{"id":3157,"date":"2025-07-09T15:56:16","date_gmt":"2025-07-09T07:56:16","guid":{"rendered":"https:\/\/www.gnn.club\/?p=3157"},"modified":"2025-07-09T15:57:06","modified_gmt":"2025-07-09T07:57:06","slug":"senet%ef%bc%9asqueeze-and-excitation-networks","status":"publish","type":"post","link":"http:\/\/gnn.club\/?p=3157","title":{"rendered":"SENet\uff1aSqueeze-and-Excitation Networks"},"content":{"rendered":"<h1>\u57fa\u672c\u4fe1\u606f<\/h1>\n<ul>\n<li>\n<p>\ud83d\udcf0\u6807\u9898: Squeeze-and-Excitation Networks<\/p>\n<\/li>\n<li>\n<p>\ud83d\udd8b\ufe0f\u4f5c\u8005: Jie Hu<\/p>\n<\/li>\n<li>\n<p>\ud83c\udfdb\ufe0f\u673a\u6784: Momenta (Beijing) \/ \u5317\u4eacMomenta<\/p>\n<\/li>\n<li>\n<p>\ud83d\udd17\u94fe\u63a5: <a href=\"https:\/\/arxiv.org\/abs\/1709.01507\">arXiv:1709.01507<\/a><\/p>\n<\/li>\n<li>\n<p>\ud83d\udd25\u5173\u952e\u8bcd: CNN, SE block, channel attention, image recognition<\/p>\n<\/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=\"width: 20%;\">\u9879\u76ee<\/th>\n<th style=\"width: 80%;\">\u5185\u5bb9<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>\ud83d\udcd6\u7814\u7a76\u80cc\u666f<\/td>\n<td>\u5377\u79ef\u795e\u7ecf\u7f51\u7edc\uff08CNN\uff09\u4e2d\u901a\u9053\u5173\u7cfb\u5efa\u6a21\u4e0d\u8db3<\/td>\n<\/tr>\n<tr>\n<td>\ud83c\udfaf\u7814\u7a76\u76ee\u7684<\/td>\n<td>\u63d0\u51fa\u8f7b\u91cf\u7ea7\u6a21\u5757\u663e\u5f0f\u5efa\u6a21\u901a\u9053\u95f4\u4f9d\u8d56\u5173\u7cfb<\/td>\n<\/tr>\n<tr>\n<td>\u270d\ufe0f\u7814\u7a76\u65b9\u6cd5<\/td>\n<td>\u8bbe\u8ba1Squeeze-and-Excitation\uff08SE\uff09\u6a21\u5757\uff1a\u5168\u5c40\u4fe1\u606f\u538b\u7f29+\u901a\u9053\u81ea\u9002\u5e94\u91cd\u6821\u51c6<\/td>\n<\/tr>\n<tr>\n<td>\ud83d\udd4a\ufe0f\u7814\u7a76\u5bf9\u8c61<\/td>\n<td>ImageNet\u7b49\u56fe\u50cf\u5206\u7c7b\u6570\u636e\u96c6<\/td>\n<\/tr>\n<tr>\n<td>\ud83d\udd0d\u7814\u7a76\u7ed3\u8bba<\/td>\n<td>SE\u6a21\u5757\u663e\u8457\u63d0\u5347\u6a21\u578b\u6027\u80fd\uff08ImageNet\u4e0atop-5\u8bef\u5dee\u964d\u4f4e25%\uff09\u4e14\u8ba1\u7b97\u4ee3\u4ef7\u6781\u5c0f<\/td>\n<\/tr>\n<tr>\n<td>\u2b50\u521b\u65b0\u70b9<\/td>\n<td>1. \u901a\u9053\u6ce8\u610f\u529b\u673a\u5236 2. \u5373\u63d2\u5373\u7528\u8bbe\u8ba1 3. \u4ec5\u589e\u52a00.5%\u8ba1\u7b97\u91cf<\/td>\n<\/tr>\n<tr>\n<td colspan=\"2\">\u6ce8\uff1aSE\u6a21\u5757\u901a\u8fc7\u7279\u5f81\u91cd\u6807\u5b9a\u5b9e\u73b0\u901a\u9053\u81ea\u9002\u5e94\uff0c\u540e\u7eed\u53d1\u5c55\u4e3a\u8ba1\u7b97\u673a\u89c6\u89c9\u57fa\u7840\u6a21\u5757\u4e4b\u4e00\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>\uff1aCNN\u5728\u89c6\u89c9\u4efb\u52a1\u4e2d\u8868\u73b0\u51fa\u8272\uff0c\u4f46\u4f20\u7edf\u5377\u79ef\u5c42\u4e3b\u8981\u5173\u6ce8\u5c40\u90e8\u7a7a\u95f4\u8fde\u63a5\u6a21\u5f0f\uff0c\u7f3a\u4e4f\u5bf9channel-wise\u5173\u7cfb\u7684\u663e\u5f0f\u5efa\u6a21\u3002Inception\u7b49\u67b6\u6784\u901a\u8fc7\u591