{"id":3199,"date":"2025-07-10T15:50:24","date_gmt":"2025-07-10T07:50:24","guid":{"rendered":"https:\/\/www.gnn.club\/?p=3199"},"modified":"2025-07-14T15:52:52","modified_gmt":"2025-07-14T07:52:52","slug":"dual-attention-network","status":"publish","type":"post","link":"http:\/\/www.gnn.club\/?p=3199","title":{"rendered":"Dual Attention Network"},"content":{"rendered":"<h1>\u57fa\u672c\u4fe1\u606f<\/h1>\n<ul>\n<li>\ud83d\udcf0\u6807\u9898: Dual Attention Network for Scene Segmentation<\/li>\n<li>\ud83d\udd8b\ufe0f\u4f5c\u8005: Jun Fu<\/li>\n<li>\ud83c\udfdb\ufe0f\u673a\u6784: \u4e2d\u79d1\u9662<\/li>\n<li>\ud83d\udd25\u5173\u952e\u8bcd: Dual Attention Network, Scene Segmentation<\/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>\u573a\u666f\u5206\u5272\u4efb\u52a1\u4e2d\uff0c\u590d\u6742\u573a\u666f\u7684\u4e0a\u4e0b\u6587\u4fe1\u606f\u5efa\u6a21\u548c\u7ec6\u8282\u6355\u6349\u5b58\u5728\u6311\u6218\u3002<\/td>\n<\/tr>\n<tr>\n<td>\ud83c\udfaf\u7814\u7a76\u76ee\u7684<\/td>\n<td>\u63d0\u51fa\u4e00\u79cd\u53cc\u6ce8\u610f\u529b\u7f51\u7edc\uff08Dual Attention Network\uff09\u4ee5\u63d0\u5347\u573a\u666f\u5206\u5272\u6027\u80fd\u3002<\/td>\n<\/tr>\n<tr>\n<td>\u270d\ufe0f\u7814\u7a76\u65b9\u6cd5<\/td>\n<td>\u7ed3\u5408\u901a\u9053\u6ce8\u610f\u529b\u6a21\u5757\u548c\u7a7a\u95f4\u6ce8\u610f\u529b\u6a21\u5757\uff0c\u81ea\u9002\u5e94\u6574\u5408\u5168\u5c40\u4e0a\u4e0b\u6587\u4e0e\u5c40\u90e8\u7279\u5f81\u3002<\/td>\n<\/tr>\n<tr>\n<td>\ud83d\udd4a\ufe0f\u7814\u7a76\u5bf9\u8c61<\/td>\n<td>\u81ea\u7136\u573a\u666f\u56fe\u50cf\uff08\u5982Cityscapes\u3001PASCAL VOC\u7b49\u6570\u636e\u96c6\uff09\u3002<\/td>\n<\/tr>\n<tr>\n<td>\ud83d\udd0d\u7814\u7a76\u7ed3\u8bba<\/td>\n<td>\u5728\u591a\u4e2a\u57fa\u51c6\u6570\u636e\u96c6\u4e0a\u8fbe\u5230SOTA\u6027\u80fd\uff0c\u9a8c\u8bc1\u4e86\u53cc\u6ce8\u610f\u529b\u673a\u5236\u7684\u6709\u6548\u6027\u3002<\/td>\n<\/tr>\n<tr>\n<td>\u2b50\u521b\u65b0\u70b9<\/td>\n<td>\u9996\u6b21\u5c06\u901a\u9053\u4e0e\u7a7a\u95f4\u6ce8\u610f\u529b\u5e76\u884c\u7ed3\u5408\uff0c\u5b9e\u73b0\u591a\u5c3a\u5ea6\u4e0a\u4e0b\u6587\u4fe1\u606f\u7684\u52a8\u6001\u5efa\u6a21\u3002<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<h1>\u80cc\u666f<\/h1>\n<h3>\u7814\u7a76\u80cc\u666f<\/h3>\n<p>\u573a\u666f\u5206\u5272\uff08Scene