{"id":2326,"date":"2024-08-06T11:30:38","date_gmt":"2024-08-06T03:30:38","guid":{"rendered":"https:\/\/setabout.cn\/?p=2326"},"modified":"2024-10-17T13:26:16","modified_gmt":"2024-10-17T05:26:16","slug":"clip","status":"publish","type":"post","link":"https:\/\/setabout.cn\/?p=2326","title":{"rendered":"CLIP"},"content":{"rendered":"\n

CLIP\u662f\u4e00\u4e2a\u795e\u7ecf\u7f51\u7edc\uff0c\u5b83\u4f7f\u7528\u81ea\u7136\u8bed\u8a00\u6765\u76d1\u7763\u89c6\u89c9\u8bad\u7ec3\uff0c\u53ef\u4ee5\u5c06\u5b83\u8fd0\u7528\u5230\u4efb\u4f55\u89c6\u89c9\u5206\u7c7b\u4efb\u52a1\uff0c\u5177\u4f53\u7684\u5c31\u662f\u5c06\u7c7b\u522b\u7684\u540d\u5b57\u7528\u81ea\u7136\u8bed\u8a00\u8fdb\u884c\u63cf\u8ff0\uff0c\u7136\u540e\u548c\u56fe\u7247\u4e00\u8d77\u8fdb\u884c\u6df7\u5408\u7f16\u7801\uff0c\u53ef\u4ee50\u6837\u672c\u5b66\u4e60\uff08zero-shot learning\uff09\u3002\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0c\u8be5\u795e\u7ecf\u7f51\u7edc\u5e76\u6ca1\u6709\u4f7f\u7528\u5230\u6807\u7b7e\uff0c\u662f\u81ea\u76d1\u7763\u5b66\u4e60\u3002<\/p>\n\n\n\n

\u80cc\u666f<\/h2>\n\n\n\n

SimCLR<\/h3>\n\n\n\n

\u81ea\u76d1\u7763\u5b66\u4e60\u4f1a\u6d89\u53ca\u5230\u5bf9\u6bd4\u5b66\u4e60\u7684\u6982\u5ff5\uff0cSimCLR [1]<\/a>(A Simple Framework for Contrastive Learning of Visual Representations)\u662f\u4e00\u79cd\u5bf9\u6bd4\u5b66\u4e60\u7f51\u7edc\uff0c\u53ef\u4ee5\u5bf9\u542b\u6709\u5c11\u91cf\u6807\u7b7e\u7684\u6570\u636e\u96c6\u8fdb\u884c\u8bad\u7ec3\u63a8\u7406\uff0c\u5b83\u5305\u542b\u65e0\u76d1\u7763\u5b66\u4e60\u548c\u6709\u76d1\u7763\u5b66\u4e60\u4e24\u4e2a\u90e8\u5206\u3002<\/p>\n\n\n\n

