4b\u6a21\u578b\uff0c\u901a\u5e38\u6839\u636e\u8bcd\u5e8f\u5217\u4e2d\u82e5\u5e72\u4e2a\u8fde\u7eed\u7684\u4e0a\u4e0b\u6587\u5355\u8bcd\u6765\u9884\u6d4b\u4e0b\u4e00\u4e2a\u8bcd\u7684\u51fa\u73b0\u6982\u7387\uff0c\u5373\u6839\u636e\u4e00\u4e2a\u56fa\u5b9a\u957f\u5ea6\u7684\u524d\u7f00\u6765\u9884\u6d4b\u76ee\u6807\u5355\u8bcd\u3002\u5177\u6709\u56fa\u5b9a\u4e0a\u4e0b\u6587\u957f\u5ea6<span class=\"katex-eq\" data-katex-display=\"false\">n<\/span>\u7684\u7edf\u8ba1\u8bed\u8a00\u6a21\u578b\u901a\u5e38\u88ab\u79f0\u4e3a<span class=\"katex-eq\" data-katex-display=\"false\">n<\/span>\u5143\uff08n-gram\uff09\u8bed\u8a00\u6a21\u578b<\/li>\n<\/ul>\n<h3>\u5927\u8bed\u8a00\u6a21\u578b\u57fa\u7840<\/h3>\n<h4>Transformer\u7ed3\u6784<\/h4>\n<p>Transformer\u7ed3\u6784\u662f\u8c37\u6b4c\u57282017\u5e74\u63d0\u51fa\u5e76\u9996\u5148\u56e0\u679c\u5e72\u9884\u673a\u5668\u7ffb\u8bd1\u7684\u795e\u7ecf\u7f51\u7edc\u6a21\u578b\u67b6\u6784\u3002\u673a\u5668\u7ffb\u8bd1\u7684\u76ee\u6807\u662f\u4ece\u6e90\u8bed\u8a00\uff08Source Language\uff09\u8f6c\u6362\u5230\u76ee\u6807\u8bed\u8a00\uff08Target Language\uff09\u3002Transformer\u7ed3\u6784\u5b8c\u5168\u901a\u8fc7\u6ce8\u610f\u529b\u673a\u5236\u5b8c\u6210\u5bf9\u6e90\u8bed\u8a00\u5e8f\u5217\u548c\u76ee\u6807\u8bed\u8a00\u5e8f\u5217\u5168\u5c40\u4f9d\u8d56\u7684\u5efa\u6a21\u3002<br \/>\n\u57fa\u4e8eTransformer\u7684\u7f16\u7801\u5668\u548c\u89e3\u7801\u5668\u7ed3\u6784\u5982\u4e0b\u56fe\u6240\u793a\uff0c\u5de6\u4fa7\u548c\u53f3\u4fa7\u5206\u522b\u5bf9\u5e94\u7740\u7f16\u7801\u5668\uff08Encoder\uff09\u548c\u89e3\u7801\u5668\uff08Decoder\uff09\u7ed3\u6784\uff0c\u4ed6\u4eec\u5747\u7531\u82e5\u5e72\u4e2a\u57fa\u672c\u7684Transformer\u5757\uff08Block\uff09\u7ec4\u6210\u3002\u5176\u4e2d<span class=\"katex-eq\" data-katex-display=\"false\">N\u00d7<\/span>\u8868\u793a\u8fdb\u884c\u4e86<span class=\"katex-eq\" data-katex-display=\"false\">N<\/span>\u6b21\u5806\u53e0\u3002\u6bcf\u4e2aTransformer\u5757\u90fd\u63a5\u6536\u4e00\u4e2a\u5411\u91cf\u5e8f\u5217<span class=\"katex-eq\" data-katex-display=\"false\">\\{x_i\\}_{i=1}^{t}<\/span>,\u5e76\u8f93\u51fa\u4e00\u4e2a\u7b49\u957f\u7684\u5411\u91cf\u5e8f\u5217\u4f5c\u4e3a\u8f93\u51fa<span class=\"katex-eq\" data-katex-display=\"false\">\\{y_i\\}_{i=1}^{t}<\/span>\u3002\u8fd9\u91cc\u7684<span class=\"katex-eq\" data-katex-display=\"false\">x_i<\/span>\u548c<span class=\"katex-eq\" data-katex-display=\"false\">y_i<\/span>\u5206\u522b\u5bf9\u5e94\u6587\u672c\u5e8f\u5217\u4e2d\u7684\u4e00\u4e2a\u8bcd\u5143\uff08Token\uff09\u7684\u8868\u793a\u3002<span class=\"katex-eq\" data-katex-display=\"false\">y_i<\/span>\u662f\u5f53\u524dTransformer\u5757\u5bf9\u8f93\u5165<span class=\"katex-eq\" data-katex-display=\"false\">x_i<\/span>\u8fdb\u4e00\u6b65\u6574\u5408\u5176\u4e0a\u4e0b\u6587\u8bed\u4e49\u540e\u5bf9\u5e94\u7684\u8f93\u51fa\u3002\u5728\u4ece\u8f93\u5165<span class=\"katex-eq\" data-katex-display=\"false\">\\{x_i\\}_{i=1}^{t}<\/span>\u5230\u8f93\u51fa<span class=\"katex-eq\" data-katex-display=\"false\">\\{y_i\\}_{i=1}^{t}<\/span>\u7684\u8bed\u4e49\u62bd\u8c61\u8fc7\u7a0b\u4e2d\uff0c\u4e3b\u8981\u6d89\u53ca\u5982\u4e0b\u51e0\u4e2a\u6a21\u5757\uff1a<\/p>\n<ul>\n<li><strong>\u6ce8\u610f\u529b\u5c42\uff1a<\/strong>\u4f7f\u7528\u591a\u5934\u6ce8\u610f\u529b\uff08Multi-Head Attention\uff09\u673a\u5236\u6574\u5408\u4e0a\u4e0b\u6587\u8bed\u4e49\u3002\u591a\u5934\u6ce8\u610f\u529b\u5e76\u884c\u8fd0\u884c\u591a\u4e2a\u72ec\u7acb\u6ce8\u610f\u529b\u673a\u5236\uff0c\u8fdb\u800c\u4ece\u591a\u7ef4\u5ea6\u6355\u6349\u5e8f\u5217\u4fe1\u606f\u3002\u4ed6\u4f7f\u5f97\u5e8f\u5217\u4e2d\u4efb\u610f\u4e24\u4e2a\u5355\u8bcd\u4e4b\u95f4\u7684\u4f9d\u8d56\u5173\u7cfb\u53ef\u4ee5\u76f4\u63a5\u88ab\u5efa\u6a21\u800c\u4e0d\u57fa\u4e8e\u4f20\u7edf\u7684\u5faa\u73af\u7ed3\u6784\uff0c\u4ece\u800c\u66f4\u597d\u5730\u89e3\u51b3\u6587\u672c\u7684\u957f\u7a0b\u4f9d\u8d56\u95ee\u9898\u3002<\/li>\n<li><strong>\u4f4d\u7f6e\u611f\u77e5\u524d\u9988\u7f51\u7edc\u5c42\uff08Position-wise Feed-Forward