a\u5c3a\u5ea6\u5904\u7406\u63d0\u5347\u6027\u80fd\uff0c\u8fd1\u671f\u7814\u7a76\u591a\u805a\u7126\u4e8e\u7a7a\u95f4\u4f9d\u8d56\u6027\u5efa\u6a21\u548c\u7a7a\u95f4\u6ce8\u610f\u529b\u673a\u5236\u3002<\/p>\n<\/li>\n<li>\n<p><strong>\u8fc7\u53bb\u65b9\u6848<\/strong>\uff1a\u73b0\u6709\u65b9\u6cd5\uff08\u5982Inception\u3001\u7a7a\u95f4\u6ce8\u610f\u529b\u6a21\u578b\uff09\u4fa7\u91cd\u4e8e\u4f18\u5316\u7a7a\u95f4\u7ef4\u5ea6\u7279\u5f81\u4ea4\u4e92\uff0c\u4f46\u672a\u7cfb\u7edf\u89e3\u51b3\u901a\u9053\u95f4\u4f9d\u8d56\u5173\u7cfb\u5efa\u6a21\u95ee\u9898\u3002\u8fd9\u7c7b\u65b9\u6cd5\u9700\u5f15\u5165\u989d\u5916\u8d85\u53c2\u6570\u6216\u590d\u6742\u7ed3\u6784\uff0c\u5de5\u7a0b\u5b9e\u73b0\u6210\u672c\u8f83\u9ad8\u3002<\/p>\n<\/li>\n<li>\n<p><strong>\u7814\u7a76\u52a8\u673a<\/strong>\uff1a\u9488\u5bf9CNN\u901a\u9053\u5173\u7cfb\u5efa\u6a21\u4e0d\u8db3\u7684\u6838\u5fc3\u95ee\u9898\uff0c\u63d0\u51fa\u8f7b\u91cf\u5316SE\u6a21\u5757\uff0c\u901a\u8fc7\u7279\u5f81\u91cd\u6807\u5b9a\u673a\u5236\u81ea\u9002\u5e94\u8c03\u6574\u901a\u9053\u7279\u5f81\u54cd\u5e94\u3002\u8be5\u65b9\u6cd5\u4ee5\u6781\u5c0f\u8ba1\u7b97\u4ee3\u4ef7\uff08\u4ec5\u589e0.5%\u53c2\u6570\u91cf\uff09\u5b9e\u73b0\u5168\u5c40\u901a\u9053\u4f9d\u8d56\u5efa\u6a21\uff0c\u5728ImageNet\u7b49\u4efb\u52a1\u4e2d\u9a8c\u8bc1\u5176\u666e\u9002\u6027\u3002<\/p>\n<\/li>\n<\/ul>\n<h1>\u65b9\u6cd5<\/h1>\n<ul>\n<li>\n<p><strong>\u7406\u8bba\u80cc\u666f<\/strong><br \/>\n\u57fa\u4e8eCNN\u7279\u5f81\u901a\u9053\u7684\u5168\u5c40\u4e0a\u4e0b\u6587\u5efa\u6a21\u9700\u6c42\uff0c\u53d7\u795e\u7ecf\u7cfb\u7edf\u6ce8\u610f\u529b\u673a\u5236\u542f\u53d1\uff0c\u63d0\u51fa\u7279\u5f81\u901a\u9053\u7684\u663e\u5f0f\u52a8\u6001\u8c03\u63a7\u7406\u8bba\u3002\u6838\u5fc3\u5047\u8bbe\uff1a\u901a\u9053\u7279\u5f81\u54cd\u5e94\u5e94\u5177\u5907\u81ea\u9002\u5e94\u91cd\u6807\u5b9a\u80fd\u529b\uff0c\u901a\u8fc7\u5efa\u6a21\u901a\u9053\u95f4\u975e\u7ebf\u6027\u4f9d\u8d56\u5173\u7cfb\u63d0\u5347\u7279\u5f81\u5224\u522b\u6027\u3002<\/p>\n<\/li>\n<li>\n<p><strong>\u6280\u672f\u8def\u7ebf<\/strong><br \/>\n<img decoding=\"async\" src=\"https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2025\/07\/20250709154509208.png\" alt=\"file\" \/><\/p>\n<\/li>\n<\/ul>\n<ol>\n<li>\n<p><strong>Squeeze\u64cd\u4f5c<\/strong>\uff1a\u5168\u5c40\u5e73\u5747\u6c60\u5316\uff08GAP\uff09\u538b\u7f29\u7a7a\u95f4\u7ef4\u5ea6\uff0c\u751f\u6210\u901a\u9053\u7ea7\u7edf\u8ba1\u63cf\u8ff0\u7b26\uff08C\u00d71\u00d71\uff09<\/p>\n<\/li>\n<li>\n<p><strong>Excitation\u64cd\u4f5c<\/strong>\uff1a<\/p>\n<ul>\n<li>\u4e24\u5c42\u5168\u8fde\u63a5\u7f51\u7edc\uff08\u542bReLU\u548cSigmoid\uff09\u5b66\u4e60\u901a\u9053\u95f4\u975e\u7ebf\u6027\u4ea4\u4e92<\/li>\n<li>\u751f\u6210\u901a\u9053\u6ce8\u610f\u529b\u6743\u91cd\uff08C\u00d71\u00d71\uff09<\/li>\n<\/ul>\n<\/li>\n<li>\n<p><strong>\u7279\u5f81\u91cd\u