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

\u5728\u7b2c\u4e00\u9636\u6bb5\u5148\u8fdb\u884c\u65e0\u76d1\u7763\u5b66\u4e60\uff0c\u5bf9\u8f93\u5165\u56fe\u50cf\u8fdb\u884c\u4e24\u6b21\u968f\u673a\u56fe\u50cf\u589e\u5f3a\uff0c\u5373\u7531\u4e00\u5e45\u56fe\u50cf\u5f97\u5230\u4e24\u4e2a\u968f\u673a\u5904\u7406\u8fc7\u540e\u7684\u56fe\u50cf\uff0c\u4f9d\u6b21\u653e\u5165\u7f51\u7edc\u8fdb\u884c\u8bad\u7ec3\uff0c\u8ba1\u7b97\u635f\u5931\u5e76\u66f4\u65b0\u68af\u5ea6\u3002\u5982\u56fe\u6240\u793a\uff0c\u5c06\u5c0f\u72d7\u8fdb\u884c\u4e86\u4e24\u6b21\u968f\u673a\u56fe\u50cf\u53d8\u6362\uff0c\u518d\u7ecf\u8fc7\u6b8b\u5dee\u7f51\u7edc\uff0c\u8fdb\u884c\u56fe\u7247\u7f16\u7801\uff0c\u4f7f\u7528\u4e00\u4e2aMLP\uff08\u591a\u5c42\u611f\u77e5\u673a\uff09\u505a\u9884\u6d4b\u6216\u8005\u6295\u5f71\u5230\u53e6\u5916\u4e00\u4e2a\u7a7a\u95f4\uff0c\u8fdb\u884c\u6a2a\u5411\u6bd4\u8f83\uff0c\u4ece\u540c\u4e00\u4e2a\u56fe\u50cf\u751f\u6210\u7684\u7ed3\u679c\u76f8\u8fd1\uff0c\u4e0e\u732b\u5bf9\u6bd4\uff0c\u5219\u76f8\u5dee\u8f83\u8fdc\uff0c\u635f\u5931\u51fd\u6570\u76ee\u7684\u5c31\u662f\u4f7f\u5f97\u4e24\u4e2a\u6b63\u6837\u672c\u8ddd\u79bb\u66f4\u8fd1\uff08\\( z_{ \\theta } \\),\\( z_{\\theta}^{0} \\)\uff09\uff0c\u4e0e\u8d1f\u6837\u672c\u4e4b\u95f4\u7684\u8ddd\u79bb\u66f4\u8fdc(\\( z_{ \\theta } {1} \\))\u3002\u7b2c\u4e8c\u9636\u6bb5\uff0c\u52a0\u8f7d\u7b2c\u4e00\u9636\u6bb5\u7684\u7279\u5f81\u63d0\u53d6\u5c42\u8bad\u7ec3\u53c2\u6570\uff0c\u7528\u5c11\u91cf\u5e26\u6807\u7b7e\u6837\u672c\u8fdb\u884c\u6709\u76d1\u7763\u5b66\u4e60\uff08\u53ea\u8bad\u7ec3\u5168\u8fde\u63a5\u5c42\uff09\u3002\u8fd9\u4e00\u9636\u6bb5\u635f\u5931\u51fd\u6570\u4e3a\u4ea4\u53c9\u71b5\u635f\u5931\u51fd\u6570CrossEntropyLoss[2]<\/a>\u3002<\/p>\n\n\n\n

BYOL<\/h3>\n\n\n\n

\u518d\u8fdb\u4e00\u6b65\u53d1\u5c55\uff0c\u63d0\u51fa\u4e86\u4e00\u79cd\u65b0\u7684\u81ea\u76d1\u7763\u5b66\u4e60\u65b9\u6cd5\uff0cBYOL\uff08Bootstrap Your Own Latent\uff09[3]<\/a>[4]<\/a>\uff0c\u548c\u4ee5\u5f80\u9700\u8981\u5927\u91cf\u8d1f\u6837\u672c\u7684\u5bf9\u6bd4\u5b66\u4e60\u65b9\u6cd5\u4e0d\u540c\uff08SimCLR\uff09\uff0cBYOL\u4e0d\u4f9d\u8d56\u4e8e\u8d1f\u6837\u672c\u5bf9\u3002<\/p>\n\n\n\n