Segmentation\uff09\u662f\u8ba1\u7b97\u673a\u89c6\u89c9\u4e2d\u7684\u57fa\u7840\u6027\u6311\u6218\u4efb\u52a1\uff0c\u65e8\u5728\u5c06\u56fe\u50cf\u5206\u5272\u4e3a\u4e0d\u540c\u8bed\u4e49\u533a\u57df\uff08\u5982&quot;\u5929\u7a7a&quot;\u3001&quot;\u9053\u8def&quot;\u7b49stuff\u7c7b\u548c&quot;\u4eba&quot;\u3001&quot;\u8f66&quot;\u7b49\u79bb\u6563\u5bf9\u8c61\uff09\u3002\u5176\u5e94\u7528\u6db5\u76d6\u81ea\u52a8\u9a7e\u9a76\u3001\u673a\u5668\u4eba\u611f\u77e5\u7b49\u9886\u57df\uff0c\u4f46\u9762\u4e34\u590d\u6742\u573a\u666f\u4e2d\u591a\u5c3a\u5ea6\u7269\u4f53\u3001\u906e\u6321\u3001\u5149\u7167\u53d8\u5316\u53ca\u76f8\u4f3c\u7c7b\u522b\u6df7\u6dc6\uff08\u5982&quot;\u8349\u5730&quot;\u4e0e&quot;\u7530\u91ce&quot;\uff09\u7b49\u6838\u5fc3\u96be\u9898\uff0c\u4e9f\u9700\u589e\u5f3a\u50cf\u7d20\u7ea7\u7279\u5f81\u8868\u793a\u7684\u5224\u522b\u80fd\u529b\u3002<\/p>\n<h3>\u8fc7\u53bb\u65b9\u6848<\/h3>\n<ol>\n<li>\n<p><strong>\u57fa\u4e8eFCN\u7684\u591a\u5c3a\u5ea6\u4e0a\u4e0b\u6587\u878d\u5408<\/strong>\uff1a<\/p>\n<ul>\n<li>\n<p>\u901a\u8fc7\u7a7a\u6d1e\u5377\u79ef\/\u6c60\u5316\u64cd\u4f5c\u805a\u5408\u591a\u5c3a\u5ea6\u7279\u5f81\uff08\u5982DeepLab\u7cfb\u5217\uff09<\/p>\n<\/li>\n<li>\n<p>\u4f7f\u7528\u5927\u6838\u5206\u89e3\u7ed3\u6784\u6216\u7f16\u7801\u5c42\u6355\u83b7\u5168\u5c40\u4e0a\u4e0b\u6587\uff08\u5982PSPNet\uff09<\/p>\n<\/li>\n<li>\n<p><strong>\u5c40\u9650\u6027<\/strong>\uff1a\u96be\u4ee5\u663e\u5f0f\u5efa\u6a21\u5168\u5c40\u7269\u4f53\/\u573a\u666f\u95f4\u7684\u7a7a\u95f4\u5173\u7cfb<\/p>\n<\/li>\n<\/ul>\n<\/li>\n<li>\n<p><strong>\u57fa\u4e8eRNN\u7684\u957f\u7a0b\u4f9d\u8d56\u5efa\u6a21<\/strong>\uff1a<\/p>\n<ul>\n<li>\n<p>\u91c7\u75282D LSTM\u6216\u56fe\u7ed3\u6784RNN\u6355\u6349\u6807\u7b7e\u95f4\u7a7a\u95f4\u4f9d\u8d56<\/p>\n<\/li>\n<li>\n<p><strong>\u5c40\u9650\u6027<\/strong>\uff1a\u4f9d\u8d56\u5faa\u73af\u795e\u7ecf\u7f51\u7edc\u7684\u9690\u5f0f\u5b66\u4e60\uff0c\u957f\u671f\u8bb0\u5fc6\u6548\u679c\u4e0d\u7a33\u5b9a<\/p>\n<\/li>\n<\/ul>\n<\/li>\n<\/ol>\n<h3>\u7814\u7a76\u52a8\u673a<\/h3>\n<p>\u9488\u5bf9\u73b0\u6709\u65b9\u6cd5\u5728<strong>\u5168\u5c40\u5173\u7cfb\u5efa\u6a21\u4e0d\u8db3<\/strong>\u548c<strong>\u591a\u5c3a\u5ea6\u9002\u5e94\u6027\u6709\u9650<\/strong>\u7684\u95ee\u9898\uff0c\u63d0\u51fa\u53cc\u6ce8\u610f\u529b\u7f51\u7edc\uff08DANet\uff09\uff1a<\/p>\n<ul>\n<li>\u901a\u8fc7\u5e76\u884c\u7a7a\u95f4\/\u901a\u9053\u6ce8\u610f\u529b\u6a21\u5757\u663e\u5f0f\u6355\u83b7\u957f\u7a0b\u4f9d\u8d56<\/li>\n<li>\u81ea\u9002\u5e94\u805a\u5408\u76f8\u4f3c\u7279\u5f81\uff0c\u89e3\u51b3\u5c0f\u76ee\u6807\u88ab\u663e\u8457\u7269\u4f53\u5e72\u6270\u3001\u591a\u5c3a\u5ea6\u7269\u4f53\u8bc6\u522b\u56f0\u96be\u7b49\u75db\u70b9<\/li>\n<li>\u9996\u6b21\u5b9e\u73b0\u7a7a\u95f4\u4e0e\u901a\u9053\u7ef4\u5ea6\u6ce8\u610f\u529b\u673a\u5236\u7684\u534f\u540c\u4f18\u5316\uff0c\u63d0\u5347\u590d\u6742\u573a\u666f\u4e0b\u7684\u5206\u5272\u9c81\u68d2\u6027<\/li>\n<\/ul>\n<h1>\u65b9\u6cd5<\/h1>\n<h3>\u7406\u8bba\u80cc\u666f<\/h3>\n<p>\u672c\u7814\u7a76\u57fa\u4e8e\u6ce8\u610f\u529b\u673a\u5236\uff08Attention