{"id":2989,"date":"2025-04-25T00:07:42","date_gmt":"2025-04-24T16:07:42","guid":{"rendered":"https:\/\/www.gnn.club\/?p=2989"},"modified":"2025-04-25T16:05:13","modified_gmt":"2025-04-25T08:05:13","slug":"albef%ef%bc%88align-before-fuse%ef%bc%89%ef%bc%9a%e6%a8%a1%e6%80%81%e5%af%b9%e9%bd%90%e4%b8%8e%e7%9f%a5%e8%af%86%e8%92%b8%e9%a6%8f","status":"publish","type":"post","link":"http:\/\/gnn.club\/?p=2989","title":{"rendered":"ALBEF\uff08Align Before Fuse\uff09\uff1a\u6a21\u6001\u5bf9\u9f50\u4e0e\u77e5\u8bc6\u84b8\u998f"},"content":{"rendered":"\n<h2 class=\"wp-block-heading\">1. \u6458\u8981<\/h2>\n\n\n\n<ol class=\"wp-block-list\">\n<li><a href=\"https:\/\/link.zhihu.com\/?target=https%3A\/\/arxiv.org\/abs\/2107.07651\" target=\"_blank\" rel=\"noreferrer noopener\">Align before Fuse: Vision and Language Representation Learning with Momentum Distillation<\/a><\/li>\n\n\n\n<li><a href=\"https:\/\/link.zhihu.com\/?target=https%3A\/\/github.com\/salesforce\/ALBEF%3Ftab%3Dreadme-ov-file\" target=\"_blank\" rel=\"noreferrer noopener\">GitHub - salesforce\/ALBEF: Code for ALBEF: a new vision-language pre-training method<\/a><\/li>\n<\/ol>\n\n\n\n<h4 class=\"wp-block-heading\">1.<strong>1  \u95ee\u9898<\/strong><\/h4>\n\n\n\n<p>\u73b0\u6709\u7684\u591a\u6a21\u6001\uff08\u56fe\u50cf+\u6587\u672c\uff09\u6a21\u578b\u901a\u5e38\u7528\u4e00\u4e2aTransformer\u7f16\u7801\u5668\u540c\u65f6\u5904\u7406\u56fe\u50cf\u7279\u5f81\u548c\u6587\u672c\u8bcd\u7b26\u3002\u4f46<strong>\u56fe\u50cf\u533a\u57df\u548c\u5355\u8bcd\u4e4b\u95f4\u6ca1\u6709\u5929\u7136\u7684\u5bf9\u9f50\u5173\u7cfb<\/strong>\uff0c\u5bfc\u81f4\u6a21\u578b\u96be\u4ee5\u51c6\u786e\u5b66\u4e60\u4e24\u8005\u4e4b\u95f4\u7684\u4ea4\u4e92\u3002\u6bd4\u5982\uff0c\u6a21\u578b\u53ef\u80fd\u4e0d\u77e5\u9053\u56fe\u7247\u4e2d\u7684\"\u72d7\"\u5bf9\u5e94\u6587\u672c\u4e2d\u7684\u54ea\u4e2a\u8bcd\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\"><strong>1.2 \u89e3\u51b3\u65b9\u6848\uff1aALBEF<\/strong><\/h4>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u5148\u5bf9\u9f50\u518d\u878d\u5408<\/strong>\uff1a\u5728\u878d\u5408\u56fe\u50cf\u548c\u6587\u672c\u7279\u5f81\u4e4b\u524d\uff0c\u5148\u7528\u4e00\u79cd<strong>\u5bf9\u6bd4\u635f\u5931\uff08contrastive loss\uff09<\/strong>\u8ba9\u56fe\u50cf\u548c\u6587\u672c\u7684\u8868\u5f81\u5728\u7279\u5f81\u7a7a\u95f4\u4e2d\u5bf9\u9f50\uff08\u6bd4\u5982\u76f8\u4f3c\u7684\u56fe\u50cf\u548c\u6587\u672c\u5728\u7279\u5f81\u7a7a\u95f4\u4e2d\u9760\u8fd1\uff09\u3002\u8fd9\u76f8\u5f53\u4e8e\u8ba9\u6a21\u578b\u5148\u5b66\u4f1a\"\u914d\u5bf9\"\uff0c\u518d\u6df1\u5165\u5b66\u4e60\u7ec6\u8282\u3002<\/li>\n\n\n\n<li><strong>\u65e0\u9700\u989d\u5916\u6807\u6ce8<\/strong>\uff1a\u4e0d\u50cf\u4e00\u4e9b\u65b9\u6cd5\u9700\u8981\u4eba\u5de5\u6807\u6ce8\u7684\u7269\u4f53\u8fb9\u754c\u6846\uff08bounding boxes\uff09\u6216\u9ad8\u6e05\u56fe\u7247\uff0cALBEF\u76f4\u63a5\u4ece\u666e\u901a\u7f51\u7edc\u6570\u636e\u5b66\u4e60\uff0c\u66f4\u5b9e\u7528\u3002<\/li>\n\n\n\n<li><strong>\u6297\u566a\u58f0\u6280\u5de7\uff1a\u52a8\u91cf\u84b8\u998f<\/strong><br>\u7f51\u7edc\u6570\u636e\uff08\u6bd4\u5982\u56fe\u7247\u914d\u9519\u8bef\u6587\u5b57\uff09\u566a\u58f0\u5927\uff0c\u4f5c\u8005\u63d0\u51fa<strong>\u52a8\u91cf\u84b8\u998f<\/strong>\uff1a\u7528\u4e00\u4e2a\u6301\u7eed\u66f4\u65b0\u7684\u201c\u52a8\u91cf\u6a21\u578b\u201d\uff08\u7c7b\u4f3c\u8001\u5e08\uff09\u751f\u6210\u4f2a\u6807\u7b7e\uff0c\u6307\u5bfc\u5b66\u751f\u6a21\u578b\u8bad\u7ec3\uff0c\u51cf\u5c11\u566a\u58f0\u5e72\u6270\u3002<\/li>\n<\/ul>\n\n\n\n<h2 class=\"wp-block-heading\">2. \u6a21\u578b\u67b6\u6784<\/h2>\n\n\n\n<p>ALBEF\uff08Align Before Fusing\uff09\u7684\u6a21\u578b\u7ed3\u6784\u901a\u8fc7<strong>\u56fe\u50cf\u7f16\u7801\u5668\u3001\u6587\u672c\u7f16\u7801\u5668\u548c\u591a\u6a21\u6001\u7f16\u7801\u5668<\/strong>\u4e09\u90e8\u5206\u534f\u540c\u5de5\u4f5c\uff0c\u5b9e\u73b0\u56fe\u50cf\u4e0e\u6587\u672c\u7684\u9ad8\u6548\u5bf9\u9f50\u548c\u878d\u5408\u3002\u4ee5\u4e0b\u662f\u8be6\u7ec6\u89e3\u6790\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"1225\" height=\"748\" src=\"https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/04\/20250421211315671.png\" alt=\"\" class=\"wp-image-2991\" srcset=\"https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/04\/20250421211315671.png 1225w, https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/04\/20250421211315671-300x183.png 300w, https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/04\/20250421211315671-1024x625.png 1024w, https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/04\/20250421211315671-768x469.png 768w\" sizes=\"auto, (max-width: 1225px) 100vw, 1225px\" \/><\/figure>\n\n\n\n<h4 class=\"wp-block-heading\"><strong>2.1 \u5404\u7ec4\u4ef6\u7ec6\u8282<\/strong><\/h4>\n\n\n\n<h5 class=\"wp-block-heading\"><strong>(1) \u56fe\u50cf\u7f16\u7801\u5668<\/strong><\/h5>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u7ed3\u6784<\/strong>\uff1a\u91c7\u752812\u5c42\u7684<strong>ViT-B\/16<\/strong>\uff08Vision Transformer\uff09\uff0c\u9884\u8bad\u7ec3\u4e8eImageNet-1k\u3002<\/li>\n\n\n\n<li><strong>\u8f93\u5165\u5904\u7406<\/strong>\uff1a\n<ul