{"id":595,"date":"2025-06-30T22:16:05","date_gmt":"2025-06-30T14:16:05","guid":{"rendered":"https:\/\/blogqian.com\/?p=595"},"modified":"2025-07-02T22:22:04","modified_gmt":"2025-07-02T14:22:04","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-ii-clip","status":"publish","type":"post","link":"https:\/\/blogqian.com\/index.php\/2025\/06\/30\/%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-ii-clip\/","title":{"rendered":"\u591a\u6a21\u6001\u7ecf\u5178\u8bba\u6587\u6574\u7406 II \u2013 CLIP"},"content":{"rendered":"\n

\ud83c\udf89 \u7ecf\u5178\u7684\u53cc\u5854\u7ed3\u6784 Learning Transferable Visual Models From Natural Language Supervision<\/p>\n\n\n\n

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

OpenAi<\/p>\n\n\n\n

\u8bba\u6587\u80cc\u666f\u548c\u7814\u7a76\u95ee\u9898<\/h1>\n\n\n\n

\u4e4b\u524d\u7684\u56fe\u7247\u5206\u7c7b\u7f51\u7edc\u53ea\u80fd\u5bf9\u4e8e\u56fa\u5b9a\u7684\u7c7b\u522b\u8fdb\u884c\u5206\u7c7b\uff08ImageNet Cifar10 Cifar100\u7b49\uff09\uff0c\u8fd9\u4e25\u91cd\u7f3a\u4e4f\u6cdb\u5316\u6027\uff0c\u800c\u4e14\u9700\u8981\u6839\u636e\u7c7b\u522b\u6765\u4eba\u5de5\u6807\u6ce8label\u3002\u6240\u4ee5\uff0c\u627e\u5230\u4e00\u4e2a\u4e0d\u9700\u8981\u5bf9\u6570\u636e\u8fdb\u884c\u6807\u6ce8\u5c31\u80fd\u8bad\u7ec3\uff0c\u4e14\u4e0d\u9700\u8981\u4e8b\u5148\u56fa\u5b9a\u597d\u5206\u7c7b\u7c7b\u522b\u7684\u89c6\u89c9\u6a21\u578b\u662f\u5f88\u91cd\u8981\u7684\u3002<\/p>\n\n\n\n

\u8fd9\u7bc7\u6587\u7ae0\u7684\u4f5c\u8005\u6536\u5230NLP\u9886\u57dfTransformer\u7684\u542f\u53d1\uff08BERT\u548cGPT\u90fd\u662f\u81ea\u76d1\u7763\u8bad\u7ec3\uff0c\u4e0d\u9700\u8981\u63d0\u524d\u6807\u6ce8\uff09\uff0c\u60f3\u5230\u7528\u81ea\u76d1\u7763\u5b66\u4e60\u7528\u5728\u56fe\u50cf\u5206\u7c7b\u9886\u57df\u6765\u5f97\u5230\u4e00\u4e2a\u4e0d\u9700\u8981\u63d0\u524d\u6807\u6ce8\u7684\u89c6\u89c9\u6a21\u578b\u3002<\/p>\n\n\n\n

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

\u8be5\u6587\u901a\u8fc7\u5c06\u81ea\u76d1\u7763\u5b66\u4e60\u5229\u7528\u5728\u56fe\u50cf\u5206\u7c7b\u9886\u57df\uff0c\u627e\u5230\u4e86\u4e00\u4e2a\u4e0d\u9700\u8981\u5bf9\u6570\u636e\u8fdb\u884c\u63d0\u524d\u6807\u6ce8\uff0c\u5c31\u80fd\u5bf9\u4e0d\u5b9a\u7c7b\u522b\u7684\u8fdb\u884c\u5206\u7c7b\u7684\u53ef\u8fc1\u79fb\u89c6\u89c9\u6a21\u578b\u3002\u5e76\u7814\u7a76\u4e86\u5b83\u76f8\u5bf9\u4e8e\u4f20\u7edfCNN\u7684\u6027\u80fd\u4f18\u52bf\u3002<\/p>\n\n\n\n

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

\u8be5\u6587\u7684\u8bad\u7ec3\u4efb\u52a1\uff0c\u662f\u4e00\u4e2a\u56fe\u7247\u548c\u6587\u672c\u914d\u5bf9\u7684\u4efb\u52a1\u3002\u6587\u672c\u901a\u8fc7\u4e00\u4e2aText Encoder\uff0c\u56fe\u50cf\u901a\u8fc7\u4e00\u4e2aImage Encoder\uff0c\u6700\u7ec8\u5f97\u5230\u4e24\u8005\u7684\u5411\u91cf\u8868\u793a\uff0c\u518d\u53d6\u5176\u70b9\u4e58\u3002\u7136\u540e\uff0c\u5728\u8fd9\u4e2a\u7a7a\u95f4\u4e2d\uff0c\u62c9\u8fd1\u914d\u5bf9\u56fe\u6587\u5bf9\u7684\u8ddd\u79bb\uff0c\u758f\u8fdc\u975e\u914d\u5bf9\u6587\u672c\u5bf9\u7684\u8ddd\u79bb\u3002<\/p>\n\n\n\n

