{"id":3239,"date":"2025-07-16T17:51:38","date_gmt":"2025-07-16T09:51:38","guid":{"rendered":"https:\/\/www.gnn.club\/?p=3239"},"modified":"2025-07-16T17:52:13","modified_gmt":"2025-07-16T09:52:13","slug":"3239","status":"publish","type":"post","link":"http:\/\/gnn.club\/?p=3239","title":{"rendered":"Swin Transformer"},"content":{"rendered":"<h1>\u57fa\u672c\u4fe1\u606f<\/h1>\n<ul>\n<li>\ud83d\udcf0\u6807\u9898: Swin Transformer: Hierarchical Vision Transformer using Shifted Windows<\/li>\n<li>\ud83d\udd8b\ufe0f\u4f5c\u8005: Ze Liu<\/li>\n<li>\ud83c\udfdb\ufe0f\u673a\u6784: Microsoft Research Asia (\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662)<\/li>\n<li>\ud83d\udd25\u5173\u952e\u8bcd: Vision Transformer, Hierarchical Architecture, Shifted Windows, Self-Attention<\/li>\n<\/ul>\n<h2>\u6458\u8981\u6982\u8ff0<\/h2>\n<table border=\"1\" cellspacing=\"0\" cellpadding=\"5\" style=\"border-collapse: collapse; width: 100%;\">\n<thead>\n<tr>\n<th style=\"border: 1px solid #ddd; padding: 8px; text-align: left;\">\u9879\u76ee<\/th>\n<th style=\"border: 1px solid #ddd; padding: 8px; text-align: left;\">\u5185\u5bb9<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td style=\"border: 1px solid #ddd; padding: 8px;\">\ud83d\udcd6\u7814\u7a76\u80cc\u666f<\/td>\n<td style=\"border: 1px solid #ddd; padding: 8px;\">Vision Transformer\uff08ViT\uff09\u5728\u56fe\u50cf\u5206\u7c7b\u4e2d\u8868\u73b0\u4f18\u5f02\uff0c\u4f46\u9ad8\u5206\u8fa8\u7387\u56fe\u50cf\u8ba1\u7b97\u590d\u6742\u5ea6\u9ad8\uff0c\u4e14\u7f3a\u4e4f\u591a\u5c3a\u5ea6\u5efa\u6a21\u80fd\u529b\u3002<\/td>\n<\/tr>\n<tr>\n<td style=\"border: 1px solid #ddd; padding: 8px;\">\ud83c\udfaf\u7814\u7a76\u76ee\u7684<\/td>\n<td style=\"border: 1px solid #ddd; padding: 8px;\">\u63d0\u51fa\u4e00\u79cd\u9ad8\u6548\u4e14\u901a\u7528\u7684\u89c6\u89c9Transformer\u9aa8\u5e72\u7f51\u7edc\uff0c\u652f\u6301\u5bc6\u96c6\u9884\u6d4b\u4efb\u52a1\uff08\u5982\u68c0\u6d4b\u548c\u5206\u5272\uff09\u3002<\/td>\n<\/tr>\n<tr>\n<td style=\"border: 1px solid #ddd; padding: 8px;\">\u270d\ufe0f\u7814\u7a76\u65b9\u6cd5<\/td>\n<td style=\"border: 1px solid #ddd; padding: 8px;\">\u8bbe\u8ba1\u5206\u5c42\u67b6\u6784\uff08Hierarchical Architecture\uff09\u548c\u79fb\u4f4d\u7a97\u53e3\u81ea\u6ce8\u610f\u529b\uff08Shifted Window Self-Attention\uff09\u3002<\/td>\n<\/tr>\n<tr>\n<td style=\"border: 1px solid #ddd; padding: 8px;\">\ud83d\udd4a\ufe0f\u7814\u7a76\u5bf9\u8c61<\/td>\n<td style=\"border: 1px solid #ddd; padding: 8px;\">\u56fe\u50cf\u5206\u7c7b\uff08ImageNet\uff09\u3001\u76ee\u6807\u68c0\u6d4b\uff08COCO\uff09\u3001\u8bed\u4e49\u5206\u5272\uff08ADE20K\uff09\u4efb\u52a1\u3002<\/td>\n<\/tr>\n<tr>\n<td style=\"border: 1px solid #ddd; padding: 8px;\">\ud83d\udd0d\u7814\u7a76\u7ed3\u8bba<\/td>\n<td style=\"border: 1px solid #ddd; padding: 8px;\">Swin Transformer\u5728\u8ba1\u7b97\u6548\u7387\u548c\u6027\u80fd\u4e0a\u4f18\u4e8eViT\u548cCNN\uff0cCOCO\u68c0\u6d4bAP\u8fbe58.7\uff08Swin-L\uff09\u3002<\/td>\n<\/tr>\n<tr>\n<td style=\"border: 1px solid #ddd; padding: 8px;\">\u2b50\u521b\u65b0\u70b9<\/td>\n<td style=\"border: 1px solid #ddd; padding: 8px;\">1. \u79fb\u4f4d\u7a97\u53e3\u673a\u5236\u5b9e\u73b0\u8de8\u7a97\u53e3\u4fe1\u606f\u4ea4\u4e92\uff1b2. \u5206\u5c42\u7ed3\u6784\u652f\u6301\u591a\u5c3a\u5ea6\u7279\u5f81\uff1b3. \u7ebf\u6027\u8ba1\u7b97\u590d\u6742\u5ea6\u3002<\/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 \/>\n\u8ba1\u7b97\u673a\u89c6\u89c9\u5efa\u6a21\u957f\u671f\u7531CNN\u4e3b\u5bfc\uff08\u5982AlexNet\uff09\uff0c\u800cNLP\u9886\u57df\u5219\u4ee5Transformer\u4e3a\u4e3b\u6d41\u3002\u5c3d\u7ba1Transformer\u5728\u8bed\u8a00\u4efb\u52a1\u4e2d\u8868\u73b0\u5353\u8d8a\uff0c\u4f46\u5176\u76f4\u63a5\u5e94\u7528\u4e8e\u89c6\u89c9\u4efb\u52a1\u9762\u4e34\u4e24\u5927\u6311\u6218\uff1a1) \u89c6\u89c9\u5143\u7d20\u5c3a\u5ea6\u591a\u53d8\uff0c\u800c\u4f20\u7edfTransformer\u91c7\u7528\u56fa\u5b9a\u5c3a\u5ea6\u7684token\uff1b2) \u9ad8\u5206\u8fa8\u7387\u56fe\u50cf\uff08\u5982\u8bed\u4e49\u5206\u5272\uff09\u5bfc\u81f4\u5168\u5c40\u81ea\u6ce8\u610f\u529b\u8ba1\u7b97\u590d\u6742\u5ea6\u5448\u5e73\u65b9\u7ea7\u589e\u957f\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<\/strong>\uff08\u5982ResNeXt\uff09\uff1a\u901a\u8fc7\u590d\u6742\u5377\u79ef\u7ed3\u6784\u63d0\u5347\u6027\u80fd\uff0c\u4f46\u7f3a\u4e4f\u5168\u5c40\u5efa\u6a21\u80fd\u529b\u3002<\/p>\n<\/li>\n<li>\n<p><strong>Vision Transformer\uff08ViT\uff09<\/strong>\uff1a\u867d\u5728\u56fe\u50cf\u5206\u7c7b\u4e2d\u8868\u73b0\u4f18\u5f02\uff0c\u4f46\u5b58\u5728\u5355\u5206\u8fa8\u7387\u7279\u5f81\u56fe\uff08\u65e0\u6cd5\u9002\u5e94\u591a\u5c3a\u5ea6\u4efb\u52a1\uff09\u548c\u4e8c\u6b21\u65b9\u8ba1\u7b97\u590d\u6742\u5ea6\uff08\u96be\u4ee5\u5904\u7406\u9ad8\u5206\u8fa8\u7387\u56fe\u50cf\uff09\u7684\u7f3a\u9677\u3002<\/p>\n<\/li>\n<li>\n<p><strong>\u6ed1\u52a8\u7a97\u53e3\u81ea\u6ce8\u610f\u529b<\/strong>\uff08\u5982\u5c40\u90e8\u6ce8\u610f\u529b\u673a\u5236\uff09\uff1a\u8ba1\u7b97\u6548\u7387\u4f4e\uff0c\u786c\u4ef6\u5ef6\u8fdf\u9ad8\u3002<\/p>\n<\/li>\n<\/ol>\n<ul>\n<li><strong>\u7814\u7a76\u52a8\u673a<\/strong>\uff1a<br \/>\n\u4e3a\u89e3\u51b3Transformer\u5728\u89c6\u89c9\u4efb\u52a1\u4e2d\u7684\u5c40\u9650\u6027\uff0c\u63d0\u51faSwin Transformer\uff1a<br \/>\n1) \u901a\u8fc7\u5206\u5c42\u67b6\u6784\uff08\u9010\u6b65\u5408\u5e76\u56fe\u50cf\u5757\uff09\u652f\u6301\u591a\u5c3a\u5ea6\u7279\u5f81\uff1b<br \/>\n2) \u91c7\u7528\u79fb\u4f4d\u7a97\u53e3\u81ea\u6ce8\u610f\u529b\uff08Shifted Window\uff09\u5b9e\u73b0\u8de8\u7a97\u53e3\u4ea4\u4e92\uff0c\u5c06\u8ba1\u7b97\u590d\u6742\u5ea6\u964d\u81f3\u7ebf\u6027\uff1b<br \/>\n3) \u6784\u5efa\u901a\u7528\u89c6\u89c9\u9aa8\u5e72\u7f51\u7edc\uff0c\u7edf\u4e00\u56fe\u50cf\u5206\u7c7b\u4e0e\u5bc6\u96c6\u9884\u6d4b\u4efb\u52a1\uff08\u5982\u68c0\u6d4b\u3001\u5206\u5272\uff09\u7684\u67b6\u6784\u9700\u6c42\u3002<\/li>\n<\/ul>\n<h1>\u65b9\u6cd5<\/h1>\n<ul>\n<li>\n<p><strong>\u7406\u8bba\u80cc\u666f<\/strong>\uff1a<br \/>\n\u672c\u7814\u7a76\u57fa\u4e8e\u89c6\u89c9Transformer\u67b6\u6784\uff0c\u9488\u5bf9\u4f20\u7edfVision Transformer\uff08ViT\uff09\u5728\u9ad8\u5206\u8fa8\u7387\u56fe\u50cf\u5904\u7406\u4e2d\u7684\u8ba1\u7b97\u590d\u6742\u5ea6\u9ad8\uff08 $O(N^2)$ \uff09\u548c\u7f3a\u4e4f\u591a\u5c3a\u5ea6\u5efa\u6a21\u80fd\u529b\u4e24\u5927\u6838\u5fc3\u95ee\u9898\uff0c\u63d0\u51fa\u6539\u8fdb\u65b9\u6848\u3002\u53d7CNN\u5206\u5c42\u67b6\u6784\u542f\u53d1\uff0c\u7ed3\u5408\u5c40\u90e8\u7a97\u53e3\u81ea\u6ce8\u610f\u529b\uff08Local Window Self-Attention\uff09\u4e0e\u8de8\u7a97\u53e3\u4ea4\u4e92\u673a\u5236\uff0c\u6784\u5efa\u9ad8\u6548\u4e14\u901a\u7528\u7684\u89c6\u89c9\u9aa8\u5e72\u7f51\u7edc\u3002<\/p>\n<\/li>\n<li>\n<p><strong>\u6280\u672f\u8def\u7ebf<\/strong>\uff1a<\/p>\n<\/li>\n<\/ul>\n<p>\u7b80\u5355\u770b\u4e0b\u539f\u8bba\u6587\u4e2d\u7ed9\u51fa\u7684\u5173\u4e8eSwin Transformer\u7f51\u7edc\u7684\u7ed3\u6784\u56fe\uff0c\u5982\u4e0b\u56fe\u6240\u793a\u3002\u9996\u5148\u5c06\u56fe\u7247\u8f93\u5165\u5230Patch Partition\u6a21\u5757\u4e2d\u8fdb\u884c\u5206\u5757\uff0c\u5373\u6bcf4x4\u76f8\u90bb\u7684\u50cf\u7d20\u4e3a\u4e00\u4e2aPatch\uff0c\u7136\u540e\u5728\u901a\u9053\u7ef4\u5ea6\u4e0a\u5c55\u5e73\u3002\u5047\u8bbe\u8f93\u5165\u7684\u662fRGB\u4e09\u901a\u9053\u56fe\u7247\uff0c\u90a3\u4e48\u6bcf\u4e2aPatch\u5c31\u6709\u4e2a4x4=16\u50cf\u7d20\uff0c\u7136\u540e\u6bcf\u4e2a\u50cf\u7d20\u6709R\u3001G\u3001B\u4e09\u4e2a\u503c\u6240\u4ee5\u5c55\u5e73\u540e\u6570\u636e\u7ef4\u5ea6\u662f16x3=48\uff0c\u6240\u4ee5\u901a\u8fc7Patch Partition\u540e\u56fe\u50cf\u5f20\u91cf\u7684\u5f62\u72b6\u7531[H,W,3]\u53d8\u6210\u4e86[H\/4,W\/4,48]\u3002\u7136\u540e\u518d\u901a\u7ebf\u6027\u5d4c\u5165\u5c42\u5bf9\u6bcf\u4e2a\u50cf\u7d20\u7684\u901a\u9053\u6570\u636e\u505a\u7ebf\u6027\u53d8\u6362\uff0c\u753148\u53d8\u6210C\uff0c\u5373\u56fe\u50cf\u5f20\u91cf\u7684\u5f62\u72b6\u518d\u7531[H\/4,W\/4,48]\u53d8\u6210\u4e86[H\/4,W\/4,C]\u3002\u5176\u5b9e\u5728\u6e90\u7801\u4e2dPatch Partition\u548c\u7ebf\u6027\u5d4c\u5165\u64cd\u4f5c\u5c31\u662f\u76f4\u63a5\u901a\u8fc7\u4e00\u4e2a\u5377\u79ef\u5c42\u5b9e\u73b0\u7684\uff0c\u548c\u4e4b\u524dViT\u6a21\u578b\u4e2d\u7684Embedding\u5c42\u7ed3\u6784\u4e00\u6a21\u4e00\u6837\u3002<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/07\/20250716174021939.png\" alt=\"file\" \/><\/p>\n<p>\u63a5\u4e0b\u6765\uff0cSwim