{"id":204,"date":"2026-10-25T19:06:40","date_gmt":"2026-10-25T11:06:40","guid":{"rendered":"https:\/\/99website.top\/blog\/?p=204"},"modified":"2026-10-25T19:06:40","modified_gmt":"2026-10-25T11:06:40","slug":"attention-is-all-you-need-4-5-tokenizer-%e4%b8%8e-embedding","status":"publish","type":"post","link":"https:\/\/99website.top\/blog\/?p=204","title":{"rendered":"Attention is all you need – #4.5 Tokenizer \u4e0e Embedding"},"content":{"rendered":"\n

Attention is all you need – #4.5 Tokenizer \u4e0e Embedding<\/h1>\n

\u524d\u8a00<\/h2>\n

OMG\u6211\u5fd8\u4e86\u8fd8\u8981\u5148\u5b66\u600e\u4e48\u5f97\u5230\u6b63\u786e\u7b54\u6848\u548c\u600e\u4e48\u628a\u8bcd\u8f6c\u6362\u6210\u5411\u91cf\u4e86\uff0c\u8fd9\u7bc7\u535a\u5ba2\u6211\u5c31\u6765\u7814\u7a76\u7814\u7a76\u563f\u563f<\/p>\n

Tokenizer<\/h2>\n

Tokenizer \u7a76\u7adf\u662f\u4ec0\u4e48\u5462\uff1f<\/p>\n

Tokenizer \u8d1f\u8d23\u5c06\u6587\u672c\u62c6\u5206\u4e3a Token\uff0c\u5e76\u628a\u8fd9\u4e9b Token \u8f6c\u6362\u4e3a\u6574\u6570 ID\u3002\u968f\u540e\uff0cEmbedding \u5c42\u6839\u636e\u8fd9\u4e9b ID\uff0c\u4ece\u53ef\u8bad\u7ec3\u7684 Embedding \u77e9\u9635\u4e2d\u53d6\u51fa\u5bf9\u5e94\u5411\u91cf\uff0c\u4f5c\u4e3a\u795e\u7ecf\u7f51\u7edc\u7684\u8f93\u5165\u3002<\/p>\n

