{"id":3220,"date":"2025-07-11T23:59:05","date_gmt":"2025-07-11T15:59:05","guid":{"rendered":"https:\/\/www.gnn.club\/?p=3220"},"modified":"2025-07-18T19:38:47","modified_gmt":"2025-07-18T11:38:47","slug":"attentional-feature-fusion","status":"publish","type":"post","link":"http:\/\/gnn.club\/?p=3220","title":{"rendered":"Attentional Feature Fusion"},"content":{"rendered":"<h1>\u57fa\u672c\u4fe1\u606f<\/h1>\n<ul>\n<li>\ud83d\udcf0\u6807\u9898: Attentional Feature Fusion<\/li>\n<li>\ud83d\udd8b\ufe0f\u4f5c\u8005: Yimian Dai<\/li>\n<li>\ud83c\udfdb\ufe0f\u673a\u6784: Nanjing University of Science and Technology (\u5357\u4eac\u7406\u5de5\u5927\u5b66)<\/li>\n<li>\ud83d\udd25\u5173\u952e\u8bcd: Feature Fusion, Attention Mechanism, Multi-Scale, Deep Learning<\/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>\u73b0\u6709\u7279\u5f81\u878d\u5408\u65b9\u6cd5\uff08\u5982\u7ebf\u6027\u76f8\u52a0\/\u62fc\u63a5\uff09\u5728\u8de8\u5c42\u548c\u591a\u5c3a\u5ea6\u573a\u666f\u4e0b\u5b58\u5728\u8bed\u4e49\u4e0d\u4e00\u81f4\u548c\u52a8\u6001\u9002\u5e94\u6027\u4e0d\u8db3\u7684\u95ee\u9898\u3002<\/td>\n<\/tr>\n<tr>\n<td>\ud83c\udfaf\u7814\u7a76\u76ee\u7684<\/td>\n<td>\u63d0\u51fa\u7edf\u4e00\u6846\u67b6\u4ee5\u4f18\u5316\u7279\u5f81\u878d\u5408\uff0c\u89e3\u51b3\u521d\u59cb\u5bf9\u9f50\u548c\u591a\u5c3a\u5ea6\u4e0a\u4e0b\u6587\u805a\u5408\u7684\u6311\u6218\u3002<\/td>\n<\/tr>\n<tr>\n<td>\u270d\ufe0f\u7814\u7a76\u65b9\u6cd5<\/td>\n<td>\u8bbe\u8ba1AFF\u6a21\u5757\uff08\u52a8\u6001\u6ce8\u610f\u529b\u878d\u5408\uff09\u548cMS-CAM\uff08\u591a\u5c3a\u5ea6\u901a\u9053\u6ce8\u610f\u529b\uff09\uff0c\u7ed3\u5408\u8fed\u4ee3\u4f18\u5316\uff08iAFF\uff09\u3002<\/td>\n<\/tr>\n<tr>\n<td>\ud83d\udd4a\ufe0f\u7814\u7a76\u5bf9\u8c61<\/td>\n<td>\u6df1\u5ea6\u795e\u7ecf\u7f51\u7edc\u4e2d\u7684\u540c\u5c42\/\u8de8\u5c42\u7279\u5f81\u878d\u5408\uff08\u5982ResNet\u3001FPN\u7b49\uff09\u3002<\/td>\n<\/tr>\n<tr>\n<td>\ud83d\udd0d\u7814\u7a76\u7ed3\u8bba<\/td>\n<td>AFF\/iAFF\u663e\u8457\u63d0\u5347\u591a\u5c3a\u5ea6\u76ee\u6807\uff08\u5c24\u5176\u5c0f\u7269\u4f53\uff09\u7684\u5224\u522b\u529b\uff0c\u4e14\u53c2\u6570\u91cf\u66f4\u4f18\uff1b\u521d\u59cb\u878d\u5408\u8d28\u91cf\u662f\u5173\u952e\u74f6\u9888\u3002<\/td>\n<\/tr>\n<tr>\n<td>\u2b50\u521b\u65b0\u70b9<\/td>\n<td>1. \u9996\u6b21\u5c06\u6ce