Transformer\u6a21\u578b\u901a\u8fc7\u56db\u4e2aStage\u6784\u5efa\u4e0d\u540c\u5c3a\u5bf8\u7684\u7279\u5f81\u56fe\u3002Stage1\u5148\u901a\u8fc7\u4e00\u4e2aLinear Embedding\u5c42\uff0c\u800c\u5269\u4e0b\u7684\u4e09\u4e2aStage\u90fd\u4f1a\u5148\u4f7f\u7528Patch Merging\u5c42\u8fdb\u884c\u4e0b\u91c7\u6837\u3002\u7136\u540e\u91cd\u590d\u5806\u53e0Swin Transformer Block\uff0c\u6ce8\u610f\u8fd9\u91cc\u7684Block\u5b9e\u9645\u4e0a\u6709\u4e24\u79cd\u7ed3\u6784\uff0c\u5982\u56fe6-28(b)\u6240\u793a\uff0c\u5b83\u4eec\u7684\u4e0d\u540c\u4e4b\u5904\u4ec5\u5728\u4e8e\u4e00\u4e2a\u4f7f\u7528W-MSA\u7ed3\u6784\uff0c\u53e6\u4e00\u4e2a\u4f7f\u7528SW-MSA\u7ed3\u6784\u3002\u8fd9\u4e24\u79cd\u7ed3\u6784\u662f\u6210\u5bf9\u4f7f\u7528\u7684\uff0c\u5148\u4f7f\u7528\u4e00\u4e2aW-MSA\u7ed3\u6784\uff0c\u7136\u540e\u4f7f\u7528\u4e00\u4e2aSW-MSA\u7ed3\u6784\uff0c\u56e0\u6b64\u5806\u53e0Swin Transformer Block\u7684\u6b21\u6570\u90fd\u662f\u5076\u6570\u3002\u6700\u540e\uff0c\u5206\u7c7b\u7f51\u7edc\u4f1a\u63a5\u4e0a\u4e00\u4e2aLN\u5c42\u3001\u5168\u5c40\u6c60\u5316\u5c42\u4ee5\u53ca\u5168\u8fde\u63a5\u5c42\uff0c\u4ee5\u5f97\u5230\u6700\u7ec8\u8f93\u51fa\u3002\u8fd9\u91cc\u5e76\u672a\u5728\u9876\u5c42\u56fe\u4e2d\u7ed9\u51fa\u3002<br \/>\n\u63a5\u4e0b\u6765\uff0c\u5206\u522b\u8be6\u7ec6\u4ecb\u7ecdPatch Merging\u3001W-MSA\u3001SW-MSA\u4ee5\u53ca\u4f7f\u7528\u5230\u7684\u76f8\u5bf9\u4f4d\u7f6e\u504f\u7f6e\uff08Relative Position Bias\uff09\u3002\u9700\u8981\u6ce8\u610f\u7684\u662f\uff0cSwin Transformer Block\u4e2d\u7684MLP\u7ed3\u6784\u548cVision Transformer\u4e2d\u7684\u7ed3\u6784\u662f\u4e00\u6837\u7684\u3002<\/p>\n<p><strong>Patch Merging\u8be6\u89e3<\/strong><br \/>\n\u5728\u6bcf\u4e2aStage\u4e2d\u9996\u5148\u8981\u901a\u8fc7\u4e00\u4e2aPatch Merging\u5c42\u8fdb\u884c\u4e0b\u91c7\u6837\uff08Stage 1\u9664\u5916\uff09\uff0c\u6b64\u64cd\u4f5c\u7684\u76ee\u7684\u662f\u5c06\u7279\u5f81\u4fe1\u606f\u4ece\u7a7a\u95f4\u7ef4\u5ea6\u8f6c\u79fb\u5230\u901a\u9053\u7ef4\u5ea6\u3002\u5047\u8bbe\u8f93\u5165Patch Merging\u7684\u662f\u4e00\u4e2a4x4\u5927\u5c0f\u7684\u5355\u901a\u9053\u7279\u5f81\u56fe\uff0c\u4ee5\u7279\u5f81\u56fe\u5de6\u4e0a\u89d2\u7684\u56db\u4e2a\u5143\u7d20\u4e3a\u8d77\u70b9\uff0c\u901a\u8fc7\u95f4\u9694\u91c7\u6837\u5f97\u5230\u56db\u4e2a\u5b50\u7279\u5f81\u56fe\uff0c\u7136\u540e\u5c06\u8fd9\u56db\u4e2a\u5b50\u7279\u5f81\u5728\u901a\u9053\u7ef4\u5ea6\u4e0a\u8fdb\u884c\u62fc\u63a5\uff0c\u518d\u901a\u8fc7\u4e00\u4e2aLN\u5c42\u3002\u6700\u540e\u901a\u8fc7\u4e00\u4e2a\u5168\u8fde\u63a5\u5c42\u5728\u7279\u5f81\u56fe\u7684\u6df1\u5ea6\u65b9\u5411\u505a\u7ebf\u6027\u53d8\u6362\uff0c\u5c06\u7279\u5f81\u56fe\u7684\u6df1\u5ea6\u7531C\u53d8\u6210C\/2\uff0c\u5982\u4e0b\u56fe\u6240\u793a\u3002\u901a\u8fc7\u8fd9\u4e2a\u7b80\u5355\u7684\u4f8b\u5b50\u53ef\u4ee5\u770b\u51fa\uff0c\u7ecf\u8fc7Patch