class=\"wp-block-list\">\n<li>\u56fe\u50cf&nbsp;I&nbsp;\u88ab\u5206\u5272\u4e3a\u82e5\u5e72\u5757\uff08patches\uff09\uff0c\u901a\u8fc7\u7ebf\u6027\u6295\u5f71\u5f97\u5230\u5d4c\u5165\u5e8f\u5217\uff0c\u4e5f\u5c31\u662f\u56fe\u50cf\u63d0\u53d6\u5230\u7684\u7279\u5f81\u5411\u91cf\u3002<\/li>\n\n\n\n<li>$v_{cls}$\u662f\u5206\u7c7b\u6807\u8bb0\uff08\u7c7b\u4f3cBERT\u7684[CLS]\uff09\uff0c\u7528\u4e8e\u6c47\u603b\u5168\u5c40\u56fe\u50cf\u4fe1\u606f\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n\n\n\n<h5 class=\"wp-block-heading\"><strong>(2) \u6587\u672c\u7f16\u7801\u5668<\/strong><\/h5>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u7ed3\u6784<\/strong>\uff1a6\u5c42Transformer\uff0c\u521d\u59cb\u5316\u81eaBERT-base\u7684\u524d6\u5c42\u3002<\/li>\n\n\n\n<li><strong>\u8f93\u5165\u5904\u7406<\/strong>\uff1a\n<ul class=\"wp-block-list\">\n<li>\u6587\u672c&nbsp;T&nbsp;\u88ab\u8f6c\u6362\u4e3a\u5d4c\u5165\u5e8f\u5217\uff0c\u5176\u4e2d $w_{c l s}$ \u662f\u6587\u672c\u7684\u5168\u5c40\u8868\u5f81\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n\n\n\n<h5 class=\"wp-block-heading\"><strong>(3) \u591a\u6a21\u6001\u7f16\u7801\u5668<\/strong><\/h5>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u7ed3\u6784<\/strong>\uff1a6\u5c42Transformer\uff0c\u521d\u59cb\u5316\u81eaBERT-base\u7684\u540e6\u5c42\u3002<\/li>\n\n\n\n<li><strong>\u878d\u5408\u673a\u5236<\/strong>\uff1a\n<ul class=\"wp-block-list\">\n<li>\u6bcf\u5c42\u901a\u8fc7<strong>\u4ea4\u53c9\u6ce8\u610f\u529b\uff08Cross Attention\uff09<\/strong>\u5c06\u56fe\u50cf\u7279\u5f81\u4e0e\u6587\u672c\u7279\u5f81\u4ea4\u4e92\uff1a\n<ul class=\"wp-block-list\">\n<li>Query\u6765\u81ea\u6587\u672c\u7684token\u7279\u5f81\uff0cKey\u548cValue\u6765\u81ea\u56fe\u50cf\u7684token\u7279\u5f81\uff0c\u5b9e\u73b0\u6587\u672c\u5230\u56fe\u50cf\u7684\u6ce8\u610f\u529b\u4ea4\u4e92\u3002<\/li>\n\n\n\n<li>\u81ea\u6ce8\u610f\u529b\uff08Self Attention\uff09\u8fdb\u4e00\u6b65\u7ec6\u5316\u6587\u672c\u5185\u90e8\u548c\u56fe\u50cf\u5185\u90e8\u7684\u5173\u7cfb\u3002<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li>\u6700\u7ec8\u8f93\u51fa\u878d\u5408\u540e\u7684\u591a\u6a21\u6001\u8868\u5f81\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n\n\n\n<h4 class=\"wp-block-heading\">2.2 <strong>\u5173\u952e\u8bad\u7ec3\u4efb\u52a1\u4e0e\u635f\u5931\u51fd\u6570<\/strong><\/h4>\n\n\n\n<h5 class=\"wp-block-heading\"><strong>(1) <\/strong>\u56fe\u50cf\uff0d\u6587\u672c\u5bf9\u6bd4\u635f\u5931\uff08Image\uff0dText Contrastive Loss\uff0cITC\uff09<\/h5>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u76ee\u7684\uff1a\u62c9\u8fd1\u5339\u914d\u7684\u56fe\u50cf\uff0d\u6587\u672c\u5bf9\u7684\u5d4c\u5165\u8868\u793a\uff0c\u63a8\u5f00\u4e0d\u5339\u914d\u7684\u5bf9\uff0c\u5b9e\u73b0\u8de8\u6a21\u6001\u5bf9\u9f50\u3002<\/li>\n\n\n\n<li>\u5b9e\u73b0\uff1a<\/li>\n\n\n\n<li>\u5206\u522b\u901a\u8fc7\u56fe\u50cf\u7f16\u7801\u5668\u548c\u6587\u672c\u7f16\u7801\u5668\u63d0\u53d6\u56fe\u50cf\u7279\u5f81 $v$ \u548c\u6587\u672c\u7279\u5f81 $t$ \u3002<\/li>\n\n\n\n<li>\u8ba1\u7b97\u6240\u6709\u56fe\u50cf\uff0d\u6587\u672c\u5bf9\u7684\u4f59\u5f26\u76f8\u4f3c\u5ea6\u77e9\u9635 $S$ \uff0c\u5176\u4e2d $S_{i, j}$ \u8868\u793a\u7b2c $i$ \u4e2a\u56fe\u50cf\u548c\u7b2c $j$ \u4e2a\u6587\u672c\u7684\u76f8\u4f3c\u5ea6\u3002<\/li>\n\n\n\n<li>\u5bf9\u76f8\u4f3c\u5ea6\u77e9\u9635\u5206\u522b\u6cbf\u56fe\u50cf\u548c\u6587\u672c\u7ef4\u5ea6\u8ba1\u7b97\u4ea4\u53c9\u6458\u635f\u5931\uff0c\u4f18\u5316\u53cc\u5411\u5bf9\u9f50\uff1a<\/li>\n<\/ul>\n\n\n\n<p>$$<br>\\mathcal{L}_{\\mathrm{ITC}}=\\frac{1}{2}\\left(\\mathrm{CE}(S, y)+\\mathrm{CE}\\left(S^{\\top}, y\\right)\\right)<br>$$<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>$y$ \u662f\u771f\u5b9e\u6807\u7b7e\uff08\u5bf9\u89d2\u7ebf\u4e3a\u5339\u914d\u5bf9\uff09\uff0cCE \u662f\u4ea4\u53c9\u6458\u635f\u5931\u3002<\/li>\n\n\n\n<li>\u4f5c\u7528\uff1a\u786e\u4fdd\u6a21\u578b\u80fd\u591f\u533a\u5206\u6b63\u8d1f\u6837\u672c\uff0c\u4e3a\u540e\u7eed\u4efb\u52a1\u63d0\u4f9b\u826f\u597d\u7684\u521d\u59cb\u5316\u7279\u5f81\u3002<\/li>\n<\/ul>\n\n\n\n<h5 class=\"wp-block-heading\"><strong>(2)<\/strong> \u56fe\u50cf\uff0d\u6587\u672c\u5339\u914d\u635f\u5931\uff08Image\uff0dText Matching Loss\uff0cITM\uff09<\/h5>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u76ee\u7684\uff1a\u5b66\u4e60\u7ec6\u7c92\u5ea6\u7684\u56fe\u50cf\uff0d\u6587\u672c\u5339\u914d\u5173\u7cfb\uff0c\u5224\u65ad\u4e8c\u8005\u662f\u5426\u8bed\u4e49\u5339\u914d\u3002<\/li>\n\n\n\n<li>\u5b9e\u73b0\uff1a<\/li>\n\n\n\n<li>\u4ece\u6279\u6b21\u4e2d\u91c7\u6837\u90e8\u5206\u56fe\u50cf\uff0d\u6587\u672c\u5bf9\uff08\u5305\u62ec\u6b63\u6837\u672c\u548c\u8d1f\u6837\u672c\uff09\uff0c\u8d1f\u6837\u672c\u901a\u8fc7ITC\u76f8\u4f3c\u5ea6\u9009\u62e9\u6700\u96be\u6837\u672c \uff08Hard Negative Mining\uff09\u3002<br>\uff0d\u5c06\u56fe\u50cf\u548c\u6587\u672c\u7684\u878d\u5408\u7279\u5f81\uff08\u901a\u8fc7\u8de8\u6a21\u6001\u7f16\u7801\u5668\uff09\u8f93\u5165\u4e00\u4e2a\u4e8c\u5206\u7c7b\u5934\uff0c\u9884\u6d4b\u5339\u914d\u6982\u7387 $p$ \uff1a<\/li>\n<\/ul>\n\n\n\n<p>$$<br>\\mathcal{L}_{\\mathrm{ITM}}=\\mathrm{BCE}(p, y)<br>$$<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>$y \\in{0,1}$ \u8868\u793a\u662f\u5426\u5339\u914d\uff0c BCE \u662f\u4e8c\u5143\u4ea4\u53c9\u6458\u635f\u5931\u3002<\/li>\n\n\n\n<li>\u4f5c\u7528\uff1a\u589e\u5f3a\u6a21\u578b\u5bf9\u8de8\u6a21\u6001\u8bed\u4e49\u4e00\u81f4\u6027\u7684\u7406\u89e3\uff0c\u5c24\u5176\u5173\u6ce8\u96be\u8d1f\u6837\u672c\u3002<\/li>\n<\/ul>\n\n\n\n<h5 class=\"wp-block-heading\"><strong>(3) <\/strong>\u63a9\u7801\u8bed\u8a00\u5efa\u6a21\u635f\u5931\uff08Masked