{"id":585,"date":"2025-06-28T22:13:35","date_gmt":"2025-06-28T14:13:35","guid":{"rendered":"https:\/\/blogqian.com\/?p=585"},"modified":"2025-07-02T22:16:02","modified_gmt":"2025-07-02T14:16:02","slug":"%e5%a4%9a%e6%a8%a1%e6%80%81%e7%bb%8f%e5%85%b8%e8%ae%ba%e6%96%87%e6%95%b4%e7%90%86-i-vit","status":"publish","type":"post","link":"https:\/\/blogqian.com\/index.php\/2025\/06\/28\/%e5%a4%9a%e6%a8%a1%e6%80%81%e7%bb%8f%e5%85%b8%e8%ae%ba%e6%96%87%e6%95%b4%e7%90%86-i-vit\/","title":{"rendered":"\u591a\u6a21\u6001\u7ecf\u5178\u8bba\u6587\u6574\u7406 I – ViT"},"content":{"rendered":"\n

\ud83c\udf89 \u8fd9\u7bc7\u5b9e\u5728\u662f\u592a\u7ecf\u5178\u4e86<\/p>\n\n\n\n

\u4f5c\u8005\u5355\u4f4d<\/h1>\n\n\n\n

Google Research<\/p>\n\n\n\n

\u8bba\u6587\u80cc\u666f\u548c\u95ee\u9898<\/h1>\n\n\n\n

Transformer\u6a21\u578b\u67b6\u6784\u5df2\u7ecf\u5728\u8bed\u8a00\u9886\u57df\u53d1\u5149\u53d1\u70ed\uff0c\u6709\u975e\u5e38\u826f\u597d\u7684\u6027\u8d28\uff08\u8bad\u7ec3\u6548\u7387\u9ad8\uff0c\u53ef\u4ee5\u901a\u8fc7\u6ce8\u610f\u529b\u673a\u5236\u6765\u7406\u89e3\u590d\u6742\u8bed\u4e49\uff0c\u4e14\u7ed3\u6784\u7b80\u5355\uff0c\u53ef\u4ee5\u81ea\u7531\u62d3\u5c55\u6a21\u578b\u5927\u5c0f\uff0c\u5343\u4ebf\u7ea7\u53c2\u6570\u4e0b\u90fd\u6ca1\u6709\u51fa\u73b0\u6027\u80fd\u9971\u548c\uff0c\u6f5c\u529b\u6781\u5927\uff09\u65e9\u5728\u8fd9\u7bc7\u8bba\u6587\u4e4b\u524d\uff0c\u6709\u8bb8\u591a\u4eba\u505a\u8fc7\u5c06CV\u7ed3\u5408\u4e0a\u6ce8\u610f\u529b\u673a\u5236[1]<\/a>[2]<\/a>\uff0c\u4f46\u8981\u4e48\u5c31\u662f\u89c4\u6a21\u592a\u5c0f\u6027\u80fd\u4e0d\u597d\uff0c\u8981\u4e48\u5c31\u662f\u4fee\u6539\u4e86transformer\u7684\u673a\u5236\u3002\u56e0\u6b64\uff0c\u5bf9\u4e8e\u539f\u7248transformer\u7684\u5728CV\u9886\u57df\u7684\u901a\u7528\u6027\u8feb\u5728\u7709\u776b\u3002<\/p>\n\n\n\n

\u8bba\u6587\u52a8\u673a\u548c\u8d21\u732e\u89e3\u8bfb<\/h1>\n\n\n\n

\u9a8c\u8bc1\u4e86\u5728\u4e0d\u5bf9transformer\u505a\u4efb\u4f55\u4fee\u6539\u7684\u60c5\u51b5\u4e0b\uff0c\u89e3\u51b3CV\u95ee\u9898\u7684\u53ef\u884c\u6027\uff1b<\/p>\n\n\n\n

\u7814\u7a76\u4e86\u5728\u4e0d\u540c\u5927\u5c0f\u7684\u6570\u636e\u96c6\u4e0a\uff0cTransformer\u548c\u5377\u79ef\u795e\u7ecf\u7f51\u7edc\u4e4b\u95f4\u7684\u6027\u80fd\u5bf9\u6bd4\uff1b<\/p>\n\n\n\n