8\u610f\u529b\u6cdb\u5316\u81f3\u5168\u573a\u666f\u878d\u5408\uff1b2. MS-CAM\u901a\u8fc7\u591a\u6c60\u5316\u589e\u5f3a\u5c3a\u5ea6\u611f\u77e5\uff1b3. \u8fed\u4ee3\u6ce8\u610f\u529b\u4f18\u5316\u521d\u59cb\u878d\u5408\u3002<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<h1>\u80cc\u666f<\/h1>\n<h3>\u7814\u7a76\u80cc\u666f\uff1a<\/h3>\n<p>\u5377\u79ef\u795e\u7ecf\u7f51\u7edc\uff08CNN\uff09\u901a\u8fc7\u589e\u52a0\u6df1\u5ea6\u3001\u5bbd\u5ea6\u3001\u57fa\u6570\u6216\u52a8\u6001\u4f18\u5316\u7279\u5f81\u663e\u8457\u63d0\u5347\u4e86\u8868\u5f81\u80fd\u529b\uff0c\u4f46\u7279\u5f81\u878d\u5408\u4f5c\u4e3a\u7f51\u7edc\u6838\u5fc3\u7ec4\u4ef6\u4ecd\u4f9d\u8d56\u7b80\u5355\u7684\u7ebf\u6027\u64cd\u4f5c\uff08\u5982\u76f8\u52a0\/\u62fc\u63a5\uff09\uff0c\u65e0\u6cd5\u9002\u5e94\u8de8\u5c42\/\u591a\u5c3a\u5ea6\u573a\u666f\u4e0b\u7684\u8bed\u4e49\u4e0d\u4e00\u81f4\u95ee\u9898\u3002\u73b0\u6709\u5de5\u4f5c\uff08\u5982InceptionNet\u3001ResNet\u3001FPN\uff09\u867d\u5e7f\u6cdb\u4f7f\u7528\u7279\u5f81\u878d\u5408\uff0c\u4f46\u591a\u805a\u7126\u4e8e\u8def\u5f84\u8bbe\u8ba1\u800c\u975e\u878d\u5408\u65b9\u6cd5\u672c\u8eab\u3002<\/p>\n<h3>\u8fc7\u53bb\u65b9\u6848\uff1a<\/h3>\n<ol>\n<li>\n<p><strong>\u4f20\u7edf\u65b9\u6cd5<\/strong>\uff1a\u7ebf\u6027\u878d\u5408\uff08\u52a0\u6cd5\/\u62fc\u63a5\uff09\u7f3a\u4e4f\u52a8\u6001\u9002\u5e94\u6027\uff0c\u96be\u4ee5\u5904\u7406\u7279\u5f81\u5c3a\u5ea6\u4e0e\u8bed\u4e49\u5dee\u5f02\u3002<\/p>\n<\/li>\n<li>\n<p><strong>\u6ce8\u610f\u529b\u6539\u8fdb<\/strong>\uff1aSKNet\u548cResNeSt\u5f15\u5165\u901a\u9053\u6ce8\u610f\u529b\u5b9e\u73b0\u540c\u5c42\u7279\u5f81\u52a8\u6001\u52a0\u6743\uff0c\u4f46\u5b58\u5728\u4e09\u5927\u5c40\u9650\uff1a<\/p>\n<ul>\n<li>\u4ec5\u9002\u7528\u4e8e\u540c\u5c42\u878d\u5408\uff0c\u8de8\u5c42\u573a\u666f\uff08\u5982skip connections\uff09\u672a\u89e3\u51b3\uff1b<\/li>\n<li>\u521d\u59cb\u878d\u5408\uff08\u5982\u7b80\u5355\u76f8\u52a0\uff09\u6210\u4e3a\u6027\u80fd\u74f6\u9888\uff1b<\/li>\n<li>\u5168\u5c40\u901a\u9053\u6ce8\u610f\u529b\u504f\u5411\u5927\u76ee\u6807\uff0c\u5ffd\u89c6\u591a\u5c3a\u5ea6\u4e0a\u4e0b\u6587\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ol>\n<h3>\u7814\u7a76\u52a8\u673a\uff1a<\/h3>\n<p>\u9488\u5bf9\u4e0a\u8ff0\u7f3a\u9677\uff0c\u672c\u6587\u63d0\u51fa\uff1a<\/p>\n<ol>\n<li>\