\u6bd4\u5982\uff1a<\/p>\n

vocab = {\n    "\u6211": 0,\n    "\u7231": 1,\n    "\u4f60": 2\n}\n\ntext = "\u6211\u7231\u4f60"\ntokens = [vocab[ch] for ch in text]\n\nprint(tokens)  # [0, 1, 2]\n<\/code><\/pre>\n
\u4e0d\u8fc7\u8fd9\u6837\u4e00\u4e2a\u4e00\u4e2a\u8bbe\u5b9a\u4e5f\u592a\u7d2f\u4e86\uff0c\u6211\u4eec\u53ef\u4ee5\u76f4\u63a5\u626b\u63cf\u8bad\u7ec3\u8bed\u6599\uff0c\u6309\u7167\u5b57\u7b26\u9996\u6b21\u51fa\u73b0\u7684\u987a\u5e8f\u5efa\u7acb\u8bcd\u8868<\/p>\n
corpus = "\u6211\u7231\u4f60\u4f60\u7231\u6211"\n\nchars = list(dict.fromkeys(corpus))\n\nvocab = {ch: i for i, ch in enumerate(chars)}\n<\/code><\/pre>\n
Embedding<\/h2>\n
Embedding \u662f\u4e00\u4e2a\u5f62\u72b6\u4e3a V\\times d<\/span> \u7684\u53ef\u8bad\u7ec3\u77e9\u9635\u3002\u524d\u5411\u4f20\u64ad\u65f6\uff0c\u6839\u636e Token ID \u67e5\u627e\u5bf9\u5e94\u884c\uff0c\u5c06\u957f\u5ea6\u4e3a T<\/span> \u7684\u5e8f\u5217\u8f6c\u6362\u4e3a T\\times d<\/span> \u7684\u5411\u91cf\u77e9\u9635\u3002<\/p>\n
\u5b83\u662f\u4e00\u4e2a\u53ef\u8bad\u7ec3\u7684\u53c2\u6570\u77e9\u9635\uff0c\u7528\u6765\u5c06\u79bb\u6563\u7684 Token ID \u6620\u5c04\u4e3a\u8fde\u7eed\u7684\u5411\u91cf\u8868\u793a<\/p>\n
\n\\boxed{E\\in\\mathbb{R}^{V\\times d_{\\text{model}}}}\n<\/div>\n
\u5176\u4e2d\uff1a<\/p>\n
    \n
  • E<\/span>\uff1aEmbedding \u77e9\u9635<\/li>\n
  • V<\/span>\uff1a\u8bcd\u8868\u5927\u5c0f\uff08Vocabulary Size\uff09<\/li>\n
  • d_{\\text{model}}<\/span>\uff1a\u6bcf\u4e2a Token \u5bf9\u5e94\u7684\u5411\u91cf\u7ef4\u5ea6<\/li>\n
  • E_{i,j}<\/span>\uff1a\u7b2c i<\/span> \u4e2a Token \u7684\u5411\u91cf\u4e2d\uff0c\u7b2c j<\/span> \u4e2a\u5206\u91cf<\/li>\n<\/ul>\n
    \u77e9\u9635\u7684\u6bcf\u4e00\u884c\u5bf9\u5e94\u4e00\u4e2a Token \u7684\u5411\u91cf\u3002<\/strong><\/p>\n
    \u5047\u8bbe 99GPT v0\uff1a<\/p>\n
      \n
    • \u8bcd\u8868\u5927\u5c0f V=4<\/span><\/li>\n
    • \u5411\u91cf\u7ef4\u5ea6 d_{\\text{model}}=3<\/span><\/li>\n<\/ul>\n
      \u90a3\u4e48\uff1a<\/p>\n
      \nE= \\begin{bmatrix} 0.2 & 0.5 & -0.1\\\\ 0.8 & -0.3 & 0.6\\\\ -0.4 & 0.9 & 0.2\\\\ 0.1 & 0.7 & -0.5 \\end{bmatrix}\n<\/div>\n
      \u5047\u8bbe Token\u300c\u6211\u300d\u7684 ID \u662f 2\uff0c\u90a3\u4e48\uff1a<\/p>\n
      \n\\operatorname{Embedding}(2)=E_{2,:}\n\n\\\\\n\n=\\begin{bmatrix}-0.4&0.9&0.2\\end{bmatrix}\n<\/div>\n
      \u5982\u679c\u6709\u591a\u4e2a Token ID \uff0c\u5047\u8bbe\u8f93\u5165 Token ID \u5e8f\u5217\u662f\uff1a<\/p>\n
      \nT=[2,3,0]\n<\/div>\n
      \u76f4\u63a5\u53d6\u51fa\u5bf9\u5e94\u7684\u4e09\u884c\uff1a<\/p>\n
      \nX=E[T]\n<\/div>\n
      \u4e8e\u662f\uff1a<\/p>\n
      \n\\boxed{X\\in\\mathbb{R}^{n\\times d_{\\text{model}}}}\n<\/div>\n
      \u5176\u4e2d n<\/span> \u662f\u8f93\u5165\u5e8f\u5217\u7684 Token \u6570\u91cf\u3002<\/p>\n