Merging\u5c42\u540e\uff0c\u7279\u5f81\u56fe\u7684\u9ad8\u548c\u5bbd\u4f1a\u51cf\u534a\uff0c\u6df1\u5ea6\u4f1a\u7ffb\u500d\u3002<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/07\/20250716174139769.png\"  style=\"display: block; margin: 0 auto;\" \/><\/p>\n<p><strong>W-MSA\u8be6\u89e3<\/strong><br \/>\n\u5f15\u5165W-MSA\u6a21\u5757\u662f\u4e3a\u4e86\u51cf\u5c11\u8ba1\u7b97\u91cf\u3002\u666e\u901a\u7684MSA\u6a21\u5757\u5982\u56fe6-30(a)\u6240\u793a\uff0c\u5bf9\u4e8e\u7279\u5f81\u56fe\u4e2d\u7684\u6bcf\u4e2a\u50cf\u7d20\uff0c\u5728\u81ea\u6ce8\u610f\u529b\u8ba1\u7b97\u8fc7\u7a0b\u4e2d\u9700\u8981\u548c\u6240\u6709\u7684\u50cf\u7d20\u53bb\u8ba1\u7b97\u3002\u5728\u4f7f\u7528W-MSA\u6a21\u5757\u65f6\uff0c\u9996\u5148\u5c06\u7279\u5f81\u56fe\u6309\u7167 MxM \u7684\u5927\u5c0f\u5212\u5206\u6210\u4e00\u4e2a\u7a97\u53e3\uff08\u4f8b\u5b50\u4e2d\u7684M=4\uff09\uff0c\u7136\u540e\u5355\u72ec\u5bf9\u6bcf\u4e2aWindows\u5185\u90e8\u8fdb\u884c\u81ea\u6ce8\u610f\u529b\u8ba1\u7b97\u3002<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/07\/20250716174226900.png\" width=\"300\" style=\"display: block; margin: 0 auto;\" \/><\/p>\n<p>\u4e24\u8005\u7684\u8ba1\u7b97\u91cf\u5177\u4f53\u5dee\u591a\u5c11\u5462\uff1f\u539f\u8bba\u6587\u4e2d\u6709\u7ed9\u51fa\u4e0b\u9762\u4e24\u4e2a\u516c\u5f0f\uff0c\u8fd9\u91cc\u5ffd\u7565\u4e86 Softmax \u51fd\u6570\u7684\u8ba1\u7b97\u590d\u6742\u5ea6\uff1a<\/p>\n<p>$$<br \/>\n\\begin{gathered}<br \/>\n\\Omega(\\mathrm{MSA})=4 h w C^2+2(h w)^2 C \\<br \/>\n\\Omega(\\mathrm{~W}-\\mathrm{MSA})=4 h w C^2+2 M^2 h w C<br \/>\n\\end{gathered}<br \/>\n$$<\/p>\n<p>\u5176\u4e2d\uff0c$h$ \u4ee3\u8868 feature map \u7684\u9ad8\u5ea6\uff0c$w$ \u4ee3\u8868 feature map \u7684\u5bbd\u5ea6\uff0c$C$ \u4ee3\u8868 feature map \u7684\u6df1\u5ea6\uff0c $M$ \u4ee3\u8868\u6bcf\u4e2a\u7a97\u53e3\u7684\u5927\u5c0f\u3002\u5047\u8bbe\u7279\u5f81\u56fe\u7684 $h \u3001 w$ \u90fd\u4e3a $64, M=4, C=96$ \u3002\u91c7\u7528\u591a\u5934\u6ce8\u610f\u529b\uff08MSA\uff09\u6a21\u5757\u7684\u8ba1\u7b97\u590d\u6742\u5ea6\u4e3a $4 \\times 64 \\times 64 \\times 962+2 \\times(64 \\times 64)^2 \\times 96=3372220416$ \u800c\u91c7\u7528 W\uff0dMSA \u6a21\u5757\u7684\u8ba1\u7b97\u590d\u6742\u5ea6\u4e3a $4 \\times 64 \\times 64 \\times 962+2 \\times 42 \\times 64 \\times 64 \\times 96=163577856$ \uff0c\u8282\u7701\u4e86 $95 \\%$ \u7684\u8ba1\u7b97\u590d\u6742\u5ea6\u3002<\/p>\n<p><strong>SW-MSA\u8be6\u89e3<\/strong><br \/>\n\u4f7f\u7528W-MSA\u6a21\u5757\u65f6\uff0c\u53ea\u4f1a\u5728\u6bcf\u4e2a\u7a97\u53e3\u5185\u8fdb\u884c\u81ea\u6ce8\u610f\u529b\u8ba1\u7b97\uff0c\u56e0\u6b64\u7a97\u53e3\u4e0e\u7a97\u53e3\u4e4b\u95f4\u65e0\u6cd5\u8fdb\u884c\u4fe1\u606f\u4f20\u9012\u3002\u4e3a\u4e86\u89e3\u51b3\u8fd9\u4e2a\u95ee\u9898\uff0c\u4f5c\u8005\u5f15\u5165\u4e86SW-MSA\u6a21\u5757\uff0c\u5373\u8fdb\u884c\u504f\u79fb\u7684W-MSA\u3002\u8ba1\u7b97\u81ea\u6ce8\u610f\u529b\u673a\u5236\u524d\uff0c\u7a97\u53e3\u53d1\u751f\u4e86\u504f\u79fb\uff0c\u53ef\u4ee5\u7406\u89e3\u6210\u7a97\u53e3\u4ece\u5de6\u4e0a\u89d2\u5206\u522b\u5411\u53f3\u4fa7\u548c\u4e0b\u65b9\u5404\u504f\u79fb\u4e86M\/2\u4e2a\u50cf\u7d20\u3002<\/p>\n<p>\u5728\u6b64\u60c5\u51b5\u4e0b\uff0c\u89c2\u5bdf\u4e0b\u56fe\u5c55\u793a\u7684Layer1+1\u5c42\u53ef\u4ee5\u53d1\u73b0\uff0c\u5bf9\u4e8e\u7b2c\u4e00\u884c\u7b2c\u4e8c\u52172x4\u7684\u7a97\u53e3\uff0c\u5b83\u53ef\u4ee5\u4f7fLayer1\u5c42\u7b2c\u4e00\u6392\u7684\u4e24\u4e2a4x4\u5927\u5c0f\u7684\u7a97\u53e3\u4e4b\u95f4\u8fdb\u884c\u4fe1\u606f\u8fdb\u884c\u4ea4\u6d41\u3002\u540c\u6837\u7684\u9053\u7406\uff0c\u5bf9\u4e8eLayer1+1\u7b2c\u4e8c\u884c\u7b2c\u4e8c\u52174x4\u7684\u7a97\u53e3\uff0c\u5b83\u53ef\u4ee5\u4fc3\u8fdbLayer1\u5c42\u4e2d\u56db\u4e2a\u7a97\u53e3\u4e4b\u95f4\u7684\u4fe1\u606f\u4ea4\u6d41\u3002\u5176\u5b83\u7a97\u53e3\u7684\u60c5\u51b5\u4e5f\u662f\u5982\u6b64\u3002\u8fd9\u89e3\u51b3\u4e86\u4e0d\u540c\u7a97\u53e3\u4e4b\u95f4\u65e0\u6cd5\u8fdb\u884c\u4fe1\u606f\u4ea4\u6d41\u7684\u95ee\u9898\u3002<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/07\/20250716174445563.png\"  style=\"display: block; margin: 0 auto;\" \/><\/p>\n<p>\u6839\u636e\u4e0a\u56fe\uff0c\u53ef\u4ee5\u53d1\u73b0\u901a\u8fc7\u5c06\u7a97\u53e3\u8fdb\u884c\u504f\u79fb\u540e\uff0c\u5c31\u5230\u8fbe\u4e86\u7a97\u53e3\u4e0e\u7a97\u53e3\u4e4b\u95f4\u7684\u76f8\u4e92\u901a\u4fe1\u3002\u867d\u7136\u5df2\u7ecf\u80fd\u8fbe\u5230\u7a97\u53e3\u4e0e\u7a97\u53e3\u4e4b\u95f4\u7684\u901a\u4fe1\uff0c\u4f46\u662f\u539f\u6765\u7684\u7279\u5f81\u56fe\u53ea\u67094\u4e2a\u7a97\u53e3\uff0c\u7ecf\u8fc7\u79fb\u52a8\u7a97\u53e3\u540e\uff0c\u5f97\u5230\u4e869\u4e2a\u7a97\u53e3\uff0c\u7a97\u53e3\u7684\u6570\u91cf\u6709\u6240\u589e\u52a0\u5e76\u4e149\u4e2a\u7a97\u53e3\u7684\u5927\u5c0f\u4e5f\u4e0d\u662f\u5b8c\u5168\u76f8\u540c\uff0c\u8fd9\u5c31\u5bfc\u81f4\u8ba1\u7b97\u96be\u5ea6\u589e\u52a0\u3002\u56e0\u6b64\uff0c\u4f5c\u8005\u53c8\u63d0\u51fa\u4e86Efficient batch computation for shifted configuration\uff0c\u4e00\u79cd\u66f4\u52a0\u9ad8\u6548\u7684\u8ba1\u7b97\u65b9\u6cd5\uff0c\u5982\u4e0b\u56fe\u6240\u793a\uff1a<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/07\/20250716174530981.png\"  style=\"display: block; margin: 0 auto;\" \/><\/p>\n<p>\u5148\u5c06 \u201c012\u201d\u533a\u57df\u79fb\u52a8\u5230\u6700\u4e0b\u65b9\uff0c\u518d\u5c06 \u201c360\u201d\u533a\u57df\u79fb\u52a8\u5230\u6700\u53f3\u65b9\uff0c\u6b64\u65f6\u79fb\u52a8\u5b8c\u540e\uff0c\u533a\u57df\u201c4\u201d\u6210\u4e3a\u4e86\u4e00\u4e2a\u5355\u72ec\u7684\u7a97\u53e3\uff0c\u518d\u5c06\u533a\u57df\u201c5\u201d\u548c\u201c3\u201d\u5408\u5e76\u6210\u4e00\u4e2a\u7a97\u53e3\uff1b\u201c7\u201d\u548c\u201c1\u201d\u5408\u5e76\u6210\u4e00\u4e2a\u7a97\u53e3\uff1b\u201c8\u201d\u3001\u201c6\u201d\u3001\u201c2\u201d\u548c\u201c0\u201d\u5408\u5e76\u6210\u4e00\u4e2a\u7a97\u53e3\u3002\u8fd9\u6837\u53c8\u548c\u539f\u6765\u4e00\u6837\u53d8\u4e3a4\u4e2a\u7684\u7a97\u53e3\u4e86\uff0c\u5c31\u80fd\u591f\u4fdd\u8bc1\u8ba1\u7b97\u91cf\u7684\u4e00\u81f4\uff0c\u7136\u540e\u5206\u522b\u5728\u6bcf\u4e2a\u7a97\u53e3\u5185\u8fdb\u884c\u81ea\u6ce8\u610f\u529b\u8ba1\u7b97\u64cd\u4f5c\u6765\u63d0\u53d6\u7279\u5f81\uff0c\u8fd9\u6837\u53ef\u4ee5\u66f4\u6709\u6548\u5730\u5229\u7528GPU\u5e76\u884c\u8ba1\u7b97\u7684\u52a0\u901f\u6548\u679c\u3002\u4f46\u662f\u628a\u4e0d\u540c\u7684\u533a\u57df\u5408\u5e76\u5728\u4e00\u8d77\uff08\u6bd4\u59825\u548c3\uff09\u8fdb\u884cMSA\uff0c\u4ece\u903b\u8f91\u4e0a\u8bb2\u662f\u4e0d\u5408\u7406\u7684\uff0c\u56e0\u4e3a\u8fd9\u4e24\u4e2a\u533a\u57df\u5e76\u4e0d\u5e94\u8be5\u76f8\u90bb\u3002\u4e3a\u4e86\u9632\u6b62\u8fd9\u4e2a\u95ee\u9898\uff0c\u5728\u5b9e\u9645\u8ba1\u7b97\u4e2d\u4f7f\u7528\u7684\u662f\u5e26\u8499\u677f\u7684MSA\uff08Masked MSA\uff09\uff0c\u8fd9\u6837\u5c31\u80fd\u591f\u901a\u8fc7\u8bbe\u7f6e\u8499\u677f\u6765\u9694\u7edd\u4e0d\u540c\u533a\u57df\u7684\u4fe1\u606f\u4e86\u3002<\/p>\n<p>\u5173\u4e8emask\u5982\u4f55\u4f7f\u7528\uff0c\u5148\u56de\u987e\u5e38\u89c4\u7684MSA\u8ba1\u7b97\u8fc7\u7a0b\uff0c\u53ef\u4ee5\u53d1\u73b0\uff0c\u5176\u5728\u6700\u540e\u8f93\u51fa\u65f6\u8981\u7ecf\u8fc7Softmax\u64cd\u4f5c\uff0cSoftmax\u5728\u8f93\u5165\u5f88\u5c0f\u65f6\uff0c\u5176\u8f93\u51fa\u51e0\u4e4e\u4e3a0\u3002\u4ee5\u4e0a\u56fe\u7684\u533a\u57df\u201c5\u201d\u548c\u533a\u57df\u201c3\u201d\u5408\u5e76\u540e\u7684\u533a\u57df\u201c53\u201d\u4e3a\u4f8b\uff0c\u5982\u679c\u67d0\u50cf\u7d20\u662f\u5c5e\u4e8e\u533a\u57df5\u7684\uff0c\u6211\u4eec\u53ea\u60f3\u8ba9\u5b83\u548c\u533a\u57df\u201c5\u201d\u5185\u7684\u50cf\u7d20\u8fdb\u884c\u5339\u914d\u3002\u4e3a\u4e86\u5b9e\u73b0\u8fd9\u4e2a\u76ee\u6807\uff0c\u53ef\u4ee5\u5c06\u8be5\u50cf\u7d20\u4e0e\u533a\u57df\u201c3\u201d\u4e2d\u7684\u6240\u6709\u50cf\u7d20\u8fdb\u884c\u6ce8\u610f\u529b\u8ba1\u7b97\u65f6\uff0c\u5c06\u8ba1\u7b97\u7ed3\u679c\u51cf\u53bb100\u3002\u8fd9\u6837\uff0c\u5728\u7ecf\u8fc7Softmax\u4e4b\u540e\uff0c\u5bf9\u5e94\u7684\u6743\u91cd\u5c31\u4f1a\u63a5\u8fd1\u4e8e0\u3002\u56e0\u6b64\uff0c\u5bf9\u4e8e\u8be5\u50cf\u7d20\u800c\u8a00\uff0c\u5b9e\u9645\u4e0a\u53ea\u4e0e\u533a\u57df\u201c5\u201d\u5185\u7684\u50cf\u7d20\u8fdb\u884c\u4e86MSA\u3002\u5bf9\u4e8e\u5176\u5b83\u50cf\u7d20\u4e5f\u662f\u540c\u7406\u3002\u9700\u8981\u6ce8\u610f\u7684\u662f\uff0c\u5728\u8ba1\u7b97\u7ed3\u675f\u540e\uff0c\u8fd8\u9700\u8981\u5c06\u6570\u636e\u79fb\u56de\u5230\u539f\u6765\u7684\u4f4d\u7f6e\u4e0a\u3002<\/p>\n<h1>\u7ed3\u8bba<\/h1>\n<ul>\n<li>\n<p>\u63d0\u51faSwin