n<p><strong>\u7edf\u4e00\u6846\u67b6\u9700\u6c42<\/strong>\uff1a\u4e9f\u9700\u4e00\u79cd\u901a\u7528\u65b9\u6cd5\u7edf\u4e00\u540c\u5c42\/\u8de8\u5c42\u7279\u5f81\u878d\u5408\uff1b<\/p>\n<\/li>\n<li>\n<p><strong>\u52a8\u6001\u4f18\u5316\u5fc5\u8981\u6027<\/strong>\uff1a\u9700\u540c\u65f6\u89e3\u51b3\u521d\u59cb\u878d\u5408\u8d28\u91cf\u4e0e\u591a\u5c3a\u5ea6\u4e0a\u4e0b\u6587\u805a\u5408\u95ee\u9898\uff1b<\/p>\n<\/li>\n<li>\n<p><strong>\u5c0f\u76ee\u6807\u654f\u611f\u5ea6<\/strong>\uff1a\u901a\u8fc7\u591a\u5c3a\u5ea6\u901a\u9053\u6ce8\u610f\u529b\uff08MS-CAM\uff09\u589e\u5f3a\u5bf9\u6781\u7aef\u5c3a\u5ea6\uff08\u5c24\u5176\u5c0f\u7269\u4f53\uff09\u7684\u5224\u522b\u529b\u3002<\/p>\n<\/li>\n<\/ol>\n<h1>\u65b9\u6cd5<\/h1>\n<ul>\n<li>\n<p><strong>\u7406\u8bba\u80cc\u666f<\/strong>:<br \/>\n\u672c\u7814\u7a76\u57fa\u4e8e\u7279\u5f81\u878d\u5408\u7684\u52a8\u6001\u4f18\u5316\u7406\u8bba\uff0c\u6307\u51fa\u4f20\u7edf\u7ebf\u6027\u878d\u5408\uff08\u5982\u76f8\u52a0\/\u62fc\u63a5\uff09\u56e0\u7f3a\u4e4f\u8bed\u4e49\u611f\u77e5\u80fd\u529b\u5bfc\u81f4\u8de8\u5c42\/\u591a\u5c3a\u5ea6\u7279\u5f81\u5bf9\u9f50\u5931\u6548\u3002\u6838\u5fc3\u7406\u8bba\u652f\u6491\u5305\u62ec\uff1a<br \/>\n\uff081\uff09\u6ce8\u610f\u529b\u673a\u5236\u7684\u7279\u5f81\u9009\u62e9\u7279\u6027\uff08\u53c2\u8003SENet\uff09\uff1b<br \/>\n\uff082\uff09\u591a\u5c3a\u5ea6\u4e0a\u4e0b\u6587\u5efa\u6a21\u7684\u5fc5\u8981\u6027\uff08\u53c2\u8003Inception\u6a21\u5757\uff09\uff1b<br \/>\n\uff083\uff09\u8fed\u4ee3\u4f18\u5316\u5bf9\u521d\u59cb\u878d\u5408\u504f\u5dee\u7684\u4fee\u6b63\u4f5c\u7528\uff08\u53c2\u8003\u6b8b\u5dee\u5b66\u4e60\u601d\u60f3\uff09\u3002<\/p>\n<\/li>\n<li>\n<p><strong>\u6280\u672f\u8def\u7ebf<\/strong>:<\/p>\n<p><strong>1. \u6846\u67b6\u8bbe\u8ba1<\/strong>\uff1a\u6784\u5efa\u901a\u7528\u6ce8\u610f\u529b\u7279\u5f81\u878d\u5408\uff08AFF\uff09\u6a21\u5757\uff0c\u901a\u8fc7\u7279\u5f81\u62fc\u63a5\u2192\u901a\u9053\u6ce8\u610f\u529b\u2192\u7a7a\u95f4\u91cd\u52a0\u6743\u5b9e\u73b0\u52a8\u6001\u878d\u5408\uff1a<\/p>\n<\/li>\n<\/ul>\n<p>AFF\u6a21\u5757\u5206\u4e3a\u4e09\u4e2a\u5173\u952e\u9636\u6bb5\uff1a<\/p>\n<p>1\uff0e\u521d\u59cb\u7279\u5f81\u6574\u5408\uff08Initial Integration\uff09<br \/>\n\u3002\u8f93\u5165\u4e24\u4e2a\u5f85\u878d\u5408\u7684\u7279\u5f81\u56fe $X$ \u548c $Y$\uff08\u4f8b\u5982\u540c\u5c42\u7684\u591a\u5c3a\u5ea6\u7279\u5f81\u6216\u8de8\u5c42\u8df3\u8dc3\u8fde\u63a5\u7279\u5f81\uff09\uff0c\u9ed8\u8ba4 $Y$ \u5177\u6709\u66f4\u5927\u7684\u611f\u53d7\u91ce\u3002<br \/>\n\uff0d\u901a\u8fc7\u5143\u7d20\u76f8\u52a0 $(\\oplus)$ \u6216\u62fc\u63a5 $(\\uplus)$ \u8fdb\u884c\u521d\u59cb\u6574\u5408\uff0c\u751f\u6210\u4e2d\u95f4\u7279\u5f81 $X \\oplus Y$\u3002<br \/>\n2\uff0e\u591a\u5c3a\u5ea6\u901a\u9053\u6ce8\u610f\u529b\uff08MS\uff0dCAM\uff09<\/p>\n<ul>\n<li>\u5bf9\u521d\u59cb\u6574\u5408\u540e\u7684\u7279\u5f81\u5e94\u7528\u591a\u5c3a\u5ea6\u901a\u9053\u6ce8\u610f\u529b\u6a21\u5757\uff08MS\uff0dCAM\uff09\uff1a<\/li>\n<li>\u5168\u5c40\u4e0a\u4e0b\u6587\uff1a\u901a\u8fc7\u5168\u5c40\u5e73\u5747\u6c60\u5316\uff08GAP\uff09\u538b\u7f29\u7a7a\u95f4\u7ef4\u5ea6\uff0c\u751f\u6210\u901a\u9053\u7ea7\u5168\u5c40\u63cf\u8ff0\u7b26 $g(X \\oplus Y)$ \u3002<\/li>\n<li>\u5c40\u90e8\u4e0a\u4e0b\u6587\uff1a\u901a\u8fc7\u70b9\u5377\u79ef\uff08PWConv\uff09\u63d0\u53d6\u9010\u50cf\u7d20\u7684\u5c40\u90e8\u901a\u9053\u4ea4\u4e92\u7279\u5f81 $L(X \\oplus Y$ \uff09\uff0c\u4fdd\u7559\u7ec6\u8282\u4fe1\u606f\u3002<\/li>\n<li>\u52a8\u6001\u6743\u91cd\u751f\u6210\uff1a\u5c06\u5168\u5c40\u4e0e\u5c40\u90e8\u4e0a\u4e0b\u6587\u76f8\u52a0\u540e\u7ecfSigmoid\u6fc0\u6d3b\uff0c\u751f\u6210\u6ce8\u610f\u529b\u6743\u91cd<\/li>\n<\/ul>\n<p>$$<br \/>\nM(X \\oplus Y) \\in[0,1]^{C \\times H \\times W} \\quad<br \/>\n$$<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/07\/20250711234713569.png\" alt=\"file\" \/><\/p>\n<p>3\uff0e\u8f6f\u9009\u62e9\u878d\u5408\uff08Soft Selection\uff09<br \/>\n\uff0d\u4f7f\u7528\u6ce8\u610f\u529b\u6743\u91cd\u5bf9\u8f93\u5165\u7279\u5f81\u8fdb\u884c\u52a0\u6743\u6c42\u548c\uff1a$Z=M(X \\oplus Y) \\otimes X+(1-M(X \\oplus Y)) \\otimes Y$ \u5176\u4e2d $\\otimes$ \u4e3a\u9010\u5143\u7d20\u4e58\u6cd5\uff0c\u5b9e\u73b0\u7279\u5f81\u7684\u81ea\u9002\u5e94\u878d\u5408\u3002<br \/>\n<img decoding=\"async\" src=\"https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/07\/20250711234821309.png\" alt=\"file\" \/><\/p>\n<p><strong>\u8fed\u4ee3\u4f18\u5316<\/strong>\uff1a\u5728AFF\u57fa\u7840\u4e0a\u5f15\u5165iAFF\u7ed3\u6784\uff0c\u901a\u8fc7\u9012\u5f52\u6ce8\u610f\u529b\u673a\u5236\u6e10\u8fdb\u4fee\u6b63\u521d\u59cb\u878d\u5408\u8bef\u5dee\uff1b<br \/>\n<img decoding=\"async\" src=\"https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/07\/20250711234847862.png\" alt=\"file\" \/><\/p>\n<ul>\n<li>\u4e24\u7ea7\u6ce8\u610f\u529b\u673a\u5236\uff1a\u7b2c\u4e00\u7ea7AFF\u6a21\u5757\u751f\u6210\u521d\u6b65\u878d\u5408\u7279\u5f81 $X \\oplus Y$ \uff0c\u7b2c\u4e8c\u7ea7AFF\u8fdb\u4e00\u6b65\u4f18\u5316\u878d\u5408\u6743\u91cd\u3002<\/li>\n<li>\u516c\u5f0f\u5316\u8868\u793a\u4e3a\uff1a<\/li>\n<\/ul>\n<p>$$<br \/>\nX \\oplus Y=M_1(X+Y) \\otimes X+\\left(1-M_1(X+Y)\\right) \\otimes Y Z=M_2(X \\oplus Y) \\otimes X+\\left(1-M_2(X \\oplus Y)\\right) \\otimes Y<br \/>\n$$<\/p>\n<p>\u5b9e\u9a8c\u8868\u660eiAFF\u5728ImageNet\u4e0a\u6bd4\u5355\u7ea7AFF\u63d0\u5347 $0.5 \\%$ \u51c6\u786e\u7387\u3002<\/p>\n<p><strong>\u517c\u5bb9\u6027\u6269\u5c55<\/strong>\uff1a\u9002\u914d\u540c\u5c42\uff08Inception\uff09\u3001\u77ed\u8df3\u8fde\uff08ResNet\uff09\u3001\u957f\u8df3\u8fde\uff08FPN\uff09\u4e09\u7c7b\u5178\u578b\u878d\u5408\u573a\u666f\u3002<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/07\/20250711234203172.png\" alt=\"file\" \/><\/p>\n<h1>\u7ed3\u8bba<\/h1>\n<ul>\n<li>\n<p>\u672c\u7814\u7a76\u5c06attention\u673a\u5236\u62d3\u5c55\u4e3a\u7279\u5f81\u878d\u5408\u7684\u901a\u7528\u52a8\u6001\u9009\u62e9\u6846\u67b6\uff0c\u89e3\u51b3\u4e86\u8de8\u5c42\/\u591a\u5c3a\u5ea6\u573a\u666f\u4e0b\u7684\u8bed\u4e49\u5bf9\u9f50\u95ee\u9898\uff0c\u4e3a\u6df1\u5ea6\u795e\u7ecf\u7f51\u7edc\u7684\u7279\u5f81\u878d\u5408\u8bbe\u8ba1\u63d0\u4f9b\u4e86\u65b0\u8303\u5f0f\u3002<\/p>\n<\/li>\n<li>\n<p><strong>\u4f18\u70b9<\/strong>\uff1a\u63d0\u51fa\u7684MS-CAM\u6a21\u5757\u901a\u8fc7\u878d\u5408\u5c40\u90e8\u4e0e\u5168\u5c40\u901a\u9053\u4e0a\u4e0b\u6587\u663e\u8457\u63d0\u5347\u591a\u5c3a\u5ea6\u611f\u77e5\u80fd\u529b\uff1biAFF\u7ed3\u6784\u9996\u6b21\u7cfb\u7edf\u89e3\u51b3\u521d\u59cb\u878d\u5408\u8d28\u91cf\u74f6\u9888\u3002<\/p>\n<\/li>\n<li>\n<p><strong>\u7f3a\u70b9<\/strong>\uff1a\u672a\u8ba8\u8bba\u8ba1\u7b97\u590d\u6742\u5ea6\u4e0e\u5b9e\u65f6\u6027\u6743\u8861\uff0c\u4e14\u5b9e\u9a8c\u4ec5\u57fa\u4e8e\u56fe\u50cf\u5206\u7c7b\u4efb\u52a1\u9a8c\u8bc1\u3002<br \/>\n<strong>\u4e3b\u8981\u7ed3\u8bba<\/strong>\uff1a<\/p>\n<\/li>\n<\/ul>\n<ol>\n<li>Attention\u673a\u5236\u53ef\u6cdb\u5316\u4e3a\u540c\u5c42\/\u8de8\u5c42\u7279\u5f81\u878d\u5408\u7684\u7edf\u4e00\u89e3\u51b3\u65b9\u6848\uff1b<\/li>\n<li>MS-CAM\u901a\u8fc7\u591a\u5c3a\u5ea6\u901a\u9053\u7edf\u8ba1\u6709\u6548\u7f13\u89e3\u8bed\u4e49\u4e0e\u5c3a\u5ea6\u4e0d\u4e00\u81f4\u6027\uff1b<\/li>\n<li>\u521d\u59cb\u878d\u5408\u8d28\u91cf\u662fattention-based\u878d\u5408\u7684\u5173\u952e\u74f6\u9888\uff0c\u8fed\u4ee3\u6ce8\u610f\u529b\uff08iAFF\uff09\u53ef\u663e\u8457\u4f18\u5316\uff1b<\/li>\n<li>\u5728CIFAR-100\/ImageNet\u4e0a\u4ee5\u66f4\u5c11\u53c2\u6570\u91cf\u8d85\u8d8aSOTA\uff0c\u9a8c\u8bc1\u4e86\u7cbe\u7ec6\u5316\u7279\u5f81\u878d\u5408\u7684\u6f5c\u529b\u3002<\/li>\n<\/ol>\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 MSCAM(nn.Module):\n    def __init__(self, channels=64, r=4):\n        super(MSCAM, self).__init__()\n        inter_channels = int(channels \/\/ r)\n\n        # \u5c40\u90e8\u6ce8\u610f\u529b\n        self.local_att = nn.Sequential(\n            nn.Conv2d(channels, inter_channels, kernel_size=1, stride=1, padding=0),\n            nn.BatchNorm2d(inter_channels),\n            nn.ReLU(inplace=True),\n            nn.Conv2d(inter_channels, channels, kernel_size=1, stride=1, padding=0),\n            nn.BatchNorm2d(channels),\n        )\n\n        # \u5168\u5c40\u6ce8\u610f\u529b - \u79fb\u9664\u4e86BatchNorm\u4ee5\u907f\u514d1x1\u8f93\u5165\u7684\u95ee\u9898\n        self.global_att = nn.Sequential(\n            nn.AdaptiveAvgPool2d(1),\n            nn.Conv2d(channels, inter_channels, kernel_size=1, stride=1, padding=0),\n            nn.ReLU(inplace=True),\n            nn.Conv2d(inter_channels, channels, kernel_size=1, stride=1, padding=0),\n        )\n\n        self.sigmoid = nn.Sigmoid()\n\n    def forward(self, x):\n        xl = self.local_att(x)\n        xg = self.global_att(x)\n        xlg = xl + xg\n        wei = self.sigmoid(xlg)\n        return wei\n\nclass AFF(nn.Module):\n    def __init__(self, channels=64, r=4):\n        super(AFF, self).__init__()\n        self.MSCAM = MSCAM(channels, r)  # \u8f93\u5165\u901a\u9053\u662fchannels\uff08\u76f8\u52a0\u540e\uff09\n\n    def forward(self, x, y):\n        xy = x + y  \n        wei = self.MSCAM(xy)\n        # \u52a0\u6743\u878d\u5408\n        xo = x * wei + y * (1 - wei)\n        return xo\n\nclass iAFF(nn.Module):\n    def __init__(self, channels=64, r=4):\n        super(iAFF, self).