      Embedding \u77e9\u9635\u4e0d\u662f\u56fa\u5b9a\u7684\u8bcd\u5178\uff0c\u800c\u662f\u901a\u8fc7\u53cd\u5411\u4f20\u64ad\u4e0d\u65ad\u5b66\u4e60\u7684\u53c2\u6570\u3002<\/strong><\/p>\n
      \u521d\u59cb\u5316\u65f6\u6bcf\u884c\u53ef\u4ee5\u662f\u968f\u673a\u5411\u91cf\uff0c\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u6839\u636e Loss \u7684\u68af\u5ea6\u66f4\u65b0\u88ab\u4f7f\u7528\u7684\u884c\uff0c\u4ece\u800c\u9010\u6e10\u5f62\u6210\u6709\u7528\u7684 Token \u8868\u793a\u3002<\/p>\n
      \u6765\u4e00\u9053\u9898~<\/p>\n
      \u5047\u8bbe 99GPT \u7684\u8bcd\u8868\u6709 100 \u4e2a Token\uff0c\u6bcf\u4e2a Token \u5bf9\u5e94 64 \u7ef4\u5411\u91cf\uff0c\u8f93\u5165\u53e5\u5b50\u88ab\u7f16\u7801\u6210 5 \u4e2a Token ID\u3002<\/strong><\/p>\n
      Embedding \u77e9\u9635 E<\/span> \u548c\u8f93\u51fa\u77e9\u9635 X<\/span> \u7684\u7ef4\u5ea6\u5206\u522b\u662f\u4ec0\u4e48\uff1f<\/p>\n
      \u7b54\u6848\u662f  E:100\\times64,\\quad X:5\\times64<\/span>  \uff01<\/p>\n
      Embedding\u7684\u53cd\u5411\u4f20\u64ad<\/h3>\n
      \u5148\u8bb0\u4f4f\u4e00\u53e5\u8bdd\uff1a<\/p>\n
      \n
      Embedding \u7684\u53cd\u5411\u4f20\u64ad\uff0c\u672c\u8d28\u4e0a\u5c31\u662f\u628a\u8f93\u5165\u5411\u91cf\u7684\u68af\u5ea6\u7d2f\u52a0\u56de Embedding \u77e9\u9635\u5bf9\u5e94\u7684\u884c\u3002<\/p>\n<\/blockquote>\n
      \u5047\u8bbe\uff1a<\/p>\n
      \nE= \\begin{bmatrix} 0.2&0.5\\\\ 0.8&0.3\\\\ 0.4&0.9 \\end{bmatrix}\n<\/div>\n
      \u8f93\u5165 Token ID \u4e3a\uff1a<\/p>\n
      \nT=[0,2]\n<\/div>\n
      \u90a3\u4e48\u524d\u5411\u4f20\u64ad\uff1a<\/p>\n
      \nX=E[T]= \\begin{bmatrix} 0.2&0.5\\\\ 0.4&0.9 \\end{bmatrix}\n<\/div>\n
      \u524d\u5411\u4f20\u64ad\u4e4b\u540e\u4e0d\u662f\u5f97\u8fd4\u56de\u68af\u5ea6\u561b\uff0c\u68af\u5ea6\u8fd9\u65f6\u5019\u56de\u6765\u4e86\uff0c\u6700\u7ec8\u4f20\u7ed9 Embedding \u8f93\u51fa X<\/span> \u7684\u68af\u5ea6\u4e3a\uff1a<\/p>\n
      \n\\frac{\\partial L}{\\partial X}=\n\\begin{bmatrix}\n0.6&-0.2\\\\\n0.1&0.4\n\\end{bmatrix}\n<\/div>\n
      \u56e0\u4e3a X<\/span> \u7684\u7b2c 0 \u884c\u6765\u81ea E<\/span> \u7684\u7b2c 0 \u884c\uff0c\u7b2c 1 \u884c\u6765\u81ea E<\/span> \u7684\u7b2c 2 \u884c\uff0c\u56e0\u6b64\u76f4\u63a5\u628a\u68af\u5ea6\u653e\u56de\u5bf9\u5e94\u4f4d\u7f6e\uff1a<\/p>\n
      \n\\boxed{\n\\frac{\\partial L}{\\partial E}=\n\\begin{bmatrix}\n0.6&-0.2\\\\\n0&0\\\\\n0.1&0.4\n\\end{bmatrix}\n}\n<\/div>\n
      \u4ece\u94fe\u5f0f\u6cd5\u5219\u6765\u7406\u89e3<\/h4>\n
      \u56de\u5fc6\u4f60\u4e4b\u524d\u5b66\u4e60\u7684\u795e\u7ecf\u7f51\u7edc\uff1a<\/p>\n
      \n\\frac{\\partial L}{\\partial w} = \\frac{\\partial L}{\\partial x} \\frac{\\partial x}{\\partial w}\n<\/div>\n
      Embedding \u4e5f\u662f\u4e00\u6837\u3002<\/p>\n
      \u5bf9\u4e8e X_{k,j}=E_{T_k,j}<\/span>\uff0c\u6709\uff1a<\/p>\n
      \n\\frac{\\partial L}{\\partial E_{i,j}}\n=\n\\sum_{k:T_k=i}\n\\frac{\\partial L}{\\partial X_{k,j}}\n<\/div>\n