Mechanism\uff09\u5728\u89c6\u89c9\u4efb\u52a1\u4e2d\u7684\u4e24\u5927\u6838\u5fc3\u4f18\u52bf\uff1a<\/p>\n<ol>\n<li>\n<p><strong>\u901a\u9053\u6ce8\u610f\u529b<\/strong>\uff08CAM\uff09\uff1a\u901a\u8fc7\u7279\u5f81\u901a\u9053\u95f4\u7684\u6743\u91cd\u91cd\u6807\u5b9a\uff0c\u589e\u5f3a\u5224\u522b\u6027\u7279\u5f81\u5e76\u6291\u5236\u5197\u4f59\u4fe1\u606f<\/p>\n<\/li>\n<li>\n<p><strong>\u7a7a\u95f4\u6ce8\u610f\u529b<\/strong>\uff08PAM\uff09\uff1a\u5efa\u7acb\u50cf\u7d20\u7ea7\u957f\u7a0b\u4f9d\u8d56\u5173\u7cfb\uff0c\u89e3\u51b3\u4f20\u7edf\u5377\u79ef\u64cd\u4f5c\u7684\u5c40\u90e8\u611f\u53d7\u91ce\u9650\u5236<br \/>\n\u7406\u8bba\u521b\u65b0\u70b9\u5728\u4e8e\u9996\u6b21\u5c06\u4e24\u79cd\u6ce8\u610f\u529b\u673a\u5236<strong>\u5e76\u884c\u6574\u5408<\/strong>\uff08\u800c\u975e\u4e32\u884c\u5806\u53e0\uff0c\u533a\u522b\u4e8eCBAM\uff09\uff0c\u5f62\u6210\u4e92\u8865\u6027\u7279\u5f81\u589e\u5f3a\uff1a<\/p>\n<\/li>\n<\/ol>\n<ul>\n<li>\u901a\u9053\u6ce8\u610f\u529b\u4f18\u5316\u7279\u5f81\u56fe\u7684\u8bed\u4e49\u533a\u5206\u5ea6<\/li>\n<li>\u7a7a\u95f4\u6ce8\u610f\u529b\u5efa\u6a21\u7269\u4f53\u95f4\u7684\u62d3\u6251\u5173\u8054\u6027<\/li>\n<\/ul>\n<h3>\u6280\u672f\u8def\u7ebf<\/h3>\n<ol>\n<li>\n<p><strong>\u57fa\u51c6\u7f51\u7edc\u67b6\u6784<\/strong>\uff1a<\/p>\n<ul>\n<li>\u4ee5Dilated FCN\u4e3a\u57fa\u7ebf\uff0c\u91c7\u7528ResNet-50\/101\u4f5c\u4e3abackbone<\/li>\n<li>\u4f7f\u7528\u7a7a\u6d1e\u5377\u79ef\uff08Dilated Convolution\uff09\u4fdd\u6301\u7279\u5f81\u56fe\u5206\u8fa8\u7387<\/li>\n<\/ul>\n<\/li>\n<\/ol>\n<p><img decoding=\"async\" src=\"https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/07\/20250710150147630.png\" width=\"600\" style=\"display: block; margin: 0 auto;\" \/><\/p>\n<ol start=\"2\">\n<li>\n<p><strong>\u53cc\u6ce8\u610f\u529b\u6a21\u5757\u8bbe\u8ba1<\/strong>\uff1a<\/p>\n<ul>\n<li>\n<p><strong>\u4f4d\u7f6e\u6ce8\u610f\u529b\u6a21\u5757\uff08PAM\uff09<\/strong>\uff1a<br \/>\n\u76ee\u6807\uff1a\u5efa\u6a21\u50cf\u7d20\u95f4\u7684\u957f\u7a0b\u7a7a\u95f4\u4f9d\u8d56\u5173\u7cfb\uff0c\u589e\u5f3a\u7279\u5f81\u56fe\u4e2d\u76ee\u6807\u7684\u7a7a\u95f4\u5173\u8054\u6027\u3002<br \/>\n\u8ba1\u7b97\u6d41\u7a0b\uff1a<\/p>\n<p>1\uff0e\u7279\u5f81\u53d8\u6362\uff1a<br \/>\n\uff0d\u8f93\u5165\u7279\u5f81\u56fe $A \\in \\mathbb{R}^{C \\times H \\times W}$ \u7ecf\u8fc7 3 \u4e2a $1 \\times 1$ \u5377\u79ef\u5206\u522b\u5f97\u5230 $B, C, D \\in \\mathbb{R}^{C \\times H \\times W}$ \u3002<br \/>\n2\uff0e\u76f8\u4f3c\u6027\u8ba1\u7b97\uff1a<br \/>\n\uff0d\u5c06 $B$ \u548c $C$ \u91cd\u5851\u4e3a $\\mathbb{R}^{C \\times N} \\quad(N=H \\times W)$ \uff0c\u8ba1\u7b97 \u7a7a\u95f4\u76f8\u5173\u6027\u77e9\u9635 $S \\in \\mathbb{R}^{N \\times N}$ \uff1a $S_{i j}=\\frac{\\exp \\left(B_i \\cdot C_j\\right)}{\\sum_{k=1}^N \\exp \\left(B_i \\cdot C_k\\right)}$ \u5176\u4e2d $S_{i j}$ \u8868\u793a\u4f4d\u7f6e $i$ \u4e0e $j$ \u7684\u76f8\u4f3c\u5ea6\uff08softmax \u5f52\u4e00\u5316\uff09\u3002<br \/>\n3\uff0e\u4e0a\u4e0b\u6587\u805a\u5408\uff1a<br \/>\n\uff0d\u5c06 $D$ \u91cd\u5851\u4e3a $\\mathbb{R}^{C \\times N}$ \uff0c\u4e0e $S$ \u76f8\u4e58\u540e\u6062\u590d\u5f62\u72b6\u5f97\u5230 \u52a0\u6743\u7279\u5f81 $E \\in \\mathbb{R}^{C \\times H \\times W}$ \u3002<br \/>\n4\uff0e\u6b8b\u5dee\u8fde\u63a5\uff1a<br \/>\n\uff0d\u6700\u7ec8\u8f93\u51fa $\\operatorname{PAM}(A)=A+E$ \uff0c\u4fdd\u7559\u539f\u59cb\u7279\u5f81\u7684\u540c\u65f6\u589e\u5f3a\u5168\u5c40\u7a7a\u95f4\u5173\u7cfb\u3002<\/p>\n<\/li>\n<li>\n<p><strong>\u901a\u9053\u6ce8\u610f\u529b\u6a21\u5757\uff08CAM\uff09<\/strong>\uff1a<br \/>\n\u76ee\u6807\uff1a\u589e\u5f3a\u7279\u5f81\u901a\u9053\u95f4\u7684\u8bed\u4e49\u533a\u5206\u5ea6\uff0c\u6291\u5236\u4e0d\u91cd\u8981\u7684\u901a\u9053\u3002<br \/>\n\u8ba1\u7b97\u6d41\u7a0b\uff1a<\/p>\n<p>1\uff0e\u7279\u5f81\u53d8\u6362\uff1a<br \/>\n\uff0d\u8f93\u5165 $A \\in \\mathbb{R}^{C \\times H \\times W}$ \u76f4\u63a5\u91cd\u5851\u4e3a $\\mathbb{R}^{C \\times N} \\quad(N=H \\times W)$ \u3002<br \/>\n2\uff0e\u901a\u9053\u76f8\u5173\u6027\u8ba1\u7b97\uff1a<br \/>\n\uff0d\u8ba1\u7b97 \u901a\u9053\u76f8\u5173\u6027\u77e9\u9635 $X \\in \\mathbb{R}^{C \\times C}: X_{i j}=\\frac{\\exp \\left(A_i \\cdot A_j\\right)}{\\sum_{k=1}^C \\exp \\left(A_i \\cdot A_k\\right)}$ \u5176\u4e2d $X_{i j}$ \u8868\u793a\u901a\u9053 $i$ \u4e0e $j$ \u7684\u76f8\u5173\u6027\u3002<br \/>\n3\uff0e\u7279\u5f81\u52a0\u6743\uff1a<br \/>\n\uff0d\u5c06 $X$ \u4e0e\u539f\u59cb\u7279\u5f81 $A$ \u76f8\u4e58\uff0c\u5f97\u5230\u901a\u9053\u589e\u5f3a\u7279\u5f81 $G \\in \\mathbb{R}^{C \\times H \\times W}$ \u3002<br \/>\n4\uff0e\u6b8b\u5dee\u8fde\u63a5\uff1a<br \/>\n\uff0d\u6700\u7ec8\u8f93\u51fa $C A M(A)=A+G$ \uff0c\u4fdd\u7559\u539f\u59cb\u901a\u9053\u4fe1\u606f\u5e76\u589e\u5f3a\u5224\u522b\u6027\u3002<\/p>\n<\/li>\n<\/ul>\n<\/li>\n<\/ol>\n<p><img decoding=\"async\" src=\"https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/07\/20250710150259364.png\" width=\"400\" style=\"display: block; margin: 0 auto;\" \/><\/p>\n<ol start=\"3\">\n<li>\n<p><strong>\u591a\u57fa\u51c6\u9a8c\u8bc1\u7b56\u7565<\/strong>\uff1a<\/p>\n<ul>\n<li>\u5728Cityscapes\/PASCAL