{"id":114,"date":"2026-05-11T16:54:37","date_gmt":"2026-05-11T08:54:37","guid":{"rendered":"http:\/\/blog.thirdbody.cn\/?p=114"},"modified":"2026-05-11T16:54:37","modified_gmt":"2026-05-11T08:54:37","slug":"%e6%96%87%e7%94%9f%e5%9b%be%e6%8a%80%e6%9c%af%e4%b8%80-vit%e6%a8%a1%e5%9e%8b%e5%92%8cclip%e6%a8%a1%e5%9e%8b","status":"publish","type":"post","link":"http:\/\/blog.thirdbody.cn\/index.php\/2026\/05\/11\/%e6%96%87%e7%94%9f%e5%9b%be%e6%8a%80%e6%9c%af%e4%b8%80-vit%e6%a8%a1%e5%9e%8b%e5%92%8cclip%e6%a8%a1%e5%9e%8b\/","title":{"rendered":"\u6587\u751f\u56fe\u6280\u672f(\u4e00)&#8211;ViT\u6a21\u578b\u548cCLIP\u6a21\u578b"},"content":{"rendered":"<h3>1. ViT(Vision Transformer)\u6a21\u578b<\/h3>\n<h4>1.1 \u80cc\u666f<\/h4>\n<p>\u57282020\u5e74\u4e4b\u524d\uff0c\u56fe\u50cf\u9886\u57df\u662fCNN\u7684&quot;\u4e00\u8a00\u5802&quot;\u2014\u2014\u4eceLeNet\u5230ResNet\uff0c\u4eceMobileNet\u5230EfficientNet\uff0c\u5377\u79ef\u64cd\u4f5c\u51ed\u501f<strong>\u5c40\u90e8\u611f\u53d7\u91ce+\u53c2\u6570\u5171\u4eab<\/strong>\u7684\u4f18\u52bf\uff0c\u5784\u65ad\u4e86\u56fe\u50cf\u5206\u7c7b\u3001\u68c0\u6d4b\u3001\u5206\u5272\u7b49\u51e0\u4e4e\u6240\u6709\u4efb\u52a1\u3002\u4f462020\u5e74Google\u63d0\u51fa\u7684Vision Transformer\uff08ViT\uff09\uff0c\u5f7b\u5e95\u6253\u7834\u4e86\u8fd9\u4e00\u683c\u5c40\uff1a\u5b83\u5b8c\u5168\u629b\u5f03\u5377\u79ef\uff0c\u4ec5\u7528Transformer\u7684<strong>\u81ea\u6ce8\u610f\u529b\u673a\u5236<\/strong>\uff0c\u5c31\u5728ImageNet\u7b49\u6570\u636e\u96c6\u4e0a\u5b9e\u73b0\u4e86\u8d85\u8d8aCNN\u7684\u6027\u80fd\uff0c\u751a\u81f3\u884d\u751f\u51faSwin Transformer\u3001ViT-L\/16\u7b49&quot;\u6027\u80fd\u602a\u517d&quot;\u3002<\/p>\n<h4>1.2 ViT\u7684\u6838\u5fc3\u67b6\u6784<\/h4>\n<p><img decoding=\"async\" src=\"http:\/\/cdn.thirdbody.cn\/blog\/202605\/vit1.jpg\" alt=\"\" \/><\/p>\n<p>\u5728\u6df1\u5165\u8bad\u7ec3\u548c\u63a8\u7406\u4e4b\u524d\uff0c\u9700\u8981\u5148\u7406\u89e3 ViT \u5982\u4f55\u5c06\u4e00\u5f20\u56fe\u50cf\u8f6c\u6362\u4e3a Transformer \u80fd\u591f\u5904\u7406\u7684\u5e8f\u5217\u3002\u8fd9\u4e2a\u8fc7\u7a0b\u662f\u6240\u6709\u540e\u7eed\u6b65\u9aa4\u7684\u57fa\u7840\u3002<\/p>\n<ul>\n<li>\n<p>\u56fe\u50cf\u5206\u5757\u4e0e\u5d4c\u5165 (Image Patching &amp; Embedding)**<\/p>\n<ul>\n<li>\n<p><strong>\u5206\u5757\uff1a<\/strong> ViT \u9996\u5148\u5c06\u4e00\u5f20\u5c3a\u5bf8\u4e3a <code>H\u00d7W\u00d7C<\/code> \u7684\u8f93\u5165\u56fe\u50cf\uff08\u4f8b\u5982 224\u00d7224\u00d73\uff09\u5206\u5272\u6210\u591a\u4e2a\u56fa\u5b9a\u5927\u5c0f\u7684\u3001\u4e0d\u91cd\u53e0\u7684\u56fe\u50cf\u5757\uff08Patches\uff09\uff0c\u4f8b\u5982 16\u00d716 \u50cf\u7d20\u3002<\/p>\n<\/li>\n<li>\n<p><strong>\u5c55\u5e73\u4e0e\u6295\u5f71\uff1a<\/strong> \u6bcf\u4e2a\u56fe\u50cf\u5757\u88ab\u5c55\u5e73\u6210\u4e00\u4e2a\u4e00\u7ef4\u5411\u91cf\u3002\u7136\u540e\uff0c\u901a\u8fc7\u4e00\u4e2a\u53ef\u5b66\u4e60\u7684\u7ebf\u6027\u6295\u5f71\u5c42\uff08Linear Projection\uff09\uff0c\u5c06\u8fd9\u4e2a\u5411\u91cf\u6620\u5c04\u5230\u4e00\u4e2a\u7edf\u4e00\u7684\u3001\u66f4\u9ad8\u7ef4\u5ea6\u7684\u5d4c\u5165\u7a7a\u95f4\uff08\u4f8b\u5982 768 \u7ef4\uff09\u3002\u8fd9\u4e2a\u64cd\u4f5c\u88ab\u79f0\u4e3a\u201c\u5757\u5d4c\u5165\u201d\uff08Patch Embedding\uff09\uff0c\u5b83\u5c06\u6bcf\u4e2a\u56fe\u50cf\u5757\u8f6c\u6362\u6210\u4e00\u4e2a\u201c\u89c6\u89c9\u8bcd\u5411\u91cf\u201d\u3002<\/p>\n<\/li>\n<\/ul>\n<p>\u4e0b\u9762\u4e3e\u4f8b\u8bf4\u660e\u8be5\u8fc7\u7a0b\uff1a<\/p>\n<ul>\n<li>\n<p>*<em>\u8f93\u5165\uff1a224 <\/em> 224**\u76843\u901a\u9053\u56fe\u50cf<\/p>\n<\/li>\n<li>\n<p>*<em>\u5206\u5757\u5927\u5c0f(Patch Size): 16 <\/em> 16** (P = 16)\u50cf\u7d20<\/p>\n<\/li>\n<li>\n<p><strong>\u9690\u85cf\u5c42\u7ef4\u5ea6\u5927\u5c0f(Hidden Size):<\/strong> 1024<\/p>\n<\/li>\n<li>\n<p>\u6b65\u9aa4\uff1a<\/p>\n<\/li>\n<li>\n<p>\u7b2c\u4e00\u6b65\uff1a\u56fe\u50cf\u5206\u5757(Patching)<\/p>\n<p>\u8ba1\u7b97\u5757\u6570\u91cf\uff1a(224 \/ 16) <em> (224 \/ 16) = 14 <\/em> 14 = 196<\/p>\n<p>\u5355\u4e2a\u5206\u5757\u7684\u5f62\u72b6\uff1a16 <em> 16 <\/em> 3 (3\u901a\u9053)<\/p>\n<\/li>\n<li>\n<p>\u7b2c\u4e8c\u6b65\uff1a\u5c55\u5e73(Flattening)<\/p>\n<p>\u5c06\u5355\u4e2a\u5206\u5757<strong>\u62c9\u76f4<\/strong>\u6210\u4e00\u4e2a\u4e00\u7ef4\u5411\u91cf\uff0c\u6b64\u65f6\u8be5\u4e00\u7ef4\u5411\u91cf\u7684\u7ef4\u5ea6\u5927\u5c0f\u662f768(16 <em> 16 <\/em> 3=768)<\/p>\n<p>\u6b64\u65f6\uff0c\u6211\u4eec\u5f97\u5230\u4e86\u4e00\u7ec4768\u7ef4\u7684\u5411\u91cf\uff1a(196,768)<\/p>\n<\/li>\n<li>\n<p>\u7ebf\u6027\u6295\u5f71(Linear Projection)<\/p>\n<p>\u6211\u4eec\u4f7f\u7528\u4e00\u4e2a\u5168\u8fde\u63a5\u5c42(\u5373\u77e9\u9635\u4e58\u6cd5\uff0c\u5f85\u8bad\u7ec3\u77e9\u9635\u53c2\u6570\u91cf\u4e3a: (768,1024))\uff0c\u5c06768\u7ef4\u7684\u5411\u91cf\u6620\u5c04\u5230\u76ee\u6807\u7ef4\u5ea6\u7a7a\u95f4(\u53731024\u7ef4\u5411\u91cf\u7a7a\u95f4)<\/p>\n<p>\u6700\u540e\uff0c\u6211\u4eec\u5f97\u5230\u4e00\u7ec41024\u7ef4\u7684\u5411\u91cf: (196,1024)<\/p>\n<\/li>\n<\/ul>\n<\/li>\n<li>\n<p><strong>\u6dfb\u52a0\u7279\u6b8a\u6807\u8bb0 (Adding Special Tokens)<\/strong><\/p>\n<ul>\n<li><strong>\u5206\u7c7b\u6807\u8bb0 ([CLS] Token)\uff1a<\/strong> ViT \u501f\u9274\u4e86 NLP \u4e2d BERT \u7684\u8bbe\u8ba1\uff0c\u5728\u56fe\u50cf\u5757\u5e8f\u5217\u7684\u5f00\u5934\u6dfb\u52a0\u4e00\u4e2a\u7279\u6b8a\u7684\u3001\u53ef\u5b66\u4e60\u7684\u5411\u91cf\uff0c\u79f0\u4e3a <code>[CLS]<\/code> \u6807\u8bb0\u3002\u8fd9\u4e2a\u6807\u8bb0\u672c\u8eab\u4e0d\u5305\u542b\u56fe\u50cf\u4fe1\u606f\uff0c\u4f46\u5b83\u7684\u4f5c\u7528\u662f\u5728\u7ecf\u8fc7 Transformer \u7f16\u7801\u5668\u7684\u5904\u7406\u540e\uff0c\u805a\u5408\u6574\u4e2a\u56fe\u50cf\u7684\u5168\u5c40\u8bed\u4e49\u4fe1\u606f\uff0c\u5176\u6700\u7ec8\u8f93\u51fa\u72b6\u6001\u5c06\u88ab\u7528\u4e8e\u5206\u7c7b\u4efb\u52a1\u3002<\/li>\n<\/ul>\n<\/li>\n<li>\n<p><strong>\u4f4d\u7f6e\u7f16\u7801 (Positional Encoding)<\/strong><\/p>\n<ul>\n<li>Transformer \u7684\u81ea\u6ce8\u610f\u529b\u673a\u5236\u672c\u8eab\u5bf9\u5e8f\u5217\u7684\u987a\u5e8f\u4e0d\u654f\u611f\u3002\u4e3a\u4e86\u4fdd\u7559\u56fe\u50cf\u4e2d\u5404\u4e2a\u5757\u7684\u7a7a\u95f4\u4f4d\u7f6e\u4fe1\u606f\uff0cViT \u4f1a\u4e3a\u6bcf\u4e2a\u5757\u5d4c\u5165\u5411\u91cf\uff08\u5305\u62ec <code>[CLS]<\/code> \u6807\u8bb0\uff09\u52a0\u4e0a\u4e00\u4e2a\u53ef\u5b66\u4e60\u7684\u4f4d\u7f6e\u7f16\u7801\u5411\u91cf(\u7c7b\u4f3c\u4e8eBERT\u4e2d\u7684\u4f4d\u7f6e\u5411\u91cf\u7f16\u7801\u65b9\u5f0f)\u3002\u8fd9\u4e2a\u4f4d\u7f6e\u7f16\u7801\u5411\u91cf\u4e0e\u5757\u5d4c\u5165\u5411\u91cf\u7684\u7ef4\u5ea6\u76f8\u540c\uff0c\u5e76\u5728\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u4e0e\u6a21\u578b\u7684\u5176\u4ed6\u53c2\u6570\u4e00\u540c\u88ab\u4f18\u5316\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<p>\u5b8c\u6210\u4ee5\u4e0a\u4e09\u6b65\u540e\uff0c\u4e00\u5f20\u4e8c\u7ef4\u7684\u56fe\u50cf\u5c31\u88ab\u8f6c\u6362\u6210\u4e86\u4e00\u4e2a\u5305\u542b\u4f4d\u7f6e\u4fe1\u606f\u7684\u3001\u4e00\u7ef4\u7684\u5411\u91cf\u5e8f\u5217\uff0c\u53ef\u4ee5\u4f5c\u4e3a Transformer \u7f16\u7801\u5668\u7684\u8f93\u5165\u3002<\/p>\n<h4>1.3 ViT\u7684\u8bad\u7ec3\u8fc7\u7a0b<\/h4>\n<ul>\n<li>\n<p>\u6570\u636e\u51c6\u5907\u4e0e\u589e\u5f3a\uff1a\u589e\u5f3a\u65b9\u6cd5\u5305\u62ec <code>RandAugment<\/code>\u3001<code>MixUp<\/code> \u548c <code>CutMix<\/code><\/p>\n<\/li>\n<li>\n<p>\u524d\u5411\u4f20\u64ad<\/p>\n<ul>\n<li>\u5c06\u9884\u5904\u7406\u540e\u7684\u56fe\u50cf\u6279\u6b21\u8f93\u5165\u5230 ViT \u6a21\u578b\u4e2d\u3002<\/li>\n<li>\u56fe\u50cf\u7ecf\u8fc7\u5206\u5757\u3001\u5d4c\u5165\u3001\u6dfb\u52a0 <code>[CLS]<\/code> \u6807\u8bb0\u548c\u4f4d\u7f6e\u7f16\u7801\u540e\uff0c\u5f62\u6210\u8f93\u5165\u5e8f\u5217\u3002<\/li>\n<li>\u8be5\u5e8f\u5217\u901a\u8fc7\u7531\u591a\u5c42 Transformer \u7f16\u7801\u5668\u5806\u53e0\u800c\u6210\u7684\u6838\u5fc3\u90e8\u5206\u3002\u6bcf\u4e00\u5c42\u90fd\u5305\u542b\u4e00\u4e2a\u591a\u5934\u81ea\u6ce8\u610f\u529b\uff08MHSA\uff09\u6a21\u5757\u548c\u4e00\u4e2a\u591a\u5c42\u611f\u77e5\u673a\uff08MLP\uff09\u6a21\u5757\uff0c\u5e76\u8f85\u4ee5\u5c42\u5f52\u4e00\u5316\uff08LayerNorm\uff09\u548c\u6b8b\u5dee\u8fde\u63a5\uff08Residual