Language Modeling Loss\uff0cMLM\uff09<\/h5>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u76ee\u7684\uff1a\u901a\u8fc7\u6587\u672c\u91cd\u5efa\u4efb\u52a1\uff0c\u5229\u7528\u56fe\u50cf\u4fe1\u606f\u8f85\u52a9\u7406\u89e3\u88ab\u63a9\u7801\u7684\u6587\u672c\uff0c\u5b66\u4e60\u4e0a\u4e0b\u6587\u611f\u77e5\u7684\u6587\u672c\u8868\u793a\u3002<\/li>\n\n\n\n<li>\u5b9e\u73b0\uff1a<\/li>\n\n\n\n<li>\u968f\u673a\u63a9\u7801\u6587\u672c\u4e2d\u7684\u90e8\u5206\u8bcd\uff08\u598215\uff05\uff09\uff0c\u7528\uff3bMASK\uff3d\u6807\u8bb0\u66ff\u6362\u3002<\/li>\n\n\n\n<li>\u7ed3\u5408\u56fe\u50cf\u7279\u5f81\u548c\u672a\u88ab\u63a9\u7801\u7684\u6587\u672c\u4e0a\u4e0b\u6587\uff0c\u9884\u6d4b\u88ab\u63a9\u7801\u7684\u8bcd\uff1a<\/li>\n<\/ul>\n\n\n\n<p>$$<br>\\mathcal{L}_{\\mathrm{MLM}}=\\mathrm{CE}(p_{mask}, y_{mask})<br>$$<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>$y_{\\text {mask }}$ \u662f\u88ab\u63a9\u7801\u8bcd\u7684\u771f\u5b9e\u6807\u7b7e\u3002<\/li>\n\n\n\n<li>\u4f5c\u7528\uff1a\u63d0\u5347\u6a21\u578b\u7684\u591a\u6a21\u6001\u63a8\u7406\u80fd\u529b\uff0c\u4f7f\u6587\u672c\u7406\u89e3\u53d7\u89c6\u89c9\u4fe1\u606f\u589e\u5f3a\u3002<\/li>\n<\/ul>\n\n\n\n<h5 class=\"wp-block-heading\"><strong>(4) \u52a8\u91cf\u84b8\u998f\u635f\u5931\uff08Momentum Distillation\uff09<\/strong><\/h5>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u76ee\u7684<\/strong>\uff1a\u7f13\u89e3\u7f51\u7edc\u6570\u636e\u7684\u566a\u58f0\u95ee\u9898\u3002<\/li>\n\n\n\n<li><strong>\u65b9\u6cd5<\/strong>\uff1a\n<ul class=\"wp-block-list\">\n<li>\u4f7f\u7528<strong>\u52a8\u91cf\u6a21\u578b<\/strong>\uff08\u57fa\u6a21\u578b\u7684\u6ed1\u52a8\u5e73\u5747\u7248\u672c\uff09\u751f\u6210\u4f2a\u6807\u7b7e\uff0c\u4f5c\u4e3a\u989d\u5916\u76d1\u7763\u4fe1\u53f7\u3002<\/li>\n\n\n\n<li>\u52a8\u91cf\u6a21\u578b\u66f4\u7a33\u5b9a\uff0c\u80fd\u63d0\u4f9b\u66f4\u53ef\u9760\u7684\u4f2a\u76ee\u6807\u3002<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li>\u52a8\u91cf\u84b8\u998f\u7684\u76f8\u5173\u4ecb\u7ecd\uff0c\u8be6\u89c1\u7b2c\u4e09\u7ae0\u8282\u3002<\/li>\n<\/ul>\n\n\n\n<h4 class=\"wp-block-heading\">2.3 <strong>\u603b\u7ed3<\/strong><\/h4>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u5148\u5bf9\u9f50\u540e\u878d\u5408<\/strong>\uff1a\u901a\u8fc7ITC\u635f\u5931\u63d0\u524d\u5bf9\u9f50\u5355\u6a21\u6001\u7279\u5f81\uff0c\u7b80\u5316\u591a\u6a21\u6001\u7f16\u7801\u5668\u7684\u5b66\u4e60\u96be\u5ea6\u3002<\/li>\n\n\n\n<li><strong>\u65e0\u9700\u989d\u5916\u6807\u6ce8<\/strong>\uff1a\u4ec5\u9700\u666e\u901a\u56fe\u6587\u5bf9\uff0c\u4e0d\u4f9d\u8d56\u7269\u4f53\u68c0\u6d4b\u6846\u6216\u9ad8\u5206\u8fa8\u7387\u56fe\u50cf\u3002<\/li>\n\n\n\n<li><strong>\u6297\u566a\u58f0\u8bbe\u8ba1<\/strong>\uff1a\u52a8\u91cf\u84b8\u998f\u5229\u7528\u52a8\u91cf\u6a21\u578b\u7684\u4f2a\u6807\u7b7e\u8fc7\u6ee4\u566a\u58f0\u6570\u636e\u3002<\/li>\n<\/ul>\n\n\n\n<p>\u901a\u8fc7\u8fd9\u79cd\u7ed3\u6784\uff0cALBEF\u5728\u51cf\u5c11\u8ba1\u7b97\u6210\u672c\u7684\u540c\u65f6\uff0c\u5b9e\u73b0\u4e86\u66f4\u9ad8\u6548\u7684\u8de8\u6a21\u6001\u7406\u89e3\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">3. <strong>\u52a8\u91cf\u84b8\u998f\uff08Momentum Distillation, MoD\uff09<\/strong><\/h2>\n\n\n\n<p>\u52a8\u91cf\u84b8\u998f\u662f\u77e5\u8bc6\u84b8\u998f\u7684\u5ef6\u7533\uff0c\u5176\u6838\u5fc3\u601d\u60f3\u662f\u901a\u8fc7\u4e00\u4e2a\u66f4\u5f3a\u5927\u7684\u201c\u6559\u5e08\u6a21\u578b\u201d\u6307\u5bfc\u201c\u5b66\u751f\u6a21\u578b\u201d\u7684\u8bad\u7ec3\uff0c\u4ece\u800c\u63d0\u5347\u5b66\u751f\u6a21\u578b\u7684\u6027\u80fd\u3002ALBEF\u9488\u5bf9\u591a\u6a21\u6001\u9884\u8bad\u7ec3\u4e2d<strong>\u7f51\u7edc\u6570\u636e\u566a\u58f0\u5927<\/strong>\u7684\u95ee\u9898\uff0c\u63d0\u51fa<strong>\u52a8\u91cf\u84b8\u998f<\/strong>\uff0c\u4ee5\u4e0b\u662f\u8be6\u7ec6\u89e3\u6790\uff1a<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">3.<strong>1  \u80cc\u666f\u4e0e\u95ee\u9898<\/strong><\/h4>\n\n\n\n<p><strong>\u4e3a\u4ec0\u4e48\u9700\u8981\u52a8\u91cf\u84b8\u998f\uff1f<\/strong><\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u6570\u636e\u566a\u58f0<\/strong>\uff1a\u7f51\u7edc\u722c\u53d6\u7684\u56fe\u6587\u5bf9\uff08Image-Text Pairs\uff09\u901a\u5e38\u5b58\u5728\u5f31\u76f8\u5173\u6027\uff08\u4f8b\u5982\u56fe\u7247\u662f\u201c\u72d7\u201d\uff0c\u6587\u672c\u63cf\u8ff0\u201c\u516c\u56ed\u6563\u6b65\u201d\u4f46\u672a\u660e\u786e\u63d0\u5230\u72d7\uff09\u3002<\/li>\n\n\n\n<li><strong>\u4f20\u7edf\u65b9\u6cd5\u7684\u5c40\u9650<\/strong>\uff1a\n<ul class=\"wp-block-list\">\n<li>ITC\uff08\u56fe\u50cf-\u6587\u672c\u5bf9\u6bd4\u5b66\u4e60\uff09\u548cMLM\uff08\u63a9\u7801\u8bed\u8a00\u5efa\u6a21\uff09\u4f7f\u7528<strong>one-hot\u6807\u7b7e<\/strong>\uff0c\u4f1a\u9519\u8bef\u60e9\u7f5a\u5b9e\u9645\u5408\u7406\u7684\u8d1f\u6837\u672c\uff08\u5982\u201c\u72d7\u201d\u7684\u8d1f\u6837\u672c\u201c\u732b\u201d\u53ef\u80fd\u4e5f\u4e0e\u56fe\u7247\u90e8\u5206\u76f8\u5173\uff09\u3002<\/li>\n\n\n\n<li>\u76f4\u63a5\u4f7f\u7528\u566a\u58f0\u6570\u636e\u8bad\u7ec3\u4f1a\u5bfc\u81f4\u6a21\u578b\u5b66\u4e60\u5230\u9519\u8bef\u7684\u5173\u8054\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n\n\n\n<h4 class=\"wp-block-heading\"><strong>3.2 \u52a8\u91cf\u84b8\u998f\u7684\u673a\u5236<\/strong><\/h4>\n\n\n\n<h5 class=\"wp-block-heading\"><strong>(1) \u52a8\u91cf\u6a21\u578b\uff08Momentum