VOC\/PASCAL Context\/COCO Stuff\u56db\u5927\u6570\u636e\u96c6\u9a8c\u8bc1<\/li>\n<li>\u91c7\u7528Mean IoU\u6838\u5fc3\u6307\u6807\uff0c\u5bf9\u6bd4DeepLab-v2\/PSPNet\/PSANet\u7b49SOTA\u6a21\u578b<\/li>\n<\/ul>\n<\/li>\n<li>\n<p><strong>\u8bad\u7ec3\u4f18\u5316<\/strong>\uff1a<\/p>\n<ul>\n<li>\u591a\u5c3a\u5ea6\u6570\u636e\u589e\u5f3a\uff08MSCOCO\u9884\u8bad\u7ec3\u2192\u76ee\u6807\u6570\u636e\u96c6\u5fae\u8c03\uff09<\/li>\n<li>\u91c7\u7528OHEM\uff08Online Hard Example Mining\uff09\u5904\u7406\u7c7b\u522b\u4e0d\u5e73\u8861\u95ee\u9898<\/li>\n<\/ul>\n<\/li>\n<\/ol>\n<h1>\u7ed3\u8bba<\/h1>\n<ul>\n<li>\n<p>\u63d0\u51faDual Attention Network (DANet)\u901a\u8fc7\u5e76\u884c\u7a7a\u95f4\/\u901a\u9053\u81ea\u6ce8\u610f\u529b\u673a\u5236\uff0c\u9996\u6b21\u5b9e\u73b0\u573a\u666f\u5206\u5272\u4e2d\u5168\u5c40\u4e0a\u4e0b\u6587\u4e0e\u5c40\u90e8\u7279\u5f81\u7684\u52a8\u6001\u81ea\u9002\u5e94\u878d\u5408\uff0c\u4e3a\u590d\u6742\u573a\u666f\u7406\u89e3\u63d0\u4f9b\u65b0\u8303\u5f0f<\/p>\n<\/li>\n<li>\n<p><strong>\u4f18\u70b9<\/strong>\uff1a\u53cc\u6ce8\u610f\u529b\u6a21\u5757\u5728Cityscapes\u7b494\u4e2a\u57fa\u51c6\u6570\u636e\u96c6\u5b9e\u73b0SOTA\u6027\u80fd\uff0c\u53ef\u89c6\u5316\u5b9e\u9a8c\u9a8c\u8bc1\u5176\u957f\u7a0b\u4f9d\u8d56\u5efa\u6a21\u80fd\u529b\uff1b<strong>\u5c40\u9650<\/strong>\uff1a\u672a\u89e3\u51b3\u8ba1\u7b97\u590d\u6742\u5ea6\u95ee\u9898\uff08\u4f5c\u8005\u660e\u786e\u5217\u4e3a\u672a\u6765\u5de5\u4f5c\u65b9\u5411\uff09<br \/>\n\u4e3b\u8981\u7ed3\u8bba:<br \/>\n(1) \u4f4d\u7f6e\u6ce8\u610f\u529b\u6a21\u5757(PAM)\u4e0e\u901a\u9053\u6ce8\u610f\u529b\u6a21\u5757(CAM)\u7684\u5e76\u884c\u67b6\u6784\u53ef\u5206\u522b\u6709\u6548\u5efa\u6a21\u7a7a\u95f4\u7ef4\u5ea6\u7269\u4f53\u5173\u8054\u6027\u548c\u901a\u9053\u7ef4\u5ea6\u8bed\u4e49\u533a\u5206\u5ea6<br \/>\n(2) \u6d88\u878d\u5b9e\u9a8c\u8bc1\u660e\u53cc\u6ce8\u610f\u529b\u673a\u5236\u76f8\u6bd4\u4f20\u7edf\u591a\u5c3a\u5ea6\u4e0a\u4e0b\u6587\u805a\u5408\u65b9\u6cd5\uff08\u5982\u7a7a\u6d1e\u5377\u79ef\/\u91d1\u5b57\u5854\u6c60\u5316\uff09\u5177\u6709\u663e\u8457\u6027\u80fd\u63d0\u5347<br \/>\n(3) \u5728Cityscapes\/PASCAL VOC\u7b49\u6570\u636e\u96c6\u7684\u5b9a\u91cf\u5206\u6790\u663e\u793amIoU\u6307\u6807\u63d0\u53471.5-3.2%\uff0c\u7279\u522b\u6539\u5584\u5c0f\u76ee\u6807\u53ca\u76f8\u4f3c\u8bed\u4e49\u533a\u57df\u7684\u5206\u5272\u7cbe\u5ea6<\/p>\n<\/li>\n<\/ul>\n<h1>Pytorch code<\/h1>\n<pre><code class=\"language-python\">import torch\nimport torch.nn as nn\nimport torch.nn.functional as F\n\nclass DualAttention(nn.Module):\n    def __init__(self, in_channels):\n        &quot;&quot;&quot;\n        Dual Attention Network \u6a21\u5757\u5b9e\u73b0\n        Args:\n            in_channels: \u8f93\u5165\u7279\u5f81\u56fe\u7684\u901a\u9053\u6570\n        &quot;&quot;&quot;\n        super(DualAttention, self).