{"id":3233,"date":"2025-07-16T17:29:46","date_gmt":"2025-07-16T09:29:46","guid":{"rendered":"https:\/\/www.gnn.club\/?p=3233"},"modified":"2025-07-16T17:52:31","modified_gmt":"2025-07-16T09:52:31","slug":"pyramid-vision-transformer","status":"publish","type":"post","link":"http:\/\/www.gnn.club\/?p=3233","title":{"rendered":"Pyramid Vision Transformer"},"content":{"rendered":"<h1>\u57fa\u672c\u4fe1\u606f<\/h1>\n<ul>\n<li>\n<p>\ud83d\udcf0\u6807\u9898: Pyramid Vision Transformer: A Versatile Backbone for Dense Prediction  without Convolutions<\/p>\n<\/li>\n<li>\n<p>\ud83d\udd8b\ufe0f\u4f5c\u8005: Wenhai Wang<\/p>\n<\/li>\n<li>\n<p>\ud83c\udfdb\ufe0f\u673a\u6784: Nanjing University\uff08\u5357\u4eac\u5927\u5b66\uff09<\/p>\n<\/li>\n<li>\n<p>\ud83d\udd17\u94fe\u63a5: <a href=\"https:\/\/github.com\/whai362\/PVT\">GitHub<\/a><\/p>\n<\/li>\n<li>\n<p>\ud83d\udd25\u5173\u952e\u8bcd: Pyramid Vision Transformer, Dense Prediction, Backbone, Convolution-free<\/p>\n<\/li>\n<\/ul>\n<h2>\u6458\u8981\u6982\u8ff0<\/h2>\n<table border=\"1\" cellspacing=\"0\" cellpadding=\"5\">\n<thead>\n<tr>\n<th>\u9879\u76ee<\/th>\n<th>\u5185\u5bb9<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>\ud83d\udcd6\u7814\u7a76\u80cc\u666f<\/td>\n<td>CNN\u9aa8\u5e72\u7f51\u7edc\u666e\u904d\u91c7\u7528\u91d1\u5b57\u5854\u7ed3\u6784\u5904\u7406\u5bc6\u96c6\u9884\u6d4b\u4efb\u52a1\uff0c\u800cViT\u4ec5\u9488\u5bf9\u56fe\u50cf\u5206\u7c7b\u8bbe\u8ba1<\/td>\n<\/tr>\n<tr>\n<td>\ud83c\udfaf\u7814\u7a76\u76ee\u7684<\/td>\n<td>\u5c06\u91d1\u5b57\u5854\u7ed3\u6784\u5f15\u5165ViT\uff0c\u6784\u5efa\u9002\u7528\u4e8e\u591a\u4efb\u52a1\u7684\u901a\u7528\u5377\u79ef-free\u9aa8\u5e72\u7f51\u7edc<\/td>\n<\/tr>\n<tr>\n<td>\u270d\ufe0f\u7814\u7a76\u65b9\u6cd5<\/td>\n<td>\u63d0\u51faPyramid Vision Transformer (PVT)\uff0c\u878d\u5408CNN\u91d1\u5b57\u5854\u7ed3\u6784\u4e0eTransformer\u7f16\u7801\u5668<\/td>\n<\/tr>\n<tr>\n<td>\ud83d\udd4a\ufe0f\u7814\u7a76\u5bf9\u8c61<\/td>\n<td>\u5bc6\u96c6\u9884\u6d4b\u4efb\u52a1\uff08\u5982\u76ee\u6807\u68c0\u6d4b\u3001\u5b9e\u4f8b\/\u8bed\u4e49\u5206\u5272\uff09<\/td>\n<\/tr>\n<tr>\n<td>\ud83d\udd0d\u7814\u7a76\u7ed3\u8bba<\/td>\n<td>PVT\u663e\u8457\u6269\u5c55ViT\u5e94\u7528\u8303\u56f4\uff0c\u53ef\u4e0eDETR\u7ed3\u5408\u6784\u5efa\u7aef\u5230\u7aef\u65e0\u5377\u79ef\u68c0\u6d4b\u7cfb\u7edf<\/td>\n<\/tr>\n<tr>\n<td>\u2b50\u521b\u65b0\u70b9<\/td>\n<td>1) \u9996\u4e2a\u91d1\u5b57\u5854\u5f0fTransformer\u9aa8\u5e72\u7f51\u7edc 