Network\uff09:<\/strong>\u901a\u8fc7\u5168\u8fde\u63a5\u5c42\u5bf9\u8f93\u5165\u6587\u672c\u5e8f\u5217\u4e2d\u7684\u6bcf\u4e2a\u5355\u8bcd\u8868\u793a\u8fdb\u884c\u66f4\u590d\u6742\u7684\u53d8\u6362\u3002<\/li>\n<li><strong>\u6b8b\u5dee\u8fde\u63a5\uff1a<\/strong>\u5bf9\u5e94\u56fe\u4e2d\u7684Add\u90e8\u5206\u3002\u5b83\u662f\u4e00\u6761\u5206\u522b\u4f5c\u7528\u5728\u4e0a\u8ff0\u4e24\u4e2a\u5b50\u5c42\u4e2d\u7684\u76f4\u8fde\u901a\u8def\uff0c\u88ab\u7528\u4e8e\u8fde\u63a5\u4e24\u4e2a\u5b50\u5c42\u7684\u8f93\u5165\u548c\u8f93\u51fa\uff0c\u4f7f\u4fe1\u606f\u6d41\u52a8\u66f4\u9ad8\u6548\uff0c\u6709\u5229\u4e8e\u6a21\u578b\u7684\u4f18\u5316\u3002<\/li>\n<li><strong>\u5c42\u5f52\u4e00\u5316\uff1a<\/strong>\u5bf9\u5e94\u56fe\u4e2d\u7684Norm\u90e8\u5206\u3002\u5b83\u4f5c\u7528\u4e8e\u4e0a\u8ff0\u4e24\u4e2a\u5b50\u5c42\u7684\u8f93\u51fa\u8868\u793a\u5e8f\u5217\uff0c\u5bf9\u8868\u793a\u5e8f\u5217\u8fdb\u884c\u5c42\u5f52\u4e00\u5316\u64cd\u4f5c\uff0c\u540c\u6837\u8d77\u5230\u7a33\u5b9a\u4f18\u5316\u7684\u4f5c\u7528\u3002<\/li>\n<\/ul>\n<p><img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2025\/07\/20250714132055833.jpg\" alt=\"\" \/><\/p>\n<h5>\u5d4c\u5165\u8868\u793a\u5c42<\/h5>\n<p>\u5bf9\u4e8e\u8f93\u5165\u7684\u6587\u672c\u5e8f\u5217\uff0c\u9996\u5148\u901a\u8fc7\u8f93\u5165\u5d4c\u5165\u5c42\uff08Input Embedding\uff09\u5c06\u6bcf\u4e2a\u5355\u8bcd\u8f6c\u6362\u4e3a\u5176\u76f8\u5bf9\u5e94\u7684\u5411\u91cf\u8868\u793a\u3002\u901a\u5e38\uff0c\u76f4\u63a5\u5bf9\u6bcf\u4e00\u4e2a\u5355\u8bcd\u521b\u5efa\u4e00\u4e2a\u5411\u91cf\u8868\u793a\u3002\u5728\u9001\u5165\u7f16\u7801\u5668\u521b\u5efa\u4e0a\u4e0b\u6587\u8bed\u4e49\u4e4b\u524d\uff0c\u9700\u8981\u5728\u8bcd\u5d4c\u5165\u4e2d\u52a0\u5165<strong>\u4f4d\u7f6e\u7f16\u7801<\/strong>\uff08Positional Encoding\uff09\u7279\u5f81\uff0c\u5373\u5e8f\u5217\u4e2d\u7684\u6bcf\u4e00\u4e2a\u8bcd\u6240\u5728\u7684\u4f4d\u7f6e\u90fd\u5bf9\u5e94\u4e00\u4e2a\u5411\u91cf\u3002\u8fd9\u4e2a\u5411\u91cf\u4f1a\u548c\u5355\u8bcd\u8868\u793a\u5bf9\u5e94\u76f8\u52a0\u5e76\u9001\u5165\u540e\u7eed\u6a21\u5757\u4e2d\u505a\u8fdb\u4e00\u6b65\u5904\u7406\u3002\u5728\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\uff0c\u6a21\u578b\u4f1a\u81ea\u52a8\u5b66\u4e60\u5982\u4f55\u5229\u7528\u4f4d\u7f6e\u4fe1\u606f\u3002<br \/>\nTransformer\u7ed3\u6784\u4f7f\u7528\u4e0d\u540c\u9891\u7387\u7684\u6b63\u4f59\u5f26\u51fd\u6570\u6765\u8868\u793a\u4e0d\u540c\u4f4d\u7f6e\u6240\u5bf9\u5e94\u7684\u7f16\u7801\u4fe1\u606f\uff1a<br \/>\n<img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2025\/07\/20250719171207729.jpg\" alt=\"\" \/><br \/>\n\u5176\u4e2d\uff0cpos\u8868\u793a\u5355\u8bcd\u6240\u5728\u7684\u4f4d\u7f6e\uff0c<span class=\"katex-eq\" data-katex-display=\"false\">2i<\/span>\u548c<span class=\"katex-eq\" data-katex-display=\"false\">2i+1<\/span>\u8868\u793a\u4f4d\u7f6e\u7f16\u7801\u5411\u91cf\u4e2d\u7684\u5bf9\u5e94\u7ef4\u5ea6\uff0c<span class=\"katex-eq\" data-katex-display=\"false\">d<\/span>\u5219\u5bf9\u5e94\u4f4d\u7f6e\u7f16\u7801\u7684\u603b\u7ef4\u5ea6\u3002\u901a\u8fc7\u8fd9\u79cd\u65b9\u5f0f\u8ba1\u7b97\u4f4d\u7f6e\u7f16\u7801\u7684\u597d\u5904\u6709\uff1a\uff081\uff09\u6b63\u4f59\u5f26\u51fd\u6570\u7684\u8303\u56f4\u662f[-1,+1]\uff0c\u5bfc\u51fa\u7684\u4f4d\u7f6e\u7f16\u7801\u4e0e\u539f\u6709\u8bcd\u5411\u91cf\u76f8\u52a0\u4e0d\u4f1a\u4f7f\u7ed3\u679c\u504f\u79bb\u8fc7\u8fdc\u800c\u7834\u574f\u539f\u6709\u5355\u8bcd\u7684\u8bed\u4e49\u4fe1\u606f\uff1b\uff082\uff09\u6839\u636e\u4e09\u89d2\u51fd\u6570\u7684\u57fa\u672c\u6027\u8d28\uff0c\u53ef\u4ee5\u5f97\u77e5\u7b2c<span class=\"katex-eq\" data-katex-display=\"false\">pos+k<\/span>\u4e2a\u4f4d\u7f6e\u7f16\u7801\u662f\u7b2c<span class=\"katex-eq\" data-katex-display=\"false\">pos<\/span>\u4e2a\u4f4d\u7f6e\u7f16\u7801\u7684\u7ebf\u6027\u7ec4\u5408\uff0c\u8fd9\u610f\u5473\u7740\u4f4d\u7f6e\u7f16\u7801\u4e2d\u8574\u542b\u7740\u5355\u8bcd\u4e4b\u95f4\u7684\u8ddd\u79bb\u4fe1\u606f\u3002<br \/>\n\u53ef\u4ee5\u4f7f\u7528pytorch\u5b8c\u6210\u4f4d\u7f6e\u7f16\u7801\uff1a<\/p>\n<pre><code class=\"language-python\">class PositionalEncoder(nn.Module):\n    def __init__(self, d_model, max_seq_len = 80):\n        super().__init__()\n        self.d_model = d_model\n        # \u6839\u636epos\u548ci\u521b\u5efa\u4e00\u4e2a\u5e38\u91cfPE\u77e9\u9635\n        pe = torch.zeros(max_seq_len, d_model)\n        for pos in range(max_seq_len):\n            for i in range(0, d_model, 2):\n                pe[pos, i] = math.sin(pos \/ (10000 ** (i\/d_model)))\n                pe[pos, i + 1] = math.cos(pos \/ (10000 ** (i\/d_model)))\n        pe = pe.unsqueeze(0)\n        self.register_buffer(&#039;pe&#039;, pe)\n\n        def forward(self, x):\n        # \u4f7f\u5f97\u5355\u8bcd\u5d4c\u5165\u8868\u793a\u76f8\u5bf9\u5927\u4e00\u4e9b\n        x = x * math.sqrt(self.d_model)\n        # \u589e\u52a0\u4f4d\u7f6e\u5e38\u91cf\u5230\u5355\u8bcd\u5d4c\u5165\u8868\u793a\u4e2d\n        seq_len = x.size(1)\n        x = x + Variable(self.pe[:,:seq_len], requires_grad=False).cuda()\n        return x<\/code><\/pre>\n<h5>\u6ce8\u610f\u529b\u5c42<\/h5>\n<p><strong>\u81ea\u6ce8\u610f\u529b<\/strong>\uff08Self-Attention\uff09\u64cd\u4f5c\u662f\u57fa\u4e8eTransformer\u7684\u673a\u5668\u7ffb\u8bd1\u6a21\u578b\u7684\u57fa\u672c\u64cd\u4f5c\uff0c\u5728\u6e90\u8bed\u8a00\u7684\u7f16\u7801\u548c\u76ee\u6807\u8bed\u8a00\u7684\u751f\u6210\u4e2d\u9891\u7e41\u7684\u88ab\u4f7f\u7528\uff0c\u4ee5\u5efa\u6a21\u6e90\u8bed\u8a00\u3001\u76ee\u6807\u8bed\u8a00\u4efb\u610f\u4e24\u4e2a\u5355\u8bcd\u4e4b\u95f4\u7684\u4f9d\u8d56\u5173\u7cfb\u3002\u5c06\u7531\u5355\u8bcd\u8bed\u4e49\u5d4c\u5165\u53ca\u5176\u4f4d\u7f6e\u7f16\u7801\u53e0\u52a0\u5f97\u5230\u8f93\u5165\u8868\u793a\u4e3a<span class=\"katex-eq\" data-katex-display=\"false\">\\{x_i \\in \\mathbb{R}^{d}\\}_{i=1}^{L}<\/span>\uff0c\u4e3a\u4e86\u5b9e\u73b0\u5bf9\u4e0a\u4e0b\u6587\u8bed\u4e49\u4f9d\u8d56\u7684\u5efa\u6a21\uff0c\u5f15\u5165\u81ea\u6ce8\u610f\u529b\u673a\u5236\u6d89\u53ca\u7684\u4e09\u4e2a\u5143\u7d20\uff1a\u67e5\u8be2<span class=\"katex-eq\" data-katex-display=\"false\">q_i<\/span>\uff08Query\uff09\u3001\u952e<span class=\"katex-eq\" data-katex-display=\"false\">k_i<\/span>(Key)\u548c\u503c<span class=\"katex-eq\" data-katex-display=\"false\">v_i<\/span>\uff08Value\uff09\u3002\u5728\u7f16\u7801\u8f93\u5165\u5e8f\u5217\u7684\u6bcf\u4e00\u4e2a\u5355\u8bcd\u7684\u8868\u793a\u4e2d\uff0c\u8fd9\u4e09\u4e2a\u5143\u7d20\u7528\u4e8e\u8ba1\u7b97\u4e0a\u4e0b\u6587\u5355\u8bcd\u5bf9\u5e94\u7684\u6743\u91cd\u5f97\u5206\u3002\u76f4\u89c2\u5730\u8bf4\uff0c\u8fd9\u4e9b\u6743\u91cd\u53cd\u6620\u4e86\u5728\u7f16\u7801\u5f53\u524d\u5355\u8bcd\u7684\u8868\u793a\u65f6\uff0c\u5bf9\u4e8e\u4e0a\u4e0b\u6587\u4e0d\u540c\u90e8\u5206\u6240\u9700\u8981\u7684\u5173\u6ce8\u7a0b\u5ea6\u3002\u5177\u4f53\u6765\u8bf4\uff0c\u5982\u4e0b\u56fe\u6240\u793a\uff0c\u901a\u8fc7\u4e09\u4e2a\u7ebf\u6027\u53d8\u5316<span class=\"katex-eq\" data-katex-display=\"false\">\\mathbf{W^{Q}} \\in \\mathbb{R}^{d\u00d7d_q}  <\/span>\uff0c<span class=\"katex-eq\" data-katex-display=\"false\">\\mathbf{W^{K}} \\in \\mathbb{R}^{d\u00d7d_k}<\/span>,<span class=\"katex-eq\" data-katex-display=\"false\">\\mathbf{W^{V}} \\in \\mathbb{R}^{d\u00d7d_v}<\/span>\uff0c\u5c06\u8f93\u5165\u5e8f\u5217\u4e2d\u7684\u6bcf\u4e00\u4e2a\u5355\u8bcd<span class=\"katex-eq\" data-katex-display=\"false\">x_i<\/span>\u8f6c\u6362\u4e3a\u5176\u5bf9\u5e94\u7684<span class=\"katex-eq\" data-katex-display=\"false\">q_i \\in \\mathbb{R^{d_q}}<\/span>,<span class=\"katex-eq\" data-katex-display=\"false\">k_i \\in \\mathbb{R^{d_k}}<\/span>,<span class=\"katex-eq\" data-katex-display=\"false\">v_i \\in \\mathbb{R^{d_v}}<\/span>\u5411\u91cf\u3002\u5bf9\u4e8e\u8f93\u5165<span class=\"katex-eq\" data-katex-display=\"false\">\\{x_i \\in \\mathbb{R^{d}}\\}^{L}_{i=1}<\/span>,<span class=\"katex-eq\" data-katex-display=\"false\">\\mathbf{Q}<\/span>\u3001<span class=\"katex-eq\" data-katex-display=\"false\">\\mathbf{K}<\/span>\u548c<span class=\"katex-eq\" data-katex-display=\"false\">\\mathbf{V}<\/span>\u77e9\u9635\u53ef\u4ee5\u901a\u8fc7\u5982\u4e0b\u516c\u5f0f\u6240\u793a\uff1a<br \/>\n<img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2025\/07\/20250719175224877.jpg\" alt=\"\" \/><br \/>\n<img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2025\/07\/20250719175314541.jpg\" alt=\"\" \/><br \/>\n\u4e3a\u4e86\u5f97\u5230\u7f16\u7801\u5355\u8bcd<span class=\"katex-eq\" data-katex-display=\"false\">x_i<\/span>\u65f6\u6240\u9700\u8981\u5173\u6ce8\u7684\u4e0a\u4e0b\u6587\u4fe1\u606f\uff0c\u901a\u8fc7\u4f4d\u7f6e<span class=\"katex-eq\" data-katex-display=\"false\">i<\/span>\u67e5\u8be2\u5411\u91cf\u4e0e\u5176\u4ed6\u4f4d\u7f6e\u7684\u952e\u5411\u91cf\u505a\u70b9\u79ef\u5f97\u5230\u5339\u914d\u5206\u6570<span class=\"katex-eq\" data-katex-display=\"false\">q_i \\cdot k_1,q_i \\cdot k_2,\\dots, q_i \\cdot k_t<\/span>\u3002\u4e3a\u4e86\u9632\u6b62\u8fc7\u5927\u7684\u5339\u914d\u5206\u6570\u5728\u540e\u7eedSoftmax\u8ba1\u7b97\u8fc7\u7a0b\u4e2d\u5bfc\u81f4\u68af\u5ea6\u7206\u70b8\u53ca\u6536\u655b\u6548\u7387\u5dee\u7684\u95ee\u9898\uff0c\u8fd9\u4e9b\u5f97\u5206\u4f1a\u9664\u4ee5\u7f29\u653e\u56e0\u5b50<span class=\"katex-eq\" data-katex-display=\"false\">\\sqrt{d}<\/span>\u4ee5\u7a33\u5b9a\u4f18\u5316\u3002\u7f29\u653e\u540e\u7684\u5f97\u5206\u7ecf\u8fc7Softmax\u5f52\u4e00\u5316\u6982\u7387\uff0c\u4e0e\u5176\u4ed6\u4f4d\u7f6e\u7684\u503c\u5411\u91cf\u76f8\u4e58\u6765\u805a\u5408\u5e0c\u671b\u5173\u6ce8\u7684\u4e0a\u4e0b\u6587\u4fe1\u606f\uff0c\u5e76\u6700\u5c0f\u5316\u4e0d\u76f8\u5173\u4fe1\u606f\u7684\u5e72\u6270\u3002\u4e0a\u8ff0\u8ba1\u7b97\u8fc7\u7a0b\u53ef\u4ee5\u88ab\u5f62\u5f0f\u5316\u7684\u8868\u793a\u4e3a\uff1a<br \/>\n<img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2025\/07\/20250719175845282.jpg\" alt=\"\" \/><br \/>\n\u5176\u4e2d\uff0c<span class=\"katex-eq\" data-katex-display=\"false\">\\mathbf{Q} \\in \\mathbb{R}^{L\u00d7d_q}\uff0c\\mathbf{K} \\in \\mathbb{R}^{L\u00d7d_k},\\mathbf{V} \\in \\mathbb{R}^{L\u00d7d_v}<\/span>\u5206\u522b\u8868\u793a\u8f93\u5165\u5e8f\u5217\u4e2d\u4e0d\u540c\u5355\u8bcd\u7684<span class=\"katex-eq\" data-katex-display=\"false\">q,k,v<\/span>\u5411\u91cf\u62fc\u63a5\u7ec4\u6210\u7684\u77e9\u9635\uff0c<span class=\"katex-eq\" data-katex-display=\"false\">L<\/span>\u8868\u793a\u5e8f\u5217\u957f\u5ea6\uff0c<span class=\"katex-eq\" data-katex-display=\"false\">\\mathbf{Z} \\in \\mathbb{R}^{L\u00d7d_v}<\/span>\u8868\u793a\u81ea\u6ce8\u610f\u529b\u64cd\u4f5c\u7684\u8f93\u51fa\u3002\u4e3a\u4e86\u8fdb\u4e00\u6b65\u589e\u5f3a\u81ea\u6ce8\u610f\u529b\u673a\u5236\u805a\u5408\u4e0a\u4e0b\u6587\u4fe1\u606f\u7684\u80fd\u529b\uff0c\u63d0\u51fa\u4e86\u591a\u5934\u6ce8\u610f\u529b\u673a\u5236\uff0c\u4ee5\u5173\u6ce8\u4e0a\u4e0b\u6587\u7684\u4e0d\u540c\u4fa7\u9762\u3002\u5177\u4f53\u6765\u8bf4\uff0c\u4e0a\u4e0b\u6587\u4e2d\u7684\u6bcf\u4e00\u4e2a\u5355\u8bcd\u8868\u793a<span class=\"katex-eq\" data-katex-display=\"false\">x_i<\/span>\u7ecf\u8fc7\u591a\u7ec4\u7ebf\u6027<span class=\"katex-eq\" data-katex-display=\"false\">\\{\\mathbf{W^{Q}_{j}},\\mathbf{W^{K}_{j}},\\mathbf{W^{V}_{j}}\\}^{N}_{j=1}<\/span>\u6620\u5c04\u5230\u4e0d\u540c\u7684\u8868\u793a\u5b50\u7a7a\u95f4\u4e2d\uff0c\u4e0b\u8ff0\u516c\u5f0f\u4f1a\u5728\u4e0d\u540c\u7684\u5b50\u7a7a\u95f4\u5206\u522b\u8ba1\u7b97\u5f97\u5230\u4e0d\u540c\u7684\u4e0a\u4e0b\u6587\u76f8\u5173\u7684\u5355\u8bcd\u5e8f\u5217\u8868\u793a<span class=\"katex-eq\" data-katex-display=\"false\">\\{\\mathbf{Z}_j\\}_{j=1}^{N}<\/span>:<br \/>\n<img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2025\/07\/20250719181119236.jpg\" alt=\"\" \/><br \/>\n\u5728\u6b64\u57fa\u7840\u4e0a\uff0c\u7ecf\u8fc7\u7ebf\u6027\u53d8\u5316<span class=\"katex-eq\" data-katex-display=\"false\">\\mathbf{W^{O}} \\in \\mathbb{R}^{(Nd_v)\u00d7d}<\/span>\u7528\u4e8e\u7efc\u5408\u4e0d\u540c\u5b50\u7a7a\u95f4\u4e2d\u7684\u4e0a\u4e0b\u6587\u8868\u793a\u5e76\u5f62\u6210\u6ce8\u610f\u529b\u5c42\u6700\u7ec8\u7684\u8f93\u51fa<span class=\"katex-eq\" data-katex-display=\"false\">\\{x_i \\in \\mathbb{R^{d}}\\}^{L}_{i=1}<\/span>\uff0c\u53ef\u4ee5\u5f97\u5230<strong>\u591a\u5934\u81ea\u6ce8\u610f\u529b<\/strong>\uff08Multi-Head Self-Attention\uff09\u8868\u793a\uff1a<br \/>\n<img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2025\/07\/20250719181455635.jpg\" alt=\"\" \/><br \/>\n\u7531\u6b64\u53ef\u89c1\uff0c\u81ea\u6ce8\u610f\u529b\u673a\u5236\u4f7f\u6a21\u578b\u80fd\u591f\u8bc6\u522b\u4e0d\u540c\u8f93\u5165\u90e8\u5206\u7684\u91cd\u8981\u6027\uff0c\u4ece\u800c\u4e0d\u53d7\u8ddd\u79bb\u7684\u5f71\u54cd\uff0c\u4ece\u800c\u80fd\u591f\u6355\u6349\u8f93\u5165\u53e5\u5b50\u4e2d\u7684\u957f\u8ddd\u79bb\u4f9d\u8d56\u5173\u7cfb\u548c\u590d\u6742\u5173\u7cfb\u3002<br \/>\n\u4f7f\u7528Pytorch\u5b9e\u73b0\u7684\u81ea\u6ce8\u610f\u529b\u5c42\u53c2\u8003\u4ee3\u7801\u5982\u4e0b\uff1a<\/p>\n<pre><code class=\"language-python\">import math\n\nclass MultiHeadAttention(nn.Moudle):\n    def __init__(self, heads, d_model, dropout=0.1):\n        super().__init__()\n\n        self.d_model = d_model\n        self.d_k = d_model \/\/ heads\n        self.h = heads\n\n        # \u7b80\u5386Q\u3001K\u3001V\u77e9\u9635\n        self.q_linear = nn.Linear(d_model, d_model)\n        self.k_linear = nn.Linear(d_model, d_model)\n        self.v_linear = nn.Linear(d_model, d_model)\n        self.dropout = nn.Dropout(dropout)\n        self.out = nn.Linear(d_model, d_model)\n\n    def attention(q, k, v, d_k, mask=None, dropout=None):\n        scores = torch.matmul(q, k.transpose(-2, -1)) \/ math.sqrt(d_k)\n\n        # \u63a9\u76d6\u90a3\u4e9b\u4e3a\u4e86\u8865\u5168\u957f\u5ea6\u800c\u589e\u52a0\u7684\u5355\u5143\uff0c\u4f7f\u5176\u901a\u8fc7Softmax\u8ba1\u7b97\u540e\u4e3a0\n        if mask is not None:\n            mask = mask.unsqueeze(1)\n            scores = scores.masked_fill(mask == 0, -1e9)\n\n        scores = F.softmax(scores, dim=-1)\n\n        # \u5bf9\u6ce8\u610f\u529b\u6743\u91cd\u505a\u526a\u679d\uff0c\u968f\u673a\u628a\u67d0\u4e9b\u6ce8\u610f\u529b\u6761\u76ee\u8bbe\u7f6e\u4e3a0\uff0c\u9632\u6b62co-adaptation\n        if dropout is not None:\n            scores = dropout(scores)\n\n        output = torch.matmul(scores, v)\n        return output\n\n    def forward(self,q,k,v,mask=None):\n        bs = q.size(0)\n\n        # \u8fdb\u884c\u7ebf\u6027\u64cd\u4f5c\uff0c\u5212\u5206\u4e3ah\u4e2a\u5934\n        k = self.k_linear(k).view(bs,-1,self.h,self.d_k)\n        q = self.q_linear(q).view(bs,-1,self.h,self.d_k)\n        v = self.v_linear(v).view(bs,-1,self.h,self.d_k)\n\n        # \u77e9\u9635\u8f6c\u7f6e\n        k = k.transpose(1,2)\n        q = q.transpose(1,2)\n        v = v.transpose(1,2)\n\n        # \u8ba1\u7b97 attention\n        scores = self.attention(q,k,v,self.d_k,mask,self.dropout)\n\n        # \u8fde\u63a5\u591a\u4e2a\u5934\u5e76\u8f93\u5165\u5230\u6700\u540e\u7684\u7ebf\u6027\u5c42\n        concat = scores.transpose(1,2).contiguous().view(bs,-1,self.d_model)\n\n        output = self.out(concat)\n\n        return output<\/code><\/pre>\n<h5>\u524d\u9988\u5c42<\/h5>\n<p>\u524d\u9988\u5c42\u63a5\u6536\u81ea\u6ce8\u610f\u5b50\u5c42\u7684\u8f93\u51fa\u4f5c\u4e3a\u8f93\u5165\uff0c\u5e76\u901a\u8fc7\u4e00\u4e2a\u5e26\u6709ReLU\u6fc0\u6d3b\u51fd\u6570\u7684\u4e24\u5c42\u5168\u8fde\u63a5\u7f51\u7edc\u5bf9\u8f93\u5165\u8fdb\u884c\u66f4\u590d\u6742\u7684\u975e\u7ebf\u6027\u53d8\u6362\u3002\u5b9e\u9a8c\u8bc1\u660e\uff0c\u8fd9\u4e00\u975e\u7ebf\u6027\u53d8\u6362\u4f1a\u5bf9\u6a21\u578b\u6700\u7ec8\u7684\u6027\u80fd\u4ea7\u751f\u91cd\u8981\u7684\u5f71\u54cd\u3002<br \/>\n<img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2025\/07\/20250720122034618.jpg\" alt=\"\" \/><br \/>\n\u5176\u4e2d\uff0c<span class=\"katex-eq\" data-katex-display=\"false\">\\mathbf{W}_1,\\mathbf{b}_1,\\mathbf{W}_2,\\mathbf{b}_2<\/span>\u8868\u793a\u524d\u9988\u5b50\u5c42\u7684\u53c2\u6570\u3002\u5b9e\u9a8c\u7ed3\u679c\u8868\u660e\uff0c\u589e\u5927\u524d\u9988\u5b50\u5c42\u9690\u72b6\u6001\u7684\u7ef4\u5ea6\u6709\u5229\u4e8e\u63d0\u9ad8\u5176\u6700\u7ec8\u7684\u7ffb\u8bd1\u7ed3\u679c\u8d28\u91cf\u3002\u56e0\u6b64\uff0c\u524d\u9988\u5b50\u5c42\u9690\u72b6\u6001\u7684\u7ef4\u5ea6\u4e00\u822c\u6bd4\u81ea\u6ce8\u610f\u529b\u5b50\u5c42\u8981\u5927\u3002<br \/>\n\u4f7f\u7528Pytorch\u5b9e\u73b0\u7684\u524d\u9988\u5c42\u53c2\u8003\u4ee3\u7801\u5982\u4e0b\uff1a<\/p>\n<pre><code class=\"language-python\">class FeedForward(nn.Module):\n    def __init__(self, d_model, d_ff=2048, dropout = 0.1):\n        super().__init__()\n        # d_ff\u9ed8\u8ba4\u8bbe\u7f6e\u4e3a2048\n        self.linear_1 = nn.Linear(d_model, d_ff)\n        self.dropout = nn.Dropout(dropout)\n        self.linear_2 = nn.Linear(d_ff, d_model)\n\n    def forward(self, x):\n        x = self.dropout(F.relu(self.linear_1(x)))\n        x = self.linear_2(x)\n        return x<\/code><\/pre>\n<h5>\u6b8b\u5dee\u8fde\u63a5\u4e0e\u5c42\u5f52\u4e00\u5316<\/h5>\n<p>\u7531Transformer\u7ed3\u6784\u7ec4\u6210\u7684\u7f51\u7edc\u7ed3\u6784\u901a\u5e38\u90fd\u975e\u5e38\u5e9e\u5927\u3002\u7f16\u7801\u5668\u548c\u89e3\u7801\u5668\u5747\u7531\u5f88\u591a\u5c42\u57fa\u672c\u7684Transformer\u5757\u7ec4\u6210\uff0c\u6bcf\u4e00\u5c42\u4e2d\u90fd\u5305\u542b\u590d\u6742\u7684\u975e\u7ebf\u6027\u6620\u5c04\uff0c\u8fd9\u5c31\u5bfc\u81f4\u6a21\u578b\u8bad\u7ec3\u6bd4\u8f83\u56f0\u96be\u3002\u56e0\u6b64\uff0c\u7814\u7a76\u4eba\u5458\u5728Transformer\u5757\u4e2d\u8fdb\u4e00\u6b65\u5f15\u5165\u4e86\u6b8b\u5dee\u94fe\u63a5\u548c\u5c42\u5f52\u4e00\u5316\u6280\u672f\uff0c\u4ee5\u8fdb\u4e00\u6b65\u63d0\u5347\u8bad\u7ec3\u7684\u7a33\u5b9a\u6027\u3002\u5177\u4f53\u6765\u8bf4\uff0c\u6b8b\u5dee\u8fde\u63a5\u4e3b\u8981\u662f\u6307\u4f7f\u7528\u4e00\u6761\u76f4\u8fde\u901a\u9053\u76f4\u63a5\u5bf9\u5bf9\u5e94\u5b50\u5c42\u7684\u8f93\u5165\u8fde\u63a5\u5230\u8f93\u51fa\uff0c\u907f\u514d\u5728\u4f18\u5316\u8fc7\u7a0b\u4e2d\u56e0\u7f51\u7edc\u8fc7\u6df1\u800c\u4ea7\u751f\u6f5c\u5728\u7684\u68af\u5ea6\u6d88\u5931\u95ee\u9898\uff1a<br \/>\n$$x^{l+1}=f(x^{l})+x^{l}$$<br \/>\n\u5176\u4e2d\uff0c<span class=\"katex-eq\" data-katex-display=\"false\">x^l<\/span>\u8868\u793a\u7b2c<span class=\"katex-eq\" data-katex-display=\"false\">l<\/span>\u5c42\u7684\u8f93\u5165\uff0c<span