__init__()\n        # \u7b2c\u4e00\u9636\u6bb5\u7279\u5f81\u878d\u5408\n        self.AFF1 = AFF(channels, r)\n        # \u7b2c\u4e8c\u9636\u6bb5\u7279\u5f81\u878d\u5408\n        self.AFF2 = AFF(channels, r)\n\n    def forward(self, x, y):\n        # \u7b2c\u4e00\u9636\u6bb5\u878d\u5408\n        z = self.AFF1(x, y)\n        # \u7b2c\u4e8c\u9636\u6bb5\u878d\u5408\n        z = self.AFF2(x, z)\n        return z\n\n# ------------------- \u7528\u6cd5\u793a\u4f8b -------------------\nif __name__ == &quot;__main__&quot;:\n    # \u521d\u59cb\u5316\n    aff = AFF(channels=64)\n    iaff = iAFF(channels=64)\n\n    # \u5047\u8bbe\u6709\u4e24\u4e2a\u7279\u5f81\u56fe\n    x = torch.randn(1, 64, 32, 32)\n    y = torch.randn(1, 64, 32, 32)\n\n    # \u4f7f\u7528AFF\n    out_aff = aff(x, y)\n\n    # \u4f7f\u7528iAFF\n    out_iaff = iaff(x, y)\n\n    print(f&quot;\u8f93\u5165\u5f62\u72b6: {x.shape}&quot;)\n    print(f&quot;\u8f93\u51fa\u5f62\u72b6: {out_iaff.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: Attentional Feature Fusion \ud83d\udd8b\ufe0f\u4f5c\u8005: Yimian Dai \ud83c\udfdb [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":3221,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[30,18],"tags":[],"class_list":["post-3220","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-30","category-18"],"_links":{"self":[{"href":"http:\/\/gnn.club\/index.php?rest_route=\/wp\/v2\/posts\/3220","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=3220"}],"version-history":[{"count":8,"href":"http:\/\/gnn.club\/index.php?rest_route=\/wp\/v2\/posts\/3220\/revisions"}],"predecessor-version":[{"id":3265,"href":"http:\/\/gnn.club\/index.php?rest_route=\/wp\/v2\/posts\/3220\/revisions\/3265"}],"wp:featuredmedia":[{"embeddable":true,"href":"http:\/\/gnn.club\/index.php?rest_route=\/wp\/v2\/media\/3221"}],"wp:attachment":[{"href":"http:\/\/gnn.club\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=3220"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/gnn.club\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=3220"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/gnn.club\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=3220"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}