﻿{"id":10,"date":"2026-03-30T18:10:23","date_gmt":"2026-03-30T10:10:23","guid":{"rendered":"http:\/\/8.134.203.102:32767\/?p=10"},"modified":"2026-03-30T18:10:23","modified_gmt":"2026-03-30T10:10:23","slug":"transformer%e5%be%b7%e8%8b%b1%e7%bf%bb%e8%af%91%e9%a1%b9%e7%9b%ae%e4%bb%8b%e7%bb%8d","status":"publish","type":"post","link":"https:\/\/www.mahui.space\/index.php\/2026\/03\/30\/transformer%e5%be%b7%e8%8b%b1%e7%bf%bb%e8%af%91%e9%a1%b9%e7%9b%ae%e4%bb%8b%e7%bb%8d\/","title":{"rendered":"Transformer\u5fb7\u82f1\u7ffb\u8bd1\u9879\u76ee\u4ecb\u7ecd"},"content":{"rendered":"\n<h1 class=\"wp-block-heading\">\u4e00\u3001\u5148\u628a\u6574\u4e2a\u9879\u76ee\u770b\u6210\u4e00\u4e2a\u6d41\u6c34\u7ebf<\/h1>\n\n\n\n<p class=\"wp-block-paragraph\">\u4f60\u8fd9\u4e2a\u9879\u76ee\u5b8c\u6574\u6d41\u7a0b\u5176\u5b9e\u662f\uff1a<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u6570\u636e\u51c6\u5907 \u2192 \u5206\u8bcd\/BPE \u2192 \u6784\u9020 batch \u2192 \u5efa\u6a21 \u2192 \u8bad\u7ec3 \u2192 \u9a8c\u8bc1 \u2192 \u63a8\u7406 \u2192 BLEU \u8bc4\u4f30<\/strong><\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4f60\u9762\u8bd5\u6216\u8005\u81ea\u5df1\u7406\u89e3\u65f6\uff0c\u4e00\u5b9a\u5148\u6293\u4f4f\u8fd9\u6761\u7ebf\uff0c\u4e0d\u7136\u4f1a\u9677\u5728\u67d0\u4e2a\u51fd\u6570\u91cc\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u6211\u5148\u7528\u6700\u6734\u7d20\u7684\u8bdd\u8bf4\u4e00\u904d\uff1a<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5047\u8bbe\u8f93\u5165\u4e00\u53e5\u5fb7\u8bed\uff1a<\/p>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p class=\"wp-block-paragraph\">ich liebe dich<\/p>\n<\/blockquote>\n\n\n\n<p class=\"wp-block-paragraph\">\u6a21\u578b\u6700\u7ec8\u60f3\u8f93\u51fa\u4e00\u53e5\u82f1\u8bed\uff1a<\/p>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p class=\"wp-block-paragraph\">i love you<\/p>\n<\/blockquote>\n\n\n\n<p class=\"wp-block-paragraph\">\u90a3\u5728\u8fd9\u4e2a\u9879\u76ee\u91cc\uff0c\u5b83\u4e0d\u662f\u76f4\u63a5\u770b\u5b57\u7b26\u4e32\uff0c\u800c\u662f\u8fd9\u6837\u8d70\u7684\uff1a<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">\u7b2c\u4e00\u6b65\uff1a\u6587\u672c\u9884\u5904\u7406<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u5148\u628a\u53e5\u5b50\u5206\u6210 token\uff0c\u6bd4\u5982\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u5fb7\u8bed\uff1a<code>[\"ich\", \"liebe\", \"dich\"]<\/code><\/li>\n\n\n\n<li>\u82f1\u8bed\uff1a<code>[\"i\", \"love\", \"you\"]<\/code><\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u518d\u53d8\u6210\u6570\u5b57 id\uff0c\u6bd4\u5982\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u5fb7\u8bed\uff1a<code>[15, 208, 77]<\/code><\/li>\n\n\n\n<li>\u82f1\u8bed\uff1a<code>[9, 51, 33]<\/code><\/li>\n<\/ul>\n\n\n\n<h3 class=\"wp-block-heading\">\u7b2c\u4e8c\u6b65\uff1a\u52a0\u7279\u6b8a\u7b26\u53f7<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e3a\u4e86\u8ba9\u6a21\u578b\u77e5\u9053\u53e5\u5b50\u5f00\u59cb\u548c\u7ed3\u675f\uff0c\u4f1a\u52a0\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><code>[BOS]<\/code>\uff1a\u5f00\u59cb<\/li>\n\n\n\n<li><code>[EOS]<\/code>\uff1a\u7ed3\u675f<\/li>\n\n\n\n<li><code>[PAD]<\/code>\uff1a\u8865\u9f50\u957f\u5ea6<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u6bd4\u5982 decoder \u7684\u8f93\u5165\u5e38\u5e38\u662f\uff1a<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><code>[BOS], i, love, you<\/code><\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u800c\u8bad\u7ec3\u6807\u7b7e\u662f\uff1a<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><code>i, love, you, [EOS]<\/code><\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u5c31\u53eb <strong>teacher forcing<\/strong>\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">\u7b2c\u4e09\u6b65\uff1aEncoder \u7f16\u7801\u6e90\u8bed\u8a00<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u5fb7\u8bed\u53e5\u5b50\u8fdb Encoder\uff0c\u7ecf\u8fc7\u591a\u5c42\u81ea\u6ce8\u610f\u529b\uff0c\u5f97\u5230\u6bcf\u4e2a\u4f4d\u7f6e\u7684\u4e0a\u4e0b\u6587\u8868\u793a\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">\u7b2c\u56db\u6b65\uff1aDecoder \u9010\u6b65\u751f\u6210\u76ee\u6807\u8bed\u8a00<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">Decoder \u4e00\u8fb9\u770b\u81ea\u5df1\u524d\u9762\u5df2\u7ecf\u751f\u6210\u7684\u82f1\u8bed\u8bcd\uff0c\u4e00\u8fb9\u770b Encoder \u8f93\u51fa\u7684\u5fb7\u8bed\u8868\u793a\uff0c\u7136\u540e\u9884\u6d4b\u4e0b\u4e00\u4e2a\u82f1\u8bed\u8bcd\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">\u7b2c\u4e94\u6b65\uff1a\u7b97\u635f\u5931\uff0c\u53cd\u5411\u4f20\u64ad<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u7528\u9884\u6d4b\u7ed3\u679c\u548c\u771f\u5b9e\u82f1\u8bed\u6807\u7b7e\u6bd4\uff0c\u7b97\u4ea4\u53c9\u71b5\u635f\u5931\uff0c\u66f4\u65b0\u53c2\u6570\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">\u7b2c\u516d\u6b65\uff1a\u8bc4\u4f30<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u8bad\u7ec3\u5b8c\u540e\uff0c\u7528 BLEU \u5206\u6570\u8bc4\u4f30\u7ffb\u8bd1\u6548\u679c\u3002<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h1 class=\"wp-block-heading\">\u4e8c\u3001\u8fd9\u4e2a\u9879\u76ee\u7684\u6838\u5fc3\u76ee\u6807\u662f\u4ec0\u4e48<\/h1>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u4e2a\u9879\u76ee\u672c\u8d28\u4e0a\u662f\u5728\u5b9e\u73b0\u8bba\u6587\u300aAttention is All You Need\u300b\u91cc\u7684\u7ecf\u5178 Transformer \u7ffb\u8bd1\u6a21\u578b\uff0c\u53ea\u4e0d\u8fc7\u505a\u4e86\u5de5\u7a0b\u5316\u5305\u88c5\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4f60\u53ef\u4ee5\u628a\u5b83\u7406\u89e3\u6210\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>Encoder<\/strong>\uff1a\u8d1f\u8d23\u201c\u8bfb\u61c2\u5fb7\u8bed\u201d<\/li>\n\n\n\n<li><strong>Decoder<\/strong>\uff1a\u8d1f\u8d23\u201c\u7ffb\u8bd1\u6210\u82f1\u8bed\u201d<\/li>\n\n\n\n<li><strong>Attention<\/strong>\uff1a\u8d1f\u8d23\u201c\u5728\u5904\u7406\u67d0\u4e2a\u8bcd\u65f6\uff0c\u53bb\u770b\u54ea\u4e9b\u8bcd\u66f4\u91cd\u8981\u201d<\/li>\n<\/ul>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h1 class=\"wp-block-heading\">\u4e09\u3001\u5148\u8bb2\u6570\u636e\u90e8\u5206\uff1a\u9879\u76ee\u662f\u600e\u4e48\u628a\u6587\u672c\u53d8\u6210\u53ef\u8bad\u7ec3\u6570\u636e\u7684<\/h1>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">1. \u6570\u636e\u96c6 <code>LangPairDataset<\/code><\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u4f60\u7684 <code>LangPairDataset<\/code> \u5e72\u7684\u4e8b\u60c5\u5f88\u7b80\u5355\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u8bfb\u53d6\u5fb7\u8bed\u6587\u4ef6 <code>*_src.bpe<\/code><\/li>\n\n\n\n<li>\u8bfb\u53d6\u82f1\u8bed\u6587\u4ef6 <code>*_trg.bpe<\/code><\/li>\n\n\n\n<li>\u4e00\u884c\u5fb7\u8bed\u5bf9\u5e94\u4e00\u884c\u82f1\u8bed<\/li>\n\n\n\n<li>\u8fc7\u6ee4\u6389\u957f\u5ea6\u592a\u957f\u7684\u53e5\u5b50<\/li>\n\n\n\n<li>\u7f13\u5b58\u6210 <code>.npy<\/code><\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u5b83\u672c\u8d28\u4e0a\u662f\uff1a<\/p>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p class=\"wp-block-paragraph\">\u6bcf\u6b21\u53d6\u4e00\u4e2a\u6837\u672c\uff0c\u5c31\u8fd4\u56de <code>(source_sentence, target_sentence)<\/code><\/p>\n<\/blockquote>\n\n\n\n<p class=\"wp-block-paragraph\">\u6bd4\u5982\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>return self.src&#91;index], self.trg&#91;index]\n<\/code><\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e5f\u5c31\u662f\u8fd4\u56de\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u6e90\u53e5\u5b50\uff08\u5fb7\u8bed\uff09<\/li>\n\n\n\n<li>\u76ee\u6807\u53e5\u5b50\uff08\u82f1\u8bed\uff09<\/li>\n<\/ul>\n\n\n\n<h3 class=\"wp-block-heading\">\u4e3a\u4ec0\u4e48\u8981\u8fc7\u6ee4\u6700\u5927\u957f\u5ea6\uff1f<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u56e0\u4e3a Transformer \u7684\u8ba1\u7b97\u590d\u6742\u5ea6\u8ddf\u5e8f\u5217\u957f\u5ea6\u5173\u7cfb\u5f88\u5927\uff0c\u957f\u5ea6\u592a\u957f\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u663e\u5b58\u5360\u7528\u5927<\/li>\n\n\n\n<li>\u8bad\u7ec3\u6162<\/li>\n\n\n\n<li>\u6ce8\u610f\u529b\u77e9\u9635\u53d8\u5927<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u6240\u4ee5\u8fd9\u4e2a\u9879\u76ee\u8bbe\u7f6e\u4e86 <code>max_length=128<\/code>\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">\u4e3a\u4ec0\u4e48\u8981\u505a\u7f13\u5b58\uff1f<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u56e0\u4e3a\u6bcf\u6b21\u90fd\u91cd\u65b0\u8bfb\u6587\u4ef6\u3001\u91cd\u65b0\u8fc7\u6ee4\u4f1a\u5f88\u6162\u3002\u7f13\u5b58\u4e4b\u540e\uff0c\u4e0b\u6b21\u76f4\u63a5\u52a0\u8f7d\u5904\u7406\u597d\u7684\u6570\u636e\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4f60\u53ef\u4ee5\u8fd9\u4e48\u8bb0\uff1a<\/p>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p