Connection\uff09\u6765\u7a33\u5b9a\u8bad\u7ec3\u3002<\/li>\n<li>MHSA \u6a21\u5757\u8ba9\u6a21\u578b\u80fd\u591f\u6355\u6349\u56fe\u50cf\u4e2d\u4efb\u610f\u4e24\u4e2a\u5757\u4e4b\u95f4\u7684\u5168\u5c40\u4f9d\u8d56\u5173\u7cfb\uff0c\u800c MLP \u6a21\u5757\u5219\u63d0\u4f9b\u975e\u7ebf\u6027\u53d8\u6362\u80fd\u529b\u3002<\/li>\n<\/ul>\n<\/li>\n<li>\n<p>\u635f\u5931\u8ba1\u7b97\u4e0e\u53cd\u5411\u4f20\u64ad<\/p>\n<p>\u6839\u636e\u4e0d\u540c\u76ee\u6807\uff0c\u901a\u5e38\u6709\u4ee5\u4e0b\u4e24\u79cd\u8bad\u7ec3\u76ee\u6807\uff1a<\/p>\n<ul>\n<li>\u5f53\u9700\u8981\u5c06ViT\u6a21\u578b\u7528\u4e8e\u56fe\u50cf\u5206\u7c7b\u65f6\uff0c\u5219\u5c06\u7ecf\u8fc7\u6240\u6709\u7f16\u7801\u5668\u5c42\u7684 <code>[CLS]<\/code> \u6807\u8bb0\u5bf9\u5e94\u7684\u6700\u7ec8\u8f93\u51fa\u5411\u91cf\uff0c\u518d\u8f93\u5165\u5230\u4e00\u4e2a\u8f7b\u91cf\u7ea7\u7684\u5206\u7c7b\u5934(\u901a\u5e38\u662f\u4e00\u4e2a\u591a\u5c42\u611f\u77e5\u673a)\uff0c\u6700\u540e\u4f7f\u7528\u4ea4\u53c9\u71b5\u635f\u5931\u51fd\u6570\u8ba1\u7b97\u6700\u7ec8\u7684\u635f\u5931<\/li>\n<li>\u5f53\u76ee\u6807\u662f\u5b66\u4e60\u56fe\u50cf\u7684\u901a\u7528\u89c6\u89c9\u7279\u5f81\uff0c\u5e76\u5c06\u5176\u5e94\u7528\u4e8e\u5404\u7c7b\u4e0b\u6e38\u4efb\u52a1\u65f6\uff0c\u901a\u5e38\u91c7\u7528\u81ea\u76d1\u7763\u5b66\u4e60\u7684\u65b9\u6cd5\u3002\u6b64\u65f6\uff0c\u6a21\u578b\u7684\u8bad\u7ec3\u76ee\u6807\u662f\u5b8c\u6210\u4e00\u4e2aMIM(Msaked Image Modeling)\u4efb\u52a1\uff0c\u7c7b\u4f3c\u4e8eBERT\u9884\u8bad\u7ec3\u65f6\u7684MLM\u4efb\u52a1\uff0c\u8be5\u4efb\u52a1\u4f1a\u8ba9\u6a21\u578b\u9884\u6d4b\u539f\u59cb\u56fe\u50cf\u4e2d\u88ab\u906e\u6321\u7684\u90e8\u5206<\/li>\n<\/ul>\n<\/li>\n<li>\n<p>\u53c2\u6570\u66f4\u65b0<\/p>\n<ul>\n<li>\u4f7f\u7528\u4f18\u5316\u5668\uff08\u5982 AdamW\uff09\u6839\u636e\u8ba1\u7b97\u51fa\u7684\u68af\u5ea6\u6765\u66f4\u65b0\u6a21\u578b\u53c2\u6570\u3002AdamW \u56e0\u5176\u80fd\u5c06\u6743\u91cd\u8870\u51cf\u4e0e L2 \u6b63\u5219\u5316\u89e3\u8026\uff0c\u66f4\u9002\u5408 ViT \u8fd9\u7c7b\u53c2\u6570\u91cf\u5927\u7684\u6a21\u578b\u3002<\/li>\n<li>\u4e3a\u4e86\u7a33\u5b9a\u8bad\u7ec3\u521d\u671f\u5e76\u8fbe\u5230\u66f4\u597d\u7684\u6536\u655b\u6548\u679c\uff0c\u901a\u5e38\u4f1a\u91c7\u7528\u201c\u7ebf\u6027\u9884\u70ed\uff08Warmup\uff09+ \u4f59\u5f26\u9000\u706b\uff08Cosine Decay\uff09\u201d\u7684\u5b66\u4e60\u7387\u8c03\u5ea6\u7b56\u7565\u3002<\/li>\n<\/ul>\n<\/li>\n<li>\n<p>\u8fed\u4ee3\u5faa\u73af<\/p>\n<ul>\n<li>\u91cd\u590d\u6267\u884c\u6b65\u9aa4 2 \u5230 