\u65b9\u6848\u8bbe\u8ba1\u5206\u6790<\/h1>\n\n\n\n

\u672c\u6587\u7684\u5929\u624d\u60f3\u6cd5\u662f\uff0c\u9996\u5148\u5c06\u4e00\u4e2a\u56fe\u7247\u6309\u7167\u56fa\u5b9a\u5927\u5c0f\u5206\u6210\u4e00\u4e2a\u4e00\u4e2apatch\uff08\u539f\u6587\u4e2d\u662f16*16\uff0c\u8fd9\u4e2a\u5730\u65b9\u5176\u5b9e\u53ef\u4ee5\u7406\u89e3\u4e3a\u521b\u7acb\u82e5\u5e72\u4e2a16 *16\u7684\u5377\u79ef\u6838\uff0c\u5176\u6b65\u957f\u4e5f\u4e3a16\uff0cpadding\u4e3avalid\uff09\uff0c\u518d\u5c06\u5176\u5c55\u5e73\u540e\u5e26\u5165\u4e00\u4e2aLinear\u5c42\u8fdb\u884cembedding\u3002<\/p>\n\n\n\n

\u7136\u540e\u518d\u7ed9\u6bcf\u4e2a\u4f4d\u7f6e\uff0c\u52a0\u4e0a\u4e00\u4e2a\u53ef\u5b66\u4e60\u7684\u4f4d\u7f6e\u7f16\u7801\u3002\u518d\u5728\u6700\u524d\u9762\u52a0\u4e0a\u4e00\u4e2a\u7528\u4e8e\u5206\u7c7b\u7684token\uff08\u5f53\u7136\u8fd9\u91cc\u4f7f\u7528\u7684\u662fencoder\u67b6\u6784\uff0c\u4e5f\u5c31\u662f\u6ca1\u6709mask\uff0c\u5e8f\u5217\u4e2d\u4efb\u4f55\u4f4d\u7f6e\u7684token\u90fd\u80fd\u770b\u5230\u6240\u6709token\uff0c\u6240\u4ee5\u8f93\u51fatoken\u653e\u5728\u5f00\u5934\u662f\u6ca1\u6709\u95ee\u9898\u7684\uff09<\/p>\n\n\n\n

\u6700\u540e\uff0c\u7ed9\u7b2c\u4e00\u4e2a\u5206\u7c7btoken\u52a0\u4e0a\u4e00\u4e2aMLP\u7684\u5934\uff0c\u8fdb\u884c\u5206\u7c7b\u3002<\/p>\n\n\n\n

\"\"<\/div><\/figure>\n\n\n\n

\u5b9e\u9a8c\u6548\u679c<\/h1>\n\n\n\n
\"\"<\/div><\/figure>\n\n\n\n

\u8fd9\u91cc\u4f5c\u8005\u91c7\u53d6\u4e86\u5f53\u65f6\u5728\u4e0d\u540c\u6570\u636e\u96c6\u4e0a\u8868\u73b0\u6700\u597d\u7684\u5206\u7c7b\u6a21\u578b\uff08ResNet\u548cEfficientNet\uff09\u8fdb\u884c\u6bd4\u8f83\uff0c\u53ef\u4ee5\u770b\u5230ViT-H\/14\u57fa\u672c\u53d6\u5f97\u4e86\u66f4\u597d\u7684\u7ed3\u679c\u3002<\/p>\n\n\n\n

\u4e14\u6700\u540e\u4e00\u884c\u8bf4\u660e\uff0cVit\u6a21\u578b\u5177\u6709\u8f83\u9ad8\u7684\u8ba1\u7b97\u6548\u7387\u4f18\u52bf\u3002<\/p>\n\n\n\n

\u63a5\u4e0b\u6765\uff0c\u4f5c\u8005\u53c8\u5bf9\u4f4d\u7f6e\u7f16\u7801\u8fdb\u884c\u4e86\u5c1d\u8bd5\uff08\u5206\u4e3a1-D\u548c2-D\u4e24\u79cd\u4f4d\u7f6e\u7f16\u7801\uff0c1-D\u5373\u5c55\u5e73\uff0c2-D\u5373\u6309\u884c\u6309\u5217\u8fdb\u884c\u5206\u5272\uff09<\/p>\n\n\n\n

\"\"<\/div><\/figure>\n\n\n\n

\u53ef\u4ee5\u770b\u5230\u6ca1\u6709\u4f4d\u7f6e\u7f16\u7801\u7684\u6548\u679c\u662f\u6700\u5dee\u7684\uff0c\u5176\u4ed6\u76841-D 2-D\u548c\u76f8\u5bf9\u4f4d\u7f6e\u7f16\u7801\u90fd\u5dee\u4e0d\u591a\u3002<\/p>\n\n\n\n