2) \u517c\u5bb9\u5bc6\u96c6\u9884\u6d4b\u4efb\u52a1 3) \u5b8c\u5168\u65e0\u5377\u79ef\u8bbe\u8ba1<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<h1>\u80cc\u666f<\/h1>\n<ul>\n<li>\n<p><strong>\u7814\u7a76\u80cc\u666f<\/strong>\uff1a<br \/>\nCNN\u5728\u8ba1\u7b97\u673a\u89c6\u89c9\u9886\u57df\u5360\u636e\u4e3b\u5bfc\u5730\u4f4d\uff0c\u4f46\u5b58\u5728\u5c40\u90e8\u611f\u53d7\u91ce\u7684\u56fa\u6709\u5c40\u9650\u3002ViT\u867d\u5728\u56fe\u50cf\u5206\u7c7b\u4e2d\u8868\u73b0\u4f18\u5f02\uff0c\u4f46\u5176\u5355\u5c3a\u5ea6\u3001\u4f4e\u5206\u8fa8\u7387\u8f93\u51fa\u53ca\u9ad8\u8ba1\u7b97\u6210\u672c\u96be\u4ee5\u9002\u914d\u5bc6\u96c6\u9884\u6d4b\u4efb\u52a1\uff08\u5982\u76ee\u6807\u68c0\u6d4b\u3001\u5206\u5272\uff09\u3002<\/p>\n<\/li>\n<li>\n<p><strong>\u8fc7\u53bb\u65b9\u6848<\/strong>\uff1a<\/p>\n<\/li>\n<\/ul>\n<ol>\n<li>\n<p><strong>CNN\u9aa8\u5e72\u7f51\u7edc<\/strong>\uff08\u5982ResNet\u3001ResNeXt\uff09\uff1a\u4f9d\u8d56\u5c40\u90e8\u611f\u53d7\u91ce\uff0c\u9700\u624b\u52a8\u8bbe\u8ba1\u951a\u6846\/NMS\u7b49\u7ec4\u4ef6\uff1b<\/p>\n<\/li>\n<li>\n<p><strong>Transformer\u6539\u8fdb\u65b9\u6848<\/strong>\uff1a<\/p>\n<ul>\n<li>\n<p><strong>\u6df7\u5408\u67b6\u6784<\/strong>\uff08\u5982ViT+CNN\uff09\uff1a\u4ecd\u4f9d\u8d56\u5377\u79ef\u64cd\u4f5c\uff0c\u975e\u7eafTransformer\u8bbe\u8ba1\uff1b<\/p>\n<\/li>\n<li>\n<p><strong>\u539f\u751fViT<\/strong>\uff1a\u56e0\u7c97\u7c92\u5ea6\u56fe\u50cf\u5206\u5757\uff0832\u00d732\u50cf\u7d20\uff09\u5bfc\u81f4\u8f93\u51fa\u5206\u8fa8\u7387\u4f4e\uff0c\u65e0\u6cd5\u652f\u6301\u50cf\u7d20\u7ea7\u9884\u6d4b\u3002\\<br \/>\n\u6838\u5fc3\u95ee\u9898\uff1a\u7f3a\u4e4f\u517c\u5177\u91d1\u5b57\u5854\u7ed3\u6784\u3001\u5168\u5c40\u611f\u53d7\u91ce\u4e0e\u8ba1\u7b97\u6548\u7387\u7684\u7eafTransformer\u9aa8\u5e72\u7f51\u7edc\u3002<\/p>\n<\/li>\n<\/ul>\n<\/li>\n<\/ol>\n<ul>\n<li><strong>\u7814\u7a76\u52a8\u673a<\/strong>\uff1a<br \/>\n\u6784\u5efa\u9996\u4e2a\u5b8c\u5168\u65e0\u5377\u79ef\u7684\u91d1\u5b57\u5854\u5f0fTransformer\u9aa8\u5e72\u7f51\u7edc\uff08PVT\uff09\uff0c\u901a\u8fc7\u7ec6\u7c92\u5ea6\u5206\u5757\uff084\u00d74\u50cf\u7d20\uff09\u3001\u6e10\u8fdb\u5f0f\u91d1\u5b57\u5854\u7f29\u51cf\u53ca\u7a7a\u95f4\u7f29\u51cf\u6ce8\u610f\u529b\uff08SRA\uff09\uff0c\u89e3\u51b3ViT\u5728\u5bc6\u96c6\u9884\u6d4b\u4efb\u52a1\u4e2d\u7684\u591a\u5c3a\u5ea6\u7279\u5f81\u7f3a\u5931\u4e0e\u8ba1\u7b97\u74f6\u9888\uff0c\u5b9e\u73b0\u7aef\u5230\u7aef\u65e0\u624b\u5de5\u7ec4\u4ef6\u7684\u68c0\u6d4b\u7cfb\u7edf\uff08\u5982PVT+DETR\uff09\u3002<\/li>\n<\/ul>\n<h1>\u65b9\u6cd5<\/h1>\n<p><img decoding=\"async\" src=\"https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/07\/20250716171341240.png\" alt=\"file\" \/><br \/>\n\u56fe1\u901a\u8fc7\u4e09\u680f\u5bf9\u6bd4\u4e86 CNN\uff08ResNet\uff09\u3001 ViT \u548c PVT \u7684\u7ed3\u6784\u5dee\u5f02\uff1a<\/p>\n<p>1\uff0eCNN\uff08\u5de6\u680f\uff09\uff1a<\/p>\n<ul>\n<li>\u5c55\u793aResNet\u76844\u9636\u6bb5\u91d1\u5b57\u5854\u7ed3\u6784\uff0c\u6bcf\u4e2a\u9636\u6bb5\u901a\u8fc7\u5377\u79ef stride $=2$ \u964d\u91c7\u6837\u3002<\/li>\n<li>\u7279\u5f81\u56fe\u5c3a\u5bf8\u4ece $\\frac{H}{4} \\times \\frac{W}{4}$ \u9010\u6b65\u964d\u81f3 $\\frac{H}{32} \\times \\frac{W}{32}$ \u3002<\/li>\n<\/ul>\n<p>2\uff0eViT\uff08\u4e2d\u680f\uff09\uff1a<\/p>\n<ul>\n<li>\u5355\u4e00\u5c3a\u5ea6\u5904\u7406\uff1a\u8f93\u5165\u56fe\u50cf\u88ab\u5206\u4e3a $16 \\times 16$ \u7684\u5757\uff0c\u901a\u8fc7Transformer\u7f16\u7801\u5668\u540e\u8f93\u51fa\u56fa\u5b9a\u5206\u8fa8\u7387\u7684\u5e8f\u5217\u3002<\/li>\n<li>\u7f3a\u4e4f\u591a\u5c3a\u5ea6\u7279\u5f81\uff0c\u9700\u989d\u5916\u4e0a\u91c7\u6837\u624d\u80fd\u7528\u4e8e\u5bc6\u96c6\u9884\u6d4b\u4efb\u52a1\u3002<\/li>\n<\/ul>\n<p>3\uff0ePVT\uff08\u53f3\u680f\uff09\uff1a<\/p>\n<ul>\n<li>\u7c7b\u4f3cCNN\u76844\u9636\u6bb5\u91d1\u5b57\u5854\u7ed3\u6784\uff0c\u4f46\u7528 Patch Embedding \u548c Transformer Encoder \u8488\u4ee3\u5377\u79ef\u3002<\/li>\n<li>\u6bcf\u4e2a\u9636\u6bb5\u6807\u6ce8\u4e86\u8f93\u51fa\u5206\u8fa8\u7387 $\\left(\\frac{H}{4}\\right.$ \u5230 $\\left.\\frac{H}{32}\\right)$ \u548c\u901a\u9053\u6570 $\\left(C_1\\right.