class=\"wp-block-paragraph\">Dataset \u8d1f\u8d23\u201c\u628a\u539f\u59cb\u6587\u672c\u6837\u672c\u51c6\u5907\u597d\u201d\u3002<\/p>\n<\/blockquote>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">2. Tokenizer\uff1a\u628a\u5355\u8bcd\u53d8\u6210\u6570\u5b57<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\"><code>Tokenizer<\/code> \u8fd9\u5757\u5f88\u5173\u952e\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5b83\u505a\u4e24\u4ef6\u4e8b\uff1a<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">\u7f16\u7801 <code>encode<\/code><\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u628a\u8bcd\u5217\u8868\u53d8\u6210 id \u5217\u8868\uff0c\u6bd4\u5982\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>&#91;\"i\", \"love\", \"you\"] -&gt; &#91;1, 9, 51, 33, 3, 0, 0]\n<\/code><\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u91cc\u53ef\u80fd\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><code>1<\/code> \u662f <code>[BOS]<\/code><\/li>\n\n\n\n<li><code>3<\/code> \u662f <code>[EOS]<\/code><\/li>\n\n\n\n<li><code>0<\/code> \u662f <code>[PAD]<\/code><\/li>\n<\/ul>\n\n\n\n<h3 class=\"wp-block-heading\">\u89e3\u7801 <code>decode<\/code><\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u628a id \u518d\u53d8\u56de\u53e5\u5b50\u3002<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">3. \u4e3a\u4ec0\u4e48 Encoder \u548c Decoder \u8f93\u5165\u4e0d\u4e00\u6837\uff1f<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u4e2a\u5730\u65b9\u5f88\u591a\u4eba\u7b2c\u4e00\u6b21\u5b66 Transformer \u4f1a\u7cca\u6d82\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4f60\u7684\u4ee3\u7801\u91cc\uff1a<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">Encoder \u8f93\u5165<\/h3>\n\n\n\n<pre class=\"wp-block-code\"><code>&#91;BOS] src &#91;EOS] &#91;PAD]\n<\/code><\/pre>\n\n\n\n<h3 class=\"wp-block-heading\">Decoder \u8f93\u5165<\/h3>\n\n\n\n<pre class=\"wp-block-code\"><code>&#91;BOS] trg &#91;PAD]\n<\/code><\/pre>\n\n\n\n<h3 class=\"wp-block-heading\">Decoder \u6807\u7b7e<\/h3>\n\n\n\n<pre class=\"wp-block-code\"><code>trg &#91;EOS] &#91;PAD]\n<\/code><\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e3a\u4ec0\u4e48\uff1f<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u56e0\u4e3a\u8bad\u7ec3\u65f6 Decoder \u662f\u201c\u62ff\u524d\u9762\u7684\u6b63\u786e\u8bcd\uff0c\u53bb\u9884\u6d4b\u4e0b\u4e00\u4e2a\u8bcd\u201d\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e3e\u4f8b\uff1a<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u76ee\u6807\u53e5\u5b50\u662f\uff1a<\/p>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p class=\"wp-block-paragraph\">i love you<\/p>\n<\/blockquote>\n\n\n\n<p class=\"wp-block-paragraph\">\u90a3\u4e48\uff1a<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>decoder_inputs<\/strong><\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>&#91;BOS] i love you\n<\/code><\/pre>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>decoder_labels<\/strong><\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>i love you &#91;EOS]\n<\/code><\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u6837\u6a21\u578b\u5728\u6bcf\u4e2a\u4f4d\u7f6e\u5b66\u7684\u662f\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u770b <code>[BOS]<\/code>\uff0c\u9884\u6d4b <code>i<\/code><\/li>\n\n\n\n<li>\u770b <code>[BOS] i<\/code>\uff0c\u9884\u6d4b <code>love<\/code><\/li>\n\n\n\n<li>\u770b <code>[BOS] i love<\/code>\uff0c\u9884\u6d4b <code>you<\/code><\/li>\n\n\n\n<li>\u770b <code>[BOS] i love you<\/code>\uff0c\u9884\u6d4b <code>[EOS]<\/code><\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u5c31\u662f\u673a\u5668\u7ffb\u8bd1\u8bad\u7ec3\u7684\u6807\u51c6\u505a\u6cd5\u3002<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h1 class=\"wp-block-heading\">\u56db\u3001\u4e3a\u4ec0\u4e48\u8fd9\u4e2a\u9879\u76ee\u8981\u4e13\u95e8\u8bbe\u8ba1 BatchSampler<\/h1>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u662f\u4f60\u9879\u76ee\u91cc\u4e00\u4e2a\u5f88\u52a0\u5206\u7684\u70b9\uff0c\u56e0\u4e3a\u8bf4\u660e\u4f60\u4e0d\u53ea\u662f\u4f1a\u8c03\u5305\uff0c\u8fd8\u8003\u8651\u4e86\u6548\u7387\u3002<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">1. \u666e\u901a batch \u7684\u95ee\u9898<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u5982\u679c\u4e00\u53e5\u8bdd\u5f88\u77ed\uff0c\u4e00\u53e5\u8bdd\u5f88\u957f\uff0c\u5f3a\u884c\u653e\u5728\u4e00\u4e2a batch \u91cc\uff0c\u5c31\u8981 pad \u5230\u540c\u6837\u957f\u5ea6\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4f8b\u5982\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u53e5\u5b50 A \u957f\u5ea6 5<\/li>\n\n\n\n<li>\u53e5\u5b50 B \u957f\u5ea6 100<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u90a3 A \u4e5f\u5f97\u8865\u5230 100\uff0c\u5f88\u6d6a\u8d39\u3002<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">2. \u4f60\u7684\u505a\u6cd5\uff1a\u6309\u957f\u5ea6\u5206\u7ec4<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\"><code>TransformerBatchSampler<\/code> \u7684\u601d\u8def\u662f\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u5148\u6309\u7167\u6e90\u53e5\u957f\u3001\u76ee\u6807\u53e5\u957f\u6392\u5e8f<\/li>\n\n\n\n<li>\u628a\u957f\u5ea6\u63a5\u8fd1\u7684\u6837\u672c\u5206\u5230\u540c\u4e00\u4e2a batch<\/li>\n\n\n\n<li>batch \u5927\u5c0f\u4e0d\u662f\u56fa\u5b9a\u201c\u6837\u672c\u6570\u201d\uff0c\u800c\u662f\u5c3d\u91cf\u63a7\u5236\u201c\u603b token \u6570\u201d<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u4ee3\u7801\u91cc\u8fd9\u4e2a\u5224\u65ad\u5f88\u5173\u952e\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>if max_len * (len(self._batch) + 1) &gt; self._batch_size:\n<\/code><\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u610f\u601d\u662f\uff1a<\/p>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p class=\"wp-block-paragraph\">\u5f53\u524d batch \u7684\u6700\u5927\u957f\u5ea6 \u00d7 \u6837\u672c\u4e2a\u6570\uff0c\u4e0d\u80fd\u8d85\u8fc7\u8bbe\u5b9a\u9608\u503c<\/p>\n<\/blockquote>\n\n\n\n<p class=\"wp-block-paragraph\">\u6240\u4ee5\u8fd9\u91cc\u7684 <code>batch_size=4096<\/code>\uff0c\u5176\u5b9e\u66f4\u50cf\u662f <strong>token budget<\/strong>\uff0c\u4e0d\u662f\u56fa\u5b9a 4096 \u6761\u6837\u672c\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">\u8fd9\u6837\u505a\u7684\u597d\u5904<\/h3>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u51cf\u5c11 padding<\/li>\n\n\n\n<li>\u63d0\u9ad8\u663e\u5b58\u5229\u7528\u7387<\/li>\n\n\n\n<li>\u63d0\u9ad8\u8bad\u7ec3\u6548\u7387<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u4f60\u53ef\u4ee5\u8fd9\u6837\u7406\u89e3\uff1a<\/p>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p class=\"wp-block-paragraph\">\u4e0d\u662f\u6bcf\u8f66\u90fd\u88c5\u56fa\u5b9a\u201c\u4eba\u6570\u201d\uff0c\u800c\u662f\u5c3d\u91cf\u6309\u201c\u603b\u4f53\u79ef\u201d\u88c5\u8d27\u3002<\/p>\n<\/blockquote>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u4e2a\u601d\u8def\u5728 NLP \u91cc\u5f88\u5e38\u89c1\uff0c\u4e5f\u5f88\u6709\u9762\u8bd5\u4ef7\u503c\u3002<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h1 class=\"wp-block-heading\">\u4e94\u3001\u73b0\u5728\u8fdb\u5165\u6838\u5fc3\uff1aTransformer \u5230\u5e95\u662f\u4ec0\u4e48<\/h1>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u91cc\u6211\u7528\u5c3d\u91cf\u201c\u8001\u5e08\u8bb2\u8bfe\u201d\u7684\u65b9\u5f0f\uff0c\u4e0d\u76f4\u63a5\u5806\u516c\u5f0f\u3002<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">1. Transformer \u6700\u6838\u5fc3\u7684\u4e00\u53e5\u8bdd<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Transformer \u7684\u672c\u8d28\uff0c\u5c31\u662f\u8ba9\u6bcf\u4e2a\u8bcd\u5728\u8868\u793a\u81ea\u5df1\u65f6\uff0c\u90fd\u80fd\u52a8\u6001\u5730\u53c2\u8003\u53e5\u5b50\u91cc\u7684\u5176\u4ed6\u8bcd\u3002<\/strong><\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u6bd4\u5982\u5fb7\u8bed\u53e5\u5b50\uff1a<\/p>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p class=\"wp-block-paragraph\">ich liebe dich<\/p>\n<\/blockquote>\n\n\n\n<p class=\"wp-block-paragraph\">\u5f53\u6a21\u578b\u5904\u7406 <code>liebe<\/code> \u65f6\uff0c\u5b83\u4e0d\u4f1a\u53ea\u770b\u8fd9\u4e2a\u8bcd\u81ea\u5df1\uff0c\u8fd8\u4f1a\u770b\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u8c01\u662f\u4e3b\u8bed\uff1a<code>ich<\/code><\/li>\n\n\n\n<li>\u8c01\u662f\u5bbe\u8bed\uff1a<code>dich<\/code><\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e8e\u662f <code>liebe<\/code> \u7684\u8868\u793a\u5c31\u4e0d\u662f\u5b64\u7acb\u7684\uff0c\u800c\u662f\u5e26\u4e0a\u4e0b\u6587\u7684\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u5c31\u662f