Model\uff09<\/strong><\/h5>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u89d2\u8272<\/strong>\uff1a\u4f5c\u4e3a\u201c\u6559\u5e08\u6a21\u578b\u201d\uff0c\u901a\u8fc7\u6ed1\u52a8\u5e73\u5747\uff08Exponential Moving Average, EMA\uff09\u66f4\u65b0\uff0c\u6bd4\u5b66\u751f\u6a21\u578b\uff08\u57fa\u6a21\u578b\uff09\u66f4\u7a33\u5b9a\u3002<\/li>\n\n\n\n<li><strong>\u66f4\u65b0\u65b9\u5f0f<\/strong>\uff1a<\/li>\n<\/ul>\n\n\n\n<p>\u6559\u5e08\u6a21\u578b\u7684\u53c2\u6570 $\\theta_t$ \u662f\u5b66\u751f\u6a21\u578b\u53c2\u6570 $\\theta_s$ \u7684\u5386\u53f2\u52a0\u6743\u5e73\u5747\uff1a<\/p>\n\n\n\n<p>$$<br>\\theta_t \\leftarrow \\lambda \\theta_t+(1-\\lambda) \\theta_s<br>$$<\/p>\n\n\n\n<p>\u5176\u4e2d $\\lambda$ \u662f\u52a8\u91cf\u7cfb\u6570\uff08\u5982 0.995 \uff09\uff0c\u63a7\u5236\u6559\u5e08\u6a21\u578b\u7684\u66f4\u65b0\u5e73\u6ed1\u5ea6\u3002<\/p>\n\n\n\n<h5 class=\"wp-block-heading\"><strong>(2) \u751f\u6210\u4f2a\u6807\u7b7e\uff08Pseudo-Targets\uff09<\/strong><\/h5>\n\n\n\n<p>\u6559\u5e08\u6a21\u578b\u5bf9\u8f93\u5165\u6570\u636e\u751f\u6210<strong>\u8f6f\u6807\u7b7e<\/strong>\uff08pseudo-target\uff09\uff0c\u800c\u975e\u786c\u6807\u7b7e\uff08one-hot\uff09\uff0c\u4ece\u800c\u6355\u6349\u66f4\u5408\u7406\u7684\u8bed\u4e49\u5173\u8054\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>ITC\u4efb\u52a1<\/strong>\uff1a<\/li>\n<\/ul>\n\n\n\n<p>\u6559\u5e08\u6a21\u578b\u8ba1\u7b97\u56fe\u50cf\uff0d\u6587\u672c\u76f8\u4f3c\u5ea6 $s^{\\prime}(I, T)$ \uff0c\u5e76\u751f\u6210\u8f6f\u76ee\u6807\u5206\u5e03 $q^{i 2 t}$ \u548c $q^{t 2 i}$ \u3002<br>\uff0d\u4f8b\u5982\uff0c\u56fe\u7247\uff02\u72d7\uff02\u53ef\u80fd\u4e0e\u6587\u672c\uff02\u5ba0\u7269\uff02\u7684\u76f8\u4f3c\u5ea6\u9ad8\u4e8e\uff02\u6c7d\u8f66\uff02\uff0c\u4f46\u4f20\u7edfone\uff0dhot\u6807\u7b7e\u4f1a\u5ffd\u7565\u8fd9\u79cd\u68af\u5ea6\u5173\u7cfb\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>MLM\u4efb\u52a1<\/strong>\uff1a<\/li>\n<\/ul>\n\n\n\n<p>\u6559\u5e08\u6a21\u578b\u9884\u6d4b\u88ab\u63a9\u7801\u8bcd\u7684\u5206\u5e03 $q^{\\mathrm{msk}}(I, \\hat{T})$ \uff0c\u5141\u8bb8\u6a21\u578b\u5b66\u4e60\u591a\u4e2a\u5408\u7406\u66ff\u6362\u8bcd\uff08\u5982\uff02\u72d7\uff02\u53ef\u88ab\uff02\u72ac\u79d1\u52a8\u7269\uff02\u66ff\u4ee3\uff09\u3002<\/p>\n\n\n\n<p>\u5bf9\u6b64\uff0c\u539f\u6587\u4e2d\u7684\u56fe2\u4e3e\u4f8b\u8be6\u7ec6\u7684\u4f8b\u5b50\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"1319\" height=\"512\" src=\"https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/04\/20250424221026887.png\" alt=\"\" class=\"wp-image-3002\" srcset=\"https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/04\/20250424221026887.png 1319w, https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/04\/20250424221026887-300x116.png 300w, https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/04\/20250424221026887-1024x397.png 1024w, https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/04\/20250424221026887-768x298.png 768w\" sizes=\"auto, (max-width: 1319px) 100vw, 1319px\" \/><\/figure>\n\n\n\n<p>\u56fe2\u4e2d\u7684\u4e94\u5f20\u56fe\u7247\u7684\u539f\u6807\u7b7e\uff08one-hot\u7684\u786c\u6807\u7b7e\uff09\u5206\u522b\u662f<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>polar bear in the [wild]<\/li>\n\n\n\n<li>a man [standing] along a road in front of nature in summer<\/li>\n\n\n\n<li>a [remote] waterfall in the deep woods<\/li>\n\n\n\n<li>breakdown of the car on the road<\/li>\n\n\n\n<li>the harbor a small village<\/li>\n<\/ul>\n\n\n\n<p>\u5bf9\u6bd4\u4e4b\u4e0b\uff0c\u6839\u636e\u6982\u7387\u5206\u5e03\u751f\u6210\u7684\u751f\u6210\u7684top-5\u4f2a\u6807\u7b7e\uff08<strong>\u8f6f\u6807\u7b7e<\/strong>\uff09\u63cf\u8ff0\u7684\u66f4\u52a0\u5408\u7406\uff0c\u4e0d\u7ba1\u662f\u7528\u201czoo\u201d\u4ee3\u66ff\u201cwild\u201d\uff1b\u201cwalks\u201d\u4ee3\u66ff\u201cstanding\u201d\u8fd8\u662f\u201csmall or beautiful\u201d\u4ee3\u66ffremote\u90fd\u66f4\u52a0\u7684\u5408\u7406\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\"><strong>3.3 \u52a8\u91cf\u84b8\u998f\u7684\u635f\u5931\u51fd\u6570\u8bbe\u8ba1<\/strong><\/h4>\n\n\n\n<p>\u52a8\u91cf\u84b8\u998f\u7684\u635f\u5931\u5728ITC\u548cMLM\u4efb\u52a1\u4e0a\u90fd\u6709\u6dfb\u52a0\uff1a<\/p>\n\n\n\n<p>\uff081\uff09ITC\u4efb\u52a1\u7684\u52a8\u91cf\u635f\u5931\uff08$\\mathbf{ITC}_{\\mathrm{MoD}}$\uff09<\/p>\n\n\n<div class=\"wp-block-image\">\n<figure class=\"aligncenter size-full is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"1327\" height=\"90\" src=\"https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/04\/20250424222716850.png\" alt=\"\" class=\"wp-image-3009\" style=\"width:700px\" srcset=\"https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/04\/20250424222716850.png 1327w, https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/04\/20250424222716850-300x20.png 300w, https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/04\/20250424222716850-1024x69.png 1024w, https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/04\/20250424222716850-768x52.png 768w\" sizes=\"auto, (max-width: 1327px) 100vw, 1327px\" \/><\/figure><\/div>\n\n\n<ul class=\"wp-block-list\">\n<li>$\\mathcal{L}_{i t c}$ \uff1a\u539f\u59cb\u5bf9\u6bd4\u635f\u5931\u3002<\/li>\n\n\n\n<li> $KL(q | p)$ \uff1a\u52a8\u91cf\u84b8\u998f\u635f\u5931\uff0c\u5b66\u751f\u6a21\u578b\u9884\u6d4b\u5206\u5e03 $p$ \u4e0e\u6559\u5e08\u6a21\u578b\u4f2a\u6807\u7b7e\u5206\u5e03 $q$ \u7684KL\u6563\u5ea6\uff0c\u8feb\u4f7f\u5b66\u751f\u6a21\u4eff\u6559\u5e08\u7684\u66f4\u5408\u7406\u5224\u65ad\u3002<\/li>\n\n\n\n<li>$\\alpha$ \uff1a\u6743\u91cd\uff08\u9ed8\u8ba40\uff0e4\uff09\uff0c\u5e73\u8861\u4e24\u8005\u8d21\u732e\u3002<\/li>\n<\/ul>\n\n\n\n<p>\uff082\uff09MLM\u4efb\u52a1\u7684\u52a8\u91cf\u635f\u5931\uff08$\\mathbf{M L M}_{\\mathrm{MoD}}$ \uff09<\/p>\n\n\n<div class=\"wp-block-image\">\n<figure class=\"aligncenter size-full is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"882\" height=\"71\" src=\"https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/04\/20250424223315906.png\" alt=\"\" class=\"wp-image-3011\" style=\"width:500px\" srcset=\"https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/04\/20250424223315906.png 882w, https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/04\/20250424223315906-300x24.png 300w, https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/04\/20250424223315906-768x62.png 768w\" sizes=\"auto, (max-width: 882px) 100vw, 882px\" \/><\/figure><\/div>\n\n\n<ul class=\"wp-block-list\">\n<li>$\\mathcal{L}_{mlm}$ \uff1a\u6807\u51c6\u7684\u63a9\u7801\u8bed\u8a00\u5efa\u6a21\u635f\u5931\uff0c\u6743\u91cd\u4e3a(1-\u03b1)\u3002<\/li>\n\n\n\n<li>$KL(q^{msk(I,T\u0302)} | p^{msk(I,T\u0302)})$ \uff1a\u52a8\u91cf\u84b8\u998f\u635f\u5931\uff0c\u540c\u6837\u662fKL\u6563\u5ea6\uff0c\u8fd9\u4e00\u9879\u9f13\u52b1\u5f53\u524d\u6a21\u578b\u5b66\u4e60\u52a8\u91cf\u6a21\u578b\u7684\u9884\u6d4b\u3002<\/li>\n\n\n\n<li>$\\alpha$ \uff1a\u6743\u91cd\uff08\u9ed8\u8ba40\uff0e4\uff09\uff0c\u5e73\u8861\u4e24\u8005\u8d21\u732e\u3002<\/li>\n<\/ul>\n\n\n\n<p>\u4e24\u4e2a\u635f\u5931\u51fd\u6570\u90fd\u91c7\u7528\u4e86\u539f\u59cb\u635f\u5931\u548c\u52a8\u91cf\u84b8\u998f\u635f\u5931\u7684\u52a0\u6743\u7ec4\u5408\uff0c\u03b1=0.4\u8868\u793a\u66f4\u503e\u5411\u4e8e\u539f\u59cb\u635f\u5931(60%)\u4f46\u52a8\u91cf\u84b8\u998f\u4e5f\u6709\u663e\u8457\u8d21\u732e(40%)\u3002\u52a8\u91cf\u6a21\u578b\u901a\u8fc7EMA\u66f4\u65b0\uff0c\u6bd4\u5f53\u524d\u6a21\u578b\u66f4\u7a33\u5b9a\uff0c\u80fd\u63d0\u4f9b\u66f4\u53ef\u9760\u7684\u76d1\u7763\u4fe1\u53f7\uff0c\u800c\u4e14\u4e0d\u540c\u4e8e\u4f20\u7edf\u4f7f\u7528\u786c\u6807\u7b7e(one-hot)\u7684\u65b9\u6cd5\uff0c\u52a8\u91cf\u6a21\u578b\u63d0\u4f9b\u7684\u662f\u6982\u7387\u5206\u5e03(\u8f6f\u76ee\u6807)\uff0c\u80fd\u66f4\u597d\u5730\u5904\u7406\u6570\u636e\u4e2d\u7684\u566a\u58f0\u3002<\/p>\n\n\n\n<p><mark style=\"background-color:rgba(0, 0, 0, 0)\" class=\"has-inline-color has-luminous-vivid-orange-color\">\u503c\u5f97\u6ce8\u610f\u7684\u662f<\/mark>\uff0c\u5728ALBEF\u6a21\u578b\u4e2d\uff0c<strong>ITM\uff08Image-Text Matching\uff09\u4efb\u52a1\u6ca1\u6709\u4f7f\u7528\u52a8\u91cf\u84b8\u998f\uff08MoD\uff09<\/strong>\u3002\u8fd9\u4e3b\u8981\u662f\u7531\u4e8e\u4ee5\u4e0b\u51e0\u4e2a\u539f\u56e0\uff1a<\/p>\n\n\n\n<p>1.&nbsp;<strong>ITM\u7684\u4efb\u52a1\u6027\u8d28\u4e0eITC\/MLM\u4e0d\u540c<\/strong><\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>ITM\u662f\u4e8c\u5206\u7c7b\u4efb\u52a1<\/strong>\uff08\u5224\u65ad\u56fe\u50cf-\u6587\u672c\u5bf9\u662f\u5426\u5339\u914d\uff09\uff0c\u800cITC\u548cMLM\u662f<strong>\u591a\u5206\u7c7b\/\u5bf9\u6bd4\u5b66\u4e60\u4efb\u52a1<\/strong>\u3002<\/li>\n\n\n\n<li>ITM\u7684\u6807\u7b7e\u662f<strong>\u786c\u6807\u7b7e\uff080\/1\uff09<\/strong>\uff0c\u800cITC\u548cMLM\u7684\u4f2a\u6807\u7b7e\u53ef\u4ee5\u662f<strong>\u8f6f\u6807\u7b7e\uff08\u6982\u7387\u5206\u5e03\uff09<\/strong>\uff0c\u66f4\u9002\u5408\u7528\u52a8\u91cf\u6a21\u578b\u751f\u6210\u66f4\u5e73\u6ed1\u7684\u76d1\u7763\u4fe1\u53f7\u3002<\/li>\n\n\n\n<li>ITM\u7684\u566a\u58f0\u95ee\u9898\u4e0d\u5982ITC\u548cMLM\u4e25\u91cd\uff08\u56e0\u4e3a\u5339\u914d\/\u4e0d\u5339\u914d\u7684\u6807\u6ce8\u76f8\u5bf9\u660e\u786e\uff09\uff0c\u800cITC\u548cMLM\u9762\u4e34\u66f4\u5927\u7684\u8bed\u4e49\u6a21\u7cca\u6027\uff08\u5982\"\u90e8\u5206\u76f8\u5173\"\u7684\u8d1f\u6837\u672c\u6216MLM\u7684\u66ff\u4ee3\u8bcd\u9009\u62e9\uff09\u3002<\/li>\n<\/ul>\n\n\n\n<p>2.&nbsp;<strong>ITM\u7684\u8d1f\u6837\u672c\u91c7\u6837\u4f9d\u8d56ITC\u7684\u56f0\u96be\u6837\u672c<\/strong><\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>ALBEF\u7684ITM\u4efb\u52a1\u4f7f\u7528<strong>ITC\u8ba1\u7b97\u7684\u76f8\u4f3c\u5ea6<\/strong>\u6765\u7b5b\u9009<strong>\u56f0\u96be\u8d1f\u6837\u672c\uff08hard negatives\uff09<\/strong>\uff0c\u800c\u4e0d\u662f\u968f\u673a\u8d1f\u6837\u672c\u3002<\/li>\n\n\n\n<li>\u5982\u679c\u5bf9ITM\u4e5f\u5e94\u7528\u52a8\u91cf\u84b8\u998f\uff0c\u53ef\u80fd\u4f1a\u5bfc\u81f4\uff1a\n<ul class=\"wp-block-list\">\n<li>\u52a8\u91cf\u6a21\u578b\u7684ITC\u5206\u6570\u5f71\u54cdITM\u7684\u6837\u672c\u9009\u62e9\uff0c\u5f15\u5165\u989d\u5916\u7684\u590d\u6742\u6027\u3002<\/li>\n\n\n\n<li>\u7834\u574f\u539f\u59cbITM\u4efb\u52a1\u7684\u5bf9\u6297\u6027\u5b66\u4e60\u76ee\u6807\uff08\u533a\u5206\u771f\u6b63\u7684\u4e0d\u5339\u914d\u5bf9\u548c\"\u4f3c\u662f\u800c\u975e\"\u7684\u8d1f\u6837\u672c\uff09\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n\n\n\n<p>3.