__init__()\n        self.in_channels = in_channels\n\n        # \u4f4d\u7f6e\u6ce8\u610f\u529b\u6a21\u5757\n        self.position_attention = PositionAttention(in_channels)\n\n        # \u901a\u9053\u6ce8\u610f\u529b\u6a21\u5757\n        self.channel_attention = ChannelAttention(in_channels)\n\n        # \u8f93\u51fa\u878d\u5408\u5377\u79ef\n        self.conv = nn.Conv2d(in_channels * 2, in_channels, kernel_size=1)\n\n    def forward(self, x):\n        # \u5e76\u884c\u8ba1\u7b97\u4e24\u79cd\u6ce8\u610f\u529b\n        pa = self.position_attention(x)\n        ca = self.channel_attention(x)\n\n        # \u62fc\u63a5\u4e24\u79cd\u6ce8\u610f\u529b\u7ed3\u679c\n        combined = torch.cat([pa, ca], dim=1)\n\n        # \u878d\u5408\u7279\u5f81\n        out = self.conv(combined)\n\n        return out + x  # \u6b8b\u5dee\u8fde\u63a5\n\nclass PositionAttention(nn.Module):\n    def __init__(self, in_channels):\n        super(PositionAttention, self).__init__()\n        self.in_channels = in_channels  # \u6dfb\u52a0\u8fd9\u884c\u521d\u59cb\u5316\n        self.query_conv = nn.Conv2d(in_channels, in_channels \/\/ 8, kernel_size=1)\n        self.key_conv = nn.Conv2d(in_channels, in_channels \/\/ 8, kernel_size=1)\n        self.value_conv = nn.Conv2d(in_channels, in_channels, kernel_size=1)\n        self.gamma = nn.Parameter(torch.zeros(1))\n\n    def forward(self, x):\n        batch_size, _, height, width = x.size()\n\n        # \u8ba1\u7b97Q, K, V\n        query = self.query_conv(x).view(batch_size, -1, height * width).permute(0, 2, 1)\n        key = self.key_conv(x).view(batch_size, -1, height * width)\n        value = self.value_conv(x).view(batch_size, -1, height * width)\n\n        # \u8ba1\u7b97\u6ce8\u610f\u529b\u56fe\n        attention = torch.bmm(query, key)  # [B, N, N]\n        attention = F.softmax(attention, dim=-1)\n\n        # \u5e94\u7528\u6ce8\u610f\u529b\n        out = torch.bmm(value, attention.permute(0, 2, 1))\n        out = out.view(batch_size, self.in_channels, height, width)\n\n        return self.gamma * out + x\n\nclass ChannelAttention(nn.Module):\n    def __init__(self, in_channels):\n        super(ChannelAttention, self).