attention \u7684\u610f\u4e49\u3002<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">2. \u4e3a\u4ec0\u4e48\u4e0d\u7528 RNN\uff0c\u800c\u7528 Transformer\uff1f<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u4f20\u7edf RNN\/LSTM \u662f\u6309\u987a\u5e8f\u8bfb\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u5148\u8bfb\u7b2c 1 \u4e2a\u8bcd<\/li>\n\n\n\n<li>\u518d\u8bfb\u7b2c 2 \u4e2a\u8bcd<\/li>\n\n\n\n<li>\u518d\u8bfb\u7b2c 3 \u4e2a\u8bcd<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u95ee\u9898\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u96be\u5e76\u884c<\/li>\n\n\n\n<li>\u957f\u8ddd\u79bb\u4f9d\u8d56\u96be\u5b66<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">Transformer \u4e0d\u4e00\u6837\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u4e00\u6574\u53e5\u540c\u65f6\u770b<\/li>\n\n\n\n<li>\u4efb\u610f\u4e24\u4e2a\u4f4d\u7f6e\u90fd\u80fd\u76f4\u63a5\u5efa\u7acb\u8054\u7cfb<\/li>\n\n\n\n<li>\u66f4\u5bb9\u6613\u5e76\u884c\u8bad\u7ec3<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u6240\u4ee5\u5b83\u7279\u522b\u9002\u5408\u7ffb\u8bd1\u4efb\u52a1\u3002<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h1 class=\"wp-block-heading\">\u516d\u3001Embedding\uff1a\u8bcd\u5411\u91cf + \u4f4d\u7f6e\u7f16\u7801<\/h1>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728\u4f60\u4ee3\u7801\u91cc\u662f <code>TransformerEmbedding<\/code>\u3002<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">1. \u4e3a\u4ec0\u4e48\u9700\u8981\u8bcd\u5411\u91cf<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u56e0\u4e3a\u6a21\u578b\u4e0d\u80fd\u76f4\u63a5\u5904\u7406\u5355\u8bcd\u5b57\u7b26\u4e32\uff0c\u5fc5\u987b\u5148\u53d8\u6210\u5411\u91cf\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u6bd4\u5982\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><code>ich<\/code> \u2192 \u4e00\u4e2a 512 \u7ef4\u5411\u91cf<\/li>\n\n\n\n<li><code>love<\/code> \u2192 \u4e00\u4e2a 512 \u7ef4\u5411\u91cf<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u90e8\u5206\u662f <code>word_embedding<\/code>\u3002<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">2. \u4e3a\u4ec0\u4e48\u8fd8\u8981\u4f4d\u7f6e\u7f16\u7801<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u6ce8\u610f\u529b\u672c\u8eab\u4e0d\u5e26\u987a\u5e8f\u6982\u5ff5\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e5f\u5c31\u662f\u8bf4\uff0c\u5982\u679c\u53ea\u770b\u8bcd\u5411\u91cf\uff0c\u6a21\u578b\u5e76\u4e0d\u77e5\u9053\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u8c01\u5728\u524d<\/li>\n\n\n\n<li>\u8c01\u5728\u540e<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u53ef\u8bed\u8a00\u91cc\u987a\u5e8f\u975e\u5e38\u91cd\u8981\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u6bd4\u5982\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><code>dog bites man<\/code><\/li>\n\n\n\n<li><code>man bites dog<\/code><\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u5355\u8bcd\u4e00\u6837\uff0c\u987a\u5e8f\u4e0d\u540c\uff0c\u610f\u601d\u5b8c\u5168\u4e0d\u540c\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u6240\u4ee5\u9700\u8981\u52a0\u5165 <strong>position embedding \/ positional encoding<\/strong>\u3002<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">3. \u4f60\u8fd9\u91cc\u7684\u4f4d\u7f6e\u7f16\u7801\u600e\u4e48\u5b9e\u73b0\u7684<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u4f60\u4ee3\u7801\u91cc\u7528\u7684\u662f\u7ecf\u5178\u7684 <strong>\u6b63\u5f26\/\u4f59\u5f26\u4f4d\u7f6e\u7f16\u7801<\/strong>\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u5076\u6570\u7ef4\u7528 <code>sin<\/code><\/li>\n\n\n\n<li>\u5947\u6570\u7ef4\u7528 <code>cos<\/code><\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u597d\u5904\u662f\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u4e0d\u9700\u8981\u5b66\u4e60\u53c2\u6570<\/li>\n\n\n\n<li>\u4e0d\u540c\u4f4d\u7f6e\u6709\u4e0d\u540c\u6a21\u5f0f<\/li>\n\n\n\n<li>\u6a21\u578b\u80fd\u611f\u77e5\u76f8\u5bf9\/\u7edd\u5bf9\u4f4d\u7f6e<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u516c\u5f0f\u672c\u8eab\u4f60\u4e0d\u7528\u6b7b\u80cc\uff0c\u91cd\u70b9\u7406\u89e3\uff1a<\/p>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p class=\"wp-block-paragraph\">\u5b83\u7ed9\u6bcf\u4e2a\u4f4d\u7f6e\u751f\u6210\u4e00\u4e2a\u72ec\u7279\u7684\u201c\u4f4d\u7f6e\u5411\u91cf\u201d\uff0c\u7136\u540e\u548c\u8bcd\u5411\u91cf\u76f8\u52a0\u3002<\/p>\n<\/blockquote>\n\n\n\n<p class=\"wp-block-paragraph\">\u6700\u7ec8\u8f93\u5165\u8868\u793a\u662f\uff1a<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u8bcd\u5411\u91cf + \u4f4d\u7f6e\u5411\u91cf<\/strong><\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u5c31\u662f\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>embeds = word_embeds + pos_embeds\n<\/code><\/pre>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h1 class=\"wp-block-heading\">\u4e03\u3001Attention \u662f\u8fd9\u4e2a\u9879\u76ee\u6700\u6838\u5fc3\u7684\u601d\u60f3<\/h1>\n\n\n\n<p class=\"wp-block-paragraph\">\u73b0\u5728\u8bb2\u6700\u5173\u952e\u7684 <code>MultiHeadAttention<\/code>\u3002<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">1. \u5148\u522b\u770b\u516c\u5f0f\uff0c\u5148\u7406\u89e3 Query \/ Key \/ Value<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u662f\u5f88\u591a\u540c\u5b66\u6700\u96be\u7406\u89e3\u7684\u5730\u65b9\uff0c\u6211\u7ed9\u4f60\u4e00\u4e2a\u201c\u67e5\u8d44\u6599\u201d\u7684\u7c7b\u6bd4\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4f60\u53bb\u56fe\u4e66\u9986\u627e\u8d44\u6599\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>Query<\/strong>\uff1a\u4f60\u73b0\u5728\u60f3\u95ee\u7684\u95ee\u9898<\/li>\n\n\n\n<li><strong>Key<\/strong>\uff1a\u6bcf\u672c\u4e66\u7684\u6807\u7b7e<\/li>\n\n\n\n<li><strong>Value<\/strong>\uff1a\u6bcf\u672c\u4e66\u771f\u6b63\u7684\u5185\u5bb9<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u6d41\u7a0b\u662f\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li>\u7528\u4f60\u7684 Query \u548c\u6bcf\u672c\u4e66\u7684 Key \u505a\u5339\u914d<\/li>\n\n\n\n<li>\u5339\u914d\u5ea6\u8d8a\u9ad8\uff0c\u8bf4\u660e\u8fd9\u672c\u4e66\u8d8a\u76f8\u5173<\/li>\n\n\n\n<li>\u518d\u628a\u76f8\u5173\u4e66\u7684 Value \u6309\u6743\u91cd\u52a0\u6743\u6c42\u548c<\/li>\n\n\n\n<li>\u5f97\u5230\u4f60\u6700\u7ec8\u9700\u8981\u7684\u4fe1\u606f<\/li>\n<\/ol>\n\n\n\n<p class=\"wp-block-paragraph\">Attention \u5c31\u662f\u5728\u5e72\u8fd9\u4ef6\u4e8b\u3002<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">2. \u5728\u53e5\u5b50\u91cc\u600e\u4e48\u7406\u89e3\uff1f<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u6bd4\u5982\u53e5\u5b50\uff1a<\/p>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p class=\"wp-block-paragraph\">The animal didn&#8217;t cross the street because it was tired.<\/p>\n<\/blockquote>\n\n\n\n<p class=\"wp-block-paragraph\">\u5904\u7406 <code>it<\/code> \u7684\u65f6\u5019\uff0c\u6a21\u578b\u4f1a\u770b\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><code>animal<\/code><\/li>\n\n\n\n<li><code>street<\/code><\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u901a\u8fc7 attention\uff0c\u5b83\u4f1a\u53d1\u73b0 <code>it<\/code> \u66f4\u53ef\u80fd\u6307\u5411 <code>animal<\/code>\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e5f\u5c31\u662f\u8bf4\uff1a<\/p>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p class=\"wp-block-paragraph\">\u5f53\u524d\u8bcd\u4f1a\u201c\u5173\u6ce8\u201d\u5bf9\u5b83\u6700\u91cd\u8981\u7684\u5176\u4ed6\u8bcd\u3002<\/p>\n<\/blockquote>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">3. \u4ee3\u7801\u91cc\u600e\u4e48\u5b9e\u73b0\u7684<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u4f60\u7684\u4ee3\u7801\u6d41\u7a0b\u662f\uff1a<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">\u7b2c\u4e00\u6b65\uff1a\u7ebf\u6027\u6620\u5c04\u5f97\u5230 Q\u3001K\u3001V<\/h3>\n\n\n\n<pre class=\"wp-block-code\"><code>self.Wq\nself.Wk\nself.Wv\n<\/code><\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u8f93\u5165 hidden states \u540e\uff0c\u6620\u5c04\u6210\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>Q<\/li>\n\n\n\n<li>K<\/li>\n\n\n\n<li>V<\/li>\n<\/ul>\n\n\n\n<h3 class=\"wp-block-heading\">\u7b2c\u4e8c\u6b65\uff1a\u5206\u5934<\/h3>\n\n\n\n<pre class=\"wp-block-code\"><code>_split_heads\n<\/code><\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u628a 512 \u7ef4\u62c6\u6210 8 \u4e2a\u5934\uff0c\u6bcf\u5934 64 \u7ef4\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e3a\u4ec0\u4e48\u8981\u591a\u5934\uff1f<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u56e0\u4e3a\u4e0d\u540c\u5934\u53ef\u4ee5\u5173\u6ce8\u4e0d\u540c\u5173\u7cfb\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u4e00\u4e2a\u5934\u770b\u4e3b\u8c13\u5173\u7cfb<\/li>\n\n\n\n<li>\u4e00\u4e2a\u5934\u770b\u4f4d\u7f6e\u5173\u7cfb<\/li>\n\n\n\n<li>\u4e00\u4e2a\u5934\u770b\u8bcd\u4e49\u76f8\u5173\u6027<\/li>\n<\/ul>\n\n\n\n<h3 class=\"wp-block-heading\">\u7b2c\u4e09\u6b65\uff1a\u7b97\u6ce8\u610f\u529b\u5206\u6570<\/h3>\n\n\n\n<pre class=\"wp-block-code\"><code>qk_logits = torch.matmul(querys, keys.mT)\n<\/code><\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u5c31\u662f Q \u548c K \u505a\u70b9\u79ef\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">\u7b2c\u56db\u6b65\uff1a\u7f29\u653e<\/h3>\n\n\n\n<pre class=\"wp-block-code\"><code>qk_logits \/ sqrt(head_dim)\n<\/code><\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e3a\u4ec0\u4e48\u7f29\u653e\uff1f<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u56e0\u4e3a\u7ef4\u5ea6\u5927\u65f6\u70b9\u79ef\u503c\u53ef\u80fd\u5f88\u5927\uff0csoftmax \u4f1a\u8fc7\u4e8e\u5c16\u9510\uff0c\u4e0d\u5229\u4e8e\u8bad\u7ec3\u7a33\u5b9a\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">\u7b2c\u4e94\u6b65\uff1amask<\/h3>\n\n\n\n<pre class=\"wp-block-code\"><code>qk_logits += attn_mask * -1e9\n<\/code><\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">mask \u7684\u4f4d\u7f6e\u52a0\u4e00\u4e2a\u5f88\u5927\u7684\u8d1f\u6570\uff0csoftmax \u540e\u5c31\u63a5\u8fd1 0\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">\u7b2c\u516d\u6b65\uff1asoftmax<\/h3>\n\n\n\n<pre class=\"wp-block-code\"><code>attn_scores = F.softmax(...)\n<\/code><\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u5f97\u5230\u201c\u6bcf\u4e2a\u8bcd\u8be5\u5173\u6ce8\u8c01\u201d\u7684\u6982\u7387\u5206\u5e03\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">\u7b2c\u4e03\u6b65\uff1a\u52a0\u6743\u6c42\u548c<\/h3>\n\n\n\n<pre class=\"wp-block-code\"><code>embeds = torch.matmul(attn_scores, values)\n<\/code><\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u7528\u6ce8\u610f\u529b\u6743\u91cd\u5bf9 V \u52a0\u6743\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">\u7b2c\u516b\u6b65\uff1a\u591a\u5934\u5408\u5e76<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u6700\u540e\u628a\u591a\u5934\u62fc\u56de\u53bb\uff0c\u518d\u8fc7\u4e00\u5c42\u7ebf\u6027\u5c42 <code>Wo<\/code>\u3002<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h1 class=\"wp-block-heading\">\u516b\u3001\u4ec0\u4e48\u662f Self-Attention\uff0c\u4ec0\u4e48\u662f Cross-Attention<\/h1>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u4e5f\u662f\u9762\u8bd5\u9ad8\u9891\u70b9\u3002<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">1. Self-Attention\uff1a\u81ea\u5df1\u770b\u81ea\u5df1\u8fd9\u53e5\u8bdd<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728 Encoder \u91cc\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>Query \u6765\u81ea\u6e90\u53e5\u5b50<\/li>\n\n\n\n<li>Key \u6765\u81ea\u6e90\u53e5\u5b50<\/li>\n\n\n\n<li>Value \u6765\u81ea\u6e90\u53e5\u5b50<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u5373\uff1a<\/p>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p class=\"wp-block-paragraph\">\u5fb7\u8bed\u53e5\u5b50\u5185\u90e8\u81ea\u5df1\u5efa\u6a21\u5185\u90e8\u5173\u7cfb<\/p>\n<\/blockquote>\n\n\n\n<p class=\"wp-block-paragraph\">\u6bd4\u5982\u5904\u7406\u5fb7\u8bed\u67d0\u4e2a\u8bcd\u65f6\uff0c\u53ef\u4ee5\u53bb\u770b\u540c\u4e00\u53e5\u91cc\u7684\u5176\u4ed6\u5fb7\u8bed\u8bcd\u3002<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">2. Cross-Attention\uff1a\u76ee\u6807\u53e5\u5b50\u53bb\u770b\u6e90\u53e5\u5b50<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728 Decoder \u91cc\u6709\u4e00\u5c42\u7279\u6b8a\u6ce8\u610f\u529b\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>Query \u6765\u81ea\u5f53\u524d Decoder \u72b6\u6001<\/li>\n\n\n\n<li>Key \u6765\u81ea Encoder \u8f93\u51fa<\/li>\n\n\n\n<li>Value \u6765\u81ea Encoder \u8f93\u51fa<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u5373\uff1a<\/p>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p class=\"wp-block-paragraph\">\u751f\u6210\u82f1\u8bed\u65f6\uff0c\u53bb\u53c2\u8003\u5fb7\u8bed\u7684\u7f16\u7801\u7ed3\u679c<\/p>\n<\/blockquote>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u5c31\u662f\u7ffb\u8bd1\u771f\u6b63\u53d1\u751f\u7684\u5730\u65b9\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4f60\u53ef\u4ee5\u8bb0\u6210\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>Self-Attention<\/strong>\uff1a\u53e5\u5185\u7406\u89e3<\/li>\n\n\n\n<li><strong>Cross-Attention<\/strong>\uff1a\u8de8\u8bed\u8a00\u5bf9\u9f50<\/li>\n<\/ul>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h1 class=\"wp-block-heading\">\u4e5d\u3001Transformer Block \u4e3a\u4ec0\u4e48\u662f\u8fd9\u4e2a\u7ed3\u6784<\/h1>\n\n\n\n<p class=\"wp-block-paragraph\">\u4f60\u7684 <code>TransformerBlock<\/code> \u7ed3\u6784\u5f88\u6807\u51c6\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li>Multi-Head Attention<\/li>\n\n\n\n<li>Add &amp; LayerNorm<\/li>\n\n\n\n<li>Feed Forward Network<\/li>\n\n\n\n<li>Add &amp; LayerNorm<\/li>\n<\/ol>\n\n\n\n<p class=\"wp-block-paragraph\">\u5982\u679c\u662f Decoder block\uff0c\u8fd8\u4f1a\u591a\u4e00\u4e2a Cross-Attention\u3002<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">1. \u6b8b\u5dee\u8fde\u63a5\u4e3a\u4ec0\u4e48\u91cd\u8981<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u4f60\u4ee3\u7801\u91cc\u50cf\u8fd9\u6837\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>hidden_states + self.self_dropout(...)\n<\/code><\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u5c31\u662f\u6b8b\u5dee\u8fde\u63a5\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4f5c\u7528\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u9632\u6b62\u6df1\u5c42\u7f51\u7edc\u8bad\u7ec3\u56f0\u96be<\/li>\n\n\n\n<li>\u4fdd\u7559\u539f\u59cb\u4fe1\u606f<\/li>\n\n\n\n<li>\u8ba9\u68af\u5ea6\u66f4\u5bb9\u6613\u4f20\u64ad<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u53ef\u4ee5\u7406\u89e3\u6210\uff1a<\/p>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p class=\"wp-block-paragraph\">\u65b0\u4fe1\u606f\u5b66\u5f97\u4e0d\u597d\uff0c\u4e5f\u522b\u628a\u8001\u4fe1\u606f\u5b8c\u5168\u4e22\u6389\u3002<\/p>\n<\/blockquote>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">2. LayerNorm \u4e3a\u4ec0\u4e48\u91cd\u8981<\/h2>\n\n\n\n<pre class=\"wp-block-code\"><code>self.self_ln(...)\n<\/code><\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u4f5c\u7528\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u7a33\u5b9a\u8bad\u7ec3<\/li>\n\n\n\n<li>\u7f13\u89e3\u6570\u503c\u6ce2\u52a8<\/li>\n\n\n\n<li>\u8ba9\u4e0d\u540c\u5c42\u8f93\u51fa\u5206\u5e03\u66f4\u5e73\u7a33<\/li>\n<\/ul>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">3. FFN \u662f\u5e72\u4ec0\u4e48\u7684<\/h2>\n\n\n\n<pre class=\"wp-block-code\"><code>nn.Linear -&gt; ReLU -&gt; nn.Linear\n<\/code><\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">Attention \u66f4\u50cf\u201c\u4fe1\u606f\u4ea4\u6362\u201d\uff0cFFN \u66f4\u50cf\u201c\u7279\u5f81\u52a0\u5de5\u201d\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u53ef\u4ee5\u7406\u89e3\u6210\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>Attention \u8d1f\u8d23\u201c\u4ece\u522b\u4eba\u90a3\u91cc\u62ff\u4fe1\u606f\u201d<\/li>\n\n\n\n<li>FFN \u8d1f\u8d23\u201c\u628a\u62ff\u5230\u7684\u4fe1\u606f\u518d\u6df1\u52a0\u5de5\u4e00\u4e0b\u201d<\/li>\n<\/ul>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h1 class=\"wp-block-heading\">\u5341\u3001Encoder \u548c Decoder \u5404\u81ea\u5e72\u4e86\u4ec0\u4e48<\/h1>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">1. Encoder<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\"><code>TransformerEncoder<\/code> \u5c31\u662f\u5806\u53e0\u591a\u4e2a Encoder block\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4f60\u7684\u53c2\u6570\u91cc\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>\"num_encoder_layers\": 6\n<\/code><\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e5f\u5c31\u662f 6 \u5c42 Encoder\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5b83\u7684\u4efb\u52a1\u662f\uff1a<\/p>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p class=\"wp-block-paragraph\">\u628a\u5fb7\u8bed\u53e5\u5b50\u7f16\u7801\u6210\u4e00\u4e32\u4e0a\u4e0b\u6587\u5316\u8868\u793a<\/p>\n<\/blockquote>\n\n\n\n<p class=\"wp-block-paragraph\">\u8f93\u51fa\u7684\u662f\u6bcf\u4e2a\u6e90\u4f4d\u7f6e\u7684 hidden state\u3002<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">2. Decoder<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\"><code>TransformerDecoder<\/code> \u4e5f\u662f\u5806 6 \u5c42\uff0c\u4f46\u6bcf\u5c42\u6bd4 Encoder \u591a\u4e86\u4e00\u4e2a cross-attention\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5b83\u505a\u4e09\u4ef6\u4e8b\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li>\u5bf9\u5df2\u751f\u6210\u7684\u76ee\u6807\u8bcd\u505a masked self-attention<\/li>\n\n\n\n<li>\u770b Encoder \u8f93\u51fa\u505a cross-attention<\/li>\n\n\n\n<li>\u8fc7 FFN<\/li>\n<\/ol>\n\n\n\n<p class=\"wp-block-paragraph\">\u6700\u7ec8\u8f93\u51fa\u6bcf\u4e2a\u76ee\u6807\u4f4d\u7f6e\u7684 hidden state\u3002<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h1 class=\"wp-block-heading\">\u5341\u4e00\u3001\u4e3a\u4ec0\u4e48 Decoder \u8981\u505a mask<\/h1>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u4e2a\u95ee\u9898\u975e\u5e38\u91cd\u8981\u3002<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">1. \u8bad\u7ec3\u65f6\u4e0d\u80fd\u5077\u770b\u672a\u6765<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u6bd4\u5982\u76ee\u6807\u53e5\u5b50\u662f\uff1a<\/p>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p class=\"wp-block-paragraph\">i love you<\/p>\n<\/blockquote>\n\n\n\n<p class=\"wp-block-paragraph\">\u5f53\u9884\u6d4b <code>love<\/code> \u65f6\uff0c\u6a21\u578b\u53ea\u80fd\u770b\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><code>[BOS]<\/code><\/li>\n\n\n\n<li><code>i<\/code><\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e0d\u80fd\u5077\u770b\u540e\u9762\u7684 <code>you<\/code>\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5426\u5219\u8bad\u7ec3\u65f6\u5b83\u5c31\u4f5c\u5f0a\u4e86\u3002<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">2. \u4f60\u4ee3\u7801\u91cc\u7684 look-ahead mask<\/h2>\n\n\n\n<pre class=\"wp-block-code\"><code>generate_square_subsequent_mask\n<\/code><\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u751f\u6210\u7684\u662f\u4e00\u4e2a\u4e0a\u4e09\u89d2\u906e\u7f69\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u610f\u601d\u662f\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u5f53\u524d\u4f4d\u7f6e\u53ea\u80fd\u770b\u81ea\u5df1\u548c\u524d\u9762<\/li>\n\n\n\n<li>\u4e0d\u80fd\u770b\u540e\u9762<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u5c31\u4fdd\u8bc1\u4e86\u81ea\u56de\u5f52\u751f\u6210\u7684\u56e0\u679c\u6027\u3002<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">3. \u8fd8\u6709 padding mask<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u53e5\u5b50\u957f\u77ed\u4e0d\u540c\uff0c\u9700\u8981 pad\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4f46 pad \u53ea\u662f\u5360\u4f4d\uff0c\u4e0d\u662f\u771f\u5b9e\u8bcd\uff0c\u6240\u4ee5\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>attention \u4e0d\u80fd\u5173\u6ce8 pad<\/li>\n\n\n\n<li>loss \u8ba1\u7b97\u4e5f\u4e0d\u80fd\u628a pad \u7b97\u8fdb\u53bb<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u6240\u4ee5\u4f60\u7684\u9879\u76ee\u91cc\u6709\u4e24\u7c7b mask\uff1a<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">attention mask<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u7528\u4e8e\u6ce8\u610f\u529b\u91cc\u5c4f\u853d\u4e0d\u8be5\u770b\u7684\u4f4d\u7f6e<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">padding mask<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u7528\u4e8e\u635f\u5931\u51fd\u6570\u91cc\u4e0d\u8ba1\u7b97 pad \u7684\u635f\u5931<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u4e2a\u4f60 notebook \u91cc\u5176\u5b9e\u5df2\u7ecf\u603b\u7ed3\u5f97\u5f88\u597d\u3002<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h1 class=\"wp-block-heading\">\u5341\u4e8c\u3001TransformerModel \u6574\u4f53\u600e\u4e48\u4e32\u8d77\u6765<\/h1>\n\n\n\n<p class=\"wp-block-paragraph\">\u4f60\u53ef\u4ee5\u628a <code>TransformerModel.forward()<\/code> \u770b\u6210\u6574\u4e2a\u6a21\u578b\u7684\u603b\u63a7\u5ba4\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5b83\u505a\u7684\u4e8b\u662f\uff1a<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">1. \u6784\u9020 mask<\/h3>\n\n\n\n<ul class=\"wp-block-list\">\n<li>encoder \u7684 padding mask<\/li>\n\n\n\n<li>decoder \u7684 look-ahead mask<\/li>\n\n\n\n<li>decoder \u7684 padding mask<\/li>\n\n\n\n<li>cross-attention mask<\/li>\n<\/ul>\n\n\n\n<h3 class=\"wp-block-heading\">2. \u505a encoder embedding<\/h3>\n\n\n\n<pre class=\"wp-block-code\"><code>encoder_inputs_embeds = self.src_embedding(encoder_inputs)\n<\/code><\/pre>\n\n\n\n<h3 class=\"wp-block-heading\">3. \u9001\u5165 encoder<\/h3>\n\n\n\n<pre class=\"wp-block-code\"><code>encoder_outputs = self.encoder(...)\n<\/code><\/pre>\n\n\n\n<h3 class=\"wp-block-heading\">4. \u505a decoder embedding<\/h3>\n\n\n\n<pre class=\"wp-block-code\"><code>decoder_inputs_embeds = self.trg_embedding(decoder_inputs)\n<\/code><\/pre>\n\n\n\n<h3 class=\"wp-block-heading\">5. \u9001\u5165 decoder<\/h3>\n\n\n\n<pre class=\"wp-block-code\"><code>decoder_outputs = self.decoder(...)\n<\/code><\/pre>\n\n\n\n<h3 class=\"wp-block-heading\">6. \u7ebf\u6027\u6620\u5c04\u5230\u8bcd\u8868<\/h3>\n\n\n\n<pre class=\"wp-block-code\"><code>logits = self.linear(...)\n<\/code><\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u6700\u540e\u5f97\u5230\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>&#91;batch_size, trg_len, vocab_size]\n<\/code><\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u5b83\u8868\u793a\uff1a<\/p>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p class=\"wp-block-paragraph\">\u6bcf\u4e2a\u4f4d\u7f6e\uff0c\u5bf9\u8bcd\u8868\u91cc\u6bcf\u4e2a\u8bcd\u7684\u9884\u6d4b\u5206\u6570\u3002<\/p>\n<\/blockquote>\n\n\n\n<p class=\"wp-block-paragraph\">\u6bd4\u5982\u7b2c 3 \u4e2a\u4f4d\u7f6e\uff0c\u5b83\u4f1a\u7ed9\u51fa\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><code>i<\/code>: 0.1<\/li>\n\n\n\n<li><code>love<\/code>: 2.8<\/li>\n\n\n\n<li><code>you<\/code>: 1.2<\/li>\n\n\n\n<li>&#8230;<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u5206\u6570\u6700\u9ad8\u7684\u8bcd\u5c31\u662f\u6a21\u578b\u5f53\u524d\u8ba4\u4e3a\u6700\u53ef\u80fd\u7684\u8f93\u51fa\u3002<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h1 class=\"wp-block-heading\">\u5341\u4e09\u3001\u4e3a\u4ec0\u4e48\u8f93\u51fa\u5c42\u8981\u63a5 Linear<\/h1>\n\n\n\n<p class=\"wp-block-paragraph\">\u56e0\u4e3a Decoder \u8f93\u51fa\u7684\u662f hidden state\uff0c\u4e0d\u662f\u5177\u4f53\u5355\u8bcd\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8981\u53d8\u6210\u201c\u9884\u6d4b\u54ea\u4e2a\u8bcd\u201d\uff0c\u5fc5\u987b\u6620\u5c04\u5230\u8bcd\u8868\u5927\u5c0f\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>hidden_size -&gt; vocab_size\n<\/code><\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u6240\u4ee5\u9700\u8981\u6700\u540e\u4e00\u4e2a\u7ebf\u6027\u5c42\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5982\u679c\u5f00\u542f embedding sharing\uff0c\u8fd8\u4f1a\u628a\u8f93\u51fa\u5c42\u548c embedding \u6743\u91cd\u5171\u4eab\uff0c\u8fd9\u6837\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u53c2\u6570\u66f4\u5c11<\/li>\n\n\n\n<li>\u6709\u65f6\u6cdb\u5316\u66f4\u597d<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u4f60\u8fd9\u91cc\u914d\u7f6e\u91cc\u662f\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>\"share_embedding\": False\n<\/code><\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u8bf4\u660e\u5f53\u524d\u5b9e\u9a8c\u6ca1\u6709\u5171\u4eab\u3002<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h1 class=\"wp-block-heading\">\u5341\u56db\u3001\u635f\u5931\u51fd\u6570\u662f\u600e\u4e48\u8bbe\u8ba1\u7684<\/h1>\n\n\n\n<p class=\"wp-block-paragraph\">\u4f60\u7684 <code>CrossEntropyWithPadding<\/code> \u5f88\u6807\u51c6\uff0c\u4e5f\u5f88\u91cd\u8981\u3002<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">1. \u4e3a\u4ec0\u4e48\u7528\u4ea4\u53c9\u71b5<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u7ffb\u8bd1\u672c\u8d28\u4e0a\u662f\uff1a<\/p>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p class=\"wp-block-paragraph\">\u6bcf\u4e2a\u4f4d\u7f6e\u505a\u4e00\u6b21\u591a\u5206\u7c7b<\/p>\n<\/blockquote>\n\n\n\n<p class=\"wp-block-paragraph\">\u56e0\u4e3a\u6bcf\u4e2a\u4f4d\u7f6e\u90fd\u8981\u4ece\u6574\u4e2a\u8bcd\u8868\u91cc\u9009\u4e00\u4e2a\u8bcd\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u6240\u4ee5\u81ea\u7136\u7528\u4ea4\u53c9\u71b5\u3002<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">2. \u4e3a\u4ec0\u4e48\u8981\u5ffd\u7565 padding<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u56e0\u4e3a <code>[PAD]<\/code> \u4e0d\u662f\u771f\u5b9e\u8bcd\uff0c\u5982\u679c\u7b97\u8fdb\u53bb\u4f1a\u8bef\u5bfc\u6a21\u578b\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u6240\u4ee5\u4f60\u505a\u4e86\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>padding_mask = 1 - padding_mask.reshape(-1)\nloss = torch.mul(loss, padding_mask).sum() \/ padding_mask.sum()\n<\/code><\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u610f\u601d\u5c31\u662f\uff1a<\/p>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p class=\"wp-block-paragraph\">\u53ea\u5bf9\u771f\u5b9e token \u6c42\u5e73\u5747\u635f\u5931\u3002<\/p>\n<\/blockquote>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">3. \u4ec0\u4e48\u662f Label Smoothing<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u4f60\u7684\u914d\u7f6e\u91cc\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>\"label_smoothing\": 0.1\n<\/code><\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u4e5f\u662f\u9762\u8bd5\u9ad8\u9891\u70b9\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u6b63\u5e38 one-hot \u6807\u7b7e\u662f\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u6b63\u786e\u7c7b\u522b\u6982\u7387 = 1<\/li>\n\n\n\n<li>\u5176\u4ed6\u7c7b\u522b = 0<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">Label smoothing \u4f1a\u628a\u5b83\u7a0d\u5fae\u201c\u62b9\u5e73\u201d\u4e00\u70b9\uff0c\u6bd4\u5982\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u6b63\u786e\u7c7b\u522b\u4e0d\u662f 1\uff0c\u800c\u662f 0.9 \u5de6\u53f3<\/li>\n\n\n\n<li>\u5176\u4f59\u7c7b\u522b\u5206\u4e00\u70b9\u5c0f\u6982\u7387<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u4f5c\u7528\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u9632\u6b62\u6a21\u578b\u8fc7\u5ea6\u81ea\u4fe1<\/li>\n\n\n\n<li>\u63d0\u9ad8\u6cdb\u5316\u80fd\u529b<\/li>\n\n\n\n<li>\u5bf9\u7ffb\u8bd1\u4efb\u52a1\u7ecf\u5e38\u6709\u5e2e\u52a9<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u4f60\u53ef\u4ee5\u7b80\u5355\u8bb2\u6210\uff1a<\/p>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p class=\"wp-block-paragraph\">\u4e0d\u8ba9\u6a21\u578b\u628a\u7b54\u6848\u770b\u5f97\u592a\u7edd\u5bf9\uff0c\u800c\u662f\u7559\u4e00\u70b9\u4e0d\u786e\u5b9a\u6027\u3002<\/p>\n<\/blockquote>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h1 class=\"wp-block-heading\">\u5341\u4e94\u3001\u5b66\u4e60\u7387\u4e3a\u4ec0\u4e48\u7528 Noam Scheduler<\/h1>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u662f Transformer \u7ecf\u5178\u8bbe\u7f6e\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4f60\u7684 <code>NoamDecayScheduler<\/code> \u662f\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u524d\u671f warmup\uff0c\u5b66\u4e60\u7387\u9010\u6e10\u5347\u9ad8<\/li>\n\n\n\n<li>\u540e\u671f\u518d\u6309\u6b65\u6570\u8870\u51cf<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u516c\u5f0f\u4e0d\u7528\u6b7b\u8bb0\uff0c\u7406\u89e3\u73b0\u8c61\u5c31\u591f\uff1a<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">\u524d\u671f\u4e3a\u4ec0\u4e48\u5347\u9ad8\uff1f<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u6a21\u578b\u521a\u5f00\u59cb\u8bad\u7ec3\u5f88\u4e0d\u7a33\u5b9a\uff0c\u76f4\u63a5\u5927\u6b65\u66f4\u65b0\u5bb9\u6613\u70b8\uff0c\u6240\u4ee5\u5148\u6162\u6162\u70ed\u8eab\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">\u540e\u671f\u4e3a\u4ec0\u4e48\u4e0b\u964d\uff1f<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u540e\u9762\u9700\u8981\u66f4\u7ec6\u81f4\u5730\u6536\u655b\uff0c\u6240\u4ee5\u9010\u6b65\u51cf\u5c0f\u5b66\u4e60\u7387\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4f60\u53ef\u4ee5\u6982\u62ec\u6210\uff1a<\/p>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p class=\"wp-block-paragraph\">\u524d\u671f\u5148\u8bd5\u63a2\u7740\u5b66\uff0c\u540e\u671f\u518d\u7a33\u7a33\u5730\u7cbe\u8c03\u3002<\/p>\n<\/blockquote>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h1 class=\"wp-block-heading\">\u5341\u516d\u3001\u8bad\u7ec3\u5faa\u73af\u5728\u505a\u4ec0\u4e48<\/h1>\n\n\n\n<p class=\"wp-block-paragraph\">\u4f60\u7684 <code>training()<\/code> \u51fd\u6570\u5c31\u662f\u6807\u51c6\u8bad\u7ec3\u6d41\u7a0b\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u6bcf\u4e2a batch\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li>\u53d6\u6570\u636e<\/li>\n\n\n\n<li>\u524d\u5411\u4f20\u64ad<\/li>\n\n\n\n<li>\u7b97 loss<\/li>\n\n\n\n<li><code>loss.backward()<\/code><\/li>\n\n\n\n<li><code>optimizer.step()<\/code><\/li>\n\n\n\n<li><code>scheduler.step()<\/code><\/li>\n\n\n\n<li>\u6bcf\u9694\u4e00\u6bb5\u65f6\u95f4\u505a\u9a8c\u8bc1<\/li>\n\n\n\n<li>\u4fdd\u5b58 checkpoint<\/li>\n<\/ol>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u5c31\u662f\u6df1\u5ea6\u5b66\u4e60\u8bad\u7ec3\u6700\u6838\u5fc3\u7684\u6a21\u677f\u3002<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h1 class=\"wp-block-heading\">\u5341\u4e03\u3001\u8fd9\u4e2a\u9879\u76ee\u4e2d\u7684\u201c\u9a8c\u8bc1\u201d\u548c\u201c\u6d4b\u8bd5\u201d\u662f\u4ec0\u4e48\u610f\u601d<\/h1>\n\n\n\n<p class=\"wp-block-paragraph\">\u5f88\u591a\u540c\u5b66\u5bb9\u6613\u6df7\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">\u8bad\u7ec3\u96c6 train<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u7528\u6765\u66f4\u65b0\u53c2\u6570<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">\u9a8c\u8bc1\u96c6 val<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u8bc4\u4f30\u6a21\u578b\u6548\u679c\uff0c\u4e0d\u66f4\u65b0\u53c2\u6570<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e3b\u8981\u7528\u4e8e\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u770b\u6a21\u578b\u6709\u6ca1\u6709\u53d8\u597d<\/li>\n\n\n\n<li>\u8c03\u53c2<\/li>\n\n\n\n<li>\u65e9\u505c<\/li>\n\n\n\n<li>\u4fdd\u5b58\u6700\u4f18\u6a21\u578b<\/li>\n<\/ul>\n\n\n\n<h3 class=\"wp-block-heading\">\u6d4b\u8bd5\u96c6 test<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u6700\u7ec8\u62a5\u544a\u6548\u679c\uff0c\u53ea\u505a\u6700\u7ec8\u68c0\u9a8c<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h1 class=\"wp-block-heading\">\u5341\u516b\u3001\u63a8\u7406\u9636\u6bb5\u548c\u8bad\u7ec3\u9636\u6bb5\u6709\u4ec0\u4e48\u533a\u522b<\/h1>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u4e2a\u662f\u7406\u89e3 Transformer \u9879\u76ee\u5fc5\u987b\u660e\u767d\u7684\u3002<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">1. \u8bad\u7ec3\u9636\u6bb5<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u8bad\u7ec3\u65f6\u4f60\u628a\u5b8c\u6574\u76ee\u6807\u53e5\u5b50\u5582\u7ed9 decoder\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u56e0\u4e3a\u6709 teacher forcing\uff0c\u6a21\u578b\u6bcf\u4e00\u6b65\u90fd\u80fd\u770b\u5230\u6b63\u786e\u524d\u7f00\u3002<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">2. \u63a8\u7406\u9636\u6bb5<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u63a8\u7406\u65f6\u6ca1\u6709\u771f\u5b9e\u76ee\u6807\u53e5\u5b50\uff0c\u53ea\u80fd\u81ea\u5df1\u4e00\u6b65\u4e00\u6b65\u751f\u6210\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li>\u5148\u8f93\u5165 <code>[BOS]<\/code><\/li>\n\n\n\n<li>\u9884\u6d4b\u7b2c\u4e00\u4e2a\u8bcd<\/li>\n\n\n\n<li>\u628a\u7b2c\u4e00\u4e2a\u8bcd\u63a5\u5230\u540e\u9762<\/li>\n\n\n\n<li>\u518d\u9884\u6d4b\u7b2c\u4e8c\u4e2a\u8bcd<\/li>\n\n\n\n<li>\u4e00\u76f4\u5faa\u73af\uff0c\u76f4\u5230 <code>[EOS]<\/code><\/li>\n<\/ol>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u53eb <strong>\u81ea\u56de\u5f52\u751f\u6210<\/strong>\u3002<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">3. \u4f60\u5f53\u524d\u6d4b\u8bd5\u4ee3\u7801\u6709\u4e2a\u9700\u8981\u6ce8\u610f\u7684\u70b9<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u4f60\u6d4b\u8bd5\u4ee3\u7801\u91cc\u8fd9\u4e00\u6bb5\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>outputs = model(\n    encoder_inputs=encoder_inputs,\n    decoder_inputs=decoder_inputs,\n    encoder_inputs_mask=encoder_inputs_mask\n)\npreds = outputs.logits.argmax(dim=-1)\n<\/code><\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u91cc\u5b9e\u9645\u4e0a\u66f4\u50cf\u662f\u5728\u505a\u201c\u5e26 teacher forcing \u7684\u9884\u6d4b\u201d\uff0c\u4e0d\u662f\u4e25\u683c\u610f\u4e49\u4e0a\u7684\u81ea\u7531\u751f\u6210\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u56e0\u4e3a\u4f60\u628a <code>decoder_inputs<\/code> \u4e5f\u4f20\u8fdb\u53bb\u4e86\uff0c\u8bf4\u660e decoder \u770b\u5230\u4e86\u771f\u5b9e\u76ee\u6807\u524d\u7f00\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u6240\u4ee5\u8fd9\u4e2a\u7ed3\u679c\u66f4\u50cf\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u6d4b\u8bd5\u96c6\u4e0a\u7684\u9010\u4f4d\u7f6e\u9884\u6d4b<\/li>\n\n\n\n<li>\u4e0d\u5b8c\u5168\u7b49\u4e8e\u771f\u5b9e\u90e8\u7f72\u65f6\u7684\u81ea\u7531\u7ffb\u8bd1<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u5982\u679c\u4f60\u8981\u4e25\u683c\u505a\u7ffb\u8bd1\u63a8\u7406\uff0c\u5e94\u8be5\u8c03\u7528\u4f60\u6a21\u578b\u91cc\u7684 <code>infer()<\/code>\uff0c\u8ba9\u5b83\u81ea\u5df1\u4e00\u6b65\u4e00\u6b65\u751f\u6210\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u4e2a\u70b9\u4f60\u5728\u590d\u8bd5\u91cc\u8bb2\u51fa\u6765\uff0c\u4f1a\u663e\u5f97\u4f60\u7406\u89e3\u5f88\u7ec6\u3002<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h1 class=\"wp-block-heading\">\u5341\u4e5d\u3001BLEU \u662f\u4ec0\u4e48\uff0c\u4e3a\u4ec0\u4e48\u7528\u5b83<\/h1>\n\n\n\n<p class=\"wp-block-paragraph\">\u673a\u5668\u7ffb\u8bd1\u91cc\u5e38\u7528 BLEU\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5b83\u8861\u91cf\u7684\u662f\uff1a<\/p>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p class=\"wp-block-paragraph\">\u6a21\u578b\u7ffb\u8bd1\u7ed3\u679c\u548c\u53c2\u8003\u7b54\u6848\uff0c\u5728 n-gram \u5c42\u9762\u6709\u591a\u63a5\u8fd1<\/p>\n<\/blockquote>\n\n\n\n<p class=\"wp-block-paragraph\">\u6bd4\u5982\u53c2\u8003\u53e5\u5b50\uff1a<\/p>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p class=\"wp-block-paragraph\">i love you<\/p>\n<\/blockquote>\n\n\n\n<p class=\"wp-block-paragraph\">\u9884\u6d4b\u53e5\u5b50\uff1a<\/p>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p class=\"wp-block-paragraph\">i like you<\/p>\n<\/blockquote>\n\n\n\n<p class=\"wp-block-paragraph\">\u867d\u7136\u4e0d\u662f\u5b8c\u5168\u4e00\u6837\uff0c\u4f46\u6709\u4e0d\u5c11\u91cd\u5408\uff0cBLEU \u4f1a\u7ed9\u4e00\u5b9a\u5206\u6570\u3002<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">\u4f60\u4ee3\u7801\u91cc\u8981\u6ce8\u610f\u7684\u4e00\u70b9<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u4f60\u8fd9\u91cc\u7528\u4e86\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>weights=(1, 0, 0, 0)\n<\/code><\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u5b9e\u9645\u4e0a\u66f4\u50cf <strong>BLEU-1<\/strong>\uff0c\u53ea\u770b 1-gram\uff0c\u4e0d\u662f\u6807\u51c6\u7684 BLEU-4\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u6240\u4ee5\u5982\u679c\u4f60\u5728\u9762\u8bd5\u91cc\u8bf4\uff1a<\/p>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p class=\"wp-block-paragraph\">\u6211\u5728 notebook \u91cc\u5148\u5b9e\u73b0\u4e86\u57fa\u4e8e unigram \u7684 BLEU \u7edf\u8ba1\uff0c\u4f5c\u4e3a\u4e00\u4e2a\u7b80\u5316\u7248\u672c\uff1b\u4e25\u683c\u6765\u8bf4\uff0c\u6807\u51c6\u673a\u5668\u7ffb\u8bd1\u8bc4\u4f30\u66f4\u5e38\u7528 BLEU-4\u3002<\/p>\n<\/blockquote>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u4f1a\u663e\u5f97\u4f60\u5f88\u8bda\u5b9e\uff0c\u4e5f\u5f88\u4e13\u4e1a\u3002<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h1 class=\"wp-block-heading\">\u4e8c\u5341\u3001\u8fd9\u4e2a\u9879\u76ee\u6700\u503c\u5f97\u4f60\u771f\u6b63\u7406\u89e3\u7684 8 \u4e2a\u70b9<\/h1>\n\n\n\n<p class=\"wp-block-paragraph\">\u5982\u679c\u4f60\u60f3\u201c\u5145\u5206\u7406\u89e3\u201d\uff0c\u6211\u5efa\u8bae\u4f60\u628a\u4e0b\u9762 8 \u70b9\u5403\u900f\u3002<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">1. \u4e3a\u4ec0\u4e48\u8981\u6709 Encoder \u548c Decoder<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u56e0\u4e3a\u7ffb\u8bd1\u662f seq2seq \u4efb\u52a1\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>Encoder \u8d1f\u8d23\u7f16\u7801\u6e90\u8bed\u8a00<\/li>\n\n\n\n<li>Decoder \u8d1f\u8d23\u751f\u6210\u76ee\u6807\u8bed\u8a00<\/li>\n<\/ul>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">2. Self-Attention \u5230\u5e95\u5728\u5e72\u4ec0\u4e48<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u8ba9\u6bcf\u4e2a\u4f4d\u7f6e\u52a8\u6001\u5173\u6ce8\u53e5\u5185\u5176\u4ed6\u4f4d\u7f6e\uff0c\u5f97\u5230\u4e0a\u4e0b\u6587\u8868\u793a\u3002<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">3. Multi-Head \u4e3a\u4ec0\u4e48\u6709\u6548<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e0d\u540c\u5934\u5b66\u4e60\u4e0d\u540c\u5173\u7cfb\uff0c\u8868\u8fbe\u80fd\u529b\u66f4\u5f3a\u3002<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">4. \u4e3a\u4ec0\u4e48\u8981\u4f4d\u7f6e\u7f16\u7801<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u6ce8\u610f\u529b\u672c\u8eab\u4e0d\u61c2\u987a\u5e8f\uff0c\u4f4d\u7f6e\u7f16\u7801\u8865\u5145\u987a\u5e8f\u4fe1\u606f\u3002<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">5. \u4e3a\u4ec0\u4e48 Decoder \u8981 mask<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u9632\u6b62\u9884\u6d4b\u65f6\u5077\u770b\u672a\u6765\u8bcd\uff0c\u4fdd\u8bc1\u81ea\u56de\u5f52\u751f\u6210\u3002<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">6. Cross-Attention \u4e3a\u4ec0\u4e48\u91cd\u8981<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u8ba9\u76ee\u6807\u8bed\u8a00\u751f\u6210\u65f6\u53c2\u8003\u6e90\u8bed\u8a00\u4fe1\u606f\uff0c\u8fd9\u662f\u7ffb\u8bd1\u6210\u7acb\u7684\u5173\u952e\u3002<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">7. \u4e3a\u4ec0\u4e48 padding \u4e0d\u80fd\u53c2\u4e0e loss<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">pad \u4e0d\u662f\u6709\u6548\u8bcd\uff0c\u4e0d\u5e94\u8be5\u5f71\u54cd\u8bad\u7ec3\u3002<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">8. \u4e3a\u4ec0\u4e48\u8981\u7528 Noam + warmup<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">Transformer \u5bf9\u5b66\u4e60\u7387\u6bd4\u8f83\u654f\u611f\uff0c\u8fd9\u79cd\u7b56\u7565\u66f4\u7a33\u5b9a\u3002<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h1 class=\"wp-block-heading\">\u4e8c\u5341\u4e00\u3001\u4f60\u53ef\u4ee5\u628a\u6574\u4e2a\u6a21\u578b\u60f3\u6210\u4e00\u4e2a\u73b0\u5b9e\u7c7b\u6bd4<\/h1>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u4e2a\u7c7b\u6bd4\u7279\u522b\u9002\u5408\u5e2e\u52a9\u4f60\u771f\u6b63\u8bb0\u4f4f\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5047\u8bbe\u4f60\u5728\u505a\u201c\u5fb7\u8bed\u542c\u5199\u7ffb\u8bd1\u201d\uff1a<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">Encoder<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u50cf\u4e00\u4e2a\u5fb7\u8bed\u9605\u8bfb\u7406\u89e3\u8001\u5e08\uff0c\u628a\u6574\u53e5\u5fb7\u8bed\u8bfb\u61c2\u5e76\u505a\u597d\u7b14\u8bb0\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">Decoder<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u50cf\u4e00\u4e2a\u82f1\u8bed\u5199\u4f5c\u8001\u5e08\uff0c\u6839\u636e\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u81ea\u5df1\u5df2\u7ecf\u5199\u51fa\u7684\u524d\u6587<\/li>\n\n\n\n<li>\u5fb7\u8bed\u8001\u5e08\u7ed9\u7684\u7b14\u8bb0<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e00\u6b65\u4e00\u6b65\u5199\u51fa\u82f1\u6587\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">Self-Attention<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u5199\u67d0\u4e2a\u8bcd\u65f6\uff0c\u56de\u5934\u770b\u4e0a\u4e0b\u6587\u91cc\u54ea\u4e9b\u8bcd\u6700\u76f8\u5173\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">Cross-Attention<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u5199\u82f1\u6587\u65f6\uff0c\u53bb\u53c2\u8003\u5fb7\u8bed\u539f\u53e5\u5bf9\u5e94\u7684\u4fe1\u606f\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">Mask<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u5199\u7b2c 3 \u4e2a\u8bcd\u65f6\uff0c\u4e0d\u80fd\u5077\u770b\u7b2c 4 \u4e2a\u8bcd\u3002<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h1 class=\"wp-block-heading\">\u4e8c\u5341\u4e8c\u3001\u4f60\u8fd9\u4e2a\u9879\u76ee\u91cc\u6709\u54ea\u4e9b\u4eae\u70b9<\/h1>\n\n\n\n<p class=\"wp-block-paragraph\">\u4ece\u201c\u9879\u76ee\u7406\u89e3\u201d\u548c\u201c\u590d\u8bd5\u8868\u8fbe\u201d\u89d2\u5ea6\u770b\uff0c\u4eae\u70b9\u6709\u8fd9\u4e9b\uff1a<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">1. \u4e0d\u662f\u76f4\u63a5\u8c03\u7528\u73b0\u6210 Transformer\uff0c\u800c\u662f\u624b\u5199\u4e86\u6838\u5fc3\u6a21\u5757<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u5305\u62ec\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>Embedding<\/li>\n\n\n\n<li>Positional Encoding<\/li>\n\n\n\n<li>Multi-Head Attention<\/li>\n\n\n\n<li>Transformer Block<\/li>\n\n\n\n<li>Encoder \/ Decoder<\/li>\n\n\n\n<li>Mask<\/li>\n\n\n\n<li>Scheduler<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u8bf4\u660e\u4f60\u662f\u201c\u7406\u89e3\u539f\u7406\u540e\u5b9e\u73b0\u201d\uff0c\u4e0d\u662f\u53ea\u4f1a\u8c03\u5e93\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">2. \u6709 batch sampler \u4f18\u5316<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u6309 token \u6570\u52a8\u6001\u5206 batch\uff0c\u4f53\u73b0\u5de5\u7a0b\u610f\u8bc6\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">3. \u6709\u5b8c\u6574\u8bad\u7ec3\u6d41\u7a0b<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u5305\u62ec\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>loss<\/li>\n\n\n\n<li>optimizer<\/li>\n\n\n\n<li>scheduler<\/li>\n\n\n\n<li>tensorboard<\/li>\n\n\n\n<li>checkpoint<\/li>\n<\/ul>\n\n\n\n<h3 