&nbsp;<strong>\u8bad\u7ec3\u6548\u7387\u4e0e\u7a33\u5b9a\u6027\u7684\u6743\u8861<\/strong><\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>ITM\u9700\u8981\u8ba1\u7b97\u56fe\u50cf-\u6587\u672c\u5bf9\u7684\u5168\u5c40\u4ea4\u4e92\uff08\u901a\u8fc7\u8de8\u6a21\u6001\u7f16\u7801\u5668\uff09\uff0c\u8ba1\u7b97\u6210\u672c\u8f83\u9ad8\u3002\u5982\u679c\u52a0\u5165\u52a8\u91cf\u84b8\u998f\uff0c\u9700\u8981\u989d\u5916\u7ef4\u62a4\u4e00\u4e2a\u52a8\u91cf\u8de8\u6a21\u6001\u7f16\u7801\u5668\uff0c\u4f1a\u589e\u52a0\u663e\u5b58\u548c\u8ba1\u7b97\u5f00\u9500\u3002<\/li>\n\n\n\n<li>ITC\u548cMLM\u4ec5\u9700\u5355\u6a21\u6001\u6216\u8f7b\u91cf\u7ea7\u8de8\u6a21\u6001\u8ba1\u7b97\uff0c\u66f4\u9002\u5408\u7528\u52a8\u91cf\u6a21\u578b\u751f\u6210\u8f6f\u76ee\u6807\u3002<\/li>\n<\/ul>\n\n\n\n<h2 class=\"wp-block-heading\">3. Downstream V+L Tasks<\/h2>\n\n\n\n<p>ALBEF \u5728\u9884\u8bad\u7ec3\u540e\u901a\u8fc7\u5fae\u8c03\uff08Fine-tuning\uff09\u53ef\u4ee5\u9002\u914d\u4e0b\u6e38\u4efb\u52a1\uff0c\u5982\u4e0b\u56fe\u4e2d\u7684&nbsp;<strong>\u89c6\u89c9\u95ee\u7b54\uff08VQA\uff09<\/strong>&nbsp;\u548c&nbsp;<strong>\u81ea\u7136\u8bed\u8a00\u89c6\u89c9\u63a8\u7406\uff08NLVR\u00b2\uff09<\/strong>\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"1198\" height=\"345\" src=\"https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/04\/20250424231904599.png\" alt=\"\" class=\"wp-image-3026\" srcset=\"https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/04\/20250424231904599.png 1198w, https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/04\/20250424231904599-300x86.png 300w, https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/04\/20250424231904599-1024x295.png 1024w, https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/04\/20250424231904599-768x221.png 768w\" sizes=\"auto, (max-width: 1198px) 100vw, 1198px\" \/><\/figure>\n\n\n\n<p>VQA\u89c6\u89c9\u95ee\u7b54\u4efb\u52a1\u76ee\u6807\uff1a\u6a21\u578b\u9700\u6839\u636e\u56fe\u50cf\u548c\u95ee\u9898\u751f\u6210\u7b54\u6848\uff08\u5f00\u653e\u5f0f\u751f\u6210\u4efb\u52a1\uff0c\u800c\u975e\u4f20\u7edf\u7684\u591a\u7b54\u6848\u5206\u7c7b\u4efb\u52a1\uff09\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u8f93\u5165\u5904\u7406<\/strong>\uff1a\n<ul class=\"wp-block-list\">\n<li>\u56fe\u50cf\u901a\u8fc7&nbsp;<strong>Image Encoder<\/strong>&nbsp;\u7f16\u7801\u4e3a\u89c6\u89c9\u7279\u5f81\u3002<\/li>\n\n\n\n<li>\u95ee\u9898\u6587\u672c\u901a\u8fc7&nbsp;<strong>Text Encoder<\/strong>&nbsp;\u7f16\u7801\u4e3a\u6587\u672c\u7279\u5f81\u3002<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li><strong>\u591a\u6a21\u6001\u878d\u5408<\/strong>\uff1a\n<ul class=\"wp-block-list\">\n<li>\u89c6\u89c9\u548c\u6587\u672c\u7279\u5f81\u8f93\u5165&nbsp;<strong>6\u5c42 Multimodal Encoder<\/strong>\uff08\u5171\u4eab\u9884\u8bad\u7ec3\u53c2\u6570\uff09\uff0c\u901a\u8fc7\u8de8\u6ce8\u610f\u529b\u673a\u5236\u4ea4\u4e92\u3002<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li><strong>\u7b54\u6848\u751f\u6210<\/strong>\uff1a\n<ul class=\"wp-block-list\">\n<li>\u4f7f\u7528&nbsp;<strong>6\u5c42 Transformer Decoder<\/strong>\uff08\u5982\u56fe&nbsp;<code>Answer Decoder<\/code>\uff09\u4ee5\u81ea\u56de\u5f52\u65b9\u5f0f\u751f\u6210\u7b54\u6848\uff1a\n<ul class=\"wp-block-list\">\n<li>\u521d\u59cb\u8f93\u5165\u4e3a&nbsp;<code>[CLS]<\/code>&nbsp;\u6807\u8bb0\uff0c\u7ed3\u675f\u6807\u8bb0\u4e3a&nbsp;<code>[SEP]<\/code>\u3002<\/li>\n\n\n\n<li>Decoder \u901a\u8fc7\u8de8\u6ce8\u610f\u529b\u63a5\u6536\u591a\u6a21\u6001\u7f16\u7801\u5668\u7684\u8f93\u51fa\uff08\u56fe\u4e2d&nbsp;<code>cross-attention input<\/code>\uff09\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n\n\n\n<p>NLVR\u81ea\u7136\u8bed\u8a00\u89c6\u89c9\u63a8\u7406\u76ee\u6807\uff1a\u5224\u65ad\u6587\u672c\u9648\u8ff0\u662f\u5426\u4e0e\u4e00\u5bf9\u56fe\u50cf\u5339\u914d<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u8f93\u5165\u5904\u7406<\/strong>\uff1a\n<ul class=\"wp-block-list\">\n<li>\u4e24\u5f20\u56fe\u50cf\uff08Image #1 \u548c Image #2\uff09\u5206\u522b\u901a\u8fc7&nbsp;<strong>Image Encoder<\/strong>&nbsp;\u7f16\u7801\u3002<\/li>\n\n\n\n<li>\u6587\u672c\u901a\u8fc7&nbsp;<strong>Text Encoder<\/strong>&nbsp;\u7f16\u7801\u3002<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li><strong>\u591a\u6a21\u6001\u7f16\u7801\u5668\u6269\u5c55<\/strong>\uff1a\n<ul class=\"wp-block-list\">\n<li>\u6bcf\u5c42 Multimodal Encoder \u88ab\u590d\u5236\u4e3a&nbsp;<strong>\u4e24\u4e2a\u8fde\u7eed\u7684 Transformer Block<\/strong>\uff08\u5982\u56fe&nbsp;<code>Multimodal Block \u2192 x6<\/code>\uff09\uff1a\n<ul class=\"wp-block-list\">\n<li>\u6bcf\u4e2a Block \u5305\u542b\u81ea\u6ce8\u610f\u529b\u5c42\u3001\u8de8\u6ce8\u610f\u529b\u5c42\uff08<code>share cross-attention layer<\/code>\uff09\u548c\u524d\u9988\u5c42\u3002<\/li>\n\n\n\n<li>\u4e24\u4e2a Block \u5171\u4eab\u9884\u8bad\u7ec3\u6743\u91cd\uff0c\u4e14\u8de8\u6ce8\u610f\u529b\u7684 Key\/Value \u6295\u5f71\u6743\u91cd\u4e00\u81f4\u3002<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li>\u4e24\u4e2a Block \u5206\u522b\u5904\u7406\u4e24\u5e45\u56fe\u50cf\u7684\u5d4c\u5165\u7279\u5f81\u3002<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li><strong>\u5206\u7c7b\u5668<\/strong>\uff1a\n<ul class=\"wp-block-list\">\n<li>\u5728\u591a\u6a21\u6001\u7f16\u7801\u5668\u7684&nbsp;<code>[CLS]<\/code>&nbsp;\u8868\u793a\u4e0a\u63a5&nbsp;<strong>MLP<\/strong>&nbsp;\u8fdb\u884c\u4e8c\u5206\u7c7b\u9884\u6d4b\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n\n\n\n<p>\u9664\u6b64\u4e4b\u5916\uff0c\u5e38\u89c1\u7684V+L\u4efb\u52a1\u8fd8\u6709\u6839\u636e\u56fe\u7247\u68c0\u7d22\u6587\u672c\uff08TR\uff09\uff0c\u6839\u636e\u6587\u672c\u68c0\u7d22\u56fe\u7247\uff08IR\uff09\uff0c\u4ee5\u53caSNLI-VE\uff08<strong>Visual