class=\"katex-eq\" data-katex-display=\"false\">f(\\cdot)<\/span>\u8868\u793a\u4e00\u4e2a\u6620\u5c04\u51fd\u6570\u3002\u6b64\u5916\uff0c\u4e3a\u4e86\u4f7f\u6bcf\u4e00\u5c42\u7684\u8f93\u5165\/\u8f93\u51fa\u7a33\u5b9a\u5728\u4e00\u4e2a\u5408\u7406\u7684\u8303\u56f4\u5185\uff0c\u5c42\u5f52\u4e00\u5316\u6280\u672f\u88ab\u8fdb\u4e00\u6b65\u5f15\u5165\u6bcf\u4e2aTransformer\u5757\u4e2d\uff1a<br \/>\n$$LN(x)=\\alpha \\cdot \\frac{x-\\mu}{\\sigma}+\\beta$$<br \/>\n\u5176\u4e2d<span class=\"katex-eq\" data-katex-display=\"false\">\\mu<\/span>\u548c<span class=\"katex-eq\" data-katex-display=\"false\">\\sigma<\/span>\u5206\u522b\u8868\u793a\u5747\u503c\u548c\u65b9\u5dee\uff0c\u7528\u4e8e\u5c06\u6570\u636e\u5e73\u79fb\u7f29\u653e\u5230\u5747\u503c\u4e3a0\uff0c\u65b9\u5dee\u4e3a1\u7684\u6807\u51c6\u5206\u5e03\uff0c<span class=\"katex-eq\" data-katex-display=\"false\">\\alpha<\/span>\u548c<span class=\"katex-eq\" data-katex-display=\"false\">\\beta<\/span>\u662f\u53ef\u5b66\u4e60\u7684\u53c2\u6570\u3002\u5c42\u5f52\u4e00\u5316\u6280\u672f\u53ef\u4ee5\u6709\u6548\u5730\u7f13\u89e3\u4f18\u5316\u8fc7\u7a0b\u4e2d\u6f5c\u5728\u7684\u4e0d\u7a33\u5b9a\u3001\u6536\u655b\u901f\u5ea6\u6162\u7b49\u95ee\u9898\u3002<br \/>\n\u4f7f\u7528Pytorch\u5b9e\u73b0\u5c42\u5f52\u4e00\u5316\u53c2\u8003\u4ee3\u7801\u5982\u4e0b:<\/p>\n<pre><code class=\"language-python\">class Norm(nn.Module):\n    def __init__(self, d_model, eps = 1e-6):\n        super().__init__()\n\n        self.size = d_model\n\n        # \u5c42\u5f52\u4e00\u5316\u5305\u542b\u4e24\u4e2a\u53ef\u4ee5\u5b66\u4e60\u7684\u53c2\u6570\n        self.alpha = nn.Parameter(torch.ones(self.size))\n        self.bias = nn.Parameter(torch.zeros(self.size))\n\n        self.eps = eps\n\n    def forward(self, x):\n        norm = self.alpha * (x - x.mean(dim=-1, keepdim=True))\n        \/ (x.std(dim=-1, keepdim=True) + self.eps) + self.bias\n        return norm<\/code><\/pre>\n<h5>\u7f16\u7801\u5668\u548c\u89e3\u7801\u5668\u7ed3\u6784<\/h5>\n<p>\u57fa\u4e8e\u4e0a\u8ff0\u6a21\u5757\uff0c\u6839\u636e\u4e0a\u56fe\u4e2d\u56fd\u5973\u7ed9\u51fa\u7684\u7f51\u7edc\u67b6\u6784\uff0c\u7f16\u7801\u5668\u7aef\u8f83\u5bb9\u6613\u5b9e\u73b0\u3002\u76f8\u6bd4\u4e8e\u7f16\u7801\u5668\u7aef\uff0c\u89e3\u7801\u5668\u7aef\u66f4\u590d\u6742\u3002\u5177\u4f53\u6765\u8bf4\uff0c\u89e3\u7801\u5668\u7684\u6bcf\u4e2aTransformer\u5757\u7684\u7b2c\u4e00\u4e2a\u81ea\u6ce8\u610f\u529b\u5b50\u5c42\u989d\u5916\u589e\u52a0\u4e86\u6ce8\u610f\u529b\u63a9\u7801\uff0c\u5bf9\u5e94\u56fe\u4e2d\u7684<strong>\u63a9\u7801\u591a\u5934\u6ce8\u610f\u529b<\/strong>\uff08Masked Multi-Head Attention\uff09\u90e8\u5206\u3002\u8fd9\u4e3b\u8981\u662f\u56e0\u4e3a\u5728\u7ffb\u8bd1\u7684\u8fc7\u7a0b\u4e2d\uff0c\u7f16\u7801\u5668\u7aef\u4e3b\u8981\u7528\u4e8e\u7f16\u7801\u6e90\u8bed\u8a00\u5e8f\u5217\u7684\u4fe1\u606f\uff0c\u800c\u8fd9\u4e2a\u5e8f\u5217\u662f\u5b8c\u5168\u5df2\u77e5\u7684\uff0c\u56e0\u800c\u7f16\u7801\u5668\u4ec5\u9700\u8981\u8003\u8651\u5982\u4f55\u878d\u5408\u4e0a\u4e0b\u6587\u8bed\u4e49\u4fe1\u606f\u3002\u89e3\u7801\u5668\u7aef\u5219\u8d1f\u8d23\u751f\u6210\u76ee\u6807\u8bed\u8a00\u5e8f\u5217\uff0c\u8fd9\u4e00\u751f\u6210\u8fc7\u7a0b\u662f\u81ea\u56de\u5f52\u7684\uff0c\u5373\u5bf9\u4e8e\u6bcf\u4e00\u4e2a\u5355\u8bcd\u7684\u751f\u6210\u8fc7\u7a0b\uff0c\u4ec5\u6709\u5f53\u524d\u5355\u8bcd\u4e4b\u524d\u7684\u76ee\u6807\u8bed\u8a00\u5e8f\u5217\u662f\u53ef\u4ee5\u88ab\u89c2\u6d4b\u7684\uff0c\u56e0\u6b64\u8fd9\u4e00\u989d\u5916\u589e\u52a0\u7684\u63a9\u7801\u662f\u7528\u6765\u63a9\u76d6\u540e\u7eed\u7684\u6587\u672c\u4fe1\u606f\u7684\uff0c\u4ee5\u9632\u6a21\u578b\u5728\u8bad\u7ec3\u9636\u6bb5\u76f4\u63a5\u770b\u5230\u540e\u7eed\u7684\u6587\u672c\u5e8f\u5217\uff0c\u8fdb\u800c\u65e0\u6cd5\u5f97\u5230\u6709\u6548\u7684\u8bad\u7ec3\u3002<br \/>\n\u6b64\u5916\uff0c\u89e3\u7801\u5668\u7aef\u989d\u5916\u589e\u52a0\u4e86\u4e00\u4e2a<strong>\u591a\u5934\u4ea4\u53c9\u6ce8\u610f\u529b<\/strong>\uff08Multi-Head Cross-Attention\uff09\u6a21\u5757\uff0c\u4f7f\u7528<strong>\u4ea4\u53c9\u6ce8\u610f\u529b<\/strong>\uff08Cross-Attention\uff09\u65b9\u6cd5\uff0c\u540c\u65f6\u63a5\u6536\u6765\u81ea\u7f16\u7801\u5668\u7aef\u7684\u8f93\u51fa\u548c\u5f53\u524d Transformer \u5757\u7684\u524d\u4e00\u4e2a\u63a9\u7801\u6ce8\u610f\u529b\u5c42\u7684\u8f93\u51fa\u3002\u67e5\u8be2\u662f\u901a\u8fc7\u89e3\u7801\u5668\u524d\u4e00\u5c42\u7684\u8f93\u51fa\u8fdb\u884c\u6295\u5f71\u7684\uff0c\u800c\u952e\u548c\u503c\u662f\u4f7f\u7528\u7f16\u7801\u5668\u7684\u8f93\u51fa\u8fdb\u884c\u6295\u5f71\u7684\u3002\u5b83\u7684\u4f5c\u7528\u662f\u5728\u7ffb\u8bd1\u7684\u8fc7\u7a0b\u4e2d\uff0c\u4e3a\u4e86\u751f\u6210\u5408\u7406\u7684\u76ee\u6807\u8bed\u8a00\u5e8f\u5217\uff0c\u89c2\u6d4b\u5f85\u7ffb\u8bd1\u7684\u6e90\u8bed\u8a00\u5e8f\u5217\u662f\u4ec0\u4e48\u3002\u57fa\u4e8e\u4e0a\u8ff0\u7f16\u7801\u5668\u548c\u89e3\u7801\u5668\u7ed3\u6784\uff0c\u5f85\u7ffb\u8bd1\u7684\u6e90\u8bed\u8a00\u6587\u672c\u7ecf\u8fc7\u7f16\u7801\u5668\u7aef\u7684\u6bcf\u4e2a