class=\"wp-block-heading\">4. \u6709\u63a8\u7406\u548c\u53ef\u89c6\u5316\u610f\u8bc6<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd8\u52a0\u5165\u4e86\u6ce8\u610f\u529b\u70ed\u529b\u56fe\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">5. \u6709 BLEU \u8bc4\u4f30<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u8bf4\u660e\u4f60\u8003\u8651\u4e86\u673a\u5668\u7ffb\u8bd1\u7684\u4efb\u52a1\u6307\u6807\u3002<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h1 class=\"wp-block-heading\">\u4e8c\u5341\u4e09\u3001\u8fd9\u4e2a\u9879\u76ee\u91cc\u4f60\u8fd8\u8981\u7279\u522b\u6ce8\u610f\u7684\u51e0\u4e2a\u7ec6\u8282<\/h1>\n\n\n\n<p class=\"wp-block-paragraph\">\u6211\u4e5f\u987a\u624b\u5e2e\u4f60\u6307\u51fa\u4e00\u4e9b\u4f60\u540e\u9762\u53ef\u4ee5\u7ee7\u7eed\u5b8c\u5584\u7684\u5730\u65b9\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">1. <code>TransformerEmbedding.forward()<\/code> \u91cc\u6709 <code>print(position_ids)<\/code><\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u8bad\u7ec3\u65f6\u4f1a\u5927\u91cf\u6253\u5370\uff0c\u5b9e\u9645\u8bad\u7ec3\u65f6\u5e94\u8be5\u5220\u6389\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">2. \u6d4b\u8bd5\u90e8\u5206\u66f4\u50cf teacher-forcing evaluation<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u5982\u679c\u8bf4\u201c\u771f\u6b63\u7684\u7ffb\u8bd1\u63a8\u7406\u201d\uff0c\u6700\u597d\u7528\u81ea\u56de\u5f52 <code>infer()<\/code>\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">3. BLEU \u8fd9\u91cc\u76ee\u524d\u66f4\u63a5\u8fd1 BLEU-1<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u5982\u679c\u8981\u66f4\u4e25\u8c28\uff0c\u6700\u597d\u7528 corpus-level BLEU \u6216\u6807\u51c6 BLEU-4\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">4. FFN \u91cc\u7528\u7684\u662f ReLU<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e5f\u53ef\u4ee5\u8003\u8651 GELU\uff0c\u4f46 ReLU \u5b8c\u5168\u6ca1\u95ee\u9898\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">5. \u5171\u4eab embedding \u7684\u5b9e\u9a8c\u8fd8\u53ef\u4ee5\u5c55\u5f00<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u56e0\u4e3a\u4f60 notebook \u91cc\u63d0\u5230\u4e86\u5171\u4eab\u548c\u4e0d\u5171\u4eab\u3002<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h1 class=\"wp-block-heading\">\u4e8c\u5341\u56db\u3001\u5982\u679c\u4f60\u8981\u5411\u8001\u5e08\u8bb2\u8fd9\u4e2a\u9879\u76ee\uff0c\u53ef\u4ee5\u8fd9\u6837\u8bf4<\/h1>\n\n\n\n<p class=\"wp-block-paragraph\">\u4f60\u53ef\u4ee5\u76f4\u63a5\u7528\u4e0b\u9762\u8fd9\u6bb5\u601d\u8def\u56de\u7b54\uff1a<\/p>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p class=\"wp-block-paragraph\">\u8fd9\u4e2a\u9879\u76ee\u662f\u4e00\u4e2a\u57fa\u4e8e Transformer \u7684\u5fb7\u8bed\u5230\u82f1\u8bed\u673a\u5668\u7ffb\u8bd1\u7cfb\u7edf\u3002\u6574\u4f53\u6d41\u7a0b\u5305\u62ec\u6570\u636e\u52a0\u8f7d\u3001BPE \u5206\u8bcd\u540e\u7684 token \u7f16\u7801\u3001\u52a8\u6001 batch \u6784\u9020\u3001Transformer \u6a21\u578b\u642d\u5efa\u3001\u8bad\u7ec3\u3001\u9a8c\u8bc1\u4ee5\u53ca\u57fa\u4e8e BLEU \u7684\u8bc4\u4f30\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u6a21\u578b\u7ed3\u6784\u4e0a\uff0c\u6211\u624b\u52a8\u5b9e\u73b0\u4e86\u8bcd\u5d4c\u5165\u548c\u4f4d\u7f6e\u7f16\u7801\u3001\u591a\u5934\u6ce8\u610f\u529b\u3001\u524d\u9988\u7f51\u7edc\u3001\u6b8b\u5dee\u8fde\u63a5\u548c\u5c42\u5f52\u4e00\u5316\uff0c\u5e76\u5206\u522b\u6784\u5efa\u4e86 Encoder \u548c Decoder\u3002Encoder \u4e3b\u8981\u901a\u8fc7 self-attention \u5bf9\u6e90\u8bed\u8a00\u53e5\u5b50\u8fdb\u884c\u4e0a\u4e0b\u6587\u5efa\u6a21\uff0cDecoder \u5219\u5728 masked self-attention \u7684\u57fa\u7840\u4e0a\uff0c\u901a\u8fc7 cross-attention \u878d\u5408\u6e90\u8bed\u8a00\u4fe1\u606f\uff0c\u9010\u6b65\u751f\u6210\u76ee\u6807\u8bed\u8a00\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8bad\u7ec3\u65f6\u91c7\u7528\u5e26 padding mask \u7684\u4ea4\u53c9\u71b5\u635f\u5931\uff0c\u5e76\u7ed3\u5408 label smoothing \u548c Noam \u5b66\u4e60\u7387\u8c03\u5ea6\u63d0\u5347\u8bad\u7ec3\u7a33\u5b9a\u6027\u3002\u5de5\u7a0b\u4e0a\uff0c\u6211\u8fd8\u5b9e\u73b0\u4e86\u6309\u53e5\u957f\u52a8\u6001\u7ec4 batch \u7684\u91c7\u6837\u5668\uff0c\u4ee5\u51cf\u5c11 padding \u5e26\u6765\u7684\u8ba1\u7b97\u6d6a\u8d39\u3002\u6700\u540e\uff0c\u5728\u6d4b\u8bd5\u9636\u6bb5\u4f7f\u7528 BLEU \u6307\u6807\u5bf9\u7ffb\u8bd1\u7ed3\u679c\u8fdb\u884c\u8bc4\u4f30\uff0c\u5e76\u652f\u6301\u6ce8\u610f\u529b\u70ed\u529b\u56fe\u53ef\u89c6\u5316\u3002<\/p>\n<\/blockquote>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u6bb5\u4f60\u57fa\u672c\u80cc\u719f\uff0c\u590d\u8bd5\u5c31\u5f88\u7a33\u3002<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h1 class=\"wp-block-heading\">\u4e8c\u5341\u4e94\u3001\u6700\u540e\u6211\u5e2e\u4f60\u505a\u4e00\u4e2a\u201c\u6700\u901a\u4fd7\u7248\u603b\u7ed3\u201d<\/h1>\n\n\n\n<p class=\"wp-block-paragraph\">\u4f60\u53ef\u4ee5\u628a\u8fd9\u4e2a\u9879\u76ee\u538b\u7f29\u6210\u4e0b\u9762\u8fd9\u51e0\u53e5\u8bdd\uff1a<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">\u6700\u901a\u4fd7\u7684\u4e00\u53e5\u8bdd<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u8fd9\u4e2a\u9879\u76ee\u5c31\u662f\u6559\u8ba1\u7b97\u673a\u628a\u5fb7\u8bed\u53e5\u5b50\u7ffb\u8bd1\u6210\u82f1\u8bed\u53e5\u5b50\u3002<\/strong><\/p>\n\n\n\n<h3 class=\"wp-block-heading\">\u5b83\u662f\u600e\u4e48\u505a\u5230\u7684<\/h3>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u5148\u628a\u5355\u8bcd\u53d8\u6210\u6570\u5b57<\/li>\n\n\n\n<li>\u518d\u8ba9\u6a21\u578b\u901a\u8fc7\u6ce8\u610f\u529b\u673a\u5236\u7406\u89e3\u53e5\u5b50\u91cc\u8bcd\u548c\u8bcd\u4e4b\u95f4\u7684\u5173\u7cfb<\/li>\n\n\n\n<li>\u7528 Encoder \u8bfb\u61c2\u5fb7\u8bed<\/li>\n\n\n\n<li>\u7528 Decoder \u4e00\u6b65\u6b65\u751f\u6210\u82f1\u8bed<\/li>\n\n\n\n<li>\u7528\u6b63\u786e\u7b54\u6848\u76d1\u7763\u8bad\u7ec3<\/li>\n\n\n\n<li>\u6700\u540e\u7528 BLEU \u770b\u7ffb\u8bd1\u5f97\u597d\u4e0d\u597d<\/li>\n<\/ul>\n\n\n\n<h3 class=\"wp-block-heading\">Transformer \u6700\u6838\u5fc3\u662f\u4ec0\u4e48<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u4e0d\u662f\u6309\u987a\u5e8f\u6b7b\u8bb0\uff0c\u800c\u662f\u8ba9\u6bcf\u4e2a\u8bcd\u90fd\u80fd\u770b\u5168\u53e5\u91cc\u6700\u91cd\u8981\u7684\u8bcd\u3002<\/strong><\/p>\n\n\n\n<h3 class=\"wp-block-heading\">\u8fd9\u4e2a\u9879\u76ee\u6700\u5173\u952e\u7684\u4ee3\u7801\u6a21\u5757<\/h3>\n\n\n\n<ul class=\"wp-block-list\">\n<li><code>LangPairDataset<\/code>\uff1a\u8bfb\u6570\u636e<\/li>\n\n\n\n<li><code>Tokenizer<\/code>\uff1a\u6587\u672c\u8f6c id<\/li>\n\n\n\n<li><code>TransformerBatchSampler<\/code>\uff1a\u52a8\u6001\u7ec4 batch<\/li>\n\n\n\n<li><code>TransformerEmbedding<\/code>\uff1a\u8bcd\u5411\u91cf + \u4f4d\u7f6e\u7f16\u7801<\/li>\n\n\n\n<li><code>MultiHeadAttention<\/code>\uff1a\u591a\u5934\u6ce8\u610f\u529b<\/li>\n\n\n\n<li><code>TransformerBlock<\/code>\uff1aattention + FFN<\/li>\n\n\n\n<li><code>TransformerModel<\/code>\uff1a\u5b8c\u6574\u6a21\u578b<\/li>\n\n\n\n<li><code>CrossEntropyWithPadding<\/code>\uff1a\u635f\u5931<\/li>\n\n\n\n<li><code>NoamDecayScheduler<\/code>\uff1a\u5b66\u4e60\u7387\u8c03\u5ea6<\/li>\n\n\n\n<li><code>training()<\/code>\uff1a\u8bad\u7ec3\u5faa\u73af<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\"><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u4e00\u3001\u5148\u628a\u6574\u4e2a\u9879\u76ee\u770b\u6210\u4e00\u4e2a\u6d41\u6c34\u7ebf \u4f60\u8fd9\u4e2a\u9879<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[],"class_list":["post-10","post","type-post","status-publish","format-standard","hentry","category-uncategorized"],"_links":{"self":[{"href":"https:\/\/www.mahui.space\/index.php\/wp-json\/wp\/v2\/posts\/10","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.mahui.space\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.mahui.space\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.mahui.space\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.mahui.space\/index.php\/wp-json\/wp\/v2\/comments?post=10"}],"version-history":[{"count":1,"href":"https:\/\/www.mahui.space\/index.php\/wp-json\/wp\/v2\/posts\/10\/revisions"}],"predecessor-version":[{"id":11,"href":"https:\/\/www.mahui.space\/index.php\/wp-json\/wp\/v2\/posts\/10\/revisions\/11"}],"wp:attachment":[{"href":"https:\/\/www.mahui.space\/index.php\/wp-json\/wp\/v2\/media?parent=10"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.mahui.space\/index.php\/wp-json\/wp\/v2\/categories?post=10"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.mahui.space\/index.php\/wp-json\/wp\/v2\/tags?post=10"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}