Entailment<\/strong>\uff09\u3002SNLI-VE\u89c6\u89c9\u8574\u542b\u7684\u4efb\u52a1\u5b9a\u4e49\u4e3a\uff1a<\/p>\n\n\n\n<p>\u7ed9\u5b9a\u4e00\u4e2a\u56fe\u50cf\uff08Image\uff09\u548c\u4e00\u6bb5\u6587\u672c\uff08Text\uff09\uff0c\u6a21\u578b\u9700\u8981\u5224\u65ad\u6587\u672c\u4e0e\u56fe\u50cf\u4e4b\u95f4\u7684\u5173\u7cfb\uff0c\u5206\u4e3a\u4e09\u7c7b\uff1a<\/p>\n\n\n\n<p><strong>\u4e2d\u7acb\uff08Neutral\uff09<\/strong>\uff1a\u56fe\u50cf\u4e0e\u6587\u672c\u65e0\u5173\u6216\u65e0\u6cd5\u786e\u5b9a\u5173\u7cfb\u3002<\/p>\n\n\n\n<p><strong>\u8574\u542b\uff08Entailment\uff09<\/strong>\uff1a\u56fe\u50cf\u4e2d\u7684\u89c6\u89c9\u5185\u5bb9<strong>\u652f\u6301<\/strong>\u6587\u672c\u63cf\u8ff0\u3002<\/p>\n\n\n\n<p><strong>\u77db\u76fe\uff08Contradiction\uff09<\/strong>\uff1a\u56fe\u50cf\u4e2d\u7684\u89c6\u89c9\u5185\u5bb9<strong>\u5426\u5b9a<\/strong>\u6587\u672c\u63cf\u8ff0\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">4. \u5b9e\u9a8c<\/h2>\n\n\n\n<h3 class=\"wp-block-heading\">4.1 \u635f\u5931\u51fd\u6570\u7684\u6d88\u878d\u5b9e\u9a8c<\/h3>\n\n\n\n<p>\u8868\u683c1\u5c55\u793a\u4e86\u4ece\u57fa\u7840\u4efb\u52a1\uff08MLM+ITM\uff09\u9010\u6b65\u589e\u52a0\u65b0\u635f\u5931\u7684\u6027\u80fd\u63d0\u5347\uff0c\u6838\u5fc3\u7ed3\u8bba\u5982\u4e0b\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"1150\" height=\"447\" src=\"https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/04\/20250424233921161.png\" alt=\"\" class=\"wp-image-3027\" srcset=\"https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/04\/20250424233921161.png 1150w, https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/04\/20250424233921161-300x117.png 300w, https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/04\/20250424233921161-1024x398.png 1024w, https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/04\/20250424233921161-768x299.png 768w\" sizes=\"auto, (max-width: 1150px) 100vw, 1150px\" \/><\/figure>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>ITC \u7684\u8d21\u732e\u6700\u5927<\/strong>\uff1a\u5728\u6240\u6709\u4efb\u52a1\u4e2d\u5e73\u5747\u63d0\u5347\u7ea6&nbsp;<strong>3%<\/strong>\uff0c\u9a8c\u8bc1\u4e86\u8de8\u6a21\u6001\u5bf9\u6bd4\u5b66\u4e60\u5bf9\u5bf9\u9f50\u56fe\u50cf-\u6587\u672c\u5168\u5c40\u7279\u5f81\u7684\u5173\u952e\u4f5c\u7528\u3002<\/li>\n\n\n\n<li><strong>\u96be\u8d1f\u6837\u672c\u6316\u6398\uff08ITM-<em>ha<\/em><em>r<\/em><em>d<\/em>\u200b\uff09<\/strong>\uff1a\u5bf9\u590d\u6742\u4efb\u52a1\uff08\u5982 NLVR\uff09\u63d0\u5347\u66f4\u660e\u663e\uff0c\u8bf4\u660e\u96be\u6837\u672c\u6709\u52a9\u4e8e\u6a21\u578b\u533a\u5206\u7ec6\u5fae\u8bed\u4e49\u5dee\u5f02\u3002<\/li>\n\n\n\n<li><strong>\u52a8\u91cf\u84b8\u998f\uff08MoD\uff09<\/strong>\uff1a\u5c0f\u5e45\u4f46\u7a33\u5b9a\u63d0\u5347\uff0c\u5c24\u5176\u5728 VQA \u4e0a\uff08+0.36\uff09\uff0c\u8868\u660e\u4f2a\u6807\u7b7e\u5e73\u6ed1\u80fd\u7f13\u89e3\u566a\u58f0\u6807\u6ce8\u7684\u5f71\u54cd\u3002<\/li>\n<\/ul>\n\n\n\n<p><strong>&nbsp;\u6570\u636e\u89c4\u6a21\u7684\u5f71\u54cd<\/strong><\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>4M \u2192 14M \u6570\u636e<\/strong>\uff1a\n<ul class=\"wp-block-list\">\n<li>NLVR\u00b2 \u63d0\u5347\u6700\u663e\u8457\uff08+2.64\uff09\uff0c\u56e0\u53cc\u56fe\u50cf\u63a8\u7406\u4efb\u52a1\u9700\u8981\u66f4\u591a\u6837\u672c\u5b66\u4e60\u590d\u6742\u5bf9\u9f50\uff1b<\/li>\n\n\n\n<li>\u68c0\u7d22\u4efb\u52a1\uff08TR\/IR\uff09\u63a5\u8fd1\u9971\u548c\uff0c\u4f46\u4ecd\u6709\u7ea6&nbsp;<strong>1.5%<\/strong>&nbsp;\u63d0\u5347\uff0c\u8bf4\u660e\u6570\u636e\u89c4\u6a21\u5bf9\u57fa\u7840\u5bf9\u9f50\u4efb\u52a1\u4ecd\u6709\u8fb9\u9645\u6548\u76ca\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n\n\n\n<h3 class=\"wp-block-heading\">4.2 \u9884\u8bad\u7ec3\u6548\u679c<\/h3>\n\n\n\n<p>\u6839\u636e\u8868\u683c2\u6570\u636e\uff0c\u9884\u8bad\u7ec3\u540e\u7684ALBEF \u5728&nbsp;<strong>Flickr30K<\/strong>&nbsp;\u548c&nbsp;<strong>MSCOCO<\/strong>&nbsp;\u6570\u636e\u96c6\u4e0a\u7684\u56fe\u50cf-\u6587\u672c\u68c0\u7d22\u4efb\u52a1\u4e2d\u8868\u73b0\u51fa\u663e\u8457\u4f18\u52bf\uff0c\u5c24\u5176\u5728\u6570\u636e\u6548\u7387\u548c\u6027\u80fd\u4e0a\u8d85\u8d8a\u73b0\u6709\u65b9\u6cd5\u3002\u4ee5\u4e0b\u662f\u8be6\u7ec6\u7ed3\u679c\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"1148\" height=\"329\" src=\"https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/04\/20250424234626380.png\" alt=\"\" class=\"wp-image-3029\" srcset=\"https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/04\/20250424234626380.png 1148w, https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/04\/20250424234626380-300x86.png 300w, https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/04\/20250424234626380-1024x293.png 1024w, https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/04\/20250424234626380-768x220.png 768w\" sizes=\"auto, (max-width: 1148px) 100vw, 1148px\" \/><\/figure>\n\n\n\n<p>\u88683\u5c55\u793a\u7684\u96f6\u6837\u672c\u7684\u68c0\u6d4b\u80fd\u529b\uff0cALBEF\uff0814M\uff09\u7684\u7cbe\u5ea6\u8981\u8fdc\u8d85CLIP\uff08400M\uff09\uff0c\u4e5f\u5c31\u662f\u8bf4\uff0c\u5176\u7528\u66f4\u5c11\u7684\u6570\u636e\uff0c\u8bad\u7ec3\u51fa\u6765\u4e86\u66f4\u9ad8\u7684\u7cbe\u5ea6\u3002<\/p>\n\n\n<div class=\"wp-block-image\">\n<figure class=\"aligncenter size-full