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

BYOL\u4f7f\u7528online\u548ctarget\u4e24\u4e2a\u795e\u7ecf\u7f51\u7edc\u6765\u5b66\u4e60\uff0conline\u7f51\u7edc\u7531\u4e00\u7ec4\u6743\u91cd\u5b9a\u4e49\u5e76\u7531\u4e09\u4e2a\u9636\u6bb5\u7ec4\u6210\uff1aencoder \u3001projector\uff08\u6295\u5f71\u5230\u540c\u4e00\u4e2a\u7a7a\u95f4\uff09 \u3001predictor \u3002target\u4e0eonline\u67b6\u6784\u76f8\u540c\uff0c\u4f46\u4f7f\u7528\u4e0d\u540c\u7684\u53c2\u6570\u3002target\u63d0\u4f9b\u4e86\u7528\u4e8e\u8bad\u7ec3online\u7684\u56de\u5f52\u76ee\u6807\uff0c\u5176\u53c2\u6570\u662fonline\u53c2\u6570\u7684\u6307\u6570\u79fb\u52a8\u5e73\u5747\u503c\u3002\u5b66\u4e60loss\u7684\u68af\u5ea6\u53ea\u4f1a\u53cd\u5411\u4f20\u64ad\u7ed9online\u7f51\u7edc\uff0c\u4e0d\u4f1a\u53cd\u5411\u4f20\u64ad\u7ed9target\u7f51\u7edc\uff0c\u7c7b\u4f3c\u4e8eMoCo[5]<\/a>\uff0ctarget\u7f51\u7edc\u4f7f\u7528\u52a8\u91cf\u66f4\u65b0\uff0c\u76f8\u5f53\u4e8etarget\u6a21\u62df\u7ed9online\u63d0\u4f9b\u4e86\u4e00\u4e2a\u8d1f\u6837\u672c\u3002\uff08\u8865\u5145\uff1a\u4e3a\u4ec0\u4e48\u4f7f\u7528\u6307\u6570\u79fb\u52a8\u5e73\u5747\u503c\uff0c\u6574\u4f53\u7f51\u7edc\u7684\u4e0d\u786e\u5b9a\u6027\u5f88\u5927\uff0ctarget\u548conline\u540c\u65f6\u5b66\u4e60\uff0c\u5f88\u5feb\u5c31\u4f1a\u6536\u655b\uff0c\u6536\u655b\u5230online\u548ctarget\u7ed3\u679c\u4e00\u81f4\uff0c\u4e5f\u5c31\u662f\u6a21\u5f0f\u5d29\u584c\uff0closs\u4e3a0\uff0c\u7136\u800c\u5e76\u4e0d\u5e0c\u671b\u51fa\u73b0\u8fd9\u6837\u7684\u7ed3\u679c\uff0c\u901a\u8fc7\u964d\u4f4e\u5b66\u4e60\u7684\u6548\u7387\u6765\u907f\u514d\u8fd9\u6837\u7684\u60c5\u51b5\uff0c\u4e5f\u5c31\u662ftarget\u4e0d\u8fdb\u884c\u5b66\u4e60\uff0ctarget\u901a\u8fc7\u5386\u53f2online\u8fdb\u884c\u6162\u6162\u7684\u66f4\u65b0\uff0conline\u9884\u6d4b\u7ed3\u679c\u548ctarget\u8fdb\u884c\u5bf9\u6bd4\uff0c\u8ba1\u7b97\u635f\u5931\uff0c\u66f4\u65b0\u8fed\u4ee3\uff09<\/p>\n\n\n\n

CLIP<\/h2>\n\n\n\n

CLIP\u6a21\u578b\u7531\u56fe\u50cf\u7f16\u7801\u5668\u548c\u6587\u672c\u7f16\u7801\u5668\u4e24\u4e2a\u90e8\u5206\u7ec4\u6210\uff0c\u56fe\u50cf\u7f16\u7801\u5668\u8d1f\u8d23\u5c06\u56fe\u50cf\u8f6c\u6362\u4e3a\u7279\u5f81\u5411\u91cf\uff0c\u53ef\u4ee5\u662f\u5377\u79ef\u795e\u7ecf\u7f51\u7edc\u6216Transformer\u6a21\u578b\uff08ViT\uff09[7]<\/a>,\u6587\u672c\u7f16\u7801\u5668\u8d1f\u8d23\u5c06\u6587\u672c\u8f6c\u6362\u4e3a\u7279\u5f81\u5411\u91cf\uff0c\u901a\u5e38\u662f\u4e00\u4e2atransformer\u6a21\u578b\uff0c\u901a\u8fc7\u5171\u4eab\u4e00\u4e2a\u5411\u91cf\u7a7a\u95f4\u6765\u5b9e\u73b0\u8de8\u6a21\u6001\u7684\u4fe1\u606f\u4ea4\u4e92\u3002<\/p>\n\n\n