$ \u5230 $\\left.C_4\\right)$ \u3002<\/li>\n<\/ul>\n<p>\u6280\u672f\u610f\u4e49<br \/>\n1\uff0e\u591a\u5c3a\u5ea6\u80fd\u529b\uff1a<br \/>\nPVT\u901a\u8fc7\u5206\u9636\u6bb5\u964d\u91c7\u6837\u4fdd\u7559\u91d1\u5b57\u5854\u7ed3\u6784\uff0c\u89e3\u51b3\u4e86ViT\u7684\u5355\u4e00\u5c3a\u5ea6\u7f3a\u9677\uff0c\u53ef\u76f4\u63a5\u7528\u4e8e\u76ee\u6807\u68c0\u6d4b\uff0f\u5206\u5272\u3002<br \/>\n2\uff0e\u7eafTransformer\u8bbe\u8ba1\uff1a<br \/>\n\u4e0eCNN\u4e0d\u540c\uff0cPVT\u5b8c\u5168\u4f9d\u8d56\u81ea\u6ce8\u610f\u529b\u673a\u5236\uff0c\u907f\u514d\u4e86\u5377\u79ef\u7684\u5c40\u90e8\u6027\u9650\u5236\uff3b1\uff0cSec\uff0e1\uff3d\u3002<br \/>\n3\uff0e\u5373\u63d2\u5373\u7528\u6027\uff1a<br \/>\n\u56fe\u793a\u8868\u660ePVT\u7684\u8f93\u51fa\u683c\u5f0f\u4e0eResNet\u5b8c\u5168\u517c\u5bb9\uff08\u591a\u5c3a\u5ea6\u7279\u5f81\u56fe\uff09\uff0c\u53ef\u76f4\u63a5\u66ff\u6362CNN\u7b49\u9aa8\u5e72\u7f51\u7edc\u3002<\/p>\n<p>\u56fe3\u5c55\u793a\u4e86PVT\u4e2d \u5355\u4e2aTransformer Encoder\u5c42 \u7684\u8be6\u7ec6\u7ed3\u6784\uff0c\u5305\u542b\u4e24\u4e2a\u6838\u5fc3\u6a21\u5757\uff1a<\/p>\n<p>1\uff0eSpatial\uff0dReduction Attention\uff08SRA\uff09\uff1a<\/p>\n<ul>\n<li>Query $(Q)$ \u4fdd\u6301\u539f\u59cb\u5206\u8fa8\u7387\uff0c\u4e0e\u6240\u6709\u4f4d\u7f6e\u4ea4\u4e92\uff0c\u4fdd\u6301\u4e86\u5168\u5c40\u5efa\u6a21\u80fd\u529b\u3002<\/li>\n<li>$\\operatorname{Key}(K)$ \u548c Value\uff08 $V$ \uff09\u901a\u8fc7 \u7a7a\u95f4\u7f29\u51cf\uff08SR\uff09\u6a21\u5757 \u964d\u91c7\u6837\uff08\u6807\u6ce8\u4e86\u7f29\u51cf\u6bd4 $R$ \uff09\u3002<\/li>\n<li>\u6ce8\u610f\u529b\u77e9\u9635\u8ba1\u7b97\u4e3a $\\operatorname{Softmax}\\left(Q K^T \/ \\sqrt{d}\\right) V$ \uff0c\u4f46 $K, V$ \u7684\u5206\u8fa8\u7387\u964d\u4f4e\u81f3 $\\frac{H}{R} \\times \\frac{W}{R}$ \u3002<\/li>\n<\/ul>\n<p>2\uff0e\u524d\u9988\u7f51\u7edc\uff08FFN\uff09\uff1a<br \/>\n\u6807\u51c6\u7684\u4e24\u5c42MLP\uff0c\u4e2d\u95f4\u6269\u5c55\u6bd4\u4e3a4\uff08\u4e0eViT\u4e00\u81f4\uff09\u3002<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/07\/20250716171319599.png\" alt=\"file\" \/><\/p>\n<p><strong>SRA\u7684\u6838\u5fc3\u52a8\u673a<\/strong><br \/>\n\u4f20\u7edfTransformer\u7684\u591a\u5934\u81ea\u6ce8\u610f\u529b\uff08MSA\uff09\u5728\u89c6\u89c9\u4efb\u52a1\u4e2d\u9762\u4e34\u4e24\u5927\u95ee\u9898\uff1a<\/p>\n<ul>\n<li>\n<p>\u8ba1\u7b97\u590d\u6742\u5ea6\u9ad8\uff1aMSA\u5bf9\u5e8f\u5217\u957f\u5ea6 $N=H \\times W$ \u7684\u590d\u6742\u5ea6\u4e3a $O\\left(N^2\\right)$ \uff0c\u9ad8\u5206\u8fa8\u7387\u8f93\u5165\uff0c\u8ba1\u7b97\u8d44\u6e90\u6d88\u8017\u8fc7\u5927\u3002<\/p>\n<\/li>\n<li>\n<p>\u7a7a\u95f4\u4fe1\u606f\u5197\u4f59\uff1a\u76f8\u90bb\u50cf\u7d20\u7684Key\uff0fValue\u9ad8\u5ea6\u76f8\u4f3c\uff0c\u76f4\u63a5\u8ba1\u7b97\u5168\u5c40\u6ce8\u610f\u529b\u5b58\u5728\u5927\u91cf\u91cd\u590d\u8fd0\u7b97\u3002<\/p>\n<\/li>\n<\/ul>\n<p><strong>SRA\u7684\u7b97\u6cd5\u5b9e\u73b0<\/strong><br \/>\nSRA\u901a\u8fc7\u964d\u91c7\u6837Key\uff0fValue\u663e\u8457\u964d\u4f4e\u8ba1\u7b97\u91cf\uff0c\u540c\u65f6\u4fdd\u7559\u5168\u5c40\u611f\u53d7\u91ce\u3002\u7ed9\u5b9a\u8f93\u5165\u7279\u5f81 $X \\in \\mathbb{R}^{H \\times W \\times C}$ \uff0cSRA\u6309\u4ee5\u4e0b\u6b65\u9aa4\u5904\u7406\uff1a<\/p>\n<ol>\n<li>\u751f\u6210 $Q \/ K \/ V$ \uff1a<br \/>\n\u3002 $Q=X W_Q \\in \\mathbb{R}^{H \\times W \\times d_k} \\quad$\uff08\u4fdd\u6301\u539f\u59cb\u5206\u8fa8\u7387\uff09<br \/>\n\uff0d$K, V$ \u5148\u901a\u8fc7 \u7a7a\u95f4\u7f29\u51cf\uff08Spatial Reduction\uff0cSR\uff09\u964d\u91c7\u6837\uff1a<\/li>\n<\/ol>\n<p>$$<br \/>\nK_{\\text {reduced }}=\\operatorname{SR}\\left(X W_K\\right), \\quad V_{\\text {reduced }}=\\operatorname{SR}\\left(X W_V\\right)<br \/>\n$$<\/p>\n<p>2\uff0e\u8ba1\u7b97\u6ce8\u610f\u529b\uff1a $\\operatorname{Attention}(Q, K, V)=\\operatorname{Softmax}\\left(\\frac{Q K_{\\text {reduced }}^T}{\\sqrt{d_k}}\\right) V_{\\text {reduced }}$<\/p>\n<p><strong>\u7a7a\u95f4\u7f29\u51cf\uff08SR\uff09\u7684\u6570\u5b66\u8868\u8fbe<\/strong><\/p>\n<p>SR\u64cd\u4f5c\u5305\u542b\u4e24\u6b65\uff1a<\/p>\n<ul>\n<li>\u5747\u5300\u6c60\u5316\u4e0b\u91c7\u6837\uff1a<br \/>\n\u5c06 $K, V \\in \\mathbb{R}^{H \\times W \\times C}$ \u6309\u7f29\u51cf\u6bd4 $R$ \u5212\u5206\u7f51\u683c\uff0c\u6bcf\u4e2a $R \\times R$ \u533a\u57df\u53d6\u5747\u503c\u6216\u6700\u5927\u503c\uff1a<br \/>\nReshape $(K) \\in \\mathbb{R}^{\\frac{H}{R} \\times R \\times \\frac{W}{R} \\times R \\times C} \\rightarrow$ Pooling $\\rightarrow \\mathbb{R}^{\\frac{H}{R} \\times \\frac{W}{R} \\times C}$<\/li>\n<li>\u7ebf\u6027\u6295\u5f71\uff1a<br \/>\n\u901a\u8fc7\u53ef\u5b66\u4e60\u6743\u91cd $W_S \\in \\mathbb{R}^{C \\times C^{\\prime}}$ \u8c03\u6574\u7ef4\u5ea6\uff1a $\\operatorname{SR}(K)=\\operatorname{Norm(Reshape(K)W_{S})\\text {\uff08Norm\u901a\u5e38\u4e3aLayer}}$<br \/>\nNormalization\uff09<\/li>\n<\/ul>\n<p><strong>\u590d\u6742\u5ea6\u5206\u6790<\/strong><br \/>\n\uff0d\u539f\u59cbMSA\uff1a$O\\left(N^2\\right)=O\\left((H W)^2\\right)$<br \/>\n\uff0dSRA\uff1a$O\\left(\\frac{(H W)^2}{R^2}\\right)$<br \/>\n\u4f8b\u5982\u5f53 $R=8$ \u65f6\uff0c\u8ba1\u7b97\u91cf\u964d\u81f3 $1 \/ 64$ \u3002<br \/>\n\uff08\u8bba\u6587\u4e2dPVT\uff0dSmall\u7684Stage 4\u8bbe\u7f6e $R=8$ \uff09<\/p>\n<h1>\u7ed3\u8bba<\/h1>\n<ul>\n<li>\n<p>\u672c\u7814\u7a76\u63d0\u51fa\u4e86\u9996\u4e2a\u7eafTransformer\u67b6\u6784\u7684\u91d1\u5b57\u5854\u9aa8\u5e72\u7f51\u7edcPVT\uff0c\u901a\u8fc7\u6e10\u8fdb\u5f0f\u6536\u7f29\u91d1\u5b57\u5854\u548c\u7a7a\u95f4\u7f29\u51cf\u6ce8\u610f\u529b\uff08SRA\uff09\u673a\u5236\uff0c\u89e3\u51b3\u4e86ViT\u5728\u5bc6\u96c6\u9884\u6d4b\u4efb\u52a1\u4e2d\u7684\u591a\u5c3a\u5ea6\u7279\u5f81\u7f3a\u5931\u4e0e\u8ba1\u7b97\u74f6\u9888\u95ee\u9898\uff0c\u4e3a\u8ba1\u7b97\u673a\u89c6\u89c9\u9886\u57df\u63d0\u4f9b\u4e86CNN\u4e4b\u5916\u7684\u66ff\u4ee3\u6027\u57fa\u7840\u67b6\u6784\u3002<\/p>\n<\/li>\n<li>\n<p><strong>\u4f18\u70b9<\/strong>\uff1a<\/p>\n<ol>\n<li>\n<p>\u5b8c\u5168\u65e0\u5377\u79ef\u8bbe\u8ba1\uff0c\u5b9e\u73b0\u7aef\u5230\u7aef\u5168\u5c40\u5efa\u6a21\uff1b<\/p>\n<\/li>\n<li>\n<p>\u5728\u53c2\u6570\u91cf\u53ef\u6bd4\u6761\u4ef6\u4e0b\u6027\u80fd\u8d85\u8d8a\u4e3b\u6d41CNN\u9aa8\u5e72\uff08\u5982ResNet\uff09\uff1b<\/p>\n<\/li>\n<li>\n<p>SRA\u673a\u5236\u663e\u8457\u964d\u4f4e\u8ba1\u7b97\u590d\u6742\u5ea6\uff08\u964d\u81f3O(N\u00b2\/R\u00b2)\uff09\u3002<\/p>\n<\/li>\n<\/ol>\n<\/li>\n<li>\n<p><strong>\u4e3b\u8981\u7ed3\u8bba<\/strong>\uff1a<\/p>\n<ol>\n<li>PVT\u901a\u8fc7\u6e10\u8fdb\u5f0f\u91d1\u5b57\u5854\u548cSRA\u5c42\uff0c\u53ef\u5728\u6709\u9650\u8ba