is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"875\" height=\"317\" src=\"https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/04\/20250424235457708.png\" alt=\"\" class=\"wp-image-3030\" style=\"width:700px\" srcset=\"https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/04\/20250424235457708.png 875w, https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/04\/20250424235457708-300x109.png 300w, https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/04\/20250424235457708-768x278.png 768w\" sizes=\"auto, (max-width: 875px) 100vw, 875px\" \/><\/figure><\/div>\n\n\n<p>\u88684\u662f\u5173\u4e8e\u591a\u6a21\u6001\u4e0b\u6e38\u4efb\u52a1\u7684\u5b9e\u9a8c:<\/p>\n\n\n<div class=\"wp-block-image\">\n<figure class=\"aligncenter size-full is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"942\" height=\"444\" src=\"https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/04\/20250424235836303.png\" alt=\"\" class=\"wp-image-3031\" style=\"width:800px\" srcset=\"https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/04\/20250424235836303.png 942w, https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/04\/20250424235836303-300x141.png 300w, https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/04\/20250424235836303-768x362.png 768w\" sizes=\"auto, (max-width: 942px) 100vw, 942px\" \/><\/figure><\/div>\n\n\n<ul class=\"wp-block-list\">\n<li>VQA\uff08\u89c6\u89c9\u95ee\u7b54\uff09ALBEF (14M)&nbsp;\u4ee5&nbsp;75.84&nbsp;\u7684\u51c6\u786e\u7387\u663e\u8457\u9886\u5148\uff0c\u6bd4\u6b64\u524d\u6700\u4f18\u65b9\u6cd5\uff08VILLA: 73.59\uff09\u9ad8\u51fa&nbsp;2.25%\uff0c\u5373\u4f7f\u5bf9\u6bd4&nbsp;4M&nbsp;\u6570\u636e\u7684 ALBEF\uff0874.54\uff09\u4e5f\u4f18\u4e8e\u5927\u591a\u6570\u65b9\u6cd5\u3002<\/li>\n\n\n\n<li>NLVR\u00b2\uff08\u89c6\u89c9\u63a8\u7406\uff09ALBEF (14M)&nbsp;\u5728\u6d4b\u8bd5\u96c6\u4e0a\u8fbe&nbsp;83.14\uff0c\u8f83\u4e4b\u524d SOTA\uff08VILLA: 79.30\uff09\u63d0\u5347&nbsp;3.84%\uff1a<\/li>\n\n\n\n<li>SNLI-VE\uff08\u89c6\u89c9\u8bed\u4e49\u63a8\u7406\uff09ALBEF (14M)&nbsp;\u8fbe\u5230&nbsp;80.80\uff0c\u8d85\u8d8a VILLA\uff0879.47\uff09\uff0c\u9a8c\u8bc1\u4e86\u5bf9\u7ec6\u7c92\u5ea6\u8bed\u4e49\u5bf9\u9f50\u7684\u4f18\u8d8a\u6027\u3002<\/li>\n<\/ul>\n\n\n\n<h2 class=\"wp-block-heading\">5. \u603b\u7ed3<\/h2>\n\n\n\n<p><strong>ALBEF<\/strong>\u2014\u2014\u4e00\u79cd\u65b0\u7684\u89c6\u89c9-\u8bed\u8a00\u8868\u793a\u5b66\u4e60\u6846\u67b6\u3002ALBEF\u7684\u6838\u5fc3\u521b\u65b0\u5728\u4e8e\u5148\u901a\u8fc7\u5355\u6a21\u6001\u7f16\u7801\u5668\u5bf9\u9f50\u56fe\u50cf\u4e0e\u6587\u672c\u7684\u72ec\u7acb\u8868\u793a\uff08\u5176\u5b9e\u5c31\u662fCLIP\uff09\uff0c\u518d\u901a\u8fc7\u591a\u6a21\u6001\u7f16\u7801\u5668\u8fdb\u884c\u878d\u5408\u3002\u4ece\u7406\u8bba\u548c\u5b9e\u9a8c\u4e0a\u9a8c\u8bc1\u4e86\u6240\u63d0\u51fa\u7684<strong>\u56fe\u50cf-\u6587\u672c\u5bf9\u6bd4\u5b66\u4e60<\/strong>\u548c<strong>\u52a8\u91cf\u84b8\u998f<\/strong>\u65b9\u6cd5\u7684\u6709\u6548\u6027\u3002\u4e0e\u73b0\u6709\u65b9\u6cd5\u76f8\u6bd4\uff0cALBEF\u5728\u591a\u79cd\u89c6\u89c9-\u8bed\u8a00\uff08V+L\uff09\u4e0b\u6e38\u4efb\u52a1\u4e2d\u5b9e\u73b0\u4e86\u66f4\u9ad8\u7684\u6027\u80fd\u4e0e\u66f4\u5feb\u7684\u63a8\u7406\u901f\u5ea6\u3002<\/p>\n\n\n\n<p><mark style=\"background-color:rgba(0, 0, 0, 0)\" class=\"has-inline-color has-luminous-vivid-orange-color\">\u503c\u5f97\u4e00\u63d0\u7684\u662f<\/mark>, \u7531\u4e8e\u5176\u9ad8\u6548\u7684\u6a21\u578b\u8bbe\u8ba1\uff0c\u548c\u8054\u5408\u635f\u5931\u7684\u4f7f\u7528\u3002ALBEF\u7684\u8bad\u7ec3\u6210\u672c\u76f8\u6bd4CLIP\u5c0f\u5c0f\u5f97\u591a\uff0c\u53ea\u7528\u4e86 8 NVIDIA A100 GPUs\uff0cCLIP\u662f64\u5f20\u3002\u8fd9\u4e2a\u8bad\u7ec3\u8d44\u6e90\u5728LLM\u9886\u57df\u4ee5\u53ca\u662f\u975e\u5e38\u53cb\u597d\u4e86\uff01<\/p>\n","protected":false},"excerpt":{"rendered":"<p>1. \u6458\u8981 1.1 \u95ee\u9898 \u73b0\u6709\u7684\u591a\u6a21\u6001\uff08\u56fe\u50cf+\u6587\u672c\uff09\u6a21\u578b\u901a\u5e38\u7528\u4e00\u4e2aTransformer\u7f16\u7801\u5668\u540c\u65f6\u5904\u7406\u56fe\u50cf\u7279\u5f81 [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":2991,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[29,18],"tags":[],"class_list":["post-2989","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-29","category-18"],"_links":{"self":[{"href":"http:\/\/gnn.club\/index.php?rest_route=\/wp\/v2\/posts\/2989","targetHints":{"allow":["GET"]}}],"collection":[{"href":"http:\/\/gnn.club\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"http:\/\/gnn.club\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"http:\/\/gnn.club\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"http:\/\/gnn.club\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=2989"}],"version-history":[{"count":35,"href":"http:\/\/gnn.club\/index.php?rest_route=\/wp\/v2\/posts\/2989\/revisions"}],"predecessor-version":[{"id":3037,"href":"http:\/\/gnn.club\/index.php?rest_route=\/wp\/v2\/posts\/2989\/revisions\/3037"}],"wp:featuredmedia":[{"embeddable":true,"href":"http:\/\/gnn.club\/index.php?rest_route=\/wp\/v2\/media\/2991"}],"wp:attachment":[{"href":"http:\/\/gnn.club\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=2989"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/gnn.club\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=2989"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/gnn.club\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=2989"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}