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

\u6a21\u578b\u7684\u8f93\u5165\u662f\u4e00\u4e2a\u56fe\u7247\u548c\u6587\u5b57\u7684\u914d\u5bf9\uff0c\u56fe\u7247\u8f93\u5165\u5230\u56fe\u7247\u7f16\u7801\u5668\uff0c\u5f97\u5230\u4e00\u4e2a\u7279\u5f81\u5411\u91cf\uff0c\u6587\u672c\u4f1a\u901a\u8fc7\u4e00\u4e2a\u6587\u672c\u7f16\u7801\u5668\uff0c\u4ece\u800c\u5f97\u5230\u5bf9\u5e94\u6587\u672c\u7684\u7279\u5f81\u5411\u91cf\u3002CLIP\u4f1a\u5728\u8fd9\u4e9b\u7279\u5f81\u4e0a\u505a\u5bf9\u6bd4\u5b66\u4e60\uff0c\u914d\u5bf9\u7684\u56fe\u7247\u6587\u672c\u5bf9\u5c31\u662f\u6b63\u6837\u672c\uff0c\u5728\u77e9\u9635\u4e2d\u5bf9\u89d2\u7ebf\u4e0a\u7684\u5373\u6b63\u6837\u672c\uff0c\u6709N\u4e2a\u6b63\u6837\u672c\uff0c\\(N^{2}-N\\)\u4e2a\u8d1f\u6837\u672c\u3002<\/p>\n\n\n

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

clip\u4e2d\u6ca1\u6709\u5206\u7c7b\u5934\uff0c\u8fdb\u884c\u9884\u6d4b\u7684\u8bdd\uff0c\u8fd9\u91cc\u505a\u4e86\u4e00\u4e2a\u5904\u7406\uff0c\u5c31\u662f\u5c06\u6570\u636e\u96c6\u4e2d\u7684\u7c7b\u522b\u63cf\u8ff0\u6210\u4e00\u4e2a\u53e5\u5b50\uff0c\u4ee5 ImageNet \u4e3a\u4f8b\uff0cCLIP \u5148\u628a ImageNet \u8fd91000\u4e2a\u7c7b\uff08\u5982\u56fe\u4e2d”plane”, \u201ccar\u201d, \u201cdog\u201d, \u2026, \u201cbrid\u201d\uff09\u53d8\u6210\u4e00\u4e2a\u53e5\u5b50\uff0c\u4e5f\u5c31\u662f\u5c06\u8fd9\u4e9b\u7c7b\u522b\u53bb\u66ff\u4ee3 \u201cA photo of a {object}\u201d \u4e2d\u7684 \u201c{object}\u201d \uff0c\u4ee5 \u201cplane\u201d \u7c7b\u4e3a\u4f8b\uff0c\u5b83\u5c31\u53d8\u6210”A photo of a plane”\uff0c\u90a3\u4e48 ImageNet \u91cc\u76841000\u4e2a\u7c7b\u522b\u5c31\u90fd\u5728\u8fd9\u91cc\u751f\u6210\u4e861000\u4e2a\u53e5\u5b50\uff0c\u7136\u540e\u901a\u8fc7\u5148\u524d\u9884\u8bad\u7ec3\u597d\u7684 Text Encoder \u5c31\u4f1a\u5f97\u52301000\u4e2a\u6587\u672c\u7684\u7279\u5f81\u3002<\/p>\n\n\n\n