6807\u5b9a<\/strong>\uff1a\u539f\u59cb\u7279\u5f81\u56fe\u4e0e\u6ce8\u610f\u529b\u6743\u91cd\u9010\u901a\u9053\u76f8\u4e58\uff0c\u5b9e\u73b0\u901a\u9053\u81ea\u9002\u5e94\u6821\u51c6<\/p>\n<\/li>\n<li>\n<p><strong>\u6a21\u5757\u90e8\u7f72<\/strong>\uff1a\u7075\u6d3b\u5d4c\u5165\u73b0\u6709\u7f51\u7edc\uff08\u5982ResNet\u5355\u5143\u540e\uff09\uff0c\u5f62\u6210SE-ResNet\u7b49\u53d8\u4f53<\/p>\n<\/li>\n<\/ol>\n<h1>\u7ed3\u8bba<\/h1>\n<ul>\n<li>\u63d0\u51faSE block\u67b6\u6784\u5355\u5143\uff0c\u901a\u8fc7\u52a8\u6001\u901a\u9053\u7279\u5f81\u91cd\u6807\u5b9a\u673a\u5236\u663e\u8457\u589e\u5f3a\u7f51\u7edc\u7684\u8868\u5f81\u80fd\u529b\uff0c\u4e3aCNN\u901a\u9053\u4f9d\u8d56\u5efa\u6a21\u63d0\u4f9b\u65b0\u8303\u5f0f<\/li>\n<li>\u4f18\u70b9\uff1a<br \/>\n1) \u5728\u591a\u4e2a\u6570\u636e\u96c6\u5b9e\u73b0SOTA\u6027\u80fd<br \/>\n2) \u6a21\u5757\u8f7b\u91cf\u5316\u4e14\u5373\u63d2\u5373\u7528<br \/>\n3) \u7279\u5f81\u91cd\u8981\u6027\u5206\u6790\u5bf9\u7f51\u7edc\u538b\u7f29\u7b49\u884d\u751f\u4efb\u52a1\u6709\u542f\u53d1\u4ef7\u503c<\/li>\n<li>\u4e3b\u8981\u7ed3\u8bba\uff1a<\/li>\n<\/ul>\n<ol>\n<li>SE\u6a21\u5757\u6709\u6548\u89e3\u51b3\u4e86\u4f20\u7edf\u67b6\u6784\u5728channel-wise\u7279\u5f81\u4f9d\u8d56\u5efa\u6a21\u4e0a\u7684\u5c40\u9650\u6027<\/li>\n<li>\u5b9e\u9a8c\u8bc1\u660eSENets\u5728\u4fdd\u6301\u8ba1\u7b97\u6548\u7387\u7684\u540c\u65f6\u63d0\u5347\u5224\u522b\u7279\u5f81\u5b66\u4e60\u80fd\u529b<\/li>\n<li>\u6a21\u5757\u8bf1\u5bfc\u7684\u7279\u5f81\u91cd\u8981\u6027\u53ef\u8fc1\u79fb\u81f3\u7f51\u7edc\u526a\u679d\u7b49\u5e94\u7528\u573a\u666f<\/li>\n<\/ol>\n<h1>pytorch\u4ee3\u7801<\/h1>\n<pre><code class=\"language-python\">import torch\nimport torch.nn as nn\nimport torch.nn.functional as F\n\nclass SEBlock(nn.Module):\n    def __init__(self, in_channels, reduction_ratio=16):\n        &quot;&quot;&quot;\n        SE Block \u5b9e\u73b0\n        Args:\n            in_channels: \u8f93\u5165\u7279\u5f81\u56fe\u7684\u901a\u9053\u6570\n            reduction_ratio: \u538b\u7f29\u6bd4\u4f8b\uff08\u7528\u4e8e\u4e2d\u95f4\u5c42\u901a\u9053\u6570\uff09\n        &quot;&quot;&quot;\n        super(SEBlock, self).__init__()\n        self.avg_pool = nn.AdaptiveAvgPool2d(1)  # \u5168\u5c40\u5e73\u5747\u6c60\u5316 (Squeeze)\n\n        # Excitation \u90e8\u5206\uff1a\u4e24\u5c42\u5168\u8fde\u63a5\u5c42 + \u6fc0\u6d3b\u51fd\u6570\n        self.fc = nn.Sequential(\n            nn.Linear(in_channels, in_channels \/\/ reduction_ratio, bias=False),\n            nn.ReLU(inplace=True),\n            nn.Linear(in_channels \/\/ reduction_ratio, in_channels, bias=False),\n            nn.Sigmoid()  # \u8f93\u51fa\u901a\u9053\u6743\u91cd [0, 1]\n        )\n\n    def forward(self, x):\n        b, c, _, _ = x.shape  # batch_size, channels, height, width\n\n        # Squeeze: \u5168\u5c40\u5e73\u5747\u6c60\u5316\u5f97\u5230\u901a\u9053\u63cf\u8ff0\u7b26\n        y = self.avg_pool(x).view(b, c)  # shape: [b, c]\n\n        # Excitation: \u751f\u6210\u901a\u9053\u6743\u91cd\n        y = self.fc(y).view(b, c, 1, 1)  # shape: [b, c, 1, 1]\n\n        # \u7279\u5f81\u56fe\u91cd\u6807\u5b9a\uff08\u901a\u9053\u4e58\u6cd5\uff09\n        return x * y.expand_as(x)\n\n# ------------------- \u7528\u6cd5\u793a\u4f8b -------------------\nif __name__ == &quot;__main__&quot;:\n    # 1. \u521d\u59cb\u5316SE Block\uff08\u8f93\u5165\u901a\u9053\u6570\u4e3a256\uff0c\u538b\u7f29\u6bd4\u4f8b16\uff09\n    se_block = SEBlock(in_channels=256, reduction_ratio=16)\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 = se_block(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<p>\u5173\u952e\u70b9\u89e3\u6790<\/p>\n<ul>\n<li>\n<p>Squeeze \u64cd\u4f5c<\/p>\n<p>\u901a\u8fc7 nn.AdaptiveAvgPool2d(1) \u5c06\u6bcf\u4e2a\u901a\u9053\u7684\u5168\u5c40\u7a7a\u95f4\u4fe1\u606f\u538b\u7f29\u4e3a\u4e00\u4e2a\u6807\u91cf\u3002<\/p>\n<\/li>\n<li>\n<p>Excitation \u64cd\u4f5c<\/p>\n<p>\u4e24\u5c42\u5168\u8fde\u63a5\u5c42\uff1a<\/p>\n<p>\u7b2c\u4e00\u5c42\u964d\u7ef4\uff08reduction_ratio \u63a7\u5236\u538b\u7f29\u6bd4\u4f8b\uff0c\u9ed8\u8ba416\uff09\u3002<\/p>\n<p>\u7b2c\u4e8c\u5c42\u6062\u590d\u539f\u59cb\u901a\u9053\u6570\uff0c\u5e76\u901a\u8fc7 Sigmoid \u8f93\u51fa\u6743\u91cd\u503c\uff080~1\uff09\u3002<\/p>\n<\/li>\n<li>\n<p>\u7279\u5f81\u91cd\u6807\u5b9a<\/p>\n<p>\u5c06\u901a\u9053\u6743\u91cd\u4e0e\u539f\u59cb\u7279\u5f81\u56fe\u9010\u901a\u9053\u76f8\u4e58\uff08x * y\uff09\uff0c\u5b9e\u73b0\u81ea\u9002\u5e94\u6821\u51c6\u3002<\/p>\n<\/li>\n<\/ul>\n","protected":false},"excerpt":{"rendered":"<p>\u57fa\u672c\u4fe1\u606f \ud83d\udcf0\u6807\u9898: Squeeze-and-Excitation Networks \ud83d\udd8b\ufe0f\u4f5c\u8005: Jie Hu  [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":3158,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[30,18],"tags":[],"class_list":["post-3157","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\/3157","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=3157"}],"version-history":[{"count":4,"href":"http:\/\/gnn.club\/index.php?rest_route=\/wp\/v2\/posts\/3157\/revisions"}],"predecessor-version":[{"id":3162,"href":"http:\/\/gnn.club\/index.php?rest_route=\/wp\/v2\/posts\/3157\/revisions\/3162"}],"wp:featuredmedia":[{"embeddable":true,"href":"http:\/\/gnn.club\/index.php?rest_route=\/wp\/v2\/media\/3158"}],"wp:attachment":[{"href":"http:\/\/gnn.club\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=3157"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/gnn.club\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=3157"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/gnn.club\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=3157"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}