4\uff0c\u904d\u5386\u6574\u4e2a\u6570\u636e\u96c6\u591a\u4e2a\u5468\u671f\uff08Epoch\uff09\uff0c\u76f4\u5230\u6a21\u578b\u6027\u80fd\u5728\u9a8c\u8bc1\u96c6\u4e0a\u4e0d\u518d\u63d0\u5347\u6216\u8fbe\u5230\u9884\u8bbe\u7684\u8bad\u7ec3\u5468\u671f\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<h4>1.4 ViT\u7684\u5e94\u7528<\/h4>\n<ul>\n<li>\u5305\u542b\u56fe\u50cf\u5206\u7c7b\u7b49\u5404\u79cd\u4f20\u7edf\u89c6\u89c9\u4efb\u52a1<\/li>\n<li>\u4ee5\u56fe\u641c\u56fe\u7b49\u573a\u666f<\/li>\n<li>\u5728\u6587\u751f\u56fe\u6a21\u578b\u4e2d\uff0c\u7528\u4e8e\u5185\u5bb9\u751f\u6210(\u5373\u53bb\u566a\u6216\u6d41\u5339\u914d)<\/li>\n<\/ul>\n<h4>1.5 Referrence<\/h4>\n<p><a href=\"http:\/\/ftp.thirdbody.cn\/study\/papers\/An%20Image%20is%20Worth%2016x16%20Words-%20Transformers%20for%20Image%20Recognition%20at%20Scale.pdf\">An Image is Worth 16x16 Words- Transformers for Image Recognition at Scale<\/a><\/p>\n<h3>2. CLIP(Contrastive Language\u2013Image Pretraining)\u6a21\u578b<\/h3>\n<h4>2.1 \u80cc\u666f<\/h4>\n<p>CLIP\uff08Contrastive Language\u2013Image Pretraining\uff0c\u5bf9\u6bd4\u8bed\u8a00-\u56fe\u50cf\u9884\u8bad\u7ec3\uff09\u662f\u7531 OpenAI \u5728 2021 \u5e74\u63d0\u51fa\u7684\u4e00\u79cd\u9769\u547d\u6027\u7684\u591a\u6a21\u6001\u6a21\u578b\u3002\u5b83\u7684\u6838\u5fc3\u601d\u60f3\u662f<strong>\u5229\u7528\u81ea\u7136\u8bed\u8a00\u4f5c\u4e3a\u76d1\u7763\u4fe1\u53f7\uff0c\u5c06\u56fe\u50cf\u548c\u6587\u672c\u6620\u5c04\u5230\u540c\u4e00\u4e2a\u8bed\u4e49\u7a7a\u95f4\u4e2d<\/strong>\uff0c\u4ece\u800c\u5b9e\u73b0\u5f3a\u5927\u7684\u96f6\u6837\u672c\uff08Zero-shot\uff09\u6cdb\u5316\u80fd\u529b\u3002<\/p>\n<p>\u7b80\u5355\u6765\u8bf4\uff0cCLIP \u5b66\u4f1a\u4e86\u7528\u201c\u8bed\u8a00\u201d\u6765\u7406\u89e3\u201c\u56fe\u50cf\u201d\uff0c\u8ba9\u8bed\u8a00\u6210\u4e3a\u4e86\u89c6\u89c9\u6a21\u578b\u7684\u201c\u63a5\u53e3\u201d\u3002<\/p>\n<h4>2.2 CLIP\u7684\u6838\u5fc3\u65b9\u6cd5\u8bba<\/h4>\n<p>CLIP \u7684\u8bad\u7ec3\u76ee\u6807\u975e\u5e38\u76f4\u89c2\uff1a\u7ed9\u5b9a\u4e00\u6279\u56fe\u50cf-\u6587\u672c\u5bf9\uff0c\u6a21\u578b\u4f1a\u62c9\u8fd1\u6b63\u786e\u914d\u5bf9\u7684\u56fe\u50cf\u548c\u6587\u672c\u5728\u5171\u4eab\u7279\u5f81\u7a7a\u95f4\u4e2d\u7684\u8ddd\u79bb\uff0c\u540c\u65f6\u63a8\u8fdc\u4e0d\u5339\u914d\u7684\u7ec4\u5408\u3002<\/p>\n<p><img decoding=\"async\" src=\"http:\/\/cdn.thirdbody.cn\/blog\/202605\/clip1.png\" alt=\"\" \/><\/p>\n<ul>\n<li><strong>\u53cc\u7f16\u7801\u5668\u67b6\u6784\uff1a<\/strong> CLIP \u7531\u4e24\u4e2a\u72ec\u7acb\u7684\u7f16\u7801\u5668\u7ec4\u6210\u3002\n<ol>\n<li><strong>\u56fe\u50cf\u7f16\u7801\u5668 (Image Encoder):<\/strong> \u8d1f\u8d23\u5c06\u56fe\u50cf\u8f6c\u6362\u4e3a\u7279\u5f81\u5411\u91cf\u3002OpenAI \u5728\u8bba\u6587\u4e2d\u5c1d\u8bd5\u4e86 ResNet \u548c Vision Transformer (ViT) \u4e24\u79cd\u67b6\u6784\uff0c\u6700\u7ec8 ViT \u8868\u73b0\u66f4\u4f73\u3002<\/li>\n<li><strong>\u6587\u672c\u7f16\u7801\u5668 (Text Encoder):<\/strong> \u8d1f\u8d23\u5c06\u6587\u672c\uff08\u5982\u201cA photo of a dog\u201d\uff09\u8f6c\u6362\u4e3a\u7279\u5f81\u5411\u91cf\u3002\u5b83\u901a\u5e38\u91c7\u7528\u7c7b\u4f3c GPT \u7684 Transformer \u67b6\u6784\u3002<\/li>\n<\/ol>\n<\/li>\n<li><strong>\u5bf9\u6bd4\u5b66\u4e60\u9884\u8bad\u7ec3\uff1a<\/strong> \u5728\u8bad\u7ec3\u65f6\uff0c\u6a21\u578b\u63a5\u6536\u4e00\u4e2a\u6279\u6b21\uff08Batch\uff09\u7684 N \u4e2a\u56fe\u50cf-\u6587\u672c\u5bf9\u3002\u5b83\u4f1a\u8ba1\u7b97 N\u00d7N \u4e2a\u6240\u6709\u53ef\u80fd\u7684\u56fe\u50cf-\u6587\u672c\u7ec4\u5408\u7684\u76f8\u4f3c\u5ea6\uff08\u5982\u4f59\u5f26\u76f8\u4f3c\u5ea6\uff09\u3002\n<ul>\n<li><strong>\u4f18\u5316\u76ee\u6807\uff1a<\/strong> \u6700\u5927\u5316 N \u4e2a\u771f\u5b9e\u914d\u5bf9\u7684\u76f8\u4f3c\u5ea6\uff08\u6b63\u6837\u672c\uff09\uff0c\u540c\u65f6\u6700\u5c0f\u5316 N\u00b2-N \u4e2a\u9519\u8bef\u914d\u5bf9\u7684\u76f8\u4f3c\u5ea6\uff08\u8d1f\u6837\u672c\uff09\u3002<\/li>\n<li><strong>\u8bad\u7ec3\u6570\u636e\uff1a<\/strong> CLIP \u7684\u6210\u529f\u79bb\u4e0d\u5f00\u5176\u89c4\u6a21\u5e9e\u5927\u7684\u8bad\u7ec3\u6570\u636e\u2014\u2014\u7ea6 4 \u4ebf\u5bf9\u4ece\u4e92\u8054\u7f51\u4e0a\u6536\u96c6\u7684\u56fe\u50cf\u53ca\u5176\u5bf9\u5e94\u7684\u6587\u672c\u63cf\u8ff0\uff0c\u8fd9\u4e3a\u6a21\u578b\u63d0\u4f9b\u4e86\u6781\u5176\u4e30\u5bcc\u7684\u81ea\u7136\u8bed\u8a00\u76d1\u7763\u4fe1\u53f7\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<h4>2.3 \u5bf9\u6bd4\u9884\u8bad\u7ec3<\/h4>\n<ul>\n<li>\n<p>\u6570\u636e\u51c6\u5907<\/p>\n<p>\u51c6\u5907\u6d77\u91cf\u7684\u201c\u56fe\u50cf-\u6587\u672c\u5bf9\u201d\u6570\u636e\u96c6\u3002OpenAI\u5728\u539f\u59cb\u8bba\u6587\u4e2d\u4f7f\u7528\u4e86\u7ea64\u4ebf\u5bf9\u4ece\u4e92\u8054\u7f51\u4e0a\u6293\u53d6\u7684\u56fe\u50cf\u53ca\u5176\u5bf9\u5e94\u7684\u6587\u672c\u63cf\u8ff0<\/p>\n<\/li>\n<li>\n<p>\u53cc\u7f16\u7801\u5668\u7279\u5f81\u63d0\u53d6<\/p>\n<p>CLIP\u91c7\u7528\u5bf9\u79f0\u7684\u201c\u53cc\u5854\u201d\u67b6\u6784\uff1a<\/p>\n<ul>\n<li><strong>\u56fe\u50cf\u7f16\u7801\u5668 (Image Encoder):<\/strong> \u5c06\u8f93\u5165\u7684\u56fe\u50cf\u8f6c\u6362\u4e3a\u4e00\u4e2a\u7279\u5f81\u5411\u91cf\u3002\u5e38\u7528\u7684\u67b6\u6784\u6709ResNet\u6216Vision Transformer (ViT)<\/li>\n<li><strong>\u6587\u672c\u7f16\u7801\u5668 (Text Encoder):<\/strong> \u5c06\u8f93\u5165\u7684\u6587\u672c\uff08\u5982\u201cA photo of a