{"id":595,"date":"2024-02-19T16:39:21","date_gmt":"2024-02-19T08:39:21","guid":{"rendered":"http:\/\/lingbo.online\/?p=595"},"modified":"2024-02-29T19:03:45","modified_gmt":"2024-02-29T11:03:45","slug":"%e5%a4%a7%e6%a8%a1%e5%9e%8b%e9%9d%a2%e8%af%95%e6%80%bb%e7%bb%93","status":"publish","type":"post","link":"https:\/\/lingbo.online\/index.php\/uncategorized\/%e5%a4%a7%e6%a8%a1%e5%9e%8b%e9%9d%a2%e8%af%95%e6%80%bb%e7%bb%93\/","title":{"rendered":"\u5927\u6a21\u578b\u9762\u8bd5\u603b\u7ed3"},"content":{"rendered":"<h1>\u57fa\u7840\u77e5\u8bc6<\/h1>\n<h2>transformer \u516b\u80a1\u6587<\/h2>\n<h3>Self-Attention\u7684\u8868\u8fbe\u5f0f<\/h3>\n<p><img decoding=\"async\" src=\"https:\/\/lingbo.online\/wp-content\/uploads\/2024\/02\/attention.jpg\" alt=\"\" \/><\/p>\n<h3>\u4e3a\u4ec0\u4e48\u5728\u4e0a\u9762\u7684\u516c\u5f0f\u4e2d\u8981\u5bf9QK\u8fdb\u884cscaling<\/h3>\n<p>sacling\u540e\u8fdb\u884csoftmax\u64cd\u4f5c\u53ef\u4ee5\u4f7f\u5f97\u8f93\u5165\u6570\u636e\u7684\u5206\u5e03\u53d8\u5f97\u5f88\u597d\uff0c\u9632\u6b62\u68af\u5ea6\u6d88\u5931\uff0c\u8ba9\u6a21\u578b\u80fd\u591f\u66f4\u5bb9\u6613\u8bad\u7ec3\u3002<\/p>\n<h3>Self-Attention\u4e00\u5b9a\u8981\u8fd9\u6837\u8868\u8fbe\u5417\uff1f<\/h3>\n<p>\u4e0d\u4e00\u5b9a\uff0c\u53ea\u8981\u80fd\u591f\u8868\u793a\u5efa\u6a21\u7684\u76f8\u5173\u6027\u5c31\u53ef\u4ee5\u3002<br \/>\n\u6700\u597d\u80fd\u591f\u9ad8\u901f\u8ba1\u7b97\uff08\u77e9\u9635\u4e58\u6cd5\uff09\uff0c\u5e76\u4e14\u8868\u8fbe\u80fd\u529b\u5f3a\uff08query\u53ef\u4ee5\u4e3b\u52a8\u5173\u6ce8\u5230\u5176\u4ed6\u7684key\u5e76\u5728value\u4e0a\u8fdb\u884c\u5f3a\u5316\uff0c\u5e76\u4e14\u5ffd\u7565\u4e0d\u76f8\u5173\u7684\u5176\u4ed6\u90e8\u5206\uff09\uff0c\u6a21\u578b\u5bb9\u91cf\u591f\uff08\u5f15\u5165\u4e86project_q\/k\/v,att_out,\u591a\u5934\uff09<\/p>\n<h3>\u6709\u5176\u4ed6\u65b9\u6cd5\u4e0d\u7528\u9664\u6839\u53f7<span class=\"katex-eq\" data-katex-display=\"false\">d_k<\/span>\u5417\uff1f<\/h3>\n<p>\u6709\uff0c\u53ea\u8981\u80fd\u7f13\u89e3\u68af\u5ea6\u6d88\u5931\u7684\u95ee\u9898\u5c31\u53ef\u4ee5<br \/>\n<span class=\"katex-eq\" data-katex-display=\"false\">d_k<\/span>\u662f\u8bcd\u5411\u91cf\/\u9690\u85cf\u5c42\u7684\u7ef4\u5ea6\uff0c\u9664\u4ee5<span class=\"katex-eq\" data-katex-display=\"false\">d_k<\/span>\u7684\u539f\u56e0\u4e3b\u8981\u7531\u4e24\u70b9\uff1a<\/p>\n<ol>\n<li>\u9996\u5148\u8981\u9664\u4ee5\u4e00\u4e2a\u6570\uff0c\u9632\u6b62\u8f93\u5165softmax\u7684\u503c\u8fc7\u5927\uff0c\u5bfc\u81f4\u504f\u5bfc\u6570\u8d8b\u8fd1\u4e8e0<\/li>\n<li>\u9009\u62e9<span