Transformer\u5757\u5bf9\u5176\u4e0a\u4e0b\u6587\u8bed\u4e49\u8fdb\u884c\u5c42\u5c42\u62bd\u8c61\uff0c\u6700\u7ec8\u8f93\u51fa\u6bcf\u4e00\u4e2a\u6e90\u8bed\u8a00\u5355\u8bcd\u4e0a\u4e0b\u6587\u76f8\u5173\u7684\u8868\u793a\u3002\u89e3\u7801\u5668\u7aef\u4ee5\u81ea\u56de\u5f52\u7684\u65b9\u5f0f\u751f\u6210\u76ee\u6807\u8bed\u8a00\u6587\u672c\uff0c\u5373\u5728\u6bcf\u4e2a\u65f6\u95f4\u6b65<span class=\"katex-eq\" data-katex-display=\"false\">t<\/span>\uff0c\u6839\u636e\u7f16\u7801\u5668\u7aef\u8f93\u51fa\u7684\u6e90\u8bed\u8a00\u6587\u672c\u8868\u793a\uff0c\u4ee5\u53ca\u524d<span class=\"katex-eq\" data-katex-display=\"false\">t-1<\/span>\u4e2a\u65f6\u523b\u751f\u6210\u7684\u76ee\u6807\u8bed\u8a00\u6587\u672c\uff0c\u751f\u6210\u5f53\u524d\u65f6\u523b\u7684\u76ee\u6807\u8bed\u8a00\u5355\u8bcd\u3002<\/p>\n<h4>\u77e5\u8bc6\u8865\u5145<\/h4>\n<h5>Dropout<\/h5>\n<p>dropout\u6307\u5728\u8bad\u7ec3\u65f6\u968f\u673a\u5207\u9664\u4e00\u5b9a\u6bd4\u4f8b\u7684\u795e\u7ecf\u5143\uff0c\u8ba9\u5b83\u4eec\u7684\u8fd9\u6b21\u524d\u5411\u4f20\u64ad\u4e0d\u8d77\u4f5c\u7528\uff0c\u4ece\u800c\u903c\u7f51\u7edc\u4e0d\u8981\u8fc7\u62df\u5408\uff0c\u63d0\u9ad8\u6a21\u578b\u7684\u6cdb\u5316\u80fd\u529b\u3002\u5728\u6d4b\u8bd5\u9a8c\u8bc1\u65f6\u518d\u5168\u90e8\u6062\u590d\u3002<br \/>\n\u4f7f\u7528Dropout\u7684\u7406\u7531\u5982\u4e0b\uff1a<\/p>\n<ul>\n<li><strong>\u89e3\u51b3\u8fc7\u62df\u5408\u95ee\u9898<\/strong>\uff1a\u5f53\u6a21\u578b\u5728\u8bad\u7ec3\u96c6\u4e0a\u8868\u73b0\u5f88\u597d\uff0c\u4f46\u662f\u5728\u6d4b\u8bd5\u96c6\u4e0a\u8868\u73b0\u5f88\u5dee\uff0c\u8fd9\u8bf4\u660e\u6a21\u578b\u8bb0\u4f4f\u4e86\u8bad\u7ec3\u6570\u636e\u7684\u566a\u58f0\uff0c\u800c\u975e\u5b66\u4e60\u901a\u7528\u6a21\u5f0f\u3002<\/li>\n<li><strong>\u795e\u7ecf\u5143\u534f\u540c\u9002\u5e94\uff08Co-adaptation\uff09<\/strong>\uff1a\u4f20\u7edf\u795e\u7ecf\u7f51\u7edc\u4e2d\uff0c\u795e\u7ecf\u5143\u53ef\u80fd\u8fc7\u5ea6\u4f9d\u8d56\u5176\u4ed6\u7279\u5b9a\u7684\u795e\u7ecf\u5143\uff0c\u5bfc\u81f4\u6a21\u578b\u8106\u5f31\u3002Dropout\u901a\u8fc7\u968f\u673a\u4e22\u5f03\u795e\u7ecf\u5143\uff0c\u6253\u7834\u8fd9\u79cd\u4f9d\u8d56\uff0c\u8feb\u4f7f\u6bcf\u4e2a\u795e\u7ecf\u5143\u72ec\u7acb\u5b66\u4e60\u6709\u7528\u7684\u7279\u5f81\u3002<\/li>\n<\/ul>\n<h5>ReLU\u6fc0\u6d3b\u51fd\u6570<\/h5>\n<p>ReLU\u6fc0\u6d3b\u51fd\u6570\u7684\u8868\u8fbe\u5f0f\u4e3a\uff1a<span class=\"katex-eq\" data-katex-display=\"false\">f(x)=\\max(0,x)<\/span>\u3002ReLU\u51fd\u6570\u7684\u56fe\u5f62\u5f62\u72b6\u5448\u73b0\u4e3a\u5206\u6bb5\u7ebf\u6027\u51fd\u6570\uff0c\u5728\u8f93\u5165\u4e3a\u8d1f\u6570\u65f6\u8f93\u51fa\u4e3a0\uff0c\u8f93\u5165\u4e3a\u6b63\u6570\u65f6\u8f93\u51fa\u4e0e\u8f93\u5165\u6210\u6b63\u6bd4\uff08\u5373<span class=\"katex-eq\" data-katex-display=\"false\">y=x<\/span>\uff09<br \/>\n<img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2025\/07\/20250720123315688.jpg\" alt=\"\" \/><\/p>\n<p>\u4f7f\u7528ReLU\u7684\u4f18\u52bf\u5305\u62ec\uff1a<\/p>\n<ul>\n<li>\u7b80\u6d01\u4e14\u9ad8\u6548\u7684\u8ba1\u7b97:ReLU\u51fd\u6570\u7684\u8ba1\u7b97\u65b9\u5f0f\u7b80\u5355\u76f4\u63a5\uff0c\u65e0\u9700\u590d\u6742\u7684\u6307\u6570\u8fd0\u7b97\uff0c\u76f8\u6bd4\u4e8eSigmiod\u6216\u8005Tanh\u7b49\u6fc0\u6d3b\u51fd\u6570\uff0cReLU\u7684\u8ba1\u7b97\u901f\u5ea6\u66f4\u5feb\u3002<\/li>\n<li>\u89e3\u51b3\u68af\u5ea6\u6d88\u5931\u95ee\u9898\uff1a\u5728\u4f20\u7edf\u7684Sigmoid\u6216\u8005Tanh\u6fc0\u6d3b\u51fd\u6570\u4e2d\uff0c\u5f53\u8f93\u5165\u503c\u975e\u5e38\u5927\u6216\u8005\u975e\u5e38\u5c0f\u65f6\uff0c\u5bfc\u6570\uff08\u68af\u5ea6\uff09\u53d8\u5f97\u975e\u5e38\u5c0f\u3002\u8fd9\u79cd\u73b0\u8c61\u79f0\u4e4b\u4e3a\u68af\u5ea6\u6d88\u5931\uff0c\u4ed6\u4f1a\u4f7f\u5f97\u53cd\u5411\u4f20\u64ad\u65f6\u7684\u68af\u5ea6\u5728\u4f20\u9012\u8fc7\u7a0b\u4e2d\u9010\u5c42\u8870\u51cf\uff0c\u5bfc\u81f4\u7f51\u7edc\u8bad\u7ec3\u56f0\u96be\uff0c\u751a\u81f3\u65e0\u6cd5\u66f4\u65b0\u53c2\u6570\u3002\u800cReLU\u7684\u5bfc\u6570\u5728\u6b63\u533a\u95f4\u4e3a\u5e38\u65701\uff0c\u8d1f\u533a\u95f4\u4e3a0\uff0c\u51e0\u4e4e\u4e0d\u53d7\u5230\u8f93\u5165\u503c\u5927\u5c0f\u7684\u5f71\u54cd\u3002<\/li>\n<li>\u975e\u7ebf\u6027\u7279\u6027\uff1aReLU\u51fd\u6570\u901a\u8fc7\u622a\u65ad\u8d1f\u503c\u533a\u57df\u5f15\u5165\u4e86\u975e\u7ebf\u6027\u7279\u6027\u3002<\/li>\n<li>\u907f\u514d\u9971\u548c\u95ee\u9898\uff1a