1\u7b97\u8d44\u6e90\u4e0b\u751f\u6210\u9ad8\u5206\u8fa8\u7387\u591a\u5c3a\u5ea6\u7279\u5f81\uff1b<\/li>\n<li>\u5728\u76ee\u6807\u68c0\u6d4b\u548c\u8bed\u4e49\u5206\u5272\u4efb\u52a1\u4e2d\uff0cPVT\u6027\u80fd\u4f18\u4e8e\u540c\u53c2\u6570\u91cfCNN\u9aa8\u5e72\uff1b<\/li>\n<li>\u4f5c\u4e3aTransformer\u5728CV\u9886\u57df\u7684\u65e9\u671f\u63a2\u7d22\uff0cPVT\u4e3aOCR\u30013D\u89c6\u89c9\u548c\u533b\u5b66\u56fe\u50cf\u7b49\u65b9\u5411\u63d0\u4f9b\u4e86\u53ef\u6269\u5c55\u7684\u57fa\u7840\u6846\u67b6\u3002<\/li>\n<\/ol>\n<\/li>\n<\/ul>\n","protected":false},"excerpt":{"rendered":"<p>\u57fa\u672c\u4fe1\u606f \ud83d\udcf0\u6807\u9898: Pyramid Vision Transformer: A Versatile Backb [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":3235,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[30,18],"tags":[31],"class_list":["post-3233","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-30","category-18","tag-transformer"],"_links":{"self":[{"href":"http:\/\/www.gnn.club\/index.php?rest_route=\/wp\/v2\/posts\/3233","targetHints":{"allow":["GET"]}}],"collection":[{"href":"http:\/\/www.gnn.club\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"http:\/\/www.gnn.club\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"http:\/\/www.gnn.club\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"http:\/\/www.gnn.club\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=3233"}],"version-history":[{"count":3,"href":"http:\/\/www.gnn.club\/index.php?rest_route=\/wp\/v2\/posts\/3233\/revisions"}],"predecessor-version":[{"id":3238,"href":"http:\/\/www.gnn.club\/index.php?rest_route=\/wp\/v2\/posts\/3233\/revisions\/3238"}],"wp:featuredmedia":[{"embeddable":true,"href":"http:\/\/www.gnn.club\/index.php?rest_route=\/wp\/v2\/media\/3235"}],"wp:attachment":[{"href":"http:\/\/www.gnn.club\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=3233"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/www.gnn.club\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=3233"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/www.gnn.club\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=3233"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}