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

\u8bad\u7ec3\u7ed3\u675f\u540e\uff0c\u5c31\u53ef\u4ee5\u8fdb\u884c\u63a8\u7406\u4e86\u3002\u6587\u7ae0\u4e2d\u7684\u65b9\u6cd5\u662f\u5148\u6309\u7167\u8fd9\u6837\u7684\u6a21\u5f0f\uff1a\u201cA photo of a (name of the object)\u201d\u7ec4\u7ec7\u8bed\u53e5\uff08\u539f\u6587\u4e2d\u7684\u5b9e\u9a8c\u6307\u51fa\uff0c\u8fd9\u6837\u7684Prompt\u53ef\u4ee5\u4f7faccuracy\u5728ImageNet\u4e0a\u7684\u7cbe\u5ea6\u63d0\u53471.3%\uff09\uff0c\u5f97\u5230n\u4e2a\u7c7b\u522b\u7684\u6587\u672c\u5411\u91cf\uff0c\u518d\u653e\u5165Text Encoder\u4e2d\u8fdb\u884cembedding\uff1b\u56fe\u50cf\u5219\u7ecf\u8fc7\u56fe\u50cf\u7f16\u7801\u5668\u8fdb\u884c\u7f16\u7801\uff0c\u4e4b\u540e\u8ba1\u7b97\u4f59\u5f26\u76f8\u4f3c\u5ea6\uff0c\u54ea\u4e2a\u76f8\u4f3c\u5ea6\u6700\u5927\uff0c\u5c31\u8f93\u51fa\u54ea\u4e2a\u7f16\u7801\u3002<\/p>\n\n\n\n

\u5b9e\u9a8c\u6548\u679c<\/h1>\n\n\n\n

CLIP\u5728\u81ea\u5df14\u4ebf\u56fe\u6587\u5bf9\u8bad\u7ec3\u540e\uff0c\u4e0d\u5728\u5176\u4ed6\u6570\u636e\u96c6\u4e0a\u505a\u4efb\u4f55\u8bad\u7ec3\uff0c\u76f4\u63a5\u8fdb\u884czero-shot\u9884\u6d4b\u3002\u518d\u5bf9\u6bd4\u5728ImageNet\u4e0a\u9884\u8bad\u7ec3\u7684ResNet50\uff08\u8fd9\u5bf9ResNet50\u80af\u5b9a\u6709\u66f4\u5927\u7684\u4f18\u52bf\uff0c\u56e0\u4e3a\u4ed6\u53ef\u4ee5\u5728\u8fd9\u4e9b\u6570\u636e\u4e0a\u505a\u5fae\u8c03\uff09<\/p>\n\n\n\n

\u7ed3\u679c\u53ef\u4ee5\u770b\u5230\uff0c\u5728\u5927\u90e8\u5206\u6570\u636e\u96c6\u4e0a\uff0c\u90fd\u662fCLIP\u9886\u5148\u3002<\/p>\n\n\n\n

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

\u90a3\u4e48\u5982\u679c\u5728CLIP\u56fe\u7247encoder\u4e5f\u52a0\u5165\u5206\u7c7b\u5934\u4f5c\u5fae\u8c03\u5462\uff0c\u53ef\u4ee5\u770b\u5230\u4e0d\u8bba\u662f\u6bcf\u4e2a\u7c7b\u522b1\u4e2a\u4f8b\u5b50\uff0c2\u4e2a\u4f8b\u5b50\uff0c4\u4e2a\u4f8b\u5b50\uff0c8\u4e2a\u4f8b\u5b50\uff0c\u90fd\u662fCLIP\u9886\u5148\u3002<\/p>\n\n\n\n