\u8fd9\u91cc\u4f5c\u8005\u8fdb\u884c\u4e86\u66f4\u6df1\u5165\u7684\u7814\u7a76\uff0c\u4e3a\u4ec0\u4e481-D 2-D\u5dee\u4e0d\u591a\uff0c\u662f\u56e0\u4e3a\u5b9e\u9645\u4e0a1-D\u4e5f\u80fd\u591f\u5728\u4e0d\u65ad\u7684\u8bad\u7ec3\u4e2d\u5305\u542b\u4e8c\u7ef4\u7684\u4f4d\u7f6e\u4fe1\u606f\uff1a<\/p>\n\n\n\n

\"\"<\/div><\/figure>\n\n\n\n

\u6b64\u5916\uff0c\u4f5c\u8005\u8fd8\u7814\u7a76\u4e86\u5934\u5728\u4e0d\u540c\u6df1\u5ea6\u4e0b\u7684\u8868\u73b0\uff0c\u53ef\u4ee5\u770b\u5230\u5728\u6d45\u5c42\uff0c\u90e8\u5206\u5934\u5173\u6ce8\u8fd1\u8ddd\u79bb\u7684\u50cf\u7d20\uff0c\u90e8\u5206\u5173\u6ce8\u8fdc\u7aef\u50cf\u7d20\uff1b\u800c\u5230\u4e86\u6df1\u5c42\u7f51\u7edc\u5219\u57fa\u672c\u5f00\u59cb\u5173\u6ce8\u8fdc\u7aef\u50cf\u7d20\u4e86\uff08\u5168\u5c40\u4fe1\u606f\uff09\u3002<\/p>\n\n\n\n

\"\"<\/div><\/figure>\n\n\n\n

\u5bf9\u4e8e\u6a21\u578b\u7ed3\u6784\u7684\u5c1d\u8bd5\uff0c\u4f5c\u8005\u9009\u62e9\u4e86\u4e09\u79cd\uff1a\u7eafCNN\uff0c\u7eafTransformer\uff0c\u4e24\u8005\u7ed3\u5408\u3002\u53ef\u4ee5\u770b\u5230\uff0c\u5728\u76f8\u540c\u7684\u9884\u8bad\u7ec3\u4ee3\u4ef7\u4e0b\uff0c\u521a\u5f00\u59cb\u6df7\u5408\u6a21\u578b\u4f1a\u5b58\u5728\u4f18\u52bf\uff0c\u4f46\u662f\u968f\u7740\u9884\u8bad\u7ec3\u4ee3\u4ef7\u7684\u589e\u52a0\uff0c\u53ef\u4ee5\u770b\u5230Transformer\u8fd8\u662f\u66f4\u4f18\u7684<\/p>\n\n\n\n

\"\"<\/div><\/figure>\n\n\n\n

\u63a5\u4e0b\u6765\uff0c\u4f5c\u8005\u7814\u7a76\u4e86\u6570\u636e\u96c6\u5927\u5c0f\u5bf9\u6a21\u578b\u7684\u5f71\u54cd\uff1a<\/p>\n\n\n\n

\"\"<\/div><\/figure>\n\n\n\n

\u53ef\u4ee5\u770b\u5230\u5728\u5c0f\u89c4\u6a21\u6570\u636e\u96c6\u4e0aResNet\u66f4\u4f18\u79c0\uff0c\u800cViT\u5219\u5728\u66f4\u5927\u7684\u6570\u636e\u96c6\u4e0a\u4f18\u4e8eResNet\u3002<\/p>\n\n\n\n

