{"id":19628,"date":"2024-05-28T16:33:11","date_gmt":"2024-05-28T08:33:11","guid":{"rendered":"https:\/\/www.aisharenet.com\/?p=19628"},"modified":"2025-01-28T16:42:10","modified_gmt":"2025-01-28T08:42:10","slug":"transformer","status":"publish","type":"post","link":"https:\/\/www.kdjingpai.com\/ja\/transformer\/","title":{"rendered":"\u4ec0\u4e48\u662f Transformer\uff1f"},"content":{"rendered":"<p><img decoding=\"async\" title=\"\u4ec0\u4e48\u662f Transformer\uff1f-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/36469c77fcfc8d0.jpg\" alt=\"\u4ec0\u4e48\u662f Transformer\uff1f-1\" \/><\/p>\n<p>&nbsp;<\/p>\n<p>Transformer \u662f\u4e00\u79cd\u7528\u4e8e\u81ea\u7136\u8bed\u8a00\u5904\u7406 (NLP) \u7684\u6df1\u5ea6\u5b66\u4e60\u6a21\u578b\u67b6\u6784, \u7531 Vaswani \u7b49\u4eba\u5728 2017 \u5e74\u63d0\u51fa. \u5b83\u4e3b\u8981\u7528\u4e8e\u5904\u7406\u5e8f\u5217\u5230\u5e8f\u5217\u7684\u4efb\u52a1, \u5982\u673a\u5668\u7ffb\u8bd1, \u6587\u672c\u751f\u6210\u7b49.<\/p>\n<p>\u7b80\u5355\u6765\u8bb2, \u6587\u672c\u751f\u6210\u7684 Transformer \u6a21\u578b\u7684\u539f\u7406\u662f\u2014\u2014&#8221;\u9884\u6d4b\u4e0b\u4e00\u4e2a\u8bcd&#8221;.<\/p>\n<p>\u7528\u6237\u7ed9\u5b9a\u7684\u6587\u672c (prompt), \u6a21\u578b\u4f1a\u9884\u6d4b\u4e0b\u4e00\u4e2a\u8bcd\u6700\u6709\u53ef\u80fd\u662f\u4ec0\u4e48. Transformer \u7684\u6838\u5fc3\u521b\u65b0\u548c\u5f3a\u5927\u4e4b\u5904\u5728\u4e8e\u5b83\u4f7f\u7528\u7684\u81ea\u6ce8\u610f\u529b\u673a\u5236\uff08self-attention mechanism\uff09, \u8fd9\u4f7f\u5f97\u5b83\u4eec\u80fd\u591f\u5904\u7406\u6574\u4e2a\u5e8f\u5217, \u5e76\u6bd4\u4e4b\u524d\u7684\u67b6\u6784 (RNN) \u66f4\u6709\u6548\u5730\u6355\u6349\u957f\u8ddd\u79bb\u4f9d\u8d56\u5173\u7cfb.<\/p>\n<p>\u53e6\u5916\u9700\u8981\u4f4f\u7684\u662f, GitHub \u4e0a\u7684 huggingface\/transformers \u662f HuggingFace \u5b9e\u73b0\u7684 Transformer \u6a21\u578b\u5e93, \u5305\u62ec\u4e86 Transformer \u7684\u5b9e\u73b0\u548c\u5927\u91cf\u7684\u9884\u8bad\u7ec3\u6a21\u578b.<\/p>\n<p>\u76ee\u524d\u7684 LLM \u57fa\u672c\u90fd\u57fa\u4e8e Transformer \u67b6\u6784, \u5e76\u5bf9\u5176\u8fdb\u884c\u4f18\u5316\u6280\u672f\u548c\u8bad\u7ec3\u65b9\u6cd5\u7684\u6539\u8fdb.<\/p>\n<p>&nbsp;<\/p>\n<h2>Transformer \u7684\u7ed3\u6784<\/h2>\n<p>\u6bcf\u4e2a\u6587\u672c\u751f\u6210 Transformer \u90fd\u7531\u4ee5\u4e0b\u4e09\u4e2a\u5173\u952e\u7ec4\u4ef6\u6784\u6210\uff1a<\/p>\n<h3><strong>\u5d4c\u5165\u5c42\uff08Embedding\uff09<\/strong>\u00a0\uff1a<\/h3>\n<ul>\n<li>\u6587\u672c\u8f93\u5165\u88ab\u5206\u5272\u6210\u79f0\u4e3a\u8bcd\u5143\uff08token\uff09\u7684\u66f4\u5c0f\u5355\u4f4d, \u53ef\u4ee5\u662f\u5355\u8bcd\u6216\u5b50\u8bcd<\/li>\n<li>\u8fd9\u4e9b\u8bcd\u5143\u88ab\u8f6c\u6362\u6210\u79f0\u4e3a\u5d4c\u5165\uff08embeddings\uff09\u7684\u6570\u503c\u5411\u91cf<\/li>\n<li>\u8fd9\u4e9b\u5d4c\u5165\u5411\u91cf\u80fd\u591f\u6355\u6349\u8bcd\u8bed\u7684\u8bed\u4e49\u542b\u4e49<\/li>\n<\/ul>\n<h3><strong>Transformer \u5757<\/strong>\u00a0\uff1a<\/h3>\n<p>\u8fd9\u662f\u6a21\u578b\u5904\u7406\u548c\u8f6c\u6362\u8f93\u5165\u6570\u636e\u7684\u57fa\u672c\u6784\u5efa\u5355\u5143. \u6bcf\u4e2a\u5757\u5305\u62ec\uff1a<\/p>\n<ul>\n<li><strong>\u6ce8\u610f\u529b\u673a\u5236\uff08Attention Mechanism\uff09<\/strong>\u00a0\uff1a\n<ul>\n<li>Transformer \u5757\u7684\u6838\u5fc3\u7ec4\u4ef6<\/li>\n<li>\u5141\u8bb8\u8bcd\u5143\u4e4b\u95f4\u76f8\u4e92\u901a\u4fe1<\/li>\n<li>\u6355\u6349\u8bcd\u8bed\u4e4b\u95f4\u7684\u4e0a\u4e0b\u6587\u4fe1\u606f\u548c\u5173\u7cfb<\/li>\n<\/ul>\n<\/li>\n<li><strong>\u591a\u5c42\u611f\u77e5\u5668\uff08MLP\uff09\u5c42<\/strong>\u00a0\uff1a\n<ul>\n<li>\u4e00\u4e2a\u524d\u9988\u7f51\u7edc, \u72ec\u7acb\u5904\u7406\u6bcf\u4e2a\u8bcd\u5143<\/li>\n<li>\u6ce8\u610f\u529b\u5c42\u7684\u76ee\u6807\u662f\u5728\u8bcd\u5143\u4e4b\u95f4\u8def\u7531\u4fe1\u606f<\/li>\n<li>MLP \u7684\u76ee\u6807\u662f\u4f18\u5316\u6bcf\u4e2a\u8bcd\u5143\u7684\u8868\u793a<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<h3><strong>\u8f93\u51fa\u6982\u7387\uff08Output Probabilities\uff09<\/strong>\u00a0\uff1a<\/h3>\n<ul>\n<li>\u6700\u7ec8\u7684\u7ebf\u6027\u5c42\u548c softmax \u5c42<\/li>\n<li>\u5c06\u5904\u7406\u540e\u7684\u5d4c\u5165\u8f6c\u6362\u4e3a\u6982\u7387<\/li>\n<li>\u4f7f\u6a21\u578b\u80fd\u591f\u9884\u6d4b\u5e8f\u5217\u4e2d\u7684\u4e0b\u4e00\u4e2a\u8bcd\u5143<\/li>\n<\/ul>\n<p>&nbsp;<\/p>\n<h2>Transformer \u7684\u4f18\u70b9\uff1a<\/h2>\n<ul>\n<li><strong>\u5e76\u884c\u5316\u5904\u7406<\/strong>\u00a0\uff1a\u4e0e RNN \u4e0d\u540c, Transformer \u4e0d\u9700\u8981\u6309\u987a\u5e8f\u5904\u7406\u6570\u636e, \u56e0\u6b64\u53ef\u4ee5\u66f4\u597d\u5730\u5229\u7528 GPU \u8fdb\u884c\u5e76\u884c\u8ba1\u7b97, \u63d0\u9ad8\u8bad\u7ec3\u901f\u5ea6.<\/li>\n<li><strong>\u957f\u8ddd\u79bb\u4f9d\u8d56<\/strong>\u00a0\uff1a\u81ea\u6ce8\u610f\u529b\u673a\u5236\u4f7f\u5f97 Transformer \u80fd\u591f\u6709\u6548\u6355\u6349\u5e8f\u5217\u4e2d\u8fdc\u8ddd\u79bb\u7684\u4f9d\u8d56\u5173\u7cfb.<\/li>\n<li><strong>\u7075\u6d3b\u6027<\/strong>\u00a0\uff1aTransformer \u53ef\u4ee5\u5f88\u5bb9\u6613\u5730\u6269\u5c55\u5230\u66f4\u5927\u7684\u6a21\u578b (\u5982BERT\u3001GPT\u7b49) , \u5e76\u5728\u591a\u79cd NLP \u4efb\u52a1\u4e2d\u8868\u73b0\u51fa\u8272.<\/li>\n<\/ul>\n<p>&nbsp;<\/p>\n<h2>Transformer \u7684\u7f3a\u70b9\uff1a<\/h2>\n<ul>\n<li><strong>\u8ba1\u7b97\u590d\u6742\u5ea6\u9ad8<\/strong>\u00a0\uff1a\u81ea\u6ce8\u610f\u529b\u673a\u5236\u7684\u8ba1\u7b97\u590d\u6742\u5ea6\u4e3aO(n^2), \u5f53\u8f93\u5165\u5e8f\u5217\u957f\u5ea6\u8f83\u957f\u65f6, \u8ba1\u7b97\u8d44\u6e90\u6d88\u8017\u8f83\u5927.<\/li>\n<li><strong>\u6570\u636e\u9700\u6c42\u5927<\/strong>\u00a0\uff1aTransformer \u901a\u5e38\u9700\u8981\u5927\u91cf\u7684\u6570\u636e\u8fdb\u884c\u8bad\u7ec3, \u4ee5\u4fbf\u5145\u5206\u53d1\u6325\u5176\u6027\u80fd.<\/li>\n<li><strong>\u7f3a\u4e4f\u5185\u5728\u7684\u5e8f\u5217\u4fe1\u606f<\/strong>\u00a0\uff1a\u7531\u4e8e\u6ca1\u6709\u5185\u7f6e\u7684\u5e8f\u5217\u5904\u7406\u673a\u5236 (\u5982 RNN \u4e2d\u7684\u65f6\u95f4\u6b65) , \u9700\u8981\u989d\u5916\u7684\u673a\u5236 (\u5982\u4f4d\u7f6e\u7f16\u7801) \u6765\u5f15\u5165\u5e8f\u5217\u4fe1\u606f.<\/li>\n<\/ul>\n<p>&nbsp;<\/p>\n<h2>Refs<\/h2>\n<p>\u8fd9\u4e24\u4e2a\u5f15\u7528\u4e2d, \u7b2c\u4e00\u4e2a\u662f\u5173\u4e8e transformer \u7684\u6700\u7ecf\u5178\u8bba\u6587, Attention Is All You Need. \u6b63\u662f\u8fd9\u7bc7\u8bba\u6587\u9996\u6b21\u63d0\u51fa\u4e86 Transformer.<\/p>\n<p>\u7b2c\u4e8c\u4e2a\u662f\u5173\u4e8e transformer \u7684\u53ef\u89c6\u5316\u89e3\u91ca, \u53ef\u4ee5\u5f88\u76f4\u89c2\u7684\u770b\u5230 transformer \u7684\u5185\u90e8\u7ed3\u6784.<\/p>\n<ul>\n<li><a href=\"https:\/\/arxiv.org\/abs\/1706.03762\">Attention Is All You Need<\/a><\/li>\n<li><a href=\"https:\/\/poloclub.github.io\/transformer-explainer\/\">Transformer Explainer<\/a><\/li>\n<li><a href=\"https:\/\/ig.ft.com\/generative-ai\/\">\u89c6\u89c9\u5316\u8bb2\u89e3<\/a><\/li>\n<\/ul>\n","protected":false},"excerpt":{"rendered":"<p>&nbsp; Transformer \u662f\u4e00\u79cd\u7528\u4e8e\u81ea\u7136\u8bed\u8a00\u5904\u7406 (NLP) \u7684\u6df1\u5ea6\u5b66\u4e60\u6a21\u578b\u67b6\u6784, \u7531 Vaswani \u7b49\u4eba\u5728 2017 \u5e74\u63d0\u51fa. \u5b83\u4e3b\u8981\u7528\u4e8e\u5904\u7406\u5e8f\u5217\u5230\u5e8f\u5217\u7684\u4efb\u52a1, \u5982\u673a\u5668\u7ffb\u8bd1, \u6587\u672c\u751f\u6210\u7b49. \u7b80\u5355\u6765\u8bb2, \u6587\u672c\u751f\u6210\u7684 Trans&#8230;<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[34],"tags":[],"class_list":["post-19628","post","type-post","status-publish","format-standard","hentry","category-knowledge"],"_links":{"self":[{"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/posts\/19628","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/comments?post=19628"}],"version-history":[{"count":0,"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/posts\/19628\/revisions"}],"wp:attachment":[{"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/media?parent=19628"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/categories?post=19628"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/tags?post=19628"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}