dog\u201d\uff09\u8f6c\u6362\u4e3a\u4e00\u4e2a\u7279\u5f81\u5411\u91cf\u3002\u5b83\u901a\u5e38\u91c7\u7528\u7c7b\u4f3cGPT\u7684Transformer\u67b6\u6784<\/li>\n<\/ul>\n<\/li>\n<li>\n<p>\u6295\u5f71\u4e0e\u5f52\u4e00\u5316<\/p>\n<ul>\n<li>\u56fe\u50cf\u548c\u6587\u672c\u7f16\u7801\u5668\u8f93\u51fa\u7684\u7279\u5f81\u5411\u91cf\u4f1a\u88ab\u5206\u522b\u901a\u8fc7\u4e00\u4e2a\u53ef\u5b66\u4e60\u7684\u7ebf\u6027\u6295\u5f71\u5c42\uff0c\u6620\u5c04\u5230\u540c\u4e00\u4e2a\u7ef4\u5ea6\uff08\u4f8b\u5982512\u7ef4\uff09\u7684\u5171\u4eab\u5d4c\u5165\u7a7a\u95f4\u3002<\/li>\n<li>\u968f\u540e\uff0c\u5bf9\u6295\u5f71\u540e\u7684\u5411\u91cf\u8fdb\u884cL2\u5f52\u4e00\u5316\uff0c\u4f7f\u5176\u6210\u4e3a\u5355\u4f4d\u5411\u91cf\u3002<\/li>\n<\/ul>\n<\/li>\n<li>\n<p>\u76f8\u4f3c\u5ea6\u548c\u635f\u5931\u8ba1\u7b97\u53ca\u53cd\u5411\u4f20\u64ad<\/p>\n<ul>\n<li>\u5728\u4e00\u4e2a\u5305\u542b <code>N<\/code> \u4e2a\u56fe\u50cf-\u6587\u672c\u5bf9\u7684\u6279\u6b21\uff08Batch\uff09\u4e2d\uff0c\u6a21\u578b\u4f1a\u8ba1\u7b97\u6240\u6709 <code>N\u00d7N<\/code> \u4e2a\u53ef\u80fd\u7684\u56fe\u6587\u7ec4\u5408\u4e4b\u95f4\u7684\u4f59\u5f26\u76f8\u4f3c\u5ea6\uff0c\u5f62\u6210\u4e00\u4e2a\u76f8\u4f3c\u5ea6\u77e9\u9635\u3002<\/li>\n<li>\u5728\u8fd9\u4e2a\u77e9\u9635\u4e2d\uff0c\u5bf9\u89d2\u7ebf\u4e0a\u7684 <code>N<\/code> \u4e2a\u5143\u7d20\u662f\u771f\u5b9e\u914d\u5bf9\u7684\u201c\u6b63\u6837\u672c\u201d\uff0c\u5176\u4f59 <code>N\u00b2-N<\/code> \u4e2a\u662f\u9519\u8bef\u914d\u5bf9\u7684\u201c\u8d1f\u6837\u672c\u201d\u3002<\/li>\n<li><strong>\u4f18\u5316\u76ee\u6807\uff1a<\/strong> \u8bad\u7ec3\u7684\u6838\u5fc3\u662f\u6700\u5927\u5316\u6b63\u6837\u672c\u5bf9\u7684\u76f8\u4f3c\u5ea6\uff0c\u540c\u65f6\u6700\u5c0f\u5316\u8d1f\u6837\u672c\u5bf9\u7684\u76f8\u4f3c\u5ea6\u3002\u8fd9\u901a\u8fc7\u4e00\u4e2a\u5bf9\u79f0\u7684\u5bf9\u6bd4\u635f\u5931\u51fd\u6570\uff08\u5982InfoNCE Loss\uff09\u6765\u5b9e\u73b0\uff0c\u5b83\u4f1a\u540c\u65f6\u4f18\u5316\u201c\u56fe\u50cf\u9884\u6d4b\u6587\u672c\u201d\u548c\u201c\u6587\u672c\u9884\u6d4b\u56fe\u50cf\u201d\u4e24\u4e2a\u65b9\u5411\u7684\u4ea4\u53c9\u71b5\u635f\u5931\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<h4>2.4 \u5e94\u7528<\/h4>\n<h5>2.4.1 Zero-Shot \u56fe\u50cf\u5206\u7c7b<\/h5>\n<ul>\n<li>\u8fd9\u662f CLIP \u6700\u7ecf\u5178\u7684\u5e94\u7528\u3002\u8981\u5bf9\u56fe\u50cf\u8fdb\u884c\u5206\u7c7b\uff0c\u53ea\u9700\u5c06\u7c7b\u522b\u540d\u79f0\uff08\u5982\u201c\u732b\u201d\u3001\u201c\u72d7\u201d\u3001\u201c\u6c7d\u8f66\u201d\uff09\u6784\u9020\u6210\u6587\u672c\u63d0\u793a\uff08Prompt\uff09\uff0c\u4f8b\u5982 \u201ca photo of a {\u7c7b\u522b}\u201d\u3002<\/li>\n<li>\u7136\u540e\uff0c\u8ba1\u7b97\u56fe\u50cf\u4e0e\u6240\u6709\u7c7b\u522b\u6587\u672c\u63d0\u793a\u7684\u76f8\u4f3c\u5ea6\uff0c\u76f8\u4f3c\u5ea6\u6700\u9ad8\u7684\u5373\u4e3a\u9884\u6d4b\u7ed3\u679c\u3002\u8fd9\u79cd\u65b9\u6cd5\u8ba9 