\u5728\u63a8\u7406\u65f6\uff0c\u53ea\u8981\u628a\u8fd9\u5f20\u56fe\u7247\u6254\u7ed9 Image Encoder\uff0c\u5f97\u5230\u56fe\u50cf\u7279\u5f81\u540e\uff0c\u5c31\u62ff\u8fd9\u4e2a\u56fe\u7247\u7279\u5f81\u53bb\u8ddf\u6240\u6709\u7684\u6587\u672c\u7279\u5f81\u53bb\u505a cosine similarity\uff08\u4f59\u5f26\u76f8\u4f3c\u5ea6\uff09\u8ba1\u7b97\u76f8\u4f3c\u5ea6\uff0c\u770b\u8fd9\u5f20\u56fe\u7247\u4e0e\u54ea\u4e2a\u6587\u672c\u6700\u76f8\u4f3c\uff0c\u5c31\u628a\u8fd9\u4e2a\u6587\u672c\u7279\u5f81\u6240\u5bf9\u5e94\u7684\u53e5\u5b50\u6311\u51fa\u6765\uff0c\u4ece\u800c\u5b8c\u6210\u5206\u7c7b\u4efb\u52a1\u3002<\/p>\n\n\n\n

\u53c2\u8003<\/h2>\n\n\n\n

[1]SimCLR\u56fe\u50cf\u5206\u7c7b\u2014\u2014pytorch\u590d\u73b0_simclr pytorch-CSDN\u535a\u5ba2<\/a><\/p>\n\n\n\n

[2]Pytorch CrossEntropyLoss() \u539f\u7406\u548c\u7528\u6cd5\u8be6\u89e3-CSDN\u535a\u5ba2<\/a><\/p>\n\n\n\n

[3]BYOL\uff08NeurIPS 2020\uff09\u539f\u7406\u89e3\u8bfb-CSDN\u535a\u5ba2<\/a><\/p>\n\n\n\n

[4]\u6df1\u5ea6\u5b66\u4e60(\u81ea\u76d1\u7763:BYOL)\u2014\u2014Bootstrap Your Own Latent A New Approach to Self-Supervised Learning_byol, bootstrap your own latent-CSDN\u535a\u5ba2<\/a><\/p>\n\n\n\n

[5]\u6df1\u5ea6\u5b66\u4e60(\u81ea\u76d1\u7763:MoCo)\u2014\u2014Momentum Contrast for Unsupervised Visual Representation Learning-CSDN\u535a\u5ba2<\/a><\/p>\n\n\n\n

[6]CLIP \u8bba\u6587\u9010\u6bb5\u7cbe\u8bfb\u3010\u8bba\u6587\u7cbe\u8bfb\u3011_\u54d4\u54e9\u54d4\u54e9_bilibili<\/a><\/p>\n\n\n\n

[7]\u6df1\u5165\u6d45\u51fa\u4e00\u6587\u56fe\u89e3Vision in Transformer\uff08ViT\uff09<\/a><\/p>\n\n\n\n

[8]https:\/\/blog.csdn.net\/h661975\/article\/details\/135116957<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"

Connect Texts and Images<\/p>\n","protected":false},"author":1,"featured_media":2338,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[8,36],"tags":[34,35,29],"class_list":["post-2326","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-8","category-36","tag-clip","tag-35","tag-29"],"views":2035,"_links":{"self":[{"href":"https:\/\/setabout.cn\/index.php?rest_route=\/wp\/v2\/posts\/2326","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/setabout.cn\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/setabout.cn\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/setabout.cn\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/setabout.cn\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=2326"}],"version-history":[{"count":7,"href":"https:\/\/setabout.cn\/index.php?rest_route=\/wp\/v2\/posts\/2326\/revisions"}],"predecessor-version":[{"id":2567,"href":"https:\/\/setabout.cn\/index.php?rest_route=\/wp\/v2\/posts\/2326\/revisions\/2567"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/setabout.cn\/index.php?rest_route=\/wp\/v2\/media\/2338"}],"wp:attachment":[{"href":"https:\/\/setabout.cn\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=2326"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/setabout.cn\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=2326"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/setabout.cn\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=2326"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}