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

\u4e3a\u4e86\u8fdb\u4e00\u6b65\u8bc1\u660eCLIP\u7684\u6cdb\u5316\u80fd\u529b\uff0c\u4f5c\u8005\u7ed9\u4e86\u4e2a\u66f4\u52a0\u6709\u8bf4\u670d\u529b\u7684\u4f8b\u5b50\uff0c\u53ef\u4ee5\u770b\u5230\uff0c\u5728\u6cdb\u5316\u7684\u6570\u636e\u96c6\uff08\u4f8b\u5982\u7d20\u63cf\uff0c\u5361\u901a\u98ce\u683c\uff09\u7b49\u7b49\uff0cResNet\u7684\u6027\u80fd\u5927\u5e45\u4e0b\u964d\uff0c\u800cCLIP\u4ecd\u7136\u575a\u633a\uff08\u53ef\u4ee5\u770b\u5230CLIP\u5177\u6709\u975e\u5e38\u4f18\u79c0 \u7684\u6cdb\u5316\u6027\u80fd\uff09\uff1a<\/p>\n\n\n\n

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

\u5b9e\u9a8c\u7ed3\u8bba<\/h1>\n\n\n\n

\u8be5\u7bc7\u8bba\u6587\u8bc1\u660e\u4e86CLIP\u6a21\u578b\u5f3a\u5927\u7684\u6cdb\u5316\u80fd\u529b\uff0c\u548c\u4f18\u79c0\u7684\u6027\u80fd\u8868\u73b0\uff08\u76f8\u8f83\u4e8e\u4f20\u7edf\u7684\u795e\u7ecf\u7f51\u7edc\u67b6\u6784\uff09\u3002 \u5176\u4f7f\u7528\u4e30\u5bcc\u7684\u8bed\u610f\u8fdb\u884c\u76d1\u7763\uff0c\u4f7f\u6a21\u578b\u80fd\u591f\u5b66\u4e60\u5230\u66f4\u591a\u7684\u7ec6\u8282\u8bed\u610f\u7279\u5f81\u3002 \u8fde\u63a5\u89c6\u89c9\u548c\u6587\u672c\u4e24\u4e2amodality\uff0c\u53ef\u4ee5\u5f15\u7528\u6587\u672c\u6765\u67e5\u8be2\u56fe\u50cf\u3002 \u4e14\u4e0d\u9700\u8981\u9884\u5148\u5b9a\u4e49\u56fa\u5b9a\u7c7b\u522b\u7684\u6807\u7b7e\uff0c\u76f4\u63a5\u7528prompt\u8fdb\u884c\u67e5\u8be2\u64cd\u4f5c\u3002<\/p>\n\n\n\n

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

\u8be5\u7bc7\u8bba\u6587\u7684\u521b\u65b0\u70b9\u5728\u4e8e\uff0c\u629b\u5f03\u4e86\u539f\u6709\u7684data+label\u7684\u7ed3\u6784\uff0c\u800c\u5e94\u7528\u7c7b\u4f3c\u4e8e\u5bf9\u6bd4\u5b66\u4e60\u7684\u6982\u5ff5\uff0c\u6ca1\u6709\u5177\u4f53\u7684\u7c7b\u522b\uff0c\u53ea\u9700\u8981\u89c4\u5b9a\u54ea\u4e9b\u662fpositive\u7684\uff0c\u54ea\u4e9b\u662fnegative\u7684\uff0c\u7528infoNCE\u4e3a\u635f\u5931\u51fd\u6570\uff0c\u62c9\u8fd1positive\u7684\u8ddd\u79bb\uff0c\u758f\u8fdcnegative\u7684\u8ddd\u79bb\u3002\uff08\u672c\u6587\u4e2d\u5219\u662f\uff0c\u5bf9\u5e94\u7684\u56fe\u6587\u5bf9\u4e3apositive\u6837\u672c\uff0c\u76f8\u53cd\u7684\u56fe\u6587\u5bf9\u4e3anegative\u6837\u672c\uff09<\/p>\n\n\n\n

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

[1]<\/a> Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision, 2021.<\/p>\n","protected":false},"excerpt":{"rendered":"

Learning Transferable Visual Models From Natural Language Supervision<\/p>\n","protected":false},"author":1,"featured_media":597,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[12],"tags":[34],"class_list":["post-595","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\/595","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=595"}],"version-history":[{"count":1,"href":"https:\/\/blogqian.com\/index.php\/wp-json\/wp\/v2\/posts\/595\/revisions"}],"predecessor-version":[{"id":600,"href":"https:\/\/blogqian.com\/index.php\/wp-json\/wp\/v2\/posts\/595\/revisions\/600"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/blogqian.com\/index.php\/wp-json\/wp\/v2\/media\/597"}],"wp:attachment":[{"href":"https:\/\/blogqian.com\/index.php\/wp-json\/wp\/v2\/media?parent=595"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/blogqian.com\/index.php\/wp-json\/wp\/v2\/categories?post=595"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/blogqian.com\/index.php\/wp-json\/wp\/v2\/tags?post=595"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}