CLIP \u53ef\u4ee5\u7075\u6d3b\u5730\u5206\u7c7b\u5b83\u5728\u8bad\u7ec3\u65f6\u4ece\u672a\u89c1\u8fc7\u7684\u7c7b\u522b\u3002<\/li>\n<\/ul>\n<h5>2.4.2 \u5728\u6587\u751f\u56fe\u6a21\u578b\u4e2d\u7684\u5e94\u7528<\/h5>\n<p>CLIP \u901a\u5e38\u4f5c\u4e3a\u6587\u672c\u7f16\u7801\u5668\uff0c\u8d1f\u8d23\u5c06\u7528\u6237\u7684\u6587\u672c\u63d0\u793a\uff08Prompt\uff09\u8f6c\u6362\u4e3a\u5bcc\u542b\u8bed\u4e49\u7684\u5d4c\u5165\u5411\u91cf\uff08Embedding\uff09\uff0c\u7528\u4ee5\u6307\u5bfc\u6269\u6563\u6a21\u578b\u751f\u6210\u4e0e\u6587\u672c\u9ad8\u5ea6\u76f8\u5173\u7684\u56fe\u50cf\u3002<\/p>\n<h5>2.4.3 \u8de8\u6a21\u6001\u68c0\u7d22<\/h5>\n<ul>\n<li>\u6587\u641c\u56fe<\/li>\n<li>\u56fe\u641c\u6587<\/li>\n<\/ul>\n<h4>2.5 Reference<\/h4>\n<p><a href=\"http:\/\/ftp.thirdbody.cn\/study\/papers\/Learning%20Transferable%20Visual%20Models%20From%20Natural%20Language%20Supervision.pdf\">Learning Transferable Visual Models From Natural Language Supervision<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>1. ViT(Vision Transformer)\u6a21\u578b 1.1 \u80cc\u666f \u57282020\u5e74\u4e4b\u524d\uff0c\u56fe\u50cf\u9886\u57df\u662fCNN\u7684&#038; [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":113,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[60],"tags":[64,63,61,62],"class_list":["post-114","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-multi_modal","tag-clip","tag-vit","tag-61","tag-62"],"_links":{"self":[{"href":"http:\/\/blog.thirdbody.cn\/index.php\/wp-json\/wp\/v2\/posts\/114","targetHints":{"allow":["GET"]}}],"collection":[{"href":"http:\/\/blog.thirdbody.cn\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"http:\/\/blog.thirdbody.cn\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"http:\/\/blog.thirdbody.cn\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"http:\/\/blog.thirdbody.cn\/index.php\/wp-json\/wp\/v2\/comments?post=114"}],"version-history":[{"count":1,"href":"http:\/\/blog.thirdbody.cn\/index.php\/wp-json\/wp\/v2\/posts\/114\/revisions"}],"predecessor-version":[{"id":115,"href":"http:\/\/blog.thirdbody.cn\/index.php\/wp-json\/wp\/v2\/posts\/114\/revisions\/115"}],"wp:featuredmedia":[{"embeddable":true,"href":"http:\/\/blog.thirdbody.cn\/index.php\/wp-json\/wp\/v2\/media\/113"}],"wp:attachment":[{"href":"http:\/\/blog.thirdbody.cn\/index.php\/wp-json\/wp\/v2\/media?parent=114"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/blog.thirdbody.cn\/index.php\/wp-json\/wp\/v2\/categories?post=114"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/blog.thirdbody.cn\/index.php\/wp-json\/wp\/v2\/tags?post=114"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}