Transformer\u4f5c\u4e3a\u65b0\u578b\u89c6\u89c9Transformer\u67b6\u6784\uff0c\u901a\u8fc7\u5206\u5c42\u7279\u5f81\u8868\u8fbe\u548c\u7ebf\u6027\u8ba1\u7b97\u590d\u6742\u5ea6\u7a81\u7834ViT\u5728\u9ad8\u5206\u8fa8\u7387\u56fe\u50cf\u5904\u7406\u4e2d\u7684\u5c40\u9650\u6027\uff0c\u4e3a\u89c6\u89c9\u4e0e\u8bed\u8a00\u4fe1\u53f7\u7684\u7edf\u4e00\u5efa\u6a21\u63d0\u4f9b\u65b0\u8303\u5f0f\u3002<\/p>\n<\/li>\n<li>\n<p>\u4f18\u70b9\uff1a<br \/>\n1) \u5206\u5c42\u7ed3\u6784\u5b9e\u73b0\u591a\u5c3a\u5ea6\u7279\u5f81\u63d0\u53d6\uff1b<br \/>\n2) \u79fb\u4f4d\u7a97\u53e3\u673a\u5236\u5728\u4fdd\u6301\u7ebf\u6027\u590d\u6742\u5ea6\u7684\u540c\u65f6\u5b9e\u73b0\u5168\u5c40\u5efa\u6a21\uff1b<br \/>\n3) \u5728COCO\u548cADE20K\u4efb\u52a1\u4e0a\u663e\u8457\u8d85\u8d8aSOTA\u65b9\u6cd5\u3002\u7f3a\u70b9\uff1a\u672a\u660e\u786e\u8ba8\u8bba\u5c0f\u6837\u672c\u573a\u666f\u4e0b\u7684\u6cdb\u5316\u80fd\u529b\u6216\u52a8\u6001\u5206\u8fa8\u7387\u9002\u914d\u6027\u3002<\/p>\n<p>\u4e3b\u8981\u7ed3\u8bba\uff1a<\/p>\n<\/li>\n<\/ul>\n<ol>\n<li>\u63d0\u51fa\u5177\u6709\u5c42\u6b21\u5316\u67b6\u6784\u7684Swin Transformer\uff0c\u5176\u8ba1\u7b97\u590d\u6742\u5ea6\u4e0e\u8f93\u5165\u56fe\u50cf\u5c3a\u5bf8\u5448\u7ebf\u6027\u5173\u7cfb\uff1b<\/li>\n<li>\u5728COCO\u76ee\u6807\u68c0\u6d4b\u548cADE20K\u8bed\u4e49\u5206\u5272\u4efb\u52a1\u4e0a\u8fbe\u5230SOTA\u6027\u80fd\uff0c\u9a8c\u8bc1\u4e86\u67b6\u6784\u901a\u7528\u6027\uff1b<\/li>\n<li>\u4e3a\u89c6\u89c9\u4e0e\u8bed\u8a00\u4fe1\u53f7\u7684\u7edf\u4e00\u5efa\u6a21\u5f00\u8f9f\u4e86\u65b0\u7814\u7a76\u65b9\u5411\u3002<\/li>\n<\/ol>\n","protected":false},"excerpt":{"rendered":"<p>\u57fa\u672c\u4fe1\u606f \ud83d\udcf0\u6807\u9898: Swin Transformer: Hierarchical Vision Transfo [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":3240,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[30,18],"tags":[31],"class_list":["post-3239","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-30","category-18","tag-transformer"],"_links":{"self":[{"href":"http:\/\/gnn.club\/index.php?rest_route=\/wp\/v2\/posts\/3239","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=3239"}],"version-history":[{"count":3,"href":"http:\/\/gnn.club\/index.php?rest_route=\/wp\/v2\/posts\/3239\/revisions"}],"predecessor-version":[{"id":3247,"href":"http:\/\/gnn.club\/index.php?rest_route=\/wp\/v2\/posts\/3239\/revisions\/3247"}],"wp:featuredmedia":[{"embeddable":true,"href":"http:\/\/gnn.club\/index.php?rest_route=\/wp\/v2\/media\/3240"}],"wp:attachment":[{"href":"http:\/\/gnn.club\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=3239"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/gnn.club\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=3239"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/gnn.club\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=3239"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}