Sigmoid\u548cTanh\u7b49\u6fc0\u6d3b\u51fd\u6570\u5bb9\u6613\u51fa\u73b0\u9971\u548c\u73b0\u8c61\uff0c\u5c24\u5176\u662f\u5728\u8f93\u5165\u503c\u5f88\u5927\u6216\u5f88\u5c0f\u65f6\uff0c\u51fd\u6570\u7684\u5bfc\u6570\u4f1a\u8d8b\u8fd1\u4e8e0\uff0c\u4ece\u800c\u5bfc\u81f4\u68af\u5ea6\u6d88\u5931\u3002\u800cReLU\u5728\u6b63\u533a\u95f4\u5185\u6ca1\u6709\u9971\u548c\u95ee\u9898\uff0c\u8f93\u51fa\u968f\u8f93\u5165\u589e\u5927\u800c\u7ebf\u6027\u589e\u52a0\u3002<\/li>\n<\/ul>\n<h5>Sigmoid\u6fc0\u6d3b\u51fd\u6570<\/h5>\n<p>Sigmoid\u51fd\u6570\u7684\u8868\u8fbe\u5f0f\u4e3a\uff1a<span class=\"katex-eq\" data-katex-display=\"false\">f(z) = \/frac{1}{(1+e^{-z})}<\/span><br \/>\n<img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2025\/07\/20250720125011210.png\" alt=\"\" \/><\/p>\n<h5>Tanh\/\u53cc\u66f2\u6b63\u5207\u6fc0\u6d3b\u51fd\u6570<\/h5>\n<p>Tanh\u662f\u4e00\u4e2a\u53cc\u66f2\u6b63\u5207\u51fd\u6570\u3002Tanh\u548csigmoid\u51fd\u6570\u7684\u66f2\u7ebf\u76f8\u5bf9\u76f8\u4f3c\u3002\u4f46\u662f\u5b83\u6bd4sigmoid\u51fd\u6570\u66f4\u6709\u4f18\u52bf\uff1a<\/p>\n<ul>\n<li>\u5f53\u8f93\u5165\u8f83\u5927\u6216\u8f83\u5c0f\u65f6\uff0c\u8f93\u51fa\u51e0\u4e4e\u662f\u5e73\u6ed1\u7684\u4e14\u68af\u5ea6\u8f83\u5c0f\uff0c\u8fd9\u4e0d\u5229\u4e8e\u6743\u91cd\u66f4\u65b0\u3002\u4e8c\u8005\u7684\u533a\u522b\u5728\u4e8e\u8f93\u51fa\u95f4\u9694\uff0ctanh\u7684\u8f93\u51fa\u95f4\u9694\u4e3a1\uff0c\u5e76\u4e14\u6574\u4e2a\u51fd\u6570\u4ee50\u4e3a\u4e2d\u5fc3\uff0c\u6bd4sigmoid\u51fd\u6570\u66f4\u597d\u3002<\/li>\n<li>\u5728tanh\u56fe\u4e2d\uff0c\u8d1f\u8f93\u5165\u5c06\u88ab\u5f3a\u6620\u5c04\u4e3a\u8d1f\uff0c\u4e14\u96f6\u8f93\u5165\u88ab\u6620\u5c04\u4e3a\u63a5\u8fd1\u96f6\u3002<\/li>\n<\/ul>\n<p>\u6ce8\u610f\uff0c\u5728\u4e00\u822c\u7684\u4e8c\u5143\u5206\u7c7b\u95ee\u9898\u4e2d\uff0ctanh\u51fd\u6570\u7528\u4e8e\u9690\u85cf\u5c42\uff0c\u800csigmoid\u51fd\u6570\u7528\u4e8e\u8f93\u51fa\u5c42\u3002\u4f46\u8fd9\u5e76\u4e0d\u662f\u56fa\u5b9a\u7684\uff0c\u9700\u8981\u6839\u636e\u7279\u5b9a\u95ee\u9898\u8fdb\u884c\u8c03\u6574\u3002<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2025\/07\/20250720125241888.png\" alt=\"\" \/><br \/>\n<img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2025\/07\/20250720125240409.png\" alt=\"\" \/><\/p>\n<h5>Softmax\u6fc0\u6d3b\u51fd\u6570<\/h5>\n<p>Softmax\u662f\u7528\u4e8e\u591a\u7c7b\u5206\u7c7b\u95ee\u9898\u7684\u6fc0\u6d3b\u51fd\u6570\u3002\u5728\u591a\u5206\u7c7b\u95ee\u9898\u4e2d\uff0c\u8d85\u8fc7\u4e24\u4e2a\u7c7b\u6807\u7b7e\u5219\u9700\u8981\u7c7b\u6210\u5458\u5173\u7cfb\u3002\u5bf9\u4e8e\u957f\u5ea6\u4e3aK\u7684\u4efb\u610f\u5b9e\u5411\u91cf\uff0cSoftmax\u53ef\u4ee5\u5c06\u5176\u538b\u7f29\u4e3a\u957f\u5ea6\u4e3aK\uff0c\u503c\u5728\uff080\uff0c1\uff09\u8303\u56f4\u5185\uff0c\u5e76\u4e14\u5411\u91cf\u4e2d\u5143\u7d20\u7684\u603b\u548c\u4e3a1\u7684\u5b9e\u5411\u91cf\u3002<br \/>\n<img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2025\/07\/20250720133257381.png\" alt=\"\" \/><br \/>\n<img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2025\/07\/20250720131904348.png\" alt=\"\" \/><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u5f15\u8a00 \u8bed\u8a00\u6a21\u578b\u7684\u53d1\u5c55\u5386\u7a0b \u7edf\u8ba1\u8bed\u8a00\u6a21\u578b\uff08Statistical Language Model,SLM\uff09\u3002\u7edf\u8ba1\u8bed\u8a00\u6a21\u578b\u4f7f\u7528\u9a6c\u5c14\u53ef\u592b\u5047 &#8230;<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"emotion":"","emotion_color":"","title_style":"","license":"","footnotes":""},"categories":[1],"tags":[],"class_list":["post-1205","post","type-post","status-publish","format-standard","hentry","category-uncategorized"],"_links":{"self":[{"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/posts\/1205","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/comments?post=1205"}],"version-history":[{"count":23,"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/posts\/1205\/revisions"}],"predecessor-version":[{"id":1296,"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/posts\/1205\/revisions\/1296"}],"wp:attachment":[{"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/media?parent=1205"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/categories?post=1205"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/tags?post=1205"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}