__init__()\n        self.in_channels = in_channels  # \u6dfb\u52a0\u8fd9\u884c\u521d\u59cb\u5316\n        self.gamma = nn.Parameter(torch.zeros(1))\n\n    def forward(self, x):\n        batch_size, channels, height, width = x.size()\n\n        # \u8ba1\u7b97\u901a\u9053\u6ce8\u610f\u529b\n        query = x.view(batch_size, channels, -1)\n        key = x.view(batch_size, channels, -1).permute(0, 2, 1)\n        value = x.view(batch_size, channels, -1)\n\n        # \u8ba1\u7b97\u6ce8\u610f\u529b\u56fe\n        attention = torch.bmm(query, key)  # [B, C, C]\n        attention = F.softmax(attention, dim=-1)\n\n        # \u5e94\u7528\u6ce8\u610f\u529b\n        out = torch.bmm(attention, value)\n        out = out.view(batch_size, self.in_channels, height, width)\n\n        return self.gamma * out + x\n\n# ------------------- \u7528\u6cd5\u793a\u4f8b -------------------\nif __name__ == &quot;__main__&quot;:\n    # 1. \u521d\u59cb\u5316Dual Attention\u6a21\u5757\uff08\u8f93\u5165\u901a\u9053\u6570\u4e3a256\uff09\n    dan = DualAttention(in_channels=256)\n\n    # 2. \u6a21\u62df\u8f93\u5165\u6570\u636e\uff08batch_size=4, \u901a\u9053=256, \u7279\u5f81\u56fe\u5c3a\u5bf8=56x56\uff09\n    dummy_input = torch.randn(4, 256, 56, 56)\n\n    # 3. \u524d\u5411\u4f20\u64ad\n    output = dan(dummy_input)\n\n    print(f&quot;\u8f93\u5165\u5f62\u72b6: {dummy_input.shape}&quot;)\n    print(f&quot;\u8f93\u51fa\u5f62\u72b6: {output.shape}&quot;)  # \u5e94\u4e0e\u8f93\u5165\u5f62\u72b6\u4e00\u81f4<\/code><\/pre>\n","protected":false},"excerpt":{"rendered":"<p>\u57fa\u672c\u4fe1\u606f \ud83d\udcf0\u6807\u9898: Dual Attention Network for Scene Segmentation [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":3200,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[30,18],"tags":[],"class_list":["post-3199","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-30","category-18"],"_links":{"self":[{"href":"http:\/\/www.gnn.club\/index.php?rest_route=\/wp\/v2\/posts\/3199","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=3199"}],"version-history":[{"count":6,"href":"http:\/\/www.gnn.club\/index.php?rest_route=\/wp\/v2\/posts\/3199\/revisions"}],"predecessor-version":[{"id":3232,"href":"http:\/\/www.gnn.club\/index.php?rest_route=\/wp\/v2\/posts\/3199\/revisions\/3232"}],"wp:featuredmedia":[{"embeddable":true,"href":"http:\/\/www.gnn.club\/index.php?rest_route=\/wp\/v2\/media\/3200"}],"wp:attachment":[{"href":"http:\/\/www.gnn.club\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=3199"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/www.gnn.club\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=3199"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/www.gnn.club\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=3199"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}