      \u8bf4\u4eba\u8bdd\u5c31\u662f<\/p>\n
      \u67d0\u4e2a Token ID \u5728\u8f93\u5165\u4e2d\u51fa\u73b0\u51e0\u6b21\uff0c\u5c31\u628a\u90a3\u4e9b\u4f4d\u7f6e\u4f20\u56de\u6765\u7684\u68af\u5ea6\u5168\u90e8\u7d2f\u52a0\u5230\u540c\u4e00\u884c\u3002<\/strong><\/p>\n
      \u6bd4\u5982\uff1a<\/p>\n
      \nT=[0,0,2]\n<\/div>\n
      \u5982\u679c\u524d\u4e24\u4e2a\u4f4d\u7f6e\u7684\u68af\u5ea6\u5206\u522b\u4e3a\uff1a<\/p>\n
      \n[0.2,0.3],\\qquad[0.5,-0.1]\n<\/div>\n
      \u90a3\u4e48E<\/span> \u7684\u7b2c 0 \u884c\u68af\u5ea6\u5c31\u662f\uff1a<\/p>\n
      \n[0.2,0.3]+[0.5,-0.1]=[0.7,0.2]\n<\/div>\n
      \u6700\u540e\u548c\u795e\u7ecf\u7f51\u7edc\u4e00\u6837\uff0c\u901a\u8fc7\u5b66\u4e60\u7387\u66f4\u65b0\uff1a<\/p>\n
      \n\\boxed{E_{\\text{new}}=E-\\eta\\frac{\\partial L}{\\partial E}}\n<\/div>\n
      \u8fd9\u5c31\u662f\u57fa\u672c\u7684Tokenizer\u548cEmbedding\uff01\uff01\uff01<\/p>\n","protected":false},"excerpt":{"rendered":"
      Attention is all you need – #4.5 Tokenizer \u4e0e Embedding \u524d\u8a00 OMG\u6211\u5fd8\u4e86\u8fd8\u8981\u5148\u5b66\u600e\u4e48\u5f97\u5230\u6b63\u786e\u7b54\u6848\u548c\u600e\u4e48\u628a\u8bcd\u8f6c\u6362\u6210\u5411\u91cf\u4e86\uff0c\u8fd9\u7bc7\u535a\u5ba2\u6211\u5c31\u6765\u7814\u7a76\u7814\u7a76\u563f\u563f Tokenizer Tokenizer \u7a76\u7adf\u662f\u4ec0\u4e48\u5462\uff1f Tokenizer \u8d1f\u8d23\u5c06\u6587\u672c\u62c6\u5206\u4e3a Token\uff0c\u5e76\u628a\u8fd9\u4e9b Token \u8f6c\u6362\u4e3a\u6574\u6570 ID\u3002\u968f\u540e\uff0cEmbedding \u5c42\u6839\u636e\u8fd9\u4e9b  […]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[25,23,26,24],"tags":[],"class_list":["post-204","post","type-post","status-publish","format-standard","hentry","category-ai","category-transformer","category-26","category-24"],"_links":{"self":[{"href":"https:\/\/99website.top\/blog\/index.php?rest_route=\/wp\/v2\/posts\/204","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/99website.top\/blog\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/99website.top\/blog\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/99website.top\/blog\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/99website.top\/blog\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=204"}],"version-history":[{"count":0,"href":"https:\/\/99website.top\/blog\/index.php?rest_route=\/wp\/v2\/posts\/204\/revisions"}],"wp:attachment":[{"href":"https:\/\/99website.top\/blog\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=204"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/99website.top\/blog\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=204"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/99website.top\/blog\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=204"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}