class=\"katex-eq\" data-katex-display=\"false\">d_k<\/span>\u662f\u56e0\u4e3a\u53ef\u4ee5\u4f7f<strong>q*k\u7684\u7ed3\u679c\u6ee1\u8db3\u671f\u671b\u4e3a0\uff0c\u65b9\u5dee\u4e3a1\u7684\u5206\u5e03\uff0c\u7c7b\u4f3c\u4e8e\u5f52\u4e00\u5316<\/strong><\/li>\n<\/ol>\n<h3>\u4e3a\u4ec0\u4e48transformer\u7528Layer Norm\uff1f\u6709\u4ec0\u4e48\u7528\uff1f<\/h3>\n<p>\u4efb\u4f55norm\u7684\u610f\u4e49\u90fd\u662f\u4e3a\u4e86\u8ba9\u4f7f\u7528norm\u7684\u7f51\u7edc\u7684\u8f93\u5165\u7684\u6570\u636e\u5206\u5e03\u53d8\u5f97\u66f4\u597d\uff0c\u4e5f\u5c31\u662f\u8f6c\u5316\u4e3a\u6807\u51c6\u6b63\u6001\u5206\u5e03\uff0c\u6570\u503c\u8fdb\u5165\u654f\u611f\u5ea6\u533a\u95f4\uff0c\u4ee5\u51cf\u7f13\u68af\u5ea6\u6d88\u5931\uff0c\u4ece\u800c\u66f4\u52a0\u5bb9\u6613\u8bad\u7ec3\u3002\u5f53\u7136\uff0c\u8fd9\u4e5f\u610f\u5473\u7740\u820d\u5f03\u4e86\u9664\u6b64\u7ef4\u5ea6\u4e4b\u5916\u7684\u5176\u4ed6\u7ef4\u5ea6\u7684\u4fe1\u606f\u3002<\/p>\n<h3>\u4e3a\u4ec0\u4e48\u4e0d\u7528BN\uff08batch normalization\uff09<\/h3>\n<p>BN\u5e7f\u6cdb\u5e94\u7528\u4e8eCV\uff0c\u9488\u5bf9\u540c\u4e00\u4e2a\u7279\u5f81\uff0c\u4ee5\u5938\u6837\u672c\u7684\u65b9\u5f0f\u5c55\u5f00\u5f52\u4e00\u5316\uff0c\u4e5f\u5c31\u662f\u5bf9\u4e0d\u540c\u6837\u672c\u7684\u540c\u4e00channel\u95f4\u7684\u5168\u90e8\u50cf\u7d20\u503c\u8fdb\u884c\u5f52\u4e00\u5316\uff0c\u4e0d\u4f1a\u7834\u574f\u4e0d\u540c\u6837\u672c\u540c\u4e00\u7279\u5f81\u4e4b\u95f4\u7684\u5173\u7cfb\u3002\u5728NLP\u4e2d\uff1a<\/p>\n<ul>\n<li>\u5bf9\u4e0d\u540c\u6837\u672c\u7684\u540c\u4e00\u7279\u5f81\u4fe1\u606f\u8fdb\u884c\u5f52\u4e00\u5316\u6ca1\u6709\u610f\u4e49\uff0c\u5982\uff08\u4e3a\u4e2d\u534e\u4e4b\u5d1b\u8d77\u800c\u8bfb\u4e66\uff1b\u6211\u7231\u4e2d\u56fd\uff1b\u6bcd\u7231\u6700\u4f1f\u5927\uff09\u4e2d\uff0c\u201c\u4e3a\u201d\u3001\u201c\u6211\u201d\u3001\u201c\u6bcd\u201d\u5f52\u4e00\u5230\u540c\u4e00\u5206\u5e03\u6ca1\u6709\u610f\u4e49\u3002<\/li>\n<li>\u820d\u5f03\u4e0d\u4e86BN\u4e2d\u820d\u5f03\u7684\u5176\u4ed6\u7ef4\u5ea6\u7684\u4fe1\u606f\uff0c\u4e5f\u5c31\u662f\u540c\u4e00\u6837\u672c\u7684\u4e0d\u540c\u7ef4\u5ea6\u7684\u4fe1\u606f\u3002\u5982\u201c\u4e3a\u201d\u3001\u201c\u6211\u201d\u3001\u201c\u6bcd\u201d\u5f52\u4e00\u5230\u540c\u4e00\u5206\u5e03\u540e\uff0c\u7b2c\u4e00\u53e5\u8bdd\u4e2d\u7684\u201c\u4e3a\u201d\u548c\u201c\u4e2d\u201d\u5c31\u6ca1\u6709\u53ef\u6bd4\u6027\u4e86\u3002<\/li>\n<\/ul>\n<h3>Bert\u4e3a\u4ec0\u4e48\u8981\u641e\u4e00\u4e2aposition embedding\uff1f<\/h3>\n<p>\u56e0\u4e3a\u53ea\u6709\u4e4b\u524d\u63d0\u5230\u7684self-attention\u65e0\u6cd5\u8868\u8fbe\u4f4d\u7f6e\u4fe1\u606f\uff08\u5bf9\u4f4d\u7f6e\u4fe1\u606f\u4e0d\u654f\u611f\uff09\uff0c\u5982\u679c\u4e0d\u6dfb\u52a0position embedding,\u201c1+1=2\u201d\u548c\u201c1+2=1\u201d\u662f\u4e00\u6837\u7684\uff0c\u56e0\u6b64\u9700\u8981\u589e\u5f3a\u6a21\u578b\u9488\u5bf9\u4f4d\u7f6e\u7684\u8868\u8fbe\u80fd\u529b\u3002<\/p>\n<h3>Bert\u4e3a\u4ec0\u4e48\u4e09\u4e2aembedding\u53ef\u4ee5\u76f8\u52a0\uff1f<\/h3>\n<p>Bert\u7684\u4e09\u4e2aembedding\u662f\u6307token embedding,segment embedding,position embedding\u3002\u52a0\u6cd5\u4e0d\u4f1a\u5bfc\u81f4\u4fe1\u606f\u635f\u5931\uff0c\u56e0\u4e3a\u672c\u8d28\u4e0a\u795e\u7ecf\u7f51\u7edc\u4e2d\u6bcf\u4e2a\u795e\u7ecf\u5143\u6536\u5230\u7684\u4fe1\u53f7\u4e5f\u662f\u201c\u6743\u91cd\u201d\u76f8\u52a0\u5f97\u5230\u7684\u3002<\/p>\n<h3>transformers\u4e3a\u4ec0\u4e48\u8981\u7528\u4e09\u4e2a\u4e0d\u4e00\u6837\u7684QKV\uff1f<\/h3>\n<p>\u4e3a\u4e86\u589e\u5f3a\u7f51\u7edc\u7684\u5bb9\u91cf\u548c\u8868\u8fbe\u80fd\u529b\u3002\u5982\u679c\u76f4\u63a5\u7528\u8f93\u5165x\u672c\u8eab\u6765\u505a\uff0c\u8868\u5f81\u80fd\u529b\u8fc7\u5f31\u3002<\/p>\n<h3>\u4e3a\u4ec0\u4e48\u8981\u591a\u5934\uff1f\u4e3e\u4f8b\u8bf4\u660e\u591a\u5934\u76f8\u6bd4\u4e8e\u5355\u5934\u6ce8\u610f\u529b\u7684\u4f18\u52bf<\/h3>\n<p>\u8fdb\u4e00\u6b65\u589e\u5f3a\u7f51\u7edc\u7684\u5bb9\u91cf\u548c\u8868\u8fbe\u80fd\u529b\uff0c\u53ef\u4ee5\u7c7b\u6bd4CV\u4e2d\u4e0d\u540c\u7684channel\uff08\u5377\u79ef\u6838\uff09\u4f1a\u5173\u6ce8\u4e0d\u540c\u7684\u4fe1\u606f\uff0c\u4e8b\u5b9e\u4e0a\u4e0d\u540c\u7684\u5934\u4e5f\u4f1a\u5173\u6ce8\u4e0d\u540c\u7684\u4fe1\u606f\u3002<br \/>\n\u5047\u8bbe\u6211\u4eec\u6709\u4e00\u4e2a\u53e5\u5b50&quot;the cat, which is black, sat on the mat&quot;\u3002\u5728\u5904\u7406&quot;sat&quot;\u8fd9\u4e2a\u8bcd\u65f6\uff0c\u4e00\u4e2a\u5934\uff08\u4e3b\u8bed\u5934\uff09\u53ef\u80fd\u4f1a\u66f4\u5173\u6ce8&quot;cat&quot;\uff0c\u56e0\u4e3a&quot;cat&quot;\u662f&quot;sat&quot;\u7684\u4e3b\u8bed\uff1b\u53e6\u4e00\u4e2a\u5934\uff08\u5bbe\u8bed\u5934\uff09\u53ef\u80fd\u4f1a\u66f4\u5173\u6ce8&quot;on the mat&quot;\uff0c\u56e0\u4e3a\u8fd9\u662f&quot;sat&quot;\u7684\u5bbe\u8bed\uff1b\u8fd8\u6709\u4e00\u4e2a\u5934\uff08\u4fee\u9970\u5934\uff09\u53ef\u80fd\u4f1a\u5173\u6ce8&quot;which is black&quot;\uff0c\u56e0\u4e3a\u8fd9\u662f\u5bf9&quot;cat&quot;\u7684\u4fee\u9970\u3002<br \/>\n\u5f53\u7136\uff0c\u8fd9\u53ea\u662f\u4e3a\u4e86\u65b9\u4fbf\u4f60\u7406\u89e3\uff0c\u4e8b\u5b9e\u4e0a\u5c31\u548c\u5377\u79ef\u6838\u4e00\u6837\uff0c\u4e0d\u540c\u5934\u5173\u6ce8\u7684\u5185\u5bb9\u662f\u5f88\u62bd\u8c61\u7684\u3002<br \/>\n\u4f60\u5f53\u7136\u53ef\u4ee5\u5c31\u7528\u4e00\u4e2a\u5934\u540c\u65f6\u505a\u8fd9\u4e2a\u4e8b\uff0c\u4f46\u662f\u8fd8\u662f\u8fd9\u4e2a\u9053\u7406\uff0c\u6211\u4eec\u7684\u76ee\u7684\u5c31\u662f\u901a\u8fc7\u589e\u52a0\u53c2\u6570\u91cf\u6765\u589e\u5f3a\u7f51\u7edc\u7684\u5bb9\u91cf\u4ece\u800c\u63d0\u5347\u7f51\u7edc\u8868\u8fbe\u80fd\u529b\u3002<br \/>\n\u7ecf\u8fc7\u591a\u5934\u4e4b\u540e\uff0c\u6211\u4eec\u8fd8\u9700\u8981att_out\u7ebf\u6027\u5c42\u6765\u505a\u7ebf\u6027\u53d8\u6362\uff0c\u4ee5\u81ea\u52a8\u51b3\u5b9a\uff08\u901a\u8fc7\u8bad\u7ec3\uff09\u5bf9\u6bcf\u4e2a\u5934\u7684\u8f93\u51fa\u8d4b\u4e88\u591a\u5927\u7684\u6743\u91cd\uff0c\u4ece\u800c\u5728\u6700\u7ec8\u7684\u8f93\u51fa\u4e2d\u5f3a\u8c03\u4e00\u4e9b\u5934\u7684\u4fe1\u606f\uff0c\u800c\u5ffd\u89c6\u5176\u4ed6\u5934\u7684\u4fe1\u606f\u3002\u8fd9\u662f\u4e00\u79cd\u81ea\u9002\u5e94\u7684\u3001\u6570\u636e\u9a71\u52a8\u7684\u65b9\u5f0f\u6765\u7ec4\u5408\u4e0d\u540c\u5934\u7684\u4fe1\u606f\u3002<\/p>\n<h1>\u53c2\u8003\u8d44\u6599<\/h1>\n<h2>transformer<\/h2>\n<ol>\n<li><a href=\"https:\/\/www.zhihu.com\/tardis\/zm\/art\/414084879?source_id=1005\" title=\"\u6ce8\u610f\u529b\u673a\u5236\u5230\u5e95\u5728\u505a\u4ec0\u4e48\uff0cQ\/K\/V\u600e\u4e48\u6765\u7684\uff1f\u4e00\u6587\u8bfb\u61c2Attention\u6ce8\u610f\u529b\u673a\u5236\" target=\"_blank\"  rel=\"nofollow\" >\u6ce8\u610f\u529b\u673a\u5236\u5230\u5e95\u5728\u505a\u4ec0\u4e48\uff0cQ\/K\/V\u600e\u4e48\u6765\u7684\uff1f\u4e00\u6587\u8bfb\u61c2Attention\u6ce8\u610f\u529b\u673a\u5236<\/a><\/li>\n<\/ol>\n","protected":false},"excerpt":{"rendered":"<p>\u57fa\u7840\u77e5\u8bc6 transformer \u516b\u80a1\u6587 Self-Attention\u7684\u8868\u8fbe\u5f0f \u4e3a\u4ec0\u4e48\u5728\u4e0a\u9762\u7684\u516c\u5f0f\u4e2d\u8981\u5bf9QK\u8fdb\u884cscaling s &#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-595","post","type-post","status-publish","format-standard","hentry","category-uncategorized"],"_links":{"self":[{"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/posts\/595","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=595"}],"version-history":[{"count":4,"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/posts\/595\/revisions"}],"predecessor-version":[{"id":640,"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/posts\/595\/revisions\/640"}],"wp:attachment":[{"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/media?parent=595"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/categories?post=595"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/tags?post=595"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}