\u7ed3\u8bba<\/h1>\n\n\n\n
    \n
  1. transformer\u6a21\u578b\u53ef\u4ee5\u4e0d\u505a\u4efb\u4f55\u4fee\u6539\u6765\u89e3\u51b3\u8ba1\u7b97\u673a\u89c6\u89c9\u95ee\u9898\u3002<\/li>\n\n\n\n
  2. \u5c0f\u89c4\u6a21\u6570\u636e\u96c6\u4e0a\u8868\u73b0\u7565\u900a\u4e8e\u5377\u79ef\u795e\u7ecf\u7f51\u7edc\uff0c\u800c\u5728\u4e2d\u7b49\u548c\u5927\u89c4\u6a21\u6570\u636e\u96c6\u4e0a\u7684\u8868\u73b0\u76f8\u5f53\u6216\u4f18\u4e8e\u5377\u79ef\u795e\u7ecf\u7f51\u7edc\u3002<\/li>\n\n\n\n
  3. \u5728\u8ba1\u7b97\u6548\u7387\u4e0a\uff0c\u8bad\u7ec3\u540c\u7b49\u7cbe\u5ea6\u7684\u6a21\u578b\uff0cTransformer\u6a21\u578b\u6bd4\u5377\u79ef\u6a21\u578b\u66f4\u6709\u4f18\u52bf\u3002<\/li>\n<\/ol>\n\n\n\n

    \u7406\u89e3<\/h1>\n\n\n\n

    \u8fd9\u7bc7\u6587\u7ae0\u7684\u5f15\u7528\u5df2\u7ecf\u63a5\u8fd160k\u4e86\uff0c\u6211\u89c9\u5f97\u4ed6\u5389\u5bb3\u7684\u70b9\u5728\u4e8e\u8bc1\u660eTransformer\u7684\u901a\u7528\u6027\uff0c\u8fd9\u4e3a\u540e\u9762\u7684\u57fa\u4e8eTransformer\u7684\u591a\u6a21\u6001\u5927\u6a21\u578b\u94fa\u5e73\u4e86\u9053\u8def\uff0c\u800c\u4e14\u8fd9\u79cd\u901a\u7528\u6027\uff0c\u4e5f\u8ba9Transformer\u7684\u4e00\u4e9b\u4f18\u5316\uff0c\u4f8b\u5982Flash Attention[3]<\/a> vLLM[4]<\/a>\u7b49\uff0c\u80fd\u591f\u76f4\u63a5\u5e94\u7528\u5728\u5305\u542bCV\u7684\u591a\u6a21\u6001\u9886\u57df\u3002<\/p>\n\n\n\n

    \u53c2\u8003\u6587\u732e<\/h1>\n\n\n\n

    [1]<\/a>Xiaolong Wang, Ross Girshick, Abhinav Gupta, and Kaiming He. Non-local neural networks. In CVPR, 2018.<\/p>\n\n\n\n

    [2]<\/a>Prajit Ramachandran, Niki Parmar, Ashish Vaswani, Irwan Bello, Anselm Levskaya, and Jon Shlens. Stand-alone self-attention in vision models. In NeurIPS, 2019.<\/p>\n\n\n\n

    [3]<\/a>Dao, Tri. “Flashattention-2: Faster attention with better parallelism and work partitioning.” arXiv preprint arXiv:2307.08691<\/em> (2023).<\/p>\n\n\n\n

    [4]<\/a>Kwon, Woosuk, et al. “vllm: Easy, fast, and cheap llm serving with pagedattention.” See https:\/\/vllm<\/a>. ai\/(accessed 9 August 2023)<\/em> (2023).<\/p>\n","protected":false},"excerpt":{"rendered":"

    AN IMAGE IS WORTH 16X16 WORDS: TRANSFORMERS FOR IMAGE RECOGNITION AT SCALE<\/p>\n","protected":false},"author":1,"featured_media":591,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[12],"tags":[34],"class_list":["post-585","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-research","tag-multimodality"],"_links":{"self":[{"href":"https:\/\/blogqian.com\/index.php\/wp-json\/wp\/v2\/posts\/585","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/blogqian.com\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/blogqian.com\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/blogqian.com\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/blogqian.com\/index.php\/wp-json\/wp\/v2\/comments?post=585"}],"version-history":[{"count":2,"href":"https:\/\/blogqian.com\/index.php\/wp-json\/wp\/v2\/posts\/585\/revisions"}],"predecessor-version":[{"id":594,"href":"https:\/\/blogqian.com\/index.php\/wp-json\/wp\/v2\/posts\/585\/revisions\/594"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/blogqian.com\/index.php\/wp-json\/wp\/v2\/media\/591"}],"wp:attachment":[{"href":"https:\/\/blogqian.com\/index.php\/wp-json\/wp\/v2\/media?parent=585"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/blogqian.com\/index.php\/wp-json\/wp\/v2\/categories?post=585"